Tag: AI Agents

AI Coding Assistant Market Share: Who Leads in 2026?

If you are choosing an AI coding assistant for yourself or your development team, the headline answer is clear: Claude Code leads the October 2026 primary-tool market at 29.4%, ahead of GitHub Copilot at 22.7%. That does not automatically make Claude Code the right purchase. The aggregate ranking hides large differences between startups and enterprises, terminal users and IDE users, and the assistant you open versus the model that actually generates the code.

The useful question is not simply which product is biggest. It is which market signal applies to your environment, what the rapid move toward coding agents changes, and how much weight market share should carry in your evaluation. Here is how to read the numbers without turning popularity into a substitute for testing.

Read primary-tool share as a competitive signal, not total adoption

The October estimate measures the percentage of professional developers who name a product as their primary AI coding assistant: the one they use most often to write, edit, or review production code. It does not count every tool a developer has tried, every installed extension, total seats, vendor revenue, or the volume of code generated.

That distinction matters because many developers use two or three assistants. A developer might rely on Claude Code for repository-wide implementation, keep GitHub Copilot enabled for inline completion, and occasionally send a background task to Codex. Only the tool used most often receives that developer’s primary-tool share.

The estimate combines an August 4 to September 26, 2026 survey of 2,350 professional developers in North America and Europe with publicly disclosed seat and usage figures. The responses were normalized into a market model. That makes the results useful for understanding competition among leading products, but they are not a worldwide census or a direct measure of software quality.

Use the rankings to build a shortlist, understand where workflows are moving, and challenge an outdated default. Do not use them alone to approve a company-wide rollout.

Key takeaways

Claude Code leads with 29.4% of primary-tool share in October 2026; GitHub Copilot follows at 22.7%, Cursor at 13.1%, and OpenAI Codex at 11.8%.
The four largest assistants hold 77.0% combined, up from 71.2% in January 2026.
Terminal and CLI agents are now the largest interface category, rising from 21.3% in January to 38.6% in October.
Company size changes the ranking: Claude Code leads among startups, while GitHub Copilot leads at companies with more than 5,000 employees.
Product share and model share are different. Claude models account for 47.3% of model-family coding usage because they are available through products beyond Claude Code.
Market share can tell you which tools deserve evaluation. Only a controlled test against your repositories, policies, and workflows can tell you which one deserves deployment.

The October leaderboard shows both concentration and disruption

Large central technology nodes and smaller fast-moving nodes compete inside a glowing circular digital arena.

The October 2026 primary-tool snapshot puts two terminal-oriented agents in the top four and shows substantial movement since January. The change column uses percentage points, not percent growth.

Rank	AI coding assistant	Developer	Primary interface	October 2026 share	Change since January
1	Claude Code	Anthropic	Terminal agent	29.4%	+10.9 points
2	GitHub Copilot	Microsoft / GitHub	IDE extension	22.7%	-7.8 points
3	Cursor	Anysphere	AI-native IDE	13.1%	-4.9 points
4	OpenAI Codex	OpenAI	Terminal and cloud agent	11.8%	+7.6 points
5	Google Antigravity and Gemini Code Assist	Google	AI-native IDE	5.6%	+1.1 points
6	JetBrains AI and Junie	JetBrains	IDE extension	4.3%	-0.6 points
7	Windsurf	Cognition	AI-native IDE	2.9%	-1.9 points
8	Amazon Q Developer and Kiro	Amazon	IDE extension	2.6%	-0.8 points
9	OpenCode	Open source	Terminal agent	2.4%	+1.6 points
10	Cline	Open source	IDE extension	1.7%	-0.5 points
–	All other tools	Various	Various	3.5%	-4.7 points

Claude Code’s lead is meaningful because it is paired with the largest gain in the table. OpenAI Codex has the second-largest increase and has nearly tripled its primary-tool share since January. Copilot and Cursor remain substantial products, but both have lost share while agent-oriented tools have gained it.

The top four products now account for 77.0% of primary-tool usage, compared with 71.2% in January. That is evidence of concentration within this definition of the market. It is not evidence that the category has settled: the order inside that concentrated group has changed quickly.

The five-quarter trajectory is more useful than a single rank

Quarterly averages smooth out the monthly movement and show that the change in leadership was not a one-month fluctuation. They also explain why the Q3 values below differ slightly from the October snapshot.

Assistant	Q3 2025	Q4 2025	Q1 2026	Q2 2026	Q3 2026
GitHub Copilot	36.2%	33.4%	30.1%	26.0%	23.1%
Claude Code	7.9%	12.6%	19.2%	24.8%	28.7%
Cursor	19.4%	19.9%	17.6%	15.2%	13.5%
OpenAI Codex	1.8%	3.1%	4.6%	8.3%	11.2%
Google	3.6%	4.1%	4.0%	4.9%	5.4%

Claude Code passed GitHub Copilot between Q2 and Q3 2026. Copilot declined in every quarter shown, while Claude Code rose in every quarter. Cursor peaked at 19.9% in Q4 2025 and then declined for three consecutive quarters. Codex accelerated most sharply after Q1 2026, moving from 4.6% to 8.3% in Q2 and 11.2% in Q3. Google’s movement was steadier, ending Q3 at 5.4%.

For a buyer, sustained direction deserves more weight than a narrow difference in one snapshot. A rising product is more likely to receive integrations, community attention, training material, and internal advocacy. A falling product can still be the best operational fit, especially when its decline reflects a changing interface preference rather than a failure of the product itself.

The decisive change is from suggestions to delegated tasks

A developer moves from receiving one code suggestion to supervising AI agents that coordinate coding, testing, and deployment tasks.

The market is not merely swapping one vendor for another. Developers are changing how they interact with coding AI. Inline completion asks an assistant to help with the next fragment of code. An agent can receive a broader goal, inspect multiple files, make coordinated edits, run commands or tests, and return a larger unit of work for review.

The interface numbers capture that shift. Tools that support several interfaces are assigned to the one each respondent uses most often, so the categories describe dominant behavior rather than permanent product boundaries.

Interface type	January 2026 share	October 2026 share	Change
Terminal / CLI agent	21.3%	38.6%	+17.3 points
IDE extension	41.2%	29.4%	-11.8 points
AI-native IDE	24.8%	17.9%	-6.9 points
Cloud / background agent	5.1%	9.2%	+4.1 points
Browser-based app builder	7.6%	4.9%	-2.7 points

Terminal and CLI agents gained 17.3 points between January and October, becoming the largest interface category at 38.6%. Cloud and background agents also gained share. IDE extensions fell from 41.2% to 29.4%, while AI-native IDEs fell from 24.8% to 17.9%.

This does not mean the IDE is disappearing. IDE extensions still represent nearly three in ten primary workflows, and many agent users review the resulting code in an editor. It means that an evaluation built entirely around autocomplete quality is now incomplete.

Your test set should include the work agents are being asked to own: a change that touches several files, a bug whose cause is not identified in the prompt, a refactor that must preserve behavior, and a review task that requires following repository conventions. Record whether the assistant finds the right context, makes coherent changes, validates them, and leaves an understandable diff. A fast completion is not useful if the developer spends longer discovering and correcting hidden mistakes.

Agent capability also changes the risk boundary. If a tool can execute commands, modify many files, access external systems, or open pull requests, test it first in a protected branch or isolated environment. Apply the least permissions it needs, keep credentials out of its context, require review before merge, and let your normal test and security controls judge the output. The specific downside is larger than a poor inline suggestion: an agent can propagate a wrong assumption across a repository or act on an unintended resource.

Your company size and model layer change the apparent winner

The overall ranking is least reliable when it is treated as though every buyer faces the same constraints. The split by employer size shows four materially different markets.

Company size	Claude Code	GitHub Copilot	Cursor	OpenAI Codex	All other tools
Startup, 1-50 employees	36.8%	9.7%	21.4%	15.2%	16.9%
Small business, 51-500	33.1%	17.5%	16.2%	13.4%	19.8%
Mid-market, 501-5,000	27.9%	25.8%	11.7%	10.9%	23.7%
Enterprise, more than 5,000	22.4%	35.6%	8.3%	9.1%	24.6%

Claude Code is strongest among startups at 36.8% and declines steadily to 22.4% at enterprises. Cursor has an even sharper segment gap, moving from 21.4% among startups to 8.3% at companies with more than 5,000 employees. Codex follows the same broad pattern, though less dramatically.

GitHub Copilot moves in the opposite direction. It holds 9.7% among startups but leads the enterprise segment at 35.6%. Existing Microsoft licensing agreements help explain why Copilot remains the default procurement route inside many large organizations. At mid-market companies, Claude Code and Copilot are much closer, at 27.9% and 25.8% respectively.

If you work at a startup, the aggregate table understates the prevalence of Claude Code, Cursor, and Codex among your peers. If you manage enterprise tooling, it understates Copilot’s position and the influence of procurement, identity, administration, and existing contracts. Use the segment closest to your organization as the starting point, then check whether your technical and governance requirements resemble that peer group.

Do not confuse the assistant with the underlying model

A product is the working environment: its interface, context handling, repository tools, permissions, integrations, and review flow. A model is the code-generating engine available inside that environment. Several assistants allow developers to choose among model families, so the product leaderboard cannot tell you which models generate the most coding output.

Model family	Developer	January 2026 share	October 2026 share
Claude	Anthropic	49.2%	47.3%
GPT	OpenAI	22.4%	28.6%
Gemini	Google	11.8%	10.2%
Open-weight models, including Qwen, DeepSeek, Kimi, and GLM	Various	10.3%	9.4%
Grok	xAI	3.4%	2.1%
All other models	Various	2.9%	2.4%

Claude models account for 47.3% of model-family coding usage, substantially more than Claude Code’s 29.4% product share. The difference exists because Claude models are also used within Cursor, GitHub Copilot, and open-source agents. GPT models gained 6.2 points between January and October, reaching 28.6% as Codex expanded. Open-weight models retained 9.4%, with their use concentrated among cost-sensitive teams and self-hosted deployments.

This separation gives you a better evaluation design. First judge whether the product fits your workflow and controls. Then compare the models available inside it on the same tasks. Keep the model name and version in your evaluation record; otherwise, a model change can be mistaken for a product improvement or regression.

Turn the market-share numbers into a defensible tool decision

Market share is useful evidence of momentum, ecosystem depth, and peer adoption. It does not directly measure correctness, security, developer satisfaction, review burden, total cost, or performance on your codebase. A defensible decision uses the market data to narrow the field and repository-level evidence to choose among the finalists.

Define the job before naming a vendor. Decide whether you mainly need inline completion, repository exploration, multi-file implementation, code review, background execution, or a combination. The interface trend shows that these are no longer interchangeable versions of the same task.
Apply your non-negotiable constraints. Check supported editors and terminals, operating environments, authentication, administrative controls, data handling, model availability, network access, auditability, and contract requirements. Remove any product that cannot meet a genuine constraint before comparing output quality.
Build a segment-aware shortlist. Include the overall leader, the leader for your company-size segment, and a credible alternative with a different interface or model strategy. An enterprise shortlist that ignores Copilot would miss the segment leader; a startup shortlist containing only Copilot would ignore how differently that segment behaves.
Use the same representative task set. Give every finalist an existing bug, a multi-file feature, a behavior-preserving refactor, and a code-review assignment drawn from the kinds of repositories it would actually encounter. Keep the prompt, starting commit, permissions, and acceptance criteria consistent.
Score the cost of reaching an acceptable result. Record whether the final change passes the relevant tests, how much developer intervention it requires, how long review and correction take, whether it follows repository conventions, and what the successful result costs. Do not reward a tool merely for producing more code or producing it faster.
Test assistant and model choices separately. When a product offers several models, rerun the important tasks with each viable model. This reveals whether the value comes from the interface and agent harness, the underlying model, or their combination.
Control the rollout and set a reassessment trigger. Begin with repositories and permissions where mistakes are detectable and reversible. Expand only after the review burden and failure modes are understood. Reassess when a major model, agent mode, pricing structure, policy requirement, or contract renewal changes the decision.

The practical choice is rarely the product with the largest number beside its name. It is the assistant that completes your representative work with the lowest combined burden of prompting, correction, review, administration, and risk. Use the 2026 leaderboard to decide what deserves a serious test, then let reproducible work in your own environment decide what your team adopts.

References

First Page Sage – AI Coding Assistant Market Share: October 2026

October 8, 2026

How to Win Visibility in Agent-Driven Search

Your page can rank first and still lose the customer. In agent-driven discovery, a person can ask an AI assistant to find, compare, book, buy, or contact a provider. The agent may evaluate several businesses and complete the task without sending that person through a familiar results page.

That changes the visibility problem. You still need to be found, but you also need to survive qualification, support verification, and offer a safe path to action. The practical goal is not merely to appear in an answer. It is to remain the best eligible choice all the way through the agent’s workflow.

Search visibility now has four separate gates

An agent commonly turns a delegated request into requirements, searches for possible candidates, evaluates each candidate against those requirements, checks important claims, and then attempts the requested action. A conventional ranking affects the candidate-gathering stage, but it does not settle the final decision.

Gate	Question the agent must resolve	What your site needs to provide	Useful metric
Retrieval	Can I find this business for the delegated task?	Indexable pages, unambiguous entities, relevant task language, and clear topical coverage	Candidate appearance rate
Qualification	Does it satisfy every non-negotiable requirement?	Explicit capabilities, limits, prices, locations, eligibility rules, integrations, and availability	Hard-requirement pass rate
Selection	Is it the best fit among the eligible choices?	Suitability guidance, evidence, differentiators, and independently verifiable claims	Selection share when retrieved
Completion	Can I safely perform the requested action?	A usable form, booking flow, checkout, approved API, or clearly defined human handoff	Successful action rate

Ranking remains important because it helps a brand enter the candidate set. It is no longer a reliable proxy for winning the decision. First Page Sage reported that, in its vendor-led analysis of 2,417 agentic commands issued from March 4 through June 10, 2026, the first-ranked result was selected 44.6% of the time, while a result ranked fourth or lower was selected 38.2% of the time. Those figures are directional rather than universal benchmarks: they come from one commercial analysis, and agent behavior can differ by platform, category, request, and user context.

The useful conclusion is narrower and more durable: rank and selection are different outcomes. If your reporting stops at impressions, positions, and clicks, you cannot tell whether an agent failed to retrieve your brand, rejected it on a requirement, distrusted a claim, or could not complete the transaction.

Give each gate its own metric. Candidate appearance rate tells you whether discovery is working. Hard-requirement pass rate exposes missing or disqualifying facts. Selection share tells you whether the agent prefers you after finding you. Successful action rate reveals whether your conversion path works for an automated assistant. A single visibility score hides all four failure modes.

Publish the facts agents need to qualify you

A central business model is connected to visual modules for location, hours, price, availability, services, accessibility, and verification.

