Category: Advertising

AI-Powered Ads on Google and Microsoft: A Control Plan

If you run paid campaigns on Google and Microsoft, the important question is no longer whether AI will touch your advertising. It already influences ad creation, query interpretation, bidding, product discovery, campaign operations, and measurement. Your real decision is which tasks to delegate and which decisions must remain under human control.

That distinction matters because the two platforms are automating different parts of the job. Google is moving ads deeper into conversational search, discovery, and commerce. Microsoft is reducing the friction of importing, bidding, and reporting across accounts. You need a control plan that reflects those differences, not one generic “AI advertising” switch.

Decide what AI may decide before you activate it

A campaign manager controls a transparent gate separating automated advertising tasks from protected human decisions, with several signals paused for review.

AI-powered advertising is not a single feature. It is a stack of decisions. An AI system can generate an asset, select an audience, adjust a bid, explain a product, recommend an account change, or predict a future outcome. Those actions do not carry the same risk.

Google’s stack now reaches from Conversational Discovery ads, Highlighted Answers, Shopping explainers, and lead-generation agents to creative production and predictive measurement. Microsoft’s stack emphasizes cross-platform imports, portfolio bidding, attribution, and more configurable reporting. Before adopting any of it, assign a human owner to the decision the system is helping make.

AI layer	Platform examples	What you should control
Customer interaction	Conversational Discovery ads, Highlighted Answers, Shopping explainers, and Business Agent for Leads	Permitted claims, qualification rules, escalation paths, and the point at which a person takes over
Creative production	Text, image, and video generation in Asset Studio	Approved facts, brand rules, legal review, asset rights, and final publication approval
Media delivery	Demand Gen optimization and Microsoft cross-account portfolio bidding	Business objective, budget boundaries, conversion values, exclusions, and stop conditions
Campaign operations	Ask Advisor and Microsoft Import Center	Which recommendations become changes, who approves them, and how changes are recorded
Measurement	Meridian, Qualified Future Conversions, data-driven attribution, and bid-strategy reporting	The definition of success, the quality of conversion data, and whether a result is predictive, attributed, or incremental

This separation prevents a common mistake: allowing the platform to define the goal while it also optimizes toward that goal. Automation can pursue an objective efficiently, but it cannot decide whether the objective represents profitable growth, a useful lead, or merely an easy conversion.

Write down the decision rights for every campaign before changing its automation. At minimum, answer these questions:

Which conversion should influence bidding, and which events are diagnostic only?
What business value is attached to each conversion?
Which claims, audiences, locations, products, or queries are outside the campaign’s scope?
Can AI-generated assets publish automatically, or must a named person approve them?
Which performance change would trigger investigation, a rollback, or a pause?

If those answers are missing, the campaign is not ready for more autonomy. The problem is governance, not a lack of AI features.

Fix the input layer before generating ads or answers

Generative systems multiply whatever you give them. Clean facts become more usable assets. Contradictory facts become more contradictory assets, produced at greater speed.

This is especially important in conversational advertising. Google’s Business Agent for Leads is designed to answer questions using information from the advertiser’s website. Its Shopping formats can add AI-generated explanations of why a product may fit a shopper’s needs. Merchant Center is also gaining Conversational Attributes and AI Performance Insights for shopping experiences across Search, Gemini, and AI Mode.

Your website and product feed are therefore operational inputs, not just destinations after the click. If a landing page, product description, promotion, and campaign brief disagree, the model cannot know which version your business intends to honor.

Prepare a compact campaign truth set before opening a generative tool:

Offer facts: the exact product or service, included features, exclusions, availability, eligibility, price conditions, and promotion terms.
Approved claims: statements the campaign may make, the evidence behind them, and wording that requires legal or compliance review.
Audience intent: the problem being solved, the questions a qualified buyer asks, and the signals that indicate poor fit.
Brand rules: tone, visual constraints, prohibited themes, required terminology, and examples of acceptable assets.
Product data: consistent titles, descriptions, attributes, images, categories, destinations, and offer details in Merchant Center.
Conversion rules: the event that counts, its value, the validation process, and the lag between an ad interaction and a confirmed business outcome.

Google’s upgraded Asset Studio is designed to interpret marketing briefs, brand guidelines, website content, and campaign goals when generating text, images, videos, and creative themes. That can remove production bottlenecks, but only if those materials are current and internally consistent.

Use generated creative as a controlled variation, not as automatically approved truth. Check every asset against the offer facts and claims list. Keep the prompt, input materials, output, reviewer, and final disposition together so that you can explain why an asset ran.

For teams working on SEO, AEO, GEO, and advertising together, align visible page copy, product-feed information, and structured data. JSON-LD cannot repair an inaccurate feed or a vague landing page, and the available platform announcements do not establish schema markup as a direct bidding signal. Its practical role here is consistency: machines and people should encounter the same entity, offer, availability, and business facts wherever those facts appear.

This becomes more consequential as commerce moves closer to the generated answer. Google has described AI-assisted checkout, Universal Cart, cross-retailer shopping, and buy-now-pay-later integrations, while its Direct Offers pilot includes AI-generated bundles and native checkout for Universal Commerce Protocol merchants. When discovery and transaction happen within the same assisted journey, inaccurate product data has fewer opportunities to be corrected later.

Give Google and Microsoft different operating roles

A split illustration contrasts an exploratory product discovery environment with a structured campaign operations room connected by a bridge.

Running both platforms does not mean cloning one campaign and calling the job complete. Their AI capabilities solve different problems, and your testing plan should reflect that.

Google is pushing further into the interaction itself. Gemini can interpret a conversational query, assemble an explanation, place a relevant offer within an AI-generated response, or support a lead conversation. Demand Gen can distribute creative and product experiences across YouTube, Discover, Maps, and Shopping. Its expanded tools include creator partnership videos, Merchant Center product videos, Maps inventory, and AI-assisted campaign setup.

Use Google when you want to test how creative, product data, and assisted discovery work together. The useful question is not merely whether a new format gets more clicks. Ask whether it helps the right user understand the offer, advances that user to a valuable action, and produces a business outcome that survives validation.

Availability should shape your plan. Conversational Discovery ads and Highlighted Answers were announced as U.S. tests on mobile and desktop. AI-powered Shopping ads and Business Agent for Leads were described for U.S. open beta, while many Demand Gen additions were expanding through open beta globally. Treat tests, pilots, and betas as learning opportunities, not guaranteed inventory in a forecast.

Microsoft is concentrating more heavily on operational leverage. Its Import Center can search and filter imports from Google Ads and Meta Ads, pause or edit imported campaigns, surface troubleshooting help, and provide recommendations after import. Cross-account portfolio bidding extends automated strategies across Search and Shopping accounts, while new reporting fields make bid targets easier to inspect.

Use Microsoft to reduce duplicated setup and coordinate related accounts, but do not confuse a successful import with an equivalent campaign. An imported structure can be technically valid while optimizing toward the wrong conversion or carrying assumptions that do not fit its new environment.

Audit every import before it spends:

Confirm campaign status, budgets, bidding strategy, and portfolio membership.
Map conversion goals and values to the business outcome you intend to optimize.
Review location, audience, product, and inventory scope.
Test landing-page URLs and tracking parameters.
Recheck negative constraints, brand exclusions, and any setting that limits where an ad can appear.
Record differences between the originating campaign and the imported version.

Cross-account portfolio bidding is most defensible when the participating accounts share compatible goals and value definitions. Pooling signals from unrelated outcomes can make the algorithm look busy without making the portfolio economically coherent.

The same discipline applies to Google’s Ask Advisor, which connects Ads, Analytics, Merchant Center, and the Google Marketing Platform to help build campaigns, analyze performance, recommend changes, and automate operational tasks. A recommendation should enter your normal approval process. The fact that an assistant can execute a task faster does not change who is accountable for the result.

Measure decisions, not just automated output

AI advertising creates more observable activity: more assets, more variations, more bid adjustments, more recommendations, and more predictions. Activity is not evidence of incremental value.