A broad category page may rank for “payroll software,” “family hotel,” or “commercial electrician” while giving an agent too little information to answer a constrained request. Real delegated tasks include conditions: company size, location, budget, dates, integrations, accessibility needs, service area, cancellation terms, or regulatory requirements.

Agents can treat those conditions differently. A hard requirement eliminates a candidate. An important requirement carries substantial weight. A nice-to-have breaks a close comparison. An optional feature may add only a small advantage. Your first content job is to discover which facts occupy each tier for the buying tasks that matter to your business.

Choose a delegated commercial task. Use a task tied to revenue, such as booking a service, selecting a product, requesting a proposal, or arranging a demonstration. Commercial requests deserve priority because delegated agent activity is more concentrated around buying, booking, and hiring than around general informational searches.
Write down the complete requirement set. Use actual sales questions, support tickets, requests for proposals, on-site searches, form responses, and objections. Separate non-negotiable conditions from preferences instead of treating every feature as equally important.
Map every hard requirement to a canonical page. The answer should be stated directly, not buried in a brochure, image, unsupported comparison chart, or sales-only conversation.
Add suitability content. Explain who the offer is for, who it is not for, which situations it supports, what prerequisites apply, and where its limits begin.
Keep consequential facts synchronized. Prices, regions, availability, policies, product names, and eligibility rules should not conflict across product pages, help content, structured data, directories, and partner profiles.

Use a suitability page pattern that answers the whole decision

A useful suitability page is not another generic “why choose us” page. It should let a machine or a person decide whether your offer fits a specific situation. A practical structure is:

Best fit: the customer, use case, location, scale, or conditions the offer is designed for.
Required conditions: prerequisites the customer must meet before buying, booking, or applying.
Supported requirements: the capabilities, integrations, service areas, configurations, or policies that satisfy common constraints.
Limitations: unsupported scenarios, exclusions, capacity boundaries, dependencies, and cases that require a different offer.
Commercial facts: visible pricing where possible, or a precise explanation of what determines price; availability; fees; cancellation terms; and what happens after submission.
Evidence: links to documentation, policies, certifications, product details, or independent material that substantiates consequential claims.
Next action: a clear route to buy, book, request a quote, schedule a demonstration, or move to a human review.

Dedicated suitability content is worth testing even if it attracts little conventional search volume. In the same vendor analysis, businesses with this kind of content were selected 2.7 times as often as equally ranked businesses without it. That multiplier should not be treated as a guaranteed result, but the mechanism is sensible: explicit fit information reduces the inference an agent must make.

Make proof machine-readable without hiding caveats

Relevant JSON-LD can express your organization, offer, product or service, availability, and other supported attributes in a consistent format. Use it to clarify facts already visible on the page. Do not use markup to introduce claims, prices, ratings, availability, or capabilities that a visitor cannot confirm in the page content.

Structured data reduces ambiguity; it does not establish truth. Agents may compare a site’s claims with what they already know and with independent material before choosing a candidate. Make important assertions easy to verify by identifying what the claim applies to, where it applies, and under which conditions. A sentence such as “integrates with accounting software” is weak. A maintained integration page that names the supported systems, required plan, setup path, and current limitations is decision-grade evidence.

Consistency matters here. Use the same business name, canonical URL, product names, locations, and core offer descriptions wherever you control the information. When a third-party profile is outdated, correct it. When a claim changes, update the visible page and its markup together. Contradictory facts force an agent to decide which version to trust, and the safest decision may be to exclude the candidate.

Remove the blockers between selection and completion

A glowing agent pathway moves through verification, availability, selection, payment, and completion while alternate routes end at digital obstacles.

A recommendation has limited commercial value if the agent cannot finish the requested job. The operational difference is whether a page is machine-actionable: can an approved agent use the interface to submit the inquiry, reserve the time, add the product, complete the purchase, or reach a defined handoff?

The vendor-led command analysis recorded 78.3% of conversions on machine-actionable pages, compared with 9.6% on pages where the agent could not act. This is not a promise that making a form accessible will produce a particular conversion rate. It is evidence that transactional usability can become a selection constraint rather than a minor conversion optimization.

Audit the complete transaction, not just the landing page

Use visible, specific field labels. “Work email,” “arrival date,” and “number of employees” are easier to interpret than placeholder-only or context-dependent fields.
State required inputs before submission. If a quote needs a postal code, account identifier, property type, budget range, or document, disclose that requirement before the agent enters the flow.
Explain validation failures precisely. Identify the affected field, preserve valid entries, and say what an acceptable value looks like.
Expose material terms before commitment. Price, fees, renewal terms, cancellation conditions, availability, and approval dependencies should not appear only after the decisive click.
Use conventional controls and stable destinations. Buttons should have meaningful labels, links should resolve predictably, and essential actions should not depend on unexplained gestures or decorative interface elements.
Return an actionable confirmation. Show what was submitted, whether it succeeded, what happens next, and any reference number or next step the user needs.
Define the human handoff. If the task cannot be automated, say which step requires a person, what information that person needs, and how the customer will be contacted.

Test the flow from a clean session using the same facts a customer would give an agent. Check every branch: unavailable dates, unsupported locations, invalid entries, expired inventory, payment failure, authentication, and confirmation. A form that works only on the happy path is not reliably actionable.

Agent-friendly does not mean unguarded. Keep authentication, fraud controls, consent, privacy safeguards, and human approval wherever the risk requires them. Do not weaken a security control to make automation easier. If automated action is allowed, provide an approved route; if it is not, provide a clear and honest handoff instead of a hidden bypass.

Use audience preference where the platform supports it

Retrieval is not driven only by topical relevance. Google Preferred Sources gives readers an explicit way to star publications in the Top Stories area so that stories from those outlets can appear more often for those readers. This is a narrow feature with a precise scope: it concerns publications and Top Stories, not every business listing, organic result, or AI-agent decision.

The feature has nevertheless become large enough for publishers to treat it as a real retention channel. Google reported that people had selected more than 600,000 unique sources, up from 200,000 in May 2026. Google has also said that users who select a preferred source are twice as likely to click. Those figures describe this specific feature; they do not establish a general ranking advantage across search or AI platforms.

If you publish news and participate in Top Stories, the implementation is straightforward:

Install Google’s Preferred Sources button using the supported implementation.
Place the prompt near a moment when the reader has received value, such as the end of a substantive story, rather than interrupting the opening.
Explain the result accurately: starring the publication can make its stories appear more often in that reader’s Top Stories experience.
Record the preferred-source user count with its reporting date so you can measure growth instead of relying on an undated total.
Compare that growth with returning readership and engagement, while keeping correlation separate from proof of causation.

Some site owners received Search Console emails showing a Preferred Source user count as of October 5, 2026. Google also surveyed recipients about future reporting methods, frequency, and metrics. Until regular reporting is established, keep your own dated record of any counts you receive.

If you are not a relevant publication, do not imitate the button or describe ordinary follows as Preferred Sources. Apply the underlying principle without inventing a platform signal: give satisfied readers a clear way to return, subscribe, follow, or search for your brand again. Explicit preference can support a durable audience, but it should not be presented as proof that an unrelated agent will select you.

Measure agent visibility as a decision path

You do not need access to an agent’s private logs to build a useful diagnostic. You need a repeatable set of realistic tasks and a disciplined record of what can be observed. Start with the commercial requests that matter most, because “explain this topic” and “choose a provider and submit an inquiry” test very different kinds of visibility.

Define the task exactly. Include the hard constraints a real buyer would provide: location, budget, timing, compatibility, eligibility, scale, or required terms.
Preserve the test context. Record the platform, date, locale, sign-in state, exact command, and any files or preferences supplied. Keep the command unchanged when comparing runs.
Capture the candidate set. Note whether your brand appeared, which page supported the appearance, what claims were surfaced, and which competing options were considered.
Score each requirement. Mark hard requirements as confirmed, failed, contradictory, or unknown. An unknown should not be counted as a pass merely because you know the answer internally.
Separate selection from retrieval. Record whether the brand was found, whether it remained eligible, whether it was selected, and the observable reasons given. Do not present an inferred reason as if the agent disclosed it.
Test the action. Where authorized, follow the process through the form, booking, cart, checkout, or handoff. Record the exact field, policy, authentication step, or interface state that prevents completion.
Fix the earliest failed gate. More suitability copy will not solve an indexing failure. More authority will not repair an unusable booking flow. Diagnose before choosing the optimization.
Repeat on a fixed cadence. Agent outputs can change, so compare patterns across repeated observations rather than treating one response as a permanent ranking.

Keep conventional SEO and analytics beside this testing. Search rankings still influence retrieval, human visitors still use results pages, and agent-driven commercial activity remains only part of search. The measurement upgrade is additive: it connects rankings and mentions to qualification, selection, and completed work.

Key takeaways

A ranking can earn entry into an agent’s candidate set without earning the final selection.
Publish explicit requirements, supported scenarios, limitations, commercial terms, and suitability guidance so the agent does not have to guess.
Use JSON-LD to clarify visible facts, not to make unsupported claims or conceal qualifications.
Make consequential claims consistent and independently verifiable.
Treat forms, booking systems, checkout, APIs, and human handoffs as part of search visibility.
Measure retrieval, qualification, selection, and completion separately so each failure receives the right fix.
Use Google Preferred Sources if its Top Stories scope fits your publication, but do not mistake it for a universal agent-ranking signal.

Choose your highest-value delegated task and trace it from discovery to completion. If your brand is absent, repair retrieval. If it appears but is rejected, expose the missing fit or proof. If it is selected but the task stalls, fix the transaction. That sequence keeps you from buying more visibility when the real leak is qualification, trust, or action.

References

October 6, 2026

AI Agent Adoption in 2026: A Practical Market Guide

If you are deciding whether to deploy an AI agent, do not start with the market leader. Start with the job you need completed, the systems the agent may touch, and the consequences when it stops halfway through.

The market is growing while its center of gravity weakens. Tracked AI agent usage rose from 142 million aggregate monthly active users in Q3 2025 to 293 million in Q3 2026, but the four largest platforms’ combined share fell from 58.6% to 49.3%. That is the environment you are buying into: rapid adoption, many credible specialists, and no safe assumption that one platform will own every workflow.

The market is expanding faster than any one leader

An AI agent is more than a chatbot with a new label. It accepts a goal, breaks that goal into subtasks, chooses actions as conditions change, and works across tools or systems until it reaches an end state. A single-turn assistant does not meet that definition. Neither does an orchestration framework such as LangGraph or Bedrock AgentCore, which helps developers build agents, nor a classification model that chooses a route without pursuing a goal of its own.

This distinction protects you from buying the wrong layer. A chat license may improve drafting without automating a process. A framework may give your engineering team control without supplying a ready-to-use worker. A fast decision model may make an agent cheaper and safer without replacing the agent itself.

The following snapshot covers selected leaders from a 40-platform market tracked between May 15 and September 10, 2026. The estimates combine company disclosures, app-store telemetry, procurement records, and account-level observations. They measure platform reach rather than unique people, so someone using several agents can appear in several platforms’ totals.

Agent	Primary use	Estimated MAUs	Q3 2026 share	Quarter-over-quarter growth
ChatGPT Agent	Multi-step research, booking, and file work	58.9M	20.1%	+16%
Microsoft 365 Copilot	Document and Office workflow agents	33.4M	11.4%	+13%
GitHub Copilot Agent	Turning bug reports into code fixes	26.7M	9.1%	+11%
Gemini Agent Mode	Browser automation and form completion	25.5M	8.7%	+19%
Claude Code	Repository-wide refactoring and test generation	19.3M	6.6%	+24%
Cursor	Multi-file changes inside the editor	13.5M	4.6%	+8%
OpenAI Atlas	Site navigation and transactional tasks	11.7M	4.0%	+27%
Perplexity Comet	Agentic browsing, comparison, and checkout	10.8M	3.7%	+22%
Salesforce Agentforce	Support deflection and CRM pipeline hygiene	9.1M	3.1%	+15%
Grok Bot	Persistent work on a cloud computer	7.9M	2.7%	New
All other agents	Vertical, open-source, and smaller platforms	45.1M	15.4%	+14%

Market-share loss does not necessarily mean user loss. ChatGPT Agent’s share declined from 24.9% in Q3 2025 to 20.1% in Q3 2026 while its estimated users increased from 35.4 million to 58.9 million. Microsoft 365 Copilot and GitHub Copilot Agent also added users while losing relative share. New entrants and expanding specialists diluted the incumbents because the total market grew faster than they did.

Use market share to assess reach, integration momentum, talent availability, and the likelihood that a product will remain supported. Do not use it as a proxy for successful task completion. The practical response to fragmentation is portability: retain task definitions, approval rules, logs, evaluation cases, and critical business data in systems you control wherever possible. Switching agents should not require rebuilding your operating knowledge from scratch.

Choose a workflow category before you choose a vendor

There is no single AI agent market in operational terms. Coding, browser automation, enterprise productivity, CRM work, personal assistance, and long-running general-purpose work have different tools, permissions, failure modes, and definitions of success.

Coding is currently the largest category, representing 24.8% of tracked agent usage. Even there, the products are not interchangeable. GitHub Copilot Agent is positioned around taking a bug report through to a finished fix. Claude Code emphasizes repository-wide changes and tests. Cursor centers work in the editor, Replit Agent spans prototype-to-deployment creation, and Amazon Q Developer focuses on cloud and coding operations.

The same specialization appears outside software development. Microsoft 365 Copilot sits inside Office workflows. Salesforce Agentforce works inside CRM processes. Gemini Agent Mode, OpenAI Atlas, and Perplexity Comet concentrate on browser actions, but their stated strengths range from form completion to transactional navigation and comparison-led checkout. A generic request for the “best agent” hides these material differences.

Write an outcome brief before requesting demonstrations

A useful evaluation begins with a workflow that has an observable finish. Document these elements before you shortlist products:

Goal: State the result the agent must produce or the action it must complete.
Starting state: Identify the request, file, ticket, record, or event that begins the run.
Permitted systems: List the applications, data, credentials, and tools the agent may use.
Definition of done: Describe the final artifact or system state precisely enough that a reviewer can mark it complete or incomplete.
Approval gates: Specify where a person must approve publishing, payment, deletion, external communication, code deployment, or another consequential action.
Stop conditions: Tell the agent what uncertainty, missing permission, policy conflict, or unexpected state requires escalation.
Recovery requirement: Define what the agent must log, preserve, or reverse when it cannot finish.

For an SEO team, “help with a content audit” is too loose to evaluate. A testable workflow identifies the properties to crawl, the fields to collect, the rule for classifying each page, the destination for the findings, and whether the agent may change a live page. The clearer the end state, the easier it becomes to compare products without being distracted by fluent demonstrations.

Adopt at the workflow level rather than declaring an organization-wide agent strategy first. A company may reasonably use one agent for repository work, another for CRM operations, and another for browser research. Fragmentation becomes manageable when every deployment has a named job and a shared governance model.

Completion rate is the buying metric that corrects popularity

An automated workflow passes through connected stations to a completed package while several alternate routes stop at incomplete handoffs.

Monthly active users tell you that people invoked a platform. They do not tell you whether it finished the job. For an autonomous workflow, the more relevant question is simple: what percentage of eligible runs reaches the defined end state without a person correcting the agent?

One standardized comparison required each platform to attempt 48 multi-step tasks across five trials, producing 240 runs per platform. A run counted as complete only when it finished end to end without human correction. Claude Code led at 72.1% unassisted completion, followed by ChatGPT Agent at 65.3% and Grok Bot at 63.7%. Gemini Agent Mode reached 59.6%, GitHub Copilot Agent 57.2%, and Cursor 55.8%.

Those figures are useful for shortlisting, not for forecasting your deployment. The task mix may not resemble your workflow, and an agent’s performance changes with tool access, permissions, data quality, integration depth, and the exact definition of completion. Claude Code’s result is especially relevant to repository work; it does not establish that a coding agent is the best choice for CRM cleanup or browser checkout.

Speed also needs context. In that benchmark, OpenAI Atlas had a median completion time of 4 minutes 51 seconds and Perplexity Comet 4 minutes 39 seconds, while ChatGPT Agent took 8 minutes 52 seconds and Grok Bot 19 minutes 14 seconds. A fast incomplete run is not efficient. A slower run may still be preferable if it completes more often, requires fewer interventions, or handles a more complex job.

Measure the run, not the demo

Your pilot dashboard should separate these outcomes instead of compressing them into a vague satisfaction score:

Unassisted completion rate: Eligible runs that reach the defined end state with no corrective intervention.
Partial completion rate: Runs that create useful progress but fail to reach the required state.
Intervention rate: Runs in which a person must clarify, repair, approve unexpectedly, or take over.
Time to successful completion: Measure completed runs separately so quick failures do not make the agent appear faster.
Cost per successful completion: Divide total run costs, including retries and supporting model calls, by completed outcomes rather than by invocations.
Recovery quality: Check whether failed runs leave clear logs, preserve work, avoid duplicate actions, and return systems to a known state.
Policy adherence: Record attempts to cross approval boundaries, use disallowed data, or invoke an unauthorized tool.

Keep every started run in the denominator. If your goal is autonomous completion, a person quietly fixing the result before it reaches the dashboard is a failed autonomous run, even when the final output looks good.

Separate the agent from the decision engines beneath it

An exploded modular AI system shows an agent above separate reasoning, memory, control, data, and tool components as a hand replaces one module.

An agent does not need a large generative model for every step. Planning, writing, summarizing, classifying, routing, policy checking, and executing an API call are different computational jobs. Treating them as one undifferentiated prompt raises latency and cost while making failures harder to diagnose.

The term System One model is being used for a model that returns a typed, calibrated decision from a predefined answer set rather than free-form prose. It can choose a ticket category, route a request to a model, select a tool, or decide whether a proposed action meets a policy. It does not independently accept a goal and pursue it, so it belongs inside an agent architecture rather than in the agent column of a market-share table.

This layer matters because structured decisions are numerous but relatively inexpensive. Across 3.1 billion production API calls observed in 1,400 applications beginning June 1, 2026, structured decision tasks represented 63.7% of calls but only 15.5% of token spend. Long-form generation showed the opposite pattern: 9.1% of calls consumed 38.4% of token spend. A specialized decision model can therefore remove a large amount of traffic from a general-purpose model without displacing a comparable share of model spending.

The best candidates have an answer space you can enumerate before the call. Binary classification led a September 2026 survey of 421 AI engineering teams, with 60.5% already piloting or planning adoption within six months. Schema extraction ranked last at 28.7% because field values are often open-ended. That gap gives you a practical rule: use a decision model when you can list all legitimate outcomes; retain a generative model when the output itself must be created.

Type safety is necessary, but it is not factual accuracy

A model can return a perfectly valid category and still choose the wrong category. Constrained decoding on a small language model achieved a 0.0% type error rate in the same benchmark as Jev, so valid output syntax is not, by itself, a differentiator. You still need labeled evaluation cases that test whether the decision is correct.

The alternatives also remain competitive. A fine-tuned encoder classifier recorded 0.09-second median latency and a $0.018 cost per million input tokens, compared with Jev at 0.14 seconds and $0.042. The tradeoff is breadth: a new classification question can require another encoder to be trained, while a broader decision endpoint can answer different predefined questions. A small language model using constrained decoding was slower at 2.1 seconds, with input priced at $0.35 per million tokens and output at $1.40.

Early demand does not prove steady-state adoption. Jev was only seven days old when launch-week estimates put it at 31,416 developers making at least one API call, while 6.2% of new accounts reached production. Treat that as evidence of interest and low integration friction, not as evidence that the architecture has already become standard.

A clean production design assigns each layer a narrow responsibility:

The agent owns the goal, task state, planning, and recovery path.
Decision models handle enumerable classifications, routing, ranking, policy checks, and tool selection.
Generative models create prose, summaries, code, and other open-ended outputs.
Deterministic tools read or change external systems under explicit permissions.
Human approval remains in front of irreversible, externally visible, or high-consequence actions.

Log the input, output, confidence or score, selected route, tool result, and final task outcome at the relevant layer. Otherwise, a failed workflow leaves you guessing whether the planner, classifier, generator, integration, or external system caused the problem.

Build an adoption plan that survives vendor churn

A durable rollout does not depend on predicting which logo will lead the next market table. It depends on preserving your workflow knowledge and measuring interchangeable components against the same definition of success.

Select one bounded workflow. Favor a repeatable job with an observable end state and enough current friction to justify integration work.
Map the action boundary. Separate read-only work, reversible internal changes, external communications, financial actions, deployments, and destructive operations. Require human approval where an error would be difficult to reverse.
Shortlist by category fit. Compare agents designed for the systems and work involved instead of beginning with overall reach.
Run identical evaluation cases. Include normal requests, missing information, ambiguous instructions, permission failures, tool errors, and requests that should trigger a refusal or escalation.
Score completed outcomes. Track unassisted completion, interventions, time, cost, policy adherence, and recovery behavior using the same denominator for every candidate.
Decompose expensive runs. Identify classification, routing, ranking, safety, and tool-selection calls that can move to a specialized decision model or deterministic rule.
Retain a migration path. Keep prompts, outcome briefs, schemas, evaluation cases, logs, and business rules outside proprietary interfaces when the platform permits it.

If customers encounter your business through agents

Agent adoption changes acquisition as well as operations. ChatGPT Agent is used for multi-step research and booking; Gemini Agent Mode handles browser automation and forms; OpenAI Atlas performs site navigation and transactions; Perplexity Comet supports comparison and checkout. If any of those journeys matter to your business, visibility alone is an incomplete success metric. The agent must be able to identify the right page, understand the offer, verify important facts, and complete or correctly hand off the next step.

Apply the same outcome-based discipline to AI SEO, AEO, and GEO work:

Put essential product, service, eligibility, policy, and contact information in visible page text rather than only in images or interactive widgets.
Give each important entity, offer, and resource a stable canonical URL with a clear page purpose.
Keep structured data consistent with the claims a visitor can see. Schema is a machine-readable consistency layer, not permission to publish contradictory or unsupported markup.
Use specific labels for links, buttons, form fields, and required inputs so an agent does not have to infer what an interface element does.
Publish dates, units, methodology, limitations, and originating evidence beside factual claims that an agent may need to evaluate or cite.
Test complete journeys from discovery to the required outcome. Record where the agent selects the wrong page, loses context, cannot operate a control, encounters conflicting facts, or reaches an unexpected approval step.

This is where agent analytics should meet search analytics. A mention in an AI answer, an agent visit, a successful product comparison, and a completed transaction are separate events. Tracking only referral traffic hides the failures between discovery and completion.

Key takeaways

AI agent usage is expanding rapidly, but market share is fragmenting rather than settling around one permanent winner.
Choose an agent for a defined workflow category and observable end state, not for overall popularity.
Use unassisted completion, intervention, recovery, time, and cost per successful outcome as the core buying metrics.
Keep goal pursuit in the agent layer while routing enumerable decisions to specialized models or deterministic rules where appropriate.
Make customer journeys explicit, structured, and testable if browser and general-purpose agents are part of your discovery or conversion path.

Your next move is deliberately small: choose one workflow whose finish you can describe in a sentence, preserve a human gate before consequential actions, and run the same cases through category-appropriate candidates. The market will keep changing. A clear outcome definition and a portable evaluation set let you benefit from that competition instead of being trapped by it.

References

September 21, 2026

AI Agent Optimization and GEO Services: A Buyer’s Guide

Your company can appear in an AI answer and still lose the buyer. The system may cite an obsolete page, combine two products, repeat an unsupported claim, or recommend your business without giving the user a workable next step. A visibility screenshot does not solve any of those failures.

If you are deciding whether to hire an AI agent optimization or generative engine optimization service, you need a more precise buying standard. The provider should make your business easier for AI systems to discover, understand, verify, represent accurately, and use during a customer task. Here is how to define that work, test the provider’s evidence, and connect the program to revenue.

AI visibility and agent readiness are separate outcomes

GEO, AEO, and AI agent optimization overlap, but they do not solve exactly the same problem.

Generative engine optimization, or GEO, improves the likelihood that your business, expertise, and content will be selected, cited, or recommended in generative search experiences.
Answer engine optimization, or AEO, makes an answer easy to extract and present directly. It emphasizes clear questions, concise answers, supporting detail, and an information structure that does not force a system to infer the main point.
AI agent optimization extends beyond the answer. It asks whether an agent can identify the right entity, retrieve current facts, understand conditions and limitations, and move the user toward an appropriate action.

This last layer is often described as agent experience, or AX. The practical test is whether an AI agent can read your information and act on it, not merely whether it can find your brand name.

Stage	What the system must resolve	Common failure	Required service output
Discovery	Whether your business is relevant to the user’s task	The brand is absent from unbranded recommendations or associated with the wrong category	A query and task map tied to markets, audiences, offers, and existing pages
Evaluation	Whether your claims are specific, current, and credible	The answer repeats vague marketing language, cites weak evidence, or confuses similar offers	A claim inventory, supporting evidence, entity cleanup, and citation-ready content
Action	What the user or agent should do next	Requirements, availability, policies, locations, or conversion paths are unclear	Explicit next steps, stable destination pages, current conditions, and safe handoff points
Measurement	Whether visibility produced a useful business result	The report counts mentions but cannot connect them to qualified demand	Versioned response logs, referral tracking, CRM fields, lead quality, customers, and cost

A provider that sells only the discovery stage is selling an AI visibility service, not a complete agent optimization program. That may still be useful, but the contract and price should reflect the narrower scope.

Structured data belongs in this system, but it is not the whole system. JSON-LD can clarify entities and relationships when it accurately describes the visible page. It cannot repair contradictory claims, create third-party authority, or guarantee that a model will cite you. Treat any promise of guaranteed placement through schema alone as a warning sign.

Turn the service label into a concrete deliverables list

Isometric illustration of a service workbench with stages for mapping a site, separating product entities, linking evidence, checking technical components, and testing an agent task path.

“GEO optimization” is too vague to approve as a statement of work. Require the provider to name the surfaces it will test, the assets it will change, the evidence it will produce, and the commercial event it will measure.

1. Establish a reproducible baseline

The baseline should contain the prompts or tasks that matter to your customers, the platforms on which they will be tested, and the result before any work begins. Each test record should preserve the exact prompt, date, market, language, interface, response, cited URLs, brand mentions, competing entities, and any factual errors.

A defensible test matrix can include ChatGPT, Gemini, Claude, Google AI Overviews, and relevant regional platforms. Do not add a platform merely to make the dashboard look comprehensive. Include it when your customers use it or when it materially influences their research environment.

Generative responses can vary between runs, so one favorable output is an observation, not a performance rate. The provider should retain successful and unsuccessful runs under the same protocol. Otherwise, you cannot tell whether a change improved repeatable visibility or merely produced a convenient screenshot.

2. Map customer tasks, not just keywords

A keyword list describes strings people type. A task map describes the decision they are trying to make. It should separate broad education, problem diagnosis, solution comparison, vendor selection, validation, and action. It should also distinguish branded from unbranded demand.

For every priority task, require a target audience, market, intended answer, relevant entity, best supporting page, evidence requirement, next action, and measurement event. This exposes gaps that ordinary keyword research can miss. You may already have a page that mentions the query while lacking the facts an AI system would need to recommend you confidently.

3. Build an entity and claim inventory

AI systems encounter your organization through many representations: service pages, product pages, profiles, interviews, directories, review sites, news coverage, partner pages, and structured data. If those representations use conflicting names, categories, capabilities, locations, or policies, the system has to resolve the conflict.

The inventory should list each material claim, where it appears, the evidence supporting it, the person responsible for it, and the condition that should trigger review. Include claims about availability, geography, pricing, certifications, integrations, performance, eligibility, and comparisons where they are relevant. Unsupported superlatives such as “best,” “leading,” and “most trusted” should not survive this process unless they have verifiable support.

4. Upgrade the content and technical layer together

Useful GEO content answers the decision question early, supports it with evidence, and then explains conditions, alternatives, and limitations. It does not bury the answer under an essay written only to occupy search-result space.

The technical work should check whether important information is available in stable, crawlable page content; whether canonical and duplicate versions create ambiguity; whether internal links express the relationship between entities and topics; and whether structured data matches what a person can see. The content and schema should be reviewed as one release. Updating one while leaving the other stale creates a new contradiction.

Do not interpret agent accessibility as permission to open every system to every crawler. Security, privacy, licensing, and infrastructure controls still apply. The provider should document which public content needs discovery, which automated access is permitted, and which sensitive or authenticated functions require a controlled interface or human confirmation.

5. Improve corroboration beyond your own domain

Your website can state what the business does. Independent references help establish whether those claims are credible. A complete service should therefore identify missing or inconsistent external evidence rather than treating on-page editing as the entire job.

This does not justify manufacturing mentions, publishing disguised endorsements, or distributing the same promotional copy across low-quality sites. The useful work is narrower: correct inaccurate profiles, align material facts, publish original evidence when you have it, make qualified experts identifiable, and earn relevant coverage or citations through legitimate public relations and reputation work.

6. Design the next action for people and agents

A recommendation has limited value if the next page does not explain how to proceed. The destination should state who the offer is for, what information is required, what happens after submission, which restrictions apply, and where the user can get help.