Build measurement at three levels:

Control quality: Did the system stay inside the approved offer, brand, audience, and budget boundaries?
Platform performance: What happened to conversions, conversion value, cost per acquisition, return on ad spend, impression share, and other campaign metrics?
Business impact: Did leads qualify, transactions hold, revenue materialize, and the campaign add outcomes that would not otherwise have occurred?

Microsoft’s reporting expansion helps with the middle layer. Advertisers can inspect average Target ROAS, average Target CPA, average Target impression share, conversion metrics in custom columns, and reports segmented by goal name. Data-driven attribution is also available for automated strategies including Maximize Conversions, Maximize Conversion Value, and Enhanced CPC.

Those fields can show how the platform allocated credit and pursued a target. They do not, by themselves, prove that advertising caused the reported outcome. Attribution distributes credit among observed interactions. Incrementality asks what changed because the campaign ran.

Google is adding tools for that broader question. Demand Gen includes Uplift Experiments and Campaign Type Attribution. Meridian, Google’s open-source marketing mix model, is being integrated into Analytics 360 to combine first-party and cross-channel data, estimate incremental performance, forecast outcomes, and support media-mix decisions.

Qualified Future Conversions add another type of evidence. The Gemini-powered metric links current advertising activity with possible future sales signals, including branded search behavior. It was announced as a restricted global pilot, with wider beta access anticipated later. A predictive future-conversion signal is useful for planning, but it is not realized revenue and should not be booked or reported as though it were.

Use a measurement ladder that matches the maturity of the campaign:

Define the validated business conversion and its value before changing bidding.
Verify that Google and Microsoft receive comparable, correctly classified conversion signals.
Inspect performance by goal so that a rise in easy secondary actions cannot hide a decline in valuable outcomes.
Compare generated assets with your established creative process using the same campaign objective and review rules.
Use controlled uplift testing where it is available to investigate causal impact.
Use marketing mix modeling for cross-channel allocation questions that campaign attribution cannot answer alone.
Treat predictive metrics as planning inputs until the predicted behavior becomes an observed business result.

Do not optimize a campaign against a forecast and then cite the same forecast as proof that the optimization worked. Separate the signal used to make a decision from the evidence used to evaluate that decision.

Key takeaways

AI-powered advertising is a stack of creative, interaction, delivery, operational, and measurement decisions. Assign human ownership at each layer.
Google’s strongest shift is toward conversational discovery, generated product explanations, integrated commerce, and creative distribution across its properties.
Microsoft’s strongest shift is toward easier cross-platform imports, coordinated portfolio bidding, attribution, and more transparent reporting.
Your website, product feed, campaign brief, brand rules, and conversion definitions must agree before you let generative systems use them.
An imported campaign needs a full settings and measurement audit; technical compatibility does not guarantee strategic equivalence.
Attributed conversions, incremental outcomes, and predicted future conversions answer different questions. Do not report them as interchangeable results.

Your next move can be deliberately small. Choose a campaign with a clear conversion, document its approved facts and decision boundaries, and activate only the AI capability whose output you can inspect. Once the measurement holds, expand the system. If the measurement does not hold, more automation will only make the uncertainty harder to unwind.

References

May 20, 2026

How to Evaluate AI-Powered Advertising Platforms

You’re probably not deciding whether AI belongs in advertising. You’re deciding how much of your budget, product catalogue and campaign analysis you can safely hand to it.

The useful question is not, “How advanced is this platform?” It is, “Which decision will this platform improve, what data will it use, and what can it change without approval?” Answer those three points before you compare features.

Key takeaways for your platform decision

Separate AI that explains performance from AI that creates or delivers ads. The second category carries more financial and brand risk.
Treat your product feed, conversion events and campaign rules as operating inputs, not setup details. Automation scales their errors as readily as their strengths.
Use prompt-driven dashboards to shorten investigation time, but verify filters, totals and metric definitions before changing spend.
Test one bounded workflow at a time. Define its inventory, budget, approval rights, primary outcome and stop condition before launch.
Judge the platform on business outcomes and control, not on how quickly it produces an ad, chart or answer.

Separate decision support from automated execution

“AI-powered advertising” describes several different jobs. Combining them into one category makes platform evaluations fuzzy and permissions unnecessarily broad.

AI role	What you provide	What it produces	Main risk to check
Reporting and interpretation	Account data, a question and reporting filters	A chart, table, breakdown or explanation	A plausible answer built on the wrong scope, filter or metric
Ad assembly	Product data, images, attributes and eligibility rules	Ads assembled from approved inputs	Incorrect or unsuitable catalogue data appearing at scale
Delivery and optimization	A budget, objective, conversion signal and constraints	Bids, placements or allocation decisions	Spend being optimized toward a weak or misconfigured signal

Google Ads’ Gemini-powered dashboards sit primarily in the first row. Advertisers can use prompts to customize views, while the dashboard presents performance through charts, graphs and tables that update with the query. That can reduce the work required to reach a useful breakdown, but it does not give the dashboard permission to define your business objective.

ChatGPT’s product-feed advertising moves further into execution. Retailers can connect catalogue data so the system can assemble sponsored product ads from names, images and other attributes. Retailers can also set rules governing which products may be featured. Here, data quality and eligibility rules directly affect what a prospective buyer can see.

Before granting access, write down four permission levels: read, recommend, create and spend. A reporting assistant may need only read access. A product-ad system needs approved data plus creation rules. A bidding system needs a tightly defined budget and a trustworthy conversion signal. Do not grant all four levels merely because one integration supports them.

This distinction also clarifies ownership. Your analyst can own reporting questions. Merchandising should own product eligibility. Marketing and finance should agree on spend limits. Whoever owns the business outcome should approve the conversion definition. “The AI team owns it” is not an operating model.

Audit the data contract before evaluating the AI

Two analysts inspect customer, product and campaign data moving through a transparent pipeline with permission, quality and verification controls.

An automated platform can only act on the facts and signals it receives. If a product is misidentified, an image is stale or a conversion fires at the wrong moment, faster automation creates a faster version of the wrong campaign.

For a feed-based commerce channel, inspect the feed as a contract between your catalogue and the advertising system. Review it in the same form the platform will receive it, not only as it appears in your storefront.

Confirm item identity. Each product and variant should be distinguishable. If two records appear identical to a machine but represent different options, ad assembly can select the wrong one.
Check customer-facing facts. Review names, images and every connected attribute for accuracy. Compare the resulting destination page with the feed record so the promise in the ad matches the page.
Define eligibility explicitly. Create rules for products that may be advertised and exclusions for products that should not be. Do not rely on someone remembering to remove an unsuitable item manually.
Assign update ownership. Name the system or person responsible for correcting catalogue facts. A feed without a clear owner becomes stale infrastructure.
Design failure handling. Decide whether questionable or incomplete records are excluded, held for review or corrected upstream. Silent substitution is a poor default when brand or pricing information is involved.
Keep an audit trail. Record which feed version, rules and approvals were active when an ad ran. Without that record, you cannot separate a platform problem from an input problem.

This matters beyond paid placement. ChatGPT’s model allows product information to support both answers and advertising, connecting organic product discovery with a paid campaign workflow. The operational lesson is larger than one channel: machine-readable product facts are becoming shared discovery infrastructure.

Your product feed and on-page structured data should therefore agree, but do not treat them as interchangeable. A channel feed supplies data to a specific system. JSON-LD describes information on a page in a machine-readable form. Keep names, product identity, images and other shared facts consistent across both, while using the integration method the advertising platform actually documents. Do not assume that publishing schema automatically enrols a product in an ad programme.

For non-commerce campaigns, the equivalent data contract is your measurement setup. Identify the event that represents the business result, the events that are merely steps toward it and the system responsible for recording each one. If the platform sees a click but not the qualified action that follows, it may become efficient at producing visits without becoming effective at producing customers.

Use conversational dashboards as an investigation layer

Prompt-driven reporting changes how you reach a view, not what makes that view trustworthy. A natural-language interface can remove report-building friction, but the underlying questions still need a metric, dimension, scope and comparison.