For higher-risk actions, build explicit confirmation points. An agent should not be encouraged to infer consent, accept legal terms, move money, expose private information, or make an irreversible change merely because the conversion path is technically available. Good AX makes safe progress easier; it does not remove necessary review.

Test a GEO provider’s evidence before you buy

A buyer examines source containers, before-and-after models, linked evidence, and repeatable agent tests while decorative glowing signals remain in the background.

The core buying question is not whether the agency understands AI vocabulary. It is whether you can reproduce its evidence and inspect the chain from optimization to business result.

Ask for a proof packet

A serious provider should be able to show a redacted example containing:

The original business objective and the unbranded customer tasks used for testing.
The baseline responses, including unfavorable results and factual errors.
The pages, structured data, entity records, or external signals that changed.
The exact prompts and testing conditions used after publication.
Raw outputs and cited URLs, not only a chart summarizing them.
The denominator behind every percentage. “Appeared in 80% of tests” is meaningful only if you know which tests qualified.
The connection between visibility, qualified leads, customers, revenue, and program cost.

Recommendation frequency is useful when the query set, platform set, market, competitor group, test conditions, and failures are disclosed. It becomes a vanity metric when a provider selects only prompts on which the client already performs well.

Score the operating model

Assess how the work will move through your organization. A technically strong plan can still fail if nobody has authority to update claims, approve schema, correct external profiles, or connect analytics to the CRM.

Method: Can the provider explain how tasks are selected, how outputs are recorded, and how it separates correlation from a plausible effect of its work?
Industry fit: Has it handled the approval burden, sales cycle, terminology, and evidence standards of a comparable category?
Regional fit: Does its platform and language coverage match your buyers rather than its standard reporting package?
Editorial control: Who checks factual accuracy, claim support, tone, and legal or compliance requirements before publication?
Technical access: Who can edit templates, structured data, internal links, rendering behavior, analytics, and consent-aware tracking?
Ownership: Do you retain the prompt set, content, schema, response logs, dashboards, and documentation when the engagement ends?
Governance: Is there a named owner for each correction, release, test, and approval?

Methodology transparency, search experience, independently cited work, and demonstrated recommendation performance can all inform due diligence. Their importance changes by context. Independent methodological validation matters more when procurement, legal, or compliance teams must defend the investment; relevant client outcomes matter more than general prestige when you need execution in a specific market.

A provider’s own agency ranking is not independent validation, even when its testing method appears thoughtful. Use vendor-published comparisons to build a shortlist and identify evaluation criteria. Verify the underlying claims separately before signing.

Reject guarantees that the provider cannot control

No agency controls a frontier model’s training data, retrieval process, product interface, citation policy, or future output. That makes guaranteed rankings, permanent citations, and universal “AI preference” claims untenable.

A responsible commitment is operational: the provider will complete named changes, test a disclosed task set, record outputs consistently, correct representation errors it can influence, and report commercial results under an agreed attribution model. That is enforceable work. A promise that ChatGPT or another platform will always recommend you is not.

Build a business case without hiding the uncertainty

GEO can be measured economically, but public benchmarks are still less mature than established paid-search or SEO benchmarks. Use external numbers to challenge your assumptions, not to replace your own baseline.

One proprietary 36-month dataset covered 341 companies across 15 industries between October 2023 and September 2026. It reported an average GEO customer acquisition cost of $581, compared with $470 for traditional SEO, a 23.6% difference. GEO received an average lead-quality score of 8.2 out of 10 and a 40-day conversion timeline, versus 7.8 and 84 days for traditional SEO.

Those averages are directional, not universal. The dataset was 64% B2B, used a minimum of eight companies per industry, and excluded paid advertising on AI platforms. Industry-level GEO CAC ranged from $265 in construction to $1,129 in higher education, while the reported conversion timelines ranged from 11 days in ecommerce to 61 days in higher education. Your sales process, margins, market, attribution method, and existing authority can move the result substantially.

The same proprietary data reported a $497 average CAC, 91% success rate, and 52-day time to results for premium agency-managed programs. In-house-only programs were reported at $947, 46%, and 203 days. The difference is large enough to make implementation quality worth investigating, but not strong enough to assume that hiring an agency automatically produces the lower figure. The data comes from an agency, the engagement models are not standardized across the market, and selection effects may account for part of the gap.

Before using any benchmark in a budget request, make the provider define “success,” “customer,” “attributed,” “program cost,” and “time to results” in terms your finance and sales teams accept. Otherwise, two dashboards can report different CACs from the same pipeline.

Measure the program at three levels

Visibility and representation: Track valid task coverage, brand inclusion, citation frequency, cited pages, competitive presence, factual error rate, and whether the answer describes your offer correctly.
Engagement and influence: Track AI-referred sessions, qualified actions, assisted conversions, CRM discovery responses, and sales notes that record meaningful AI-assisted research.
Commercial efficiency: Track qualified leads, new customers, attributable revenue, total program cost, CAC, conversion time, and payback under a documented attribution rule.

Keep direct and influenced performance separate. Direct GEO CAC divides program cost by customers assigned directly to an AI referral under your agreed model. Influenced GEO CAC uses customers with documented AI involvement. Combining the two produces a cleaner-looking number but destroys its meaning.

Set the attribution window from your real sales cycle rather than from a generic analytics default. Preserve the pre-change baseline, annotate every release, and segment branded from unbranded tasks. A rise in branded mentions may reflect demand created elsewhere; stronger performance on unbranded vendor-selection tasks is more persuasive evidence that the GEO program affected discovery.

Your allowable CAC should come from unit economics and the payback period your finance team can support. Do not approve a budget simply because it is below a published industry average. A benchmark cannot tell you whether the acquired customer’s margin, retention, or implementation cost makes the investment sensible for your business.

Key takeaways for your first operating cycle

Start with a stable set of customer tasks, target markets, platforms, and conversion outcomes. Do not begin with content production.
Capture the baseline before changing pages, structured data, profiles, or external evidence.
Require an entity and claim inventory so that every material fact has evidence, an owner, and a review trigger.
Treat GEO, AEO, technical access, reputation, and agent experience as connected workstreams with separate deliverables.
Require raw response logs and failed tests. A gallery of favorable screenshots cannot establish recommendation frequency.
Measure visibility, representation accuracy, qualified demand, customers, and cost as separate layers.
Keep direct attribution distinct from documented influence, and use your own sales cycle and unit economics.
Retain ownership of the content, structured data, task set, dashboards, logs, and implementation documentation.

Your first move should be to write the test and evidence requirements, not to choose an agency. Give each shortlisted provider the same business tasks and ask how it would baseline them, what it would change, what proof it would return, and how the result would enter your CRM. The provider that can make that operating chain concrete is worth deeper diligence. The one selling unspecified “AI visibility” is asking you to buy the label.

References

September 19, 2026

Agentic Ecommerce: A Playbook for Discovery and Advertising

If your product pages rank and your ads are live, but your products still disappear from AI-guided shopping conversations, the missing layer is usually not more promotional copy. It is decision-ready product data: facts an agent can retrieve, compare, explain, and carry into checkout.

Your goal is no longer just to win a click. You need to help an AI determine whether a specific product fits a specific buyer’s constraints, answer the next question accurately, and make the handoff to your store without changing the facts along the way.

The shopping funnel now contains a conversation

A conventional product ad asks the shopper to click before learning much. A conversational ad can answer questions about fit, compatibility, features, availability, or policies inside the discovery surface. ChatGPT is testing clearly labeled Sponsored Agents that open a separate brand conversation, while Google’s Business Agent is being tested inside YouTube ads for eligible U.S. retailers.

That changes the intermediate step, not the buyer’s underlying job. People still need to eliminate unsuitable choices, understand tradeoffs, and trust the terms of the purchase. The difference is that an agent may now perform part of that evaluation before the shopper reaches your product page.

Do not collapse every appearance in AI into one visibility metric. There are three distinct outcomes:

Citation: your content supplies an explanation or fact used in an answer.
Recommendation: your brand enters the suggested set for a category or use case.
Selection: a particular product is matched to the shopper’s stated requirements and advanced toward purchase.

Each outcome requires different work. Clear, retrievable content helps with citation. Consistent brand context supports recommendation. Complete product attributes, current commercial data, and a usable transaction path support selection. This is why LLM readability, brand context, and agentic commerce are separate optimization disciplines, even when one team owns all three.

Do not fund this shift by abandoning traditional search. An Ahrefs-based measurement found AI Overviews on 24% of shopping queries on Sept. 3, 2026, but a Datos panel of more than 10 million desktop users measured dedicated AI Mode at only about 0.13% of web traffic. A separate panel of 75 ecommerce stores, mostly producing $1 million to $20 million in annual revenue, still placed non-branded organic search second only to paid search for revenue. The practical response is a parallel search and AI strategy, not a wholesale channel migration.

Build a product record an agent can safely choose

An unbranded hiking shoe is surrounded by organized visual layers representing its materials, size, fit, availability, shipping, and return details.

An agent cannot reliably recommend what it cannot distinguish. A polished category description will not compensate for missing variant measurements, ambiguous compatibility, stale availability, or different prices in the feed and on the page.

For every product and variant you want an agent to select, create one canonical record with five layers:

Identity: product name, brand, category, model, SKU or other applicable identifiers, plus the exact relationship between parent products and variants.
Transaction truth: price, currency, condition, availability, fulfillment choices, shipping terms, returns, warranty, and any eligibility rules for discounts or member pricing.
Decision attributes: dimensions, materials, fit, capacity, supported devices or systems, care requirements, included components, and other facts buyers use to rule products in or out.
Evidence and instructions: manuals, size charts, compatibility tables, policy pages, certifications when applicable, and factual answers to recurring pre-purchase questions.
Destinations: the correct product page, variant URL, cart action, policy page, or support handoff for each answer.

Publish the same facts through the channels machines use: visible page content, merchant feeds, platform catalog integrations, and Product and Offer structured data where applicable. JSON-LD should be generated from the same commerce data as the page and feed. Treating schema as a separate copywriting exercise creates exactly the contradictions an agent should not have to resolve.

Run a variant-level consistency check before activating an agent or campaign. Compare title, identifier, price, currency, availability, shipping, return terms, and the primary decision attributes across the page, feed, structured data, and commerce API. If a field is genuinely unknown, leave it unknown and define a safe fallback. Do not let the agent infer compatibility, delivery, or warranty coverage from adjacent products.

Product copy still matters, but it should answer rather than decorate. Put the direct answer first, then the explanation, supporting evidence, and relevant conditions. Keep each FAQ block focused on one buyer question so it can be retrieved without unrelated text changing its meaning.

The commercial case for this cleanup is promising but should not be overstated. Google reports that merchants following its core Merchant Center feed practices see an average 5% conversion increase in the following month. In a Lululemon test, retailer-supplied conversational attributes were incorporated in 50% of relevant AI Mode product recommendations. These are platform-reported results, not guaranteed lifts. Their useful lesson is narrower: attributes that exist as maintained data can participate in recommendations; facts trapped in campaign copy cannot be depended on in the same way.

Design conversational ads around the next unanswered question

A shopper and an abstract AI guide exchange symbol-filled bubbles while narrowing several coffee machines to one suitable choice.

A conversational ad should not be a chat-shaped version of a display ad. Its job is to resolve the next material uncertainty and route the shopper to the correct action. Build an answer map before you generate creative.

Buyer question	Required data	Safe handoff
Will this fit?	Variant measurements, sizing method, and size-chart rules	The selected variant and relevant size guide
Will it work with what I own?	Supported models, exclusions, required accessories, and version limits	The compatible variant or compatibility table
What will I actually pay?	Current price, currency, shipping terms, and applicable member benefits	A cart with the same disclosed terms
Can I get it when and where I need it?	Live inventory and available fulfillment methods	The available purchase or pickup path
What if it is unsuitable?	Return window, condition requirements, exclusions, and warranty terms	The relevant policy section or support route

For each row, define an answer contract: the approved system of record, the claims the agent may make, the data that must be checked live, the fallback when data is unavailable, and the destination that preserves context. A useful fallback is specific: state which fact cannot be confirmed and direct the shopper to the place or person that can confirm it. A confident guess is not customer service.

AI can also compress campaign production. ChatGPT Work’s Ads Manager plugin can create, update, and analyze campaigns from natural-language instructions; its assistance can propose copy and imagery from a landing page and campaign objective. Optional text customization can adapt headlines and descriptions to the conversation or translate them into the user’s preferred language. U.S. Shopify merchants can also use a ChatGPT Ads app to manage campaigns, while Shopify Catalog data supports more accurate product appearances in shopping conversations. These workflow and catalog integrations reduce interface work, but they do not remove the need for review.

Review generated copy against the canonical product record, not just the landing page’s marketing language.
Validate translated claims, units, policies, and variant names before enabling localized customization.
Require a live lookup for price, stock, delivery, and personalized benefits when those values can change.
Send every answer to a landing state that preserves the chosen product or variant. Do not make the shopper repeat the conversation.
Log unsupported questions and corrected answers as product-data defects, then fix the underlying record.

Keep paid and independent answers conceptually separate. OpenAI says Sponsored Agent conversations are labeled and separated from the original ChatGPT conversation, advertising does not influence ChatGPT’s independent answers, and advertisers do not receive users’ private conversations. Plan your measurement around the signals the platform legitimately exposes; do not design a campaign that assumes access to private prompt history.

Measure the path from question to profitable order

Click-through rate cannot describe the whole experience when a conversation performs part of the product-page job. It may produce fewer but better-qualified visits, expose missing information, or assist a purchase completed through another surface. Build a measurement chain that distinguishes those outcomes.

Visibility: eligible ad exposure, AI share of voice, recommendation coverage across a fixed set of target shopping prompts, and the products most often surfaced.
Conversation: conversation starts, qualified question rate, common question categories, answer failure rate, and the share of conversations that reach a site handoff.
Selection: variant views, product comparisons, cart additions, and checkout starts originating from the agent experience.
Transaction: completed orders, revenue, margin where available, assisted conversions, and member-benefit usage.
Outcome quality: cancellations, returns, exchanges, and support contacts attached to agent-assisted orders.

Define the denominators before launch. Conversation start rate is starts divided by eligible ad exposures when the platform supplies both values. Qualified question rate is conversations containing a decision question divided by starts. Answer failure rate is unsupported, corrected, or escalated answers divided by starts. If a platform withholds a denominator, mark the rate unavailable instead of combining unrelated proxies.

Use distinct campaign identifiers and landing URLs for each agent surface, preserve product and variant context in the handoff, and record launch dates in your analytics annotations. Compare performance with a suitable unactivated product, market, or campaign group where possible. Keep budget, promotion, inventory, and seasonal differences visible so a lift is not automatically credited to the agent.