The Gemini-powered Google Ads dashboard is designed to show metrics including impressions, clicks, video views and costs across devices, audiences and campaign types. Those combinations are useful because they let you move from “performance changed” to “where did it change?”

Use prompts that describe a reporting operation. The following are question shapes to adapt, not guaranteed platform commands:

Show impressions, clicks and cost by device for the selected campaign type.
Break down video views and cost by audience, using the same campaign scope.
Compare clicks and cost across campaign types, then isolate the segment responsible for the largest difference.
Keep the same metrics and change only the device breakdown so the two views remain comparable.

The discipline is in changing one analytical dimension at a time. If you alter the metric, campaign scope and audience definition in the same prompt, you may get an attractive chart without knowing which change produced the result.

Build a short verification routine around every consequential finding:

Read back the date range, campaign scope, filters and dimensions shown in the resulting view.
Check the displayed total against the corresponding native account report before moving budget.
Confirm that compared views use the same definitions and aggregation.
Save the prompt or question alongside the resulting filters. Natural-language wording is part of the analysis and should be reproducible.
Translate the observation into a testable hypothesis. “Mobile cost increased” is an observation; it is not yet an instruction to reduce mobile spend.

Prompted reporting is most valuable when it shortens the path from a broad symptom to a precise segment. It is less useful when it becomes a substitute for measurement definitions or causal testing.

Access and exact behaviour also need verification. The dashboard rollout was introduced with further details still expected at Google Marketing Live. Check what is available in your own account before retiring a custom report or external analytics workflow on the assumption that every required capability has arrived.

Run a bounded pilot before expanding authority

A campaign manager monitors a small AI advertising pilot enclosed by a transparent boundary, with human controls separating it from a larger campaign network.

A good pilot answers a decision, not merely whether the software works. “The platform generated ads” proves that the integration ran. It does not prove that the ads reached appropriate buyers, produced incremental value or justified broader automation.

Name one workflow. Test prompt-driven account diagnosis, feed-based ad assembly or automated delivery separately. Combining them makes failures hard to locate.
Write the decision statement. Specify what you will expand, change or stop if the test succeeds or fails.
Capture the existing process. Record its inputs, human effort, approval path and outcome metrics. Otherwise, “faster” and “better” have no comparison point.
Limit exposure. Use a defined campaign or approved product subset, a controlled budget and explicit permissions. Automated advertising can spend real money or expose incorrect catalogue information, so set pause conditions before activation rather than during an incident.
Lock the measurement contract. Choose one primary business outcome and document the conversion event, reporting source and attribution configuration used to evaluate it. Keep clicks, impressions, views and cost as diagnostic metrics rather than automatically treating them as success.
Log human intervention. Record feed corrections, prompt revisions, exclusions, bid changes and manual pauses. A result that depends on constant rescue is not evidence of autonomous performance.
Decide explicitly. Scale, revise, hold or stop. Do not let a pilot become permanent simply because nobody scheduled the decision.

Match the test to capabilities that exist, not capabilities on a roadmap. ChatGPT’s advertising direction includes cost-per-click bidding and conversion tracking, while cost-per-action models were reported as still in development. A future buying model should not be included in the business case for a current pilot.

Ask vendors and internal owners the same practical questions before you approve expansion:

Which source fields and conversion signals drive the system’s decisions?
Can you exclude products, audiences or campaign types without rebuilding the workflow?
Which actions require human approval, and which occur automatically?
Can you export the underlying data and reproduce a reported result outside the conversational interface?
How are sponsored placements distinguished from organic recommendations? In ChatGPT’s current product-ad format, the units appear beneath responses and remain labelled as sponsored.
What happens when feed data, conversion tracking or an integration becomes incomplete?
Can you pause execution without losing the configuration and evidence needed for review?

Your next move should be narrow. If you manage a catalogue, audit one approved feed segment and its page-level structured data. If you manage campaigns, choose one recurring reporting question and test whether a prompted dashboard answers it accurately and reproducibly. Write the outcome, permissions and stop condition first. Broader authority should follow evidence, not the ease of the interface.