Google’s AI performance insights in Merchant Center are generally available in Australia, Canada, India, New Zealand, and the U.S., including comparisons of brand share of voice across AI Mode and AI Overviews. Its Universal Commerce Protocol integration can also support cart transfers to merchant sites and expanded checkout testing. Loyalty data can surface member-specific pricing and benefits. These discovery, checkout, and personalization capabilities make segmentation essential: report new and returning customers, members and non-members, and agent-assisted and conventional journeys separately.

Key takeaways: use this launch sequence

Choose one decision-heavy category. Start where buyers repeatedly ask about fit, compatibility, delivery, or policy terms, because those questions reveal whether the agent adds real value.
Separate your goals. Decide whether each activity is intended to earn a citation, a brand recommendation, a product selection, or a paid conversation.
Repair the product record first. Align variant identity, decision attributes, price, inventory, policies, page content, feed data, and JSON-LD before generating campaigns.
Create the answer map. Pair each common buyer question with an approved data field, a safe fallback, and a destination that preserves the selected product.
Apply campaign guardrails. Human-review generated claims and translations, require live checks for changing commercial facts, and prohibit unsupported inference.
Instrument the whole path. Track visibility, dialogue, selection, checkout, and post-purchase quality rather than using clicks as the sole success signal.
Feed failures back into operations. Repeated unanswered questions belong in the catalog backlog; frequent returns after an agent interaction may indicate that an answer or attribute is misleading.

Start with the category where a wrong answer would most often block or spoil a purchase. Make that category reliably answerable across organic discovery, conversational ads, and checkout. Scale only after the same facts survive every handoff.

References

September 16, 2026

Profound’s $180M Funding: What Marketing Teams Should Test

If you are deciding whether Profound’s funding makes its platform a safer strategic bet, separate two questions immediately: Does the company have more capacity to pursue its vision, and can the product remove work from your marketing operation? The first is supported by the raise. The second still requires proof inside a workflow that matters to you.

That distinction will keep a large funding number from becoming a substitute for product, governance, and commercial due diligence. It also gives you a practical way to evaluate AI Marketer without either dismissing the platform or buying the story before testing the system.

What Profound has actually committed to

Profound has raised $180 million to build an AI platform for marketing. Its stated premise is that AI is generating additional work for marketers, not simply automating existing tasks. AI Marketer is positioned as the response: a system that brings company context and agents together so marketing teams can get that work done.

Those points establish capital, direction, and a product thesis. They do not establish the return a customer will receive. A funding total cannot tell you whether the platform fits your data, integrates with your operating stack, produces reliable outputs, shortens approval cycles, or reduces the total cost of a workflow.

The stated goal also indicates a broad platform ambition rather than a single-purpose feature. That can be valuable when your work crosses research, analysis, content, brand governance, and execution. It can also increase implementation scope. The more jobs a platform is expected to coordinate, the more important permissions, source quality, handoffs, and ownership become.

Use the announcement as a reason to ask better questions, not as the answer to them. Do not add unconfirmed details about valuation, investors, product allocation, delivery dates, or business performance to your internal brief. If one of those details affects your decision, request it directly and distinguish a written commitment from a forward-looking plan.

Why more AI can create more marketing work

A marketing team sorts and reviews a growing flow of campaign materials produced by several automated machines.

AI reduces the cost of producing an output, but output generation is only one part of marketing. Every new model, answer surface, automated campaign, and content variant can create additional monitoring, interpretation, validation, approval, and measurement work. Faster production can therefore move the constraint downstream rather than remove it.

You can see that effect by mapping the full chain around an AI-assisted task:

Inputs: Someone must select the relevant brand rules, product facts, audience assumptions, performance data, and prior decisions.
Generation: A model or agent produces an analysis, recommendation, brief, campaign asset, or other deliverable.
Verification: A person checks factual accuracy, source quality, brand fit, compliance, and whether the output answers the original question.
Execution: The approved output must reach the correct channel, owner, or system without losing its context.
Learning: Results must return to the process so that the next action reflects what changed.

A platform can make generation faster while leaving every other stage intact. It can even increase review work if it produces more material than your team can verify. That is why prompts completed, agents deployed, and assets generated are weak measures of operating value on their own.

Before watching a demonstration, draw one real workflow from request to approved outcome. Mark every system, human handoff, approval, wait state, and rework loop. Record the elapsed time, active working time, and recurring errors using evidence you already have. You now have a baseline against which automation can be judged.

If your remit includes AI search visibility or generative engine optimization, a suitable workflow might begin with a visibility finding and end with an approved content or entity-data change. The test should include the analysis, supporting evidence, assignment, revision, publication approval, and follow-up measurement. Automating only the first step does not automate the workflow.

What company context and agents must prove

The combination of company context and agents is the central idea behind AI Marketer’s positioning. Those terms can sound complete while hiding the hardest implementation questions. Treat them as two systems to test separately.

Test context as a governed source of truth

Company context should do more than place files near a model. It should help the system select current, authorized information and show you what influenced an output. Ask for a live demonstration that answers these questions:

Which repositories, pages, records, and instructions can the system use for this task?
How does it decide which source is authoritative when two sources conflict?
How quickly does a changed product fact, policy, or brand rule become available?
Can access be limited by team, role, market, client, or workspace?
Can a reviewer trace an output back to the facts and instructions that shaped it?
What happens when the required evidence is missing, stale, or ambiguous?

Do not test this with a polished sample library. Bring a controlled set of realistic material that includes one outdated item, one conflict, and one fact the system should not expose to every user. Designate the correct source in advance. A useful context layer should handle the conflict predictably, respect access boundaries, and make its reasoning inspectable enough for a reviewer to catch a mistake.

Test agents as bounded operators

An agent is valuable when it can advance work without gaining more authority than the task requires. Evaluate its operating boundaries, not only the quality of its final output:

What triggers the agent, and who can change that trigger?
Which data can it read, and which systems can it alter?
Which steps require human approval before the agent proceeds?
Can you stop a run immediately and prevent it from retrying?
Does the audit history preserve inputs, actions, outputs, approvals, and failures?
How does the agent behave when a dependency is unavailable or the evidence is inconclusive?
Can its work be exported, reassigned, or completed manually?

Run the same task after changing a canonical input, revoking a permission, and withholding a required fact. You are looking for controlled behavior: the output should update when the approved context changes, access should disappear when permission is removed, and the agent should stop or escalate when it cannot support an answer.

Do not grant autonomous publishing or campaign-changing permissions merely to make a pilot look complete. An opaque error can create public misinformation, brand damage, or avoidable spend. Start with read access, draft outputs, explicit approval gates, and a visible audit trail. Expand authority only after the failure behavior is understood.

Turn the funding story into a procurement test

A cross-functional team evaluates an AI agent in a transparent test chamber using visual checkpoints for quality, security, time savings, and commercial value.

New capital can support product development, infrastructure, implementation, hiring, or market expansion, but the amount alone does not tell you which customer outcomes will improve. Ask Profound to connect its funded platform direction to the operating requirements in your evaluation.

Use a short, evidence-based process:

Separate product from roadmap. Mark every required capability as available, configurable, dependent on services, planned, or unsupported. Ask for written confirmation of anything that affects the purchase.
Select one costly workflow. Choose a process with a clear owner, recurring inputs, an observable outcome, and enough friction to justify change. Do not begin with a broad goal such as improving marketing productivity.
Run your material through the system. Use representative company context, normal approval requirements, and the systems the production workflow would need. A vendor-curated example cannot expose your integration or governance problems.
Measure total work. Compare active effort, waiting, handoffs, corrections, and review demand with the baseline. Count work displaced to administrators, analysts, agencies, or implementation teams.
Test failure and exit paths. Introduce stale context, a conflicting instruction, a denied permission, and an unavailable dependency. Then verify how you export outputs, retrieve records, remove data, and continue the workflow if the platform is unavailable.

A pass-or-fail scorecard keeps the evaluation focused when a demonstration is visually impressive:

Dimension	Evidence to request	Reason to pause
Workflow value	A proof run showing less total effort, delay, or rework	The claimed value depends mainly on future features
Context integrity	Source traceability, conflict handling, freshness controls, and scoped access	The system cannot explain which facts governed an output
Agent control	Least-privilege permissions, approvals, stop controls, and audit history	Agents require broad access or take opaque actions
Operational fit	Working integrations, clear ownership, administration, and support paths	Manual bridges recreate the work you intended to remove
Commercial durability	Written terms for current capabilities, service levels, support, and pricing	The funding total is used in place of contractual commitments
Exit safety	Documented export, deletion, access removal, and offboarding procedures	Your data or workflow history cannot leave cleanly

Funding matters most where it changes the risk of relying on the platform. Ask which capabilities exist now, which dependencies require professional services or third-party systems, what support is included, and how roadmap changes are communicated. For every answer, identify the proof: a live control, a technical document, a contractual term, or merely an intention.

Data handling deserves the same precision. Confirm what information the system stores, where it is processed, who can access it, how long it is retained, whether it is used to improve models, and how deletion is verified. If your marketing context contains customer, partner, employee, or confidential product information, involve the people responsible for security, privacy, and legal review before production access is granted.

Key takeaways

Profound’s $180 million raise supports its ability to pursue an AI platform for marketing, but it does not prove customer outcomes.
AI can create work after generation, especially in verification, approval, execution, governance, and measurement. Evaluate the whole workflow.
Company context must demonstrate source authority, freshness, traceability, conflict handling, and permission boundaries.
Agents must demonstrate limited authority, approval controls, predictable failure behavior, auditability, and a safe manual path.
Your decision should depend on production-like evidence and written commitments, not funding momentum or a curated demonstration.

For your next step, take one workflow into the evaluation meeting and bring its real inputs, permissions, exceptions, and approval rules. Ask Profound to show what AI Marketer does at each stage, what remains human work, and which capabilities are available now.

A platform is worth adopting when it reduces the total burden of producing a trustworthy marketing outcome while preserving control. The funding gives Profound room to pursue that standard. Your proof run should determine whether the product meets it for you.

References

Try Profound Blog – Building the AI platform for marketing

September 15, 2026

AI Marketing Agent Safety: A Practical Oversight Framework

Your marketing agent can draft a campaign, diagnose performance, or prepare a site update. The risk changes the moment it can spend money, suppress traffic, publish claims, email customers, or overwrite a working configuration.

You don’t need a binary verdict on whether the model is trustworthy. You need an operating system around it: complete enough context, narrowly scoped permissions, enforceable policies, approval before consequential actions, and a record that lets you reconstruct what happened.

Replace abstract trust with three control questions

The safer question is not whether you trust an AI model in the abstract. Ask what the agent can see, what it is structurally allowed to do, and who must approve its work before production. Those questions turn trust into controls you can inspect and test.

What can it see? List every account, dataset, field, date range, customer-data class, and external tool available to the agent. Record important gaps as carefully as available data.
What can it do? Separate reading, analysis, drafting, recommendation, and execution. A prompt describing what the agent should do is not a permission boundary.
Who signs off? Name the role that must approve each protected action. Reviewing a change log afterward is auditing, not approval.

Use those answers to assign every workflow an operating mode. Do not give an entire agent one blanket risk label; the same agent may be safe to query campaign data and unsafe to change a budget.

Operating mode	What the agent may do	Minimum control
Observe	Read approved data and explain findings	No production write credential; disclose data scope and gaps
Propose	Prepare copy, settings, or recommended changes	Policy validation; no direct route from proposal to production
Limited execution	Create drafts, apply labels, or act inside a designated sandbox	Named resources, hard action limits, result verification, and a tested recovery path
Protected execution	Change spend, bids, targeting, negative keywords, live content, customer communications, access, or destructive settings	Explicit approval for the exact change before execution

Reversible does not necessarily mean low risk. You can unpause a campaign, but you cannot recover traffic and opportunities lost while it was paused. You can restore a previous page version, but not necessarily retract a claim already seen by customers or answer engines. Classify risk by consequence and exposure, not merely by whether the interface has an Undo button.

Scope each permission across several dimensions:

Environment: sandbox, draft workspace, or production.
Identity: the brands, business units, clients, and accounts included.
Resource: campaigns, pages, audiences, feeds, schemas, or customer records.
Action: read, create, edit, publish, pause, archive, or delete.
Magnitude: the amount of spend, number of entities, or audience size the action can affect under your existing internal limits.
Time: when permission begins, when it expires, and whether approval can be reused.

The resulting permission register should be readable by marketing, security, and the workflow owner. If nobody can state an agent’s maximum possible action without opening its prompt, the boundary is not yet clear enough.

Ground the agent before you evaluate its reasoning

A fluent answer can still be built on an incomplete account view. The model may not know that a missing dataset contains the decisive explanation, so its tone will not reliably reveal the gap. Treat grounding as a safety control that reduces confidently wrong diagnoses, not as an optional convenience.

Write a grounding contract

A grounding contract defines the context a workflow requires before the agent may answer or act. It should record:

The systems, accounts, entities, fields, and historical periods the agent can access.
Excluded or inaccessible systems that could materially change the conclusion.
Data freshness, timezone, attribution settings, and the time of the last successful refresh.
The identifiers used to join advertising, analytics, CRM, commerce, and content data.
Which connectors are read-only and which can write.
What the workflow must do when a query fails, a join is ambiguous, or required context is stale.

For a Google Ads agent, a strong PPC grounding baseline extends well beyond a packaged performance summary:

Full Google Ads query access through GAQL for the resources, fields, segments, and metrics needed by the question.
GA4 data alongside ad data when the diagnosis depends on what happened after the click.
Complete change history across interface edits, scripts, agents, and other connected tools.
Negative keywords assembled across account-level negatives, shared lists, campaigns, and ad groups, including a deterministic check of whether a query is already blocked.
Auction Insights and an inspectable view of the keywords shared with a competitor when making competitive claims.
Relevant vertical benchmarks whose cohort and calculation are visible, rather than an unexplained generic average.

The same principle applies outside paid search. A content agent diagnosing lost visibility needs the relevant page versions, publication history, analytics context, and technical state. A schema agent needs the live markup and the page content it describes. A lead-nurture agent needs the current consent and suppression state available to the workflow. The exact systems differ; the requirement to expose material gaps does not.

Make missing context part of every answer

Require an input manifest with each recommendation. It should list the datasets queried, account and entity IDs, date ranges, filters, refresh times, failed queries, and inaccessible dependencies. When required context is absent, the agent should return an incomplete-data state instead of filling the gap with a causal story.

This also improves review. The approver can challenge the evidence itself instead of judging polished prose with no way to see what sits underneath it.

Enforce policy outside the model

An abstract AI core is surrounded by separate layers of permissions, rule gates, rate controls, and a locked execution chamber that block risky actions.

A system prompt can explain policy, but it should not be the component that enforces policy. Instructions can be misunderstood, displaced by conflicting context, or applied inconsistently. A control implemented in credentials, an action gateway, or workflow code can refuse an operation regardless of the text the model produces.