References

May 12, 2026

How to Build the Data Foundation for AI-Powered Ads

You’ve connected your ad accounts to an AI system, and it can see every impression, click, conversion and campaign change. That may look like a strong data foundation. It isn’t. The system still can’t tell whether a lead became a customer, whether an order was profitable or whether operations can fulfill the demand it creates.

Before you let AI move budget or restructure campaigns, you need a business outcome layer between the advertising platforms and the agent. Build that layer well, and automation can pursue results your company actually values. Skip it, and the agent will optimize the numbers it can see – even when those numbers point away from profit.

Give the AI an optimization contract before giving it data

An ad platform knows what happened inside its own boundary. It can report delivery, interactions and the conversions attributed to its ads. It usually doesn’t know the quality of a sales lead, the margin on a product, the value of a renewed account or the amount of work your team can fulfill. An agent using only those platform signals operates inside a closed optimization loop.

More integrations won’t fix that problem until you define what the agent is supposed to optimize. Write an optimization contract that answers six questions:

What is the business outcome? Name the final result, such as closed-won revenue, a completed order or contribution margin. Don’t use a platform conversion label as the definition.
Which outcomes are eligible? State whether cancellations, invalid leads, duplicate orders, returning customers or other disqualified records should count.
How is an outcome valued? Identify the field that carries realized revenue, margin or an approved stage value. Document its currency and whether the value is gross, net or estimated.
When is the result mature enough to use? A form submission arrives quickly; a qualified opportunity or completed sale may arrive later. Define the lifecycle point at which the business accepts the result.
What constraints outrank performance? Inventory, sales capacity, service availability, geographic coverage and fulfillment limits can all make additional conversions undesirable.
What may the AI change? Separate analysis, recommendations and account changes. Specify allowed actions, approval requirements, financial limits and rollback conditions.

This contract prevents a proxy from quietly becoming the objective. In lead generation, a form submission is an early signal, not proof of revenue. Map the progression from submission to qualification, opportunity and closed business. If only the submission reaches the ad platform, call it a proxy in reporting and keep the later CRM result on the business scorecard.

For ecommerce, order revenue is still incomplete when products have different margins or fulfillment constraints. A campaign can improve reported return on ad spend by selling more of a low-margin product or promoting something the business cannot readily fulfill. That is why CRM outcomes, product economics and operational signals belong in the decision model.

Do not ask the model to invent missing business values. If sales has not agreed on what a qualified opportunity is, or finance cannot identify the value field to use, the agent should expose the gap rather than manufacture a score. In that state, it can still draft creative, summarize performance and recommend investigations. It is not ready to control spend autonomously.

Build a business outcome layer across five data domains

Five symbolic data domains for customers, advertising, sales, transactions, and operations connect to one central business outcome hub.

A useful advertising data model keeps different kinds of evidence separate. Platform delivery data, customer outcomes and operational constraints answer different questions. Flattening them into a single conversion column destroys the distinctions the agent needs.

Data domain	What it tells the AI	Records and fields to connect	How it should affect decisions
Advertising platforms	What was delivered and what the platform attributed	Campaign, ad, creative, audience, click, conversion, timestamp and platform-reported value	Diagnose delivery and compare tactics inside the platform
Web or app analytics	What happened during observable visits	Session, landing page, traffic source, on-site events and consent state	Explain journeys and identify experience or measurement problems
CRM or order system	What became a valid lead, customer, order or realized revenue	Lead, customer or order ID; lifecycle status; outcome value; new or returning status; cancellation or invalidation state	Anchor business reporting and train toward genuine downstream outcomes
Product economics	Which sales create business value	Product or SKU, margin measure and the date for which that value applies	Prefer valuable demand rather than revenue alone
Operations	What the business can sell and fulfill	Availability, capacity, service area and fulfillment constraint	Suppress or limit spend when additional demand would create an operational problem

Competitive intelligence can sit beside these five domains, but it should not become the outcome label. Adthena says its ChatGPT advertising product monitors more than 300,000 daily prompts to surface brands, placements, messages and share of voice. That kind of market visibility can help you form targeting and creative hypotheses. It cannot tell you whether your own acquired customer was profitable or incremental.

The next job is making the records joinable. Your data contract should specify:

A stable lead, customer or order identifier in the business system.
Platform click, campaign, ad and creative identifiers where collection and use are permitted.
Separate timestamps for the interaction, conversion, lifecycle update and data ingestion.
A controlled vocabulary for statuses such as qualified, won, cancelled and invalid.
The owner, currency, unit and calculation method for every monetary field.
The system that originated each field and the last time it was refreshed.
Identity-matching rules, including what the pipeline does when it cannot safely match a person or order.
Retention, access and consent rules appropriate to the data you are permitted to use.

Those details are not housekeeping. They determine whether the same customer becomes one outcome or several apparent outcomes, whether last month’s campaign receives credit for this month’s sale and whether a stale margin value drives a current budget decision.

Time deserves special treatment because the systems do not necessarily place the same conversion in the same period. Ad platforms may credit a conversion to the day of the ad interaction, while analytics and CRM reporting commonly place it on the day the conversion occurred. This difference in attribution dates can make two accurate reports disagree at a daily or monthly boundary. Preserve both the event date and the platform credit date instead of overwriting one with the other.

Build the pipeline from the business result backward. First identify the accepted outcome in the CRM or order system. Then attach identity and campaign metadata, enrich the outcome with product and operational values, and only then send an approved signal back to the ad platform through offline conversion tracking or a direct connection. Keep the unmodified business record as well. You will need it when you reconcile totals or change the value logic later.

Reconcile the systems without forcing their numbers to match

Google Ads, Meta Ads, analytics and a CRM can all be working as designed while showing different conversion totals. They observe different parts of the journey, use different attribution rules and handle identity, privacy gaps and modeled conversions differently. Treating disagreement as proof that one tool is broken sends teams into endless tracking rebuilds.

Consider a buyer who clicks a Meta ad, encounters YouTube retargeting, searches for the brand and then buys within a week. Meta and Google may each report a conversion because neither platform has the complete cross-platform path. Analytics and the CRM may record one sale and credit the final paid-search visit. The platform conversions are not two additional customers; they are different claims on the same customer journey.

Your reporting model should therefore preserve three views:

Business outcomes: valid customers, orders, deals and revenue recorded by the CRM, commerce platform or finance system.
Attributed outcomes: conversions and value claimed by each advertising platform under its own rules.
Journey evidence: observable sessions, touchpoints and on-site behavior captured by analytics.

Never add attributed outcomes across platforms and present the sum as company revenue. Use the business system to answer how much happened. Use platform and analytics data to explain which interactions were observed and where performance changed.

A practical reconciliation process looks like this:

Choose the CRM, order system or finance record that defines the total business outcome. Document why it is authoritative and which statuses it includes.
Align time zones, currencies, conversion definitions and reporting dates before comparing systems.
Break the comparison down by outcome type, campaign group, new versus returning customer and lifecycle stage where those fields are available.
Compare platform-attributed results with business outcomes, but do not demand equality. Record the ratio between them for each stable reporting segment.
Investigate abrupt ratio changes. A jump can indicate a tagging failure, a changed attribution setting, a new sales lag, missing offline imports or a real shift in the customer journey.
Annotate known changes to schemas, consent behavior, campaigns and operational availability so the AI does not interpret a measurement change as a performance change.

Ratios are especially useful because the normal gap between systems can be more informative than an impossible attempt at perfect agreement. If a platform usually reports more attributed orders than the order system and that relationship remains stable, you have a usable baseline. If the relationship suddenly changes, investigate before the agent moves budget.

Attribution still cannot answer the causal question: would the customer have converted without the ad? Attribution allocates credit after a conversion exists. Incrementality estimates the conversions that would not have happened without the campaign. Keep those jobs separate in your data model.

When the budget and data volume can support a meaningful control group, you can test incrementality through geographic holdouts, audience holdouts or carefully designed pauses. Time-based pauses are vulnerable to seasonality and other concurrent changes, while any test with an indistinct control group can produce an inconclusive result. These methods are different from attribution reporting; do not let an agent treat an attributed conversion as proof of incremental impact.