A practical enforcement path has four parts:

Separate agent identity. Give the agent its own credentials so its activity is distinguishable from a person’s work.
Least-privilege access. Where the platform supports granular scopes, issue only the read and write capabilities required for the approved workflow.
Action gateway. Route every proposed write through one controlled service rather than allowing the model to call production tools directly.
Workflow states. Move work through proposed, validated, approved, executed, and verified states. Do not let the model skip a state.

The policy layer should inspect the actual operation, not merely the agent’s description of it. Evaluate the destination account, object IDs, current values, proposed values, batch size, credential, policy version, and approval record before the write is sent.

Start with rules you can test

Deny production writes by default and allow only named actions on named resources.
Treat drafting and publishing as different permissions.
Protect changes to budgets, bidding, targeting, conversion definitions, negative keywords, customer-facing messages, user access, and billing behind the appropriate internal approver.
Set an internal maximum for entities affected in one execution. A request above that limit must be split or separately approved.
Block execution when required data is unavailable, stale under your policy, or inconsistent across systems.
Prefer drafts and archives to deletion. If deletion is required, identify what cannot be restored before approval.
Fail closed when the policy service or approval store is unavailable. An outage in the safety layer must not silently become permission to proceed.
Log blocked attempts and policy exceptions as well as successful actions.

Use your organization’s existing budget authority and publishing ownership to set thresholds. A generic dollar limit copied from another company cannot express your margins, account size, customer commitments, or tolerance for interruption.

Test the boundary, not just the happy path

Before granting production access, deliberately submit requests that should fail:

A valid action aimed at the wrong client or brand.
A batch larger than the configured action limit.
A protected change with no approval.
A request based on missing or stale required data.
A connected document containing instructions that conflict with the workflow policy.
A proposal altered after approval.
An execution in which the platform accepts some changes and rejects others.

For every test, verify the operation was blocked or contained, the event was recorded, and the right owner was notified. If success depends on the model deciding to behave, the test has exposed a prompt preference rather than a hard control.

Make human approval an exact, usable decision

A campaign operator reviews a website publication package, audience envelope, spending token, and rollback component before choosing between separate approval and rejection controls.

Human approval is valuable only when it happens before the consequential action and gives the reviewer enough evidence to make a decision. Grounding makes proposals more useful to review, while policy filtering removes obvious non-starters before they reach the queue. That combination keeps human attention focused on judgment rather than basic cleanup.

Build a proposal packet, not a chat transcript

Every approval request should contain:

The exact account, campaign, page, audience, feed, schema, or record affected.
A before-and-after representation of every proposed value.
The business reason for the change and the evidence used, with its date range and refresh time.
The expected effect, known uncertainty, and any plausible downside.
The policies evaluated, including passes, blocks, warnings, and requested exceptions.
The total number of entities and the maximum spend, reach, or publication surface exposed under the proposal.
The recovery procedure, including anything that cannot be reversed.
The person or role responsible for approval and the time at which that approval expires.

Show this information in the marketing system reviewers already understand when possible. A technically complete payload is not enough if the person accountable for the campaign cannot see the practical effect.

Bind approval to the exact proposal version, destination IDs, and values. If the agent edits the proposal, the underlying account state changes, or the approval expires, require validation and approval again. Never treat approval of an idea as standing permission for whatever implementation the agent later chooses.

Verify the write and prepare for partial failure

Recheck the destination, current state, data freshness, policy version, and approval immediately before execution.
Apply only the approved delta. Do not let execution broaden into related cleanup that was absent from the proposal.
Read the affected resources back from the platform and compare them with the approved values.
Record the request, approval, actor, platform response, successful entities, failed entities, and verification result.
If only part of a batch succeeds, stop the remaining work and send the exact partial state to the owner. Do not improvise a rollback whose consequences have not been reviewed.

A rollback plan should be tested against the real platform before you rely on it. Some operations can be restored from a known previous value; others create exposure that restoration cannot undo. Keep a kill switch that can revoke the agent’s write path independently of the model and document who is authorized to use it.

Monitor adoption, safety, and outcomes separately

A central view is useful because unregistered agents become invisible operational dependencies. At minimum, maintain an agent registry with the owner, purpose, connected systems, permissions, policy set, approver, current status, and kill-switch owner for each workflow.

Management dashboards can help expose usage patterns. For example, one vendor describes a command center that shows how teams use marketing agents, the hours their work returns, and adoption relative to peers. Those are adoption and capacity signals. They do not, by themselves, prove that the work was safe, accurate, or commercially valuable.

Organize oversight metrics into three lenses:

Adoption and capacity: active agents, active users, workflow frequency, proposals created, actions executed, and estimated hours returned. Document how any time-return estimate is calculated.
Safety and control: missing-context responses, policy blocks, exception requests, rejected proposals, stale approvals, out-of-scope attempts, partial executions, failed verification, rollbacks, incidents, and near misses.
Business outcomes: the marketing measures the workflow was intended to influence, alongside cost, error, complaint, and rework signals. Do not attribute an outcome to the agent merely because the two appeared in the same reporting period.

Configure immediate alerts for attempted protected actions, unavailable policy enforcement, writes to an unregistered destination, changes to agent credentials, partial execution, and failed post-write verification. A weekly dashboard cannot contain an agent that is actively writing to the wrong account.

During rollout, inspect every attempted production write and every policy block. Once the controls have behaved correctly under real workload, choose a recurring review cadence based on action frequency and consequence, while keeping event-driven alerts for protected operations.

Read metrics in context. Zero policy blocks can mean that workflows are well designed, that nobody is using them, or that enforcement is not recording failures. High approval rates can indicate good proposals or automatic rubber-stamping. Pair each number with sample-level review and an accountable owner.

Key takeaways

Trust is the result of inspectable controls, not a personality judgment about the model.
Give agents enough context to reason well, and force them to expose material gaps.
Enforce permissions and policies outside prompts.
Require approval before actions that can affect money, traffic, customers, access, or live content.
Bind approval to an exact, time-limited proposal and verify the resulting platform state.
Measure adoption, safety, and business outcomes as separate questions.

Start with the highest-consequence agent workflow you already use. Write its grounding contract, remove every unnecessary permission, and force its next production change through proposal, policy validation, exact approval, execution, and verification. Expand only one permission or action class at a time after that path works as designed.

References

September 15, 2026

Commercial AI Token Costs: Budgeting Beyond List Price

Your spreadsheet says one model is cheaper. Your invoice says otherwise. The gap appears because the spreadsheet priced the prompt and final answer, while production also paid for reasoning, repeated instructions, failed tool calls, retries, discarded drafts, and cache behavior.

If you are choosing a commercial AI model or defending an AI budget, compare cost per accepted outcome, not cost per million tokens. That change turns a rate card into a forecast you can actually use.

A token price is only one layer of your production cost

Published input and output prices tell you the rate applied to certain tokens. They do not tell you how many tokens the model will consume before your application gets an acceptable result. A useful cost model therefore has three layers:

Unit rates: the applicable prices for input, output, reasoning, cache reads, cache writes, and any long-context tier.
Consumption: the number of tokens used by the prompt, retrieved context, system instructions, tool definitions, intermediate reasoning, and response.
Completion efficiency: how many attempts, revisions, and tool calls you pay for before the result passes your acceptance checks.

The third layer causes many budget misses. A cheap attempt is not a cheap task if the attempt is rejected and repeated. Nor is a successful API response necessarily a completed business task. A coding agent that returns malformed code, a content model that produces an unusable draft, or a schema generator that fails validation has consumed tokens without delivering the outcome you intended to buy.

In measured 2026 production usage, the categories commonly omitted from simple estimates represented 52.5% of billed tokens and added 70.4% above a list-price-only estimate. These percentages are not universal overhead rates. They are a practical checklist of what your own logging needs to capture.

Cost commonly missed	Share of billed tokens	Added cost versus list-price estimate	What to inspect
Invisible reasoning tokens	22.4%	38.6%	Whether reasoning usage is returned separately from visible output
Re-sent system prompts and tool schemas	11.9%	9.4%	How much fixed context is transmitted on every model call
Retried and discarded generations	7.8%	8.1%	Every failed, rejected, or superseded attempt
Long-context pricing above 200K tokens	3.1%	6.2%	Requests crossing a provider’s long-context pricing boundary
Failed tool calls and malformed structured output	4.6%	5.3%	Calls that return successfully but fail downstream validation
Unrecovered cache-write premium	2.7%	2.8%	Cache entries written without enough subsequent reuse

Do not solve this by applying one generic markup to every vendor quote. Instrument each category instead. A reasoning-heavy model, a tool-using agent, and a short classification call can have radically different overhead even when their visible prompts look similar.

Falling rate-card prices do not remove this problem. Within a constant-capability mid-tier series from Q1 2023 through Q3 2026, the list-price index fell 91.4%, but real cost per completed task fell only 62.9%. Token consumption per completed task rose 4.3 times. The completed-task cost reached its low point in Q4 2024 and then increased 80% by Q3 2026 even as published rates generally continued downward. More capable reasoning behavior can consume part of the saving advertised on the price sheet.

Compare models by accepted task, not by token rate

Three abstract AI processing stations turn identical inputs into rejected fragments and one finished object that fits a quality-check fixture.

A model comparison becomes useful only after the denominator represents something your business accepts. From May 4 through August 21, 2026, a standardized set of 14 production tasks was run across 11 commercial models. The resulting cost included billed reasoning, prompt repetition, cache activity, retries, and discarded output. The September 2026 prices and measured completed-task costs show why rate-card ranking and production ranking can diverge.

Model	Input per 1M tokens	Output per 1M tokens	Measured cost per completed task
GPT-5.4 nano	$0.20	$1.25	$0.0219
Gemini 3.1 Flash-Lite	$0.25	$1.50	$0.0288
Claude Haiku 4.5	$1.00	$5.00	$0.0474
GPT-5.6 Luna	$1.00	$6.00	$0.0607
GPT-5.4 mini	$0.75	$4.50	$0.0627
Claude Sonnet 5	$2.00	$10.00	$0.0848
Gemini 3.6 Flash	$1.50	$7.50	$0.1040
GPT-5.6 Terra	$2.50	$15.00	$0.1662
Gemini 3.1 Pro	$2.00	$12.00	$0.1683
Claude Opus 5	$5.00	$25.00	$0.2131
GPT-5.6 Sol	$5.00	$30.00	$0.3447

Several reversals matter when you shortlist a model. GPT-5.4 mini had lower published input and output prices than Claude Haiku 4.5, yet its measured task cost was $0.0627 versus $0.0474. Claude Sonnet 5 had higher published rates than Gemini 3.6 Flash but completed the task set for $0.0848 instead of $0.1040. At the frontier end, GPT-5.6 Sol and Claude Opus 5 shared the same $5.00 input price, but Sol cost 62% more per completed task, with the difference driven almost entirely by output volume.

These results do not make one model universally cheaper. Your prompts, tools, input-to-output ratio, quality threshold, and retry policy may reverse the ranking again. Use published comparisons to choose candidates, then reproduce the comparison on your own workflow.

Define completion before testing. For JSON-LD, completion might require parsable output that passes your validation checks. For a content brief, it might require every mandatory field and entity. An HTTP success code is not an acceptance criterion.
Freeze a representative task set. Give every candidate the same source material, system instructions, tools, output requirements, and acceptance tests.
Record every billable attempt. Keep rejected generations, malformed output, repair prompts, tool-call failures, and fallback calls in the numerator.
Separate visible output from total usage. Store every usage field the provider exposes, including reasoning and cache categories where available.
Compare only models that meet the quality gate. A low-cost result that cannot be used is a failed attempt, not a bargain.
Divide total model spend by accepted completions. That figure is your effective task cost and the basis for a credible monthly forecast.

Content costs multiply after the first draft

Content teams often estimate AI spend from the tokens in one draft. That calculation stops before the expensive part: revisions, replacement drafts, citation repair, structural fixes, and output that never reaches publication.

For 1,000 words of finished, publishable copy, the measured token cost included revision rounds and discarded generations. The difference between first-draft and finished cost was substantial across every tested model.

Model	First-draft cost	Average revision rounds	Discarded drafts	Finished cost per 1,000 words	Finished versus first draft
GPT-5.6 Sol	$0.086	1.6	14%	$0.207	2.4x
Claude Opus 5	$0.079	1.2	9%	$0.164	2.1x
GPT-5.6 Terra	$0.043	1.8	17%	$0.114	2.7x
Gemini 3.1 Pro	$0.036	1.9	19%	$0.101	2.8x
Gemini 3.6 Flash	$0.024	2.4	26%	$0.084	3.5x
Claude Sonnet 5	$0.032	1.5	13%	$0.074	2.3x
GPT-5.6 Luna	$0.017	2.3	24%	$0.058	3.4x
GPT-5.4 mini	$0.013	2.9	31%	$0.054	4.2x
Claude Haiku 4.5	$0.016	2.1	22%	$0.051	3.2x
Gemini 3.1 Flash-Lite	$0.0041	4.1	45%	$0.024	5.9x
GPT-5.4 nano	$0.0034	4.4	48%	$0.021	6.2x

The cheapest and most expensive first drafts were separated by roughly 25 to 1. After revisions and discards, finished costs were separated by about 10 to 1. Draft rejection narrowed the apparent advantage of the cheapest models.

Discard rate was also more useful than list price for anticipating finished cost. Claude Sonnet 5 started at $0.032 per 1,000 words, above Gemini 3.6 Flash at $0.024. Sonnet finished lower, at $0.074 versus $0.084, because its discarded-draft rate was 13% rather than 26%.

Build that distinction into your content operations. Give every generated asset a final status such as accepted, revised, or discarded, and associate all attempts with the same job identifier. Then calculate finished token cost from all spend attached to accepted copy, divided by accepted word count and multiplied by 1,000. Counting only the last successful generation erases the waste you are trying to manage.

Keep the quality gate explicit. For an SEO or GEO workflow, your requirements may cover factual accuracy, source support, search intent, entity coverage, structure, brand constraints, and valid structured output. The exact rubric is yours, but it must be stable across models. Otherwise, a permissive review process can make a weak model look artificially inexpensive.

The figures above cover model-token spend. They do not represent a fully loaded content cost. Your internal budget should add editorial review, fact-checking, workflow infrastructure, monitoring, and any human repair work rather than treating a low token figure as the total cost of publication.

Budget by workload, then route each job to the right tier

Different task objects move through a central routing hub toward small, medium, and large processing machines, with one path passing through a cache chamber.

A single company-wide average hides the workflows most likely to break your budget. Agentic coding, customer support, retrieval-based research, document processing, sales personalization, and content production have different volumes, context sizes, output patterns, and failure modes.

For a modeled 50-person company, the same mix of 157,400 monthly tasks cost $6,610 at the economy tier, $19,150 at the mid tier, and $48,670 at the frontier tier. That is a 7.4-times spread before changing the workload itself.

Workload	Monthly tasks	Frontier tier	Mid tier	Economy tier
Coding agent, 20-developer team	14,800	$18,350	$7,140	$2,510
Customer support automation	62,000	$9,610	$3,720	$1,240
Internal RAG research tool	21,500	$7,290	$2,940	$1,020
Document and contract processing	9,700	$6,410	$2,580	$890
Sales outreach personalization	46,000	$4,830	$1,910	$640
Content marketing, 8-person team	3,400	$2,180	$860	$310
All workloads	157,400	$48,670	$19,150	$6,610

Volume alone does not reveal the expensive workflow. The coding agent ranked fourth by task count but was the largest monthly cost. At the frontier tier, it cost $1.24 per completed task, compared with $0.16 for customer support. Agentic workflows repeatedly call models, tools, and validation steps, so a task can contain much more billable activity than one support interaction.

Build your forecast from accepted workload volume

Your budget sheet should have one row per distinct workflow, not one row per provider. Separate content briefs from finished drafts, retrieval answers from document ingestion, and schema generation from schema repair. They may use the same API while having different cost behavior.

Workload identity: team, application, task type, model, and model version.
Demand: expected completed tasks, not merely API requests.
Usage: input, output, reasoning, cache-read, and cache-write tokens where exposed.
Workflow overhead: attempts, tool calls, validation failures, fallback calls, and discarded results.
Outcome: accepted, repaired, rejected, or abandoned.
Unit economics: total billed spend divided by accepted completions.

Forecast monthly model spend by multiplying expected accepted-task volume by your measured cost per accepted task. Keep the rate-card calculation beside it as a reconciliation check, not as the primary forecast. A widening gap between the two tells you to investigate prompt growth, longer retrieved context, increased reasoning, lower cache reuse, tool failures, or a rising retry rate.

Recalculate after changes to the model version, system prompt, tool schema, context strategy, output format, or acceptance threshold. Each can alter consumption or completion efficiency even when the published token rate stays fixed.

Use routing instead of choosing one model for everything

Model tier should be a workload decision. Economy models are strongest candidates when the task is constrained, output can be checked automatically, and failure is cheap to retry. Mid-tier models suit broader production work where reliability and cost both matter. Frontier models deserve the jobs whose ambiguity or quality requirement produces a measurable improvement worth their higher completed-task cost.

That does not require moving every workflow downmarket. In the modeled company, moving only the two highest-volume workloads – customer support and sales personalization – to economy models while leaving the other four at the frontier tier reduced total monthly spend by 26%. Selective routing captured savings without imposing one capability tier on every task.

Put a quality gate after the lower-cost route and send only failed or uncertain cases to a stronger model. Count both calls when escalation occurs. Otherwise, the first model appears cheaper in your dashboard while the fallback cost disappears into another service or team.

Key takeaways

Published cost per million tokens is a unit rate. Your actionable metric is total billed spend per accepted task.
Log reasoning, repeated system context, cache activity, retries, discarded output, tool failures, and long-context pricing instead of hiding them in a generic contingency.
For content, calculate cost per 1,000 accepted words from every draft and revision associated with the finished asset.
Benchmark candidates on the same tasks and acceptance criteria. Compare costs only among models that clear the required quality threshold.
Route by workload. High-volume, tightly validated tasks may justify an economy model, while ambiguous or high-impact work may justify a more capable tier.
Refresh the forecast whenever the model, prompt, tools, context, output contract, or quality gate changes.

Start with one workflow that already generates meaningful volume. Attach every billable attempt to an accepted or rejected outcome, calculate its effective cost, and use that result to challenge the rate-card estimate. Once the accounting works for one workflow, extend the same measurement to the rest of your AI stack and route each task on evidence rather than model reputation.

References

First Page Sage Blog – AI Token Usage Cost by Model: OpenAI vs. Anthropic vs. Google

September 9, 2026

How to Make Your Brand and Pricing Visible in AI Search

Your brand can appear in an AI answer and still lose the buyer. The assistant may recognize your name but misstate your category, omit your price, surface an expired offer, or recommend you to someone your product was never designed to serve. You get exposure, but the buying facts do not survive.

The practical goal is not to make every model repeat your messaging. It is to make the answers that influence discovery and evaluation accurate, specific, and verifiable. That requires a clear source of commercial truth, pricing content that can be interpreted without guesswork, matching structured data, and an audit process built around real buyer questions.

AI visibility must preserve the commercial decision

AI discovery compresses several stages of research into one response. A buyer can ask which products fit a use case, what they cost, how their plans differ, and which option has a particular constraint. If your brand is mentioned but the answer cannot resolve those questions, visibility has not yet become commercial visibility.

One vendor dataset is enough to justify taking this channel seriously, though not to forecast your own results. A Semrush study reported that more than a third of consumers start searching with AI and customers from AI search channels convert 4.4 times better than organic-search visitors. Treat that conversion figure as directional: channel definitions, attribution, audience, and purchase cycle can all affect the result.

The competitive field also appears unsettled. In a dataset covering 1,094 categories, only 15.2% had a clear owner. That indicates room for brands to establish category associations, not a guarantee that publishing more content will produce ownership.

Measure AI visibility against the questions a buyer needs answered:

Identity: Does the answer identify the correct company, product, and official website?
Category fit: Does it explain what you offer and which audience or use case it suits?
Commercial clarity: Does it state the price accurately or explain how the price is determined?
Qualification: Does it preserve material limits, required commitments, availability, and exclusions?
Verifiability: Can the buyer follow a citation to a page that supports the answer?

These are separate outcomes. A branded query may show that an assistant recognizes you, while a category query reveals that it does not associate you with the market you serve. A correct plan name does not prove that it understands the billing unit. A citation does not make an outdated price correct.

Pricing therefore deserves its own audit. The growing focus on what AI agents understand about pricing reflects an important distinction: recognizing a brand and understanding its commercial model are not the same task.

Build a canonical commercial truth layer

Your website needs an unambiguous source of record for every fact an assistant might use in a recommendation. Canonical does not mean putting everything on one enormous page. It means that each important question has an authoritative URL and that supporting pages do not contradict it.

Start by assigning an official page to each type of commercial fact:

Fact to establish	What the canonical page should resolve	Common failure to remove
Brand identity	Official name, website, product names, and the relationship between the company and its products	Old names, inconsistent capitalization, or several pages describing the same entity differently
Category and audience	What the offer is, who it is for, the problem it solves, and meaningful limits on fit	Brand slogans that never state the category in plain language
Offer structure	Plans, editions, services, add-ons, and how they relate to each other	Plan names without an explanation of what changes between them
Pricing mechanics	Currency, billing cadence, billing unit, included usage, additional fees, and overage treatment	A price displayed without enough context to interpret it
Qualification	Market availability, eligibility, minimum commitments, exclusions, and when a custom quote is required	Important conditions hidden in a tooltip, checkout flow, or sales conversation
Freshness	Whether the information is current and where changed or retired offers now live	Expired campaign pages and old documentation remaining discoverable

Write the central facts in visible HTML text. A calculator, toggle, configurator, or comparison widget can help a buyer, but it should not be the only place where the billing model is explained. If the critical answer appears only after a login or interaction, any system that cannot reach that state will have an incomplete record.

Use literal language before persuasive language. Your category statement should name the category, audience, and primary use case. Your pricing statement should connect the amount to its currency, unit, cadence, and conditions. Headlines such as “built to scale with you” can support positioning, but they cannot carry these facts.

Maintain a commercial-facts inventory alongside your content calendar. For each important claim, record its approved wording, canonical URL, content owner, structured-data location, last review, and every supporting page that repeats it. When a plan or policy changes, this inventory tells you what must be updated instead of leaving old claims scattered across the site.

A safe publishing sequence is:

Update the canonical product or pricing page.
Update the matching JSON-LD in the same release.
Revise comparison pages, FAQs, documentation, and relevant market-specific pages.
Replace, redirect, or clearly mark obsolete offer pages.
Check external profiles you control for conflicting descriptions or prices.
Retest the buyer questions affected by the change.

Make every pricing model answerable without inventing certainty

Price visibility does not require every company to publish a universal amount. It requires you to explain the commercial model as far as you truthfully can. The right treatment depends on whether your offer has public list pricing, negotiated pricing, or a mixture of fixed and variable charges.

Public list pricing

A bare amount is not a complete price fact. Write a sentence that remains accurate when removed from the surrounding design: “The [plan] costs [amount] in [currency] per [billing unit] when billed [cadence].” Then state the conditions that materially change what a buyer pays.

Name the billing unit, such as an account, user, location, project, transaction, or usage quantity.
Distinguish recurring charges from onboarding, implementation, service, or usage charges.
Explain what is included and how additional usage is handled.
State required commitments or minimum purchases where they apply.
Identify the market and currency when pricing differs by region.
Separate standard pricing from temporary promotions and eligibility-based discounts.
Place material conditions near the amount instead of relying on distant fine print.

If annual billing changes the effective rate, do not let a monthly-looking amount imply month-to-month availability. Connect the displayed amount to the actual cadence and commitment in the same sentence. If taxes or mandatory fees are excluded, say so where the price is presented.

Quote-based pricing

“Contact sales” is a conversion action, not a pricing explanation. If the final amount must be negotiated, publish the mechanics that determine it. This gives an assistant a truthful answer without forcing your team to disclose a range it cannot support.

State what is being priced: access, usage, seats, locations, services, outcomes, or a combination.
Name the variables that change the quote, such as scale, scope, support, integrations, service level, or contract structure.
Clarify whether implementation, migration, training, or support is priced separately.
Explain what information a buyer must provide to receive a quote.
Publish minimum commitments only when they are approved, current, and generally applicable.
Describe which offers require a custom agreement and which can be purchased directly.

Do not publish a speculative “typical” price merely to fill the gap. A false anchor can be repeated without the negotiation context that would have corrected it. If commercial or legal constraints prevent disclosure, be explicit about what remains variable and give the buyer a direct path to the current answer.

Hybrid and usage-based pricing

Hybrid offers are especially easy to misread because a real starting amount can coexist with required variable charges. Bind every “starts at” claim to the scope it actually covers.

Identify the base charge and what it includes.
Name the event that creates a variable charge.
Explain whether usage resets, rolls over, or is measured across a longer contract period.
Separate optional add-ons from charges required for the represented use case.
Show where a published tier ends and custom pricing begins.
Explain whether displayed examples are illustrative or purchasable configurations.

Do not use a low starting price as the headline if the represented customer cannot buy a functional version at that price without mandatory additions. The issue is not only conversion ethics. An assistant can detach the amount from its qualifier and present it as the price of the whole offer.

Use JSON-LD to confirm the visible truth, not replace it

Structured data is a clarification layer. It can name entities, connect products to offers, and make commercial fields easier to interpret. It cannot turn missing, inaccessible, or contradictory page copy into a reliable claim.

Model the smallest set of facts you can keep correct:

Give the organization or brand a stable @id, official name, canonical url, and carefully selected sameAs references.
Represent the actual subject of the page as a Product or Service when appropriate, and connect it to the organization that provides it.
Use an Offer only for a real offer. Its price, currency, availability, and URL must agree with visible content.
Use AggregateOffer only when the page presents a genuine range composed of real offers. Do not manufacture a range from unrelated packages.
Use pricing specifications only when they accurately express the billing unit, recurrence, or other commercial structure shown to the visitor.
For quote-based services, describe the service and quote path without encoding a placeholder as though it were a purchasable price.
Keep entity identifiers stable when URLs or templates change so that your own markup does not imply several disconnected brands or products.

Validate syntax and meaning separately. A parser can confirm that the JSON is well formed, but it cannot decide whether the amount is current or whether the offer actually includes what the page implies. Have a reviewer compare each commercial property with the visible sentence that supports it. If no sentence supports a property, either add the explanation or remove the property.

Make pricing content and pricing schema part of the same publishing event. Updating the page now and leaving the markup for a later ticket creates two versions of the truth. The same rule applies to currency, availability, plan names, and retired offers.

Structured data can reduce ambiguity, but it does not guarantee that an assistant will retrieve, cite, or repeat the page. Treat JSON-LD as useful redundancy inside a wider evidence system: clear visible copy, consistent owned pages, stable URLs, accurate external profiles, and independent corroboration where it naturally exists.

Audit AI answers as a buyer journey, then fix the costly gaps

A useful AI visibility audit starts with prompts, not brand mentions. Build a fixed set from the questions customers ask during discovery, evaluation, pricing, and comparison. Preserve the wording so that later tests remain comparable.

Your prompt set should cover:

Category discovery: “Which [category] options fit [audience and use case]?”
Constraint discovery: “Which [category] options support [required capability, market, or buying constraint]?”
Brand understanding: “What does [brand] offer, and who is it designed for?”
Price retrieval: “What does [brand or product] cost for [defined scenario]?”
Price mechanics: “Does [brand] charge by [possible unit], and what additional charges apply?”
Comparison: “Compare [brand] with [alternative] for [specific use case and constraint].”
Verification: “Where can I confirm [brand’s] current plans, pricing, or availability?”

Use the same scenario details that materially affect a real quote. A generic “What does it cost?” prompt may test brand recognition, but it cannot reveal whether the assistant understands seats, usage, locations, contract structure, or implementation charges.

Run the set across the assistants your audience uses, including ChatGPT, Claude, and Perplexity when they are relevant to your market. Record enough context to make the observation interpretable:

The exact prompt and scenario variables
The assistant, product surface, and model name when exposed
The market, language, signed-in state, and personalization conditions
The complete answer rather than a paraphrased note
Every cited URL and whether it supports the attached claim
Whether the brand is absent, merely mentioned, described, compared, or recommended
Whether each material price fact is correct, partial, wrong, or unverifiable
The canonical page that contains the approved answer

Do not collapse this into a single visibility percentage. An uncited but accurate mention, a cited false price, and a correct recommendation for the wrong audience create different problems. Classify the failure before choosing the fix.

Observed answer	Likely gap to investigate	Next action
Your brand is absent from non-branded category prompts	The category relationship may be weak, ambiguous, or poorly corroborated	Strengthen the canonical category statement, relevant use-case pages, internal links, and truthful third-party descriptions
Your brand appears but is assigned to the wrong audience	Positioning language is broad or inconsistent across pages	Name the intended audience, use cases, and exclusions in plain language on the canonical product page
The answer says pricing is unavailable	The price or pricing model may be hidden behind interaction, vague copy, or a sales form	Publish an accessible pricing summary or a concrete explanation of quote variables
The answer gives an old price or retired plan	Obsolete pages or conflicting structured data remain discoverable	Update the canonical page and schema, then replace, redirect, or mark outdated URLs
The amount is correct but the unit or commitment is wrong	The qualifier is separated from the amount or expressed only in interface controls	Put amount, currency, unit, cadence, and commitment in the same visible statement
The answer is accurate but cites another site	Your page may not provide a concise, stable, directly supporting passage	Add a clear answer on the canonical URL and make its evidence easy to verify
Different assistants produce conflicting answers	The evidence may be inconsistent, stale, unavailable to some systems, or interpreted differently	Trace each claim to its cited URL and repair the conflicting facts instead of assuming one universal cause

Prioritize by consequence. Correct false current prices, fabricated fees, wrong availability, and misleading commitments before pursuing more mentions. Then repair missing answers on high-intent pricing and comparison prompts. Category breadth and uncited awareness can follow once the buying facts are safe.

Keep evidence from each audit because generated answers can vary with product surface, context, and time. A saved answer, prompt, citation set, and test conditions let you distinguish a persistent information problem from an isolated response. Do not promise that a page edit will deterministically change every assistant; test again after the updated information has had a reasonable opportunity to become discoverable.

Key takeaways

Commercial AI visibility means that a buyer can identify your brand, understand its fit, interpret its pricing, and verify the answer.
Give every important brand and pricing fact a canonical URL, then remove contradictions from supporting pages and profiles.
If pricing is negotiated, publish the pricing model and quote variables instead of inventing a representative amount.
Make JSON-LD match visible content exactly; valid syntax does not rescue stale or misleading commercial data.
Measure real discovery and buying prompts, not mention volume alone.
Fix incorrect price, availability, and commitment claims before trying to expand category reach.

Start with the commercial question most likely to block your next buyer. Run it across the relevant assistants, capture exactly what is missing or wrong, and repair the canonical page that should own the answer. Once that answer is accurate and verifiable, move to the next decision in the journey. The first meaningful gain is not a larger mention count. It is fewer opportunities for an AI system to make your offer wrong, vague, or impossible to evaluate.

References

September 4, 2026

How to Build Trustworthy AI Agents for Marketing Operations

You have an agent that can inspect ad accounts overnight, draft a content brief before stand-up, or flag a broken funnel. The uncomfortable question arrives just after the demo: what, exactly, are you willing to let it do without asking?

If your answer is “we’ll review it,” you don’t yet have a control system. You have an intention. A trustworthy marketing agent needs a bounded job, owned data, explicit permissions, evidence attached to its conclusions, a release gate, and a way to stop or reverse its actions. Here is how to put that operating model in place.

A trustworthy agent is a controlled workflow, not a clever model

A model generates an answer. An agent combines a model with data, instructions, tools, scheduled triggers, and permission to take or prepare actions. That surrounding system determines whether a plausible mistake becomes a harmless draft, a misleading alert, or a customer-facing incident.

Trustworthiness therefore isn’t the promise that an agent will never be wrong. It is your ability to see what the agent observed, understand why it reached a conclusion, constrain what it can do, route uncertain cases to the right person, and recover when something fails. In production, reliability is decided by governance, realistic testing, and named review paths at least as much as by model capability.

The most useful mental model is a new employee with unusual speed. You wouldn’t give a new marketing analyst unrestricted CRM access, authority to change pricing, and permission to email customers on the first morning. You would define the role, grant only the access it needs, review early work, and expand responsibility after the work proves dependable. An AI agent needs the same management discipline, encoded in the workflow rather than left in a manager’s head.

Before deployment, make sure every agent has clear answers to these questions:

What specific decision or task does the agent own?
Which systems, records, fields, and time periods may it inspect?
Which facts and business rules must it know before making a judgment?
What evidence must accompany each conclusion or recommendation?
When must it abstain, escalate, or ask for missing information?
Who reviews consequential work, and what counts as approval?
Which actions can it take, and how can those actions be stopped or reversed?
Which version of the model, instructions, tools, and data definitions produced the result?

If any answer is “it depends,” write down what it depends on. That conditional logic is part of the product. It cannot remain tribal knowledge if the agent is expected to make repeatable decisions.

Begin with one bounded decision, not a general marketing assistant

“Monitor our marketing” sounds like a useful assignment, but it contains dozens of hidden jobs. Does monitoring mean detecting a tracking outage, explaining a CPA change, checking whether campaigns are serving, judging lead quality, finding off-brand copy, or recommending budget shifts? Each job needs different data, context, freshness rules, and escalation paths.

Start with a task whose input and acceptable output can be described precisely. Read-only analysis is usually the safest entry point because the agent can create value without changing the underlying system. Examples include investigating an ad-delivery alert, identifying content briefs with missing source material, finding inconsistent campaign naming, or preparing a proposed JSON-LD correction for validation and human review.

Write a short job card for the workflow:

Trigger: State what starts the run, such as a scheduled account check or an anomaly from an existing monitoring rule.
Question: Express the decision in one sentence. For example: “Has campaign delivery stopped during comparable business hours?”
Inputs: Name the approved systems, fields, reporting windows, business rules, and account notes.
Output: Define the required finding, supporting evidence, uncertainty, and proposed next step.
Prohibited behavior: State what the agent must not infer, retrieve, publish, send, or change.
Escalation: List the conditions that require abstention or human judgment.
Reviewer: Assign a role or person responsible for accepting consequential recommendations.
Success and failure: Describe both a useful result and an unsafe result. A fluent explanation without adequate evidence belongs in the failure column.

Pay special attention to time. Marketing data often arrives on different schedules, so “recent” does not necessarily mean “complete.” A production ad-management agent once interpreted conversions that had not arrived yet as a severe performance decline. Making its analysis dependable required safe comparison windows, conversion-maturity rules, uncertainty ranges, and refusal when the lag could not be modeled reliably.

Apply that lesson beyond paid media. A CRM agent should not label a campaign unproductive before the normal sales cycle has elapsed. A content agent should not declare a page unsuccessful before the chosen reporting period is complete. An SEO agent should not turn a partial crawl or delayed analytics import into a confident diagnosis. Freshness and maturity are different properties, and the agent needs rules for both.

Refusal is not a defect when the evidence is immature, contradictory, or missing. A trustworthy response may be: “I cannot distinguish a real decline from reporting delay with the approved data.” That is more useful than an elaborate guess because it tells the operator what information is needed next.

Give the agent a data contract and a business context pack

Connecting an agent to more systems does not automatically make it better informed. It can instead create several conflicting versions of revenue, conversion, customer status, or campaign ownership. The agent will still produce coherent prose even when the underlying records disagree.

A data contract tells the agent what it may use and how each input should be interpreted. Create one before refining the prompt. For every permitted input, record:

The system and field that hold the data.
The business owner responsible for its meaning and quality.
Whether it is the authoritative value or a convenience copy.
How frequently it updates and when it becomes mature enough for judgment.
The unit, attribution rule, time zone, status definition, and other interpretation rules.
Known gaps, exclusions, and failure signals.
What the agent must do when the input is absent, stale, or inconsistent.
Whether the field contains personal, confidential, regulated, or otherwise restricted information.

Then create a separate context pack for facts that do not live cleanly in reporting tables. Include the products the business actually sells, excluded services, target locations, budget constraints, active promotions, sales-cycle expectations, conversion-lag patterns, campaign goals, approved claims, brand restrictions, and known tracking limitations. Without this context, an agent can correctly calculate the numbers and still reach the wrong business conclusion. A paid-media agent, for example, cannot identify an irrelevant pet-insurance keyword for a business-insurance advertiser unless it knows what the business sells and can access the operational context used by human analysts.

Keep the context pack owned and maintainable. Each rule should have an owner, a status, and a replacement path when the business changes. Otherwise an old promotion, discontinued service, or superseded approval rule can remain active inside the agent long after people have moved on.

Use least-privilege access. If the task requires campaign totals, do not expose raw customer records. If the agent only prepares a content update, give it draft access rather than publishing rights. If it reads a CRM status, restrict it to the approved fields rather than the full contact object. Governed implementations can limit access to approved data, mask immature conversion information, and require evidence for recommendations.

Trace where the data goes as well as what the agent can retrieve. Before customer, prospect, health, or financial information reaches a third-party AI service, determine where it is processed, what the provider may retain or reuse, and which internal policy governs that transfer. Marketing data deserves the same boundary-setting applied to other sensitive operational systems; convenient access is not the same as necessary access.

If the team cannot identify the owner or meaning of an important field, stop at read-only experimentation. A better prompt cannot resolve a disputed definition of revenue, repair missing conversion data, or decide which system is authoritative.

Set autonomy by consequence and reversibility

An AI device faces three increasingly restricted action zones, from reversible draft tasks to guarded campaign controls and a locked high-consequence mechanism.

Teams often treat autonomy as a switch: either the agent acts or a person does. A safer design separates observation, recommendation, preparation, and execution. The agent can then earn broader permissions without receiving blanket authority.

Operating level	Marketing example	Default permission	Release condition
Observe	Check reporting data and surface a possible anomaly	Read approved fields; create an internal record	Freshness checks pass and evidence is attached
Recommend	Explain a performance change or propose a content correction	No external change	Assumptions, uncertainty, affected assets, and reviewer are explicit
Prepare	Build a draft ad, email, brief, metadata edit, or schema patch	Write only to a draft or sandbox	Validation passes and a named person approves publication
Act	Pause a campaign, move budget, publish content, change pricing, or send a message	Off by default	The action is narrowly pre-approved, policy-compliant, observable, and safely reversible; otherwise human approval remains mandatory

Two variables should control the level: consequence and reversibility. A duplicate internal alert is annoying but recoverable. An incorrect customer email, pricing change, destructive CRM update, or large budget movement can create brand, financial, privacy, or legal exposure. Work carrying that weight needs a human checkpoint; letting an unreviewed agent send customer communications or make consequential commercial decisions is not an acceptable starting posture.

For high-impact recommendations, add an independent check before the decision reaches the approver. That check should evaluate the evidence and policy conditions, not merely ask another model whether the prose sounds convincing. It can verify that the reporting window is mature, the cited records exist, the requested action is permitted, and contradictory data has been surfaced. Higher-stakes analysis benefits from a separate review path before a person is asked to act.

Require an evidence packet for every recommendation. It should contain:

The conclusion in plain language.
The period, comparison, account, page, campaign, or record under review.
The approved inputs actually used.
Missing, stale, masked, or contradictory inputs.
Assumptions and relevant business rules.
The agent’s uncertainty or reason for abstaining.
The proposed action and assets it would affect.
The required approval and available rollback path.

Do not allow the agent to hide uncertainty inside polished prose. Evidence must be inspectable by the person making the decision. If a recommendation cannot be traced back to permitted inputs, it should fail the release gate regardless of how reasonable it sounds.

Release, monitor, and stop the agent like production software

Human operators monitor an AI agent moving from testing through a gated deployment lane, with health sensors, an evidence trail, an emergency stop, and a rollback track.

Test safe behavior, not just good answers

A handful of impressive demo prompts proves very little. Build an evaluation set from the situations the agent will face after release: routine work, different ways users phrase the same request, incomplete data, delayed conversions, stale account notes, conflicting systems, out-of-scope requests, and cases where the correct response is escalation.

For each case, define the expected behavior rather than one perfect paragraph. Should the agent answer, flag uncertainty, request information, refuse, or escalate? Which evidence must appear? Which tools may it call? Which actions must remain blocked? This makes the evaluation durable even when wording varies.

Add simple pass-or-fail checks around important invariants:

A read-only agent cannot invoke a write operation.
A draft-only content agent cannot publish.
Restricted fields never appear in retrieved context or output.
A performance judgment cannot use a reporting window marked immature.
A recommendation cannot pass without evidence identifiers and required assumptions.
A missing authoritative input triggers the prescribed abstention or escalation.
An action outside the job card is rejected even when a user asks persuasively.

Run the agent in shadow mode before granting action rights. Let it inspect real work and produce results without changing external systems. Compare its findings with the decisions made through the existing process, examine both disagreements and omissions, and update the job card, data contract, context pack, and evaluation set. Only then consider expanding its operating level.

Version every component that can change behavior

The prompt is not the whole agent. Store the system instructions, policy rules, model identifier, provider settings, tool definitions, data-field mappings, business definitions, context-pack version, evaluation results, approval decision, and release date as one traceable configuration.

This matters because behavior can drift even when your team changes nothing visible. A provider can update the model underneath a workflow, while a data field, tool response, or business rule can change independently. Unversioned models and prompts make it difficult to explain why customer-facing behavior changed or recreate how the system acted earlier. Marketing teams need release discipline and behavior monitoring around model and prompt changes, just as they do around application changes.

Rerun the relevant evaluations whenever any behavioral component changes. If the provider does not expose a fixed model version, record the identifier it does provide and use recurring evaluation results to detect observed changes. Do not assume unchanged prompts guarantee unchanged behavior.

Monitor usefulness, silence, and operator burden

Accuracy on answered cases is not enough. Monitor unsupported conclusions, inappropriate certainty, policy violations, reviewer overrides, action reversals, duplicate alerts, unnecessary escalations, and cases where a reviewer had to retrieve evidence the agent should have supplied.

Review non-alerts as well as alerts. An agent can look quiet because nothing is wrong, because its thresholds are sensible, or because it missed the problem. Sample runs where it concluded that no action was needed and verify that the underlying data supports that silence.

Noise is an operational failure. If people repeatedly dismiss duplicate, untimely, or low-value alerts, they will stop treating the agent as a useful colleague. A working ad-management agent had to remove duplicate notifications and messages that could wait because convincing the team to pay attention depended on reducing noise as well as improving analysis.

Give operators a visible stop path. When the agent behaves unexpectedly, they should be able to pause scheduled runs, revoke write credentials, preserve the decision trace, identify the changed component, rerun evaluations, and restore a known configuration. Re-enable a smaller scope before returning full permissions.

Rollback has limits. You can restore a campaign setting or draft, but you cannot make a sent email unread or erase a public impression of an incorrect claim. Keep human approval in front of actions whose consequences cannot be meaningfully reversed.

Key takeaways

Trust is a property of the whole workflow: data, context, permissions, evidence, review, monitoring, and recovery.
Start with a bounded, read-only decision whose correct inputs and safe output can be described precisely.
Treat data freshness, data maturity, and business meaning as separate requirements.
Grant the minimum fields and tools needed for the job; broad access is not a substitute for context.
Increase autonomy according to consequence and reversibility, not model fluency.
Make abstention, escalation, and evidence-bearing recommendations part of the success criteria.
Version every component that can change behavior, then retest and monitor real-world use.

Pick the smallest marketing decision that currently consumes repeated human attention. Write its job card and data contract before connecting an agent. If you cannot define the evidence, permissions, reviewer, and stop path, the workflow is not ready for autonomy. If you can, you have a foundation that can earn broader responsibility instead of merely requesting trust.

References

August 31, 2026