The decision hierarchy is simple: business records tell you how much happened, attribution tools describe the credit assigned to observed interactions, and controlled experiments provide evidence about what caused additional outcomes. Your AI should preserve that hierarchy rather than collapse it into one synthetic score.

Expand the agent’s permissions only after the data proves reliable

A glowing AI core passes through sequential security gates as validated data signals unlock access to advertising controls.

Generating headlines or summarizing a dashboard is not the same as running an advertising account. A true agent can adjust budgets, bids, targeting or campaign structure. That power also accelerates mistakes when business data is missing or misaligned. Because those actions spend real money, enforce limits in the surrounding system rather than relying on a prompt to remember them.

Stage 1: Observe in read-only mode

Let the agent read platform, CRM, product and operational data without changing an account. Run this stage through a period long enough to include the normal delay between an ad interaction and the business outcome you care about.

Review whether it joins the correct records, respects lifecycle updates and explains discrepancies without summing incompatible numbers. Every conclusion should identify the metric definition, originating system and data timestamp it used. If the agent cannot show that lineage, you cannot reliably audit its reasoning.

Stage 2: Produce structured recommendations

Require each recommendation to contain the proposed action, business objective, evidence, applicable constraint, estimated exposure and rollback condition. A person should approve the action while you compare recommendations with actual downstream outcomes.

This stage exposes a common failure early: the model may recommend scaling a campaign because platform return improved even though CRM quality, product margin or capacity deteriorated. Rejecting that proposal is not a prompt-tuning exercise. It means the optimization contract, data mapping or decision rule still needs work.

Stage 3: Allow bounded execution

Once recommendations are consistently traceable to accepted business outcomes, allow only a narrow set of reversible actions. Put the following controls outside the model:

An allowlist of accounts, campaigns and action types the agent may touch.
Per-action and cumulative financial limits over a defined period.
A freshness gate that blocks changes when CRM, margin or operational data is late.
A completeness gate that blocks optimization when essential outcome fields are missing.
A cooldown that prevents repeated changes before delayed results can arrive.
A before-and-after audit record containing the input data version, decision, approver and resulting account state.
A rollback procedure and kill switch that do not depend on the agent remaining available.

Fail closed when the business context disappears. If the inventory feed stops updating, the CRM import fails or a margin table changes schema, the safe response is to pause autonomous changes and alert an operator. Continuing with platform-only data recreates the closed loop you built the foundation to avoid.

Keep experimentation separate from routine optimization as well. Mark campaigns, regions or audiences participating in a holdout so the agent cannot erase the control group in pursuit of short-term attributed performance. An autonomous optimizer should execute the experiment design, not silently rewrite it.

Key takeaways: your AI advertising readiness check

Your foundation is ready for controlled automation when you can answer yes to every item below:

The optimization objective maps to an accepted CRM, order or finance outcome rather than a platform conversion label alone.
Early proxies such as clicks, form submissions and attributed conversions are clearly distinguished from realized business results.
Outcome values have documented owners, currencies, units, calculation methods and validity dates.
Campaign, customer and order records can be joined without counting one business outcome as several customers.
Interaction, conversion, attribution and ingestion timestamps remain separate.
Product margin and operational constraints reach the decision layer before the agent allocates budget.
CRM totals, analytics journeys and platform attribution remain separate views, with normal discrepancies monitored rather than erased.
Incrementality evidence is labeled separately from attribution evidence.
Missing or stale business data automatically blocks account changes.
Every permitted action has an enforced limit, audit trail, rollback path and independent kill switch.

If any essential item fails, keep the system in read-only or recommendation mode. That is still useful automation. It becomes unsafe automation only when the authority to spend grows faster than the quality of the data underneath it.

Start with one campaign group and one downstream outcome that sales, finance or commerce operations already recognizes. Connect that result, reconcile it against platform reporting and let the AI recommend changes before it executes them. Expand to more campaigns and wider permissions only after the outcome remains traceable from ad interaction to business record.

References

May 11, 2026

AdSense Vignette Ads No Longer Trigger on Browser Back

Your AdSense implementation can be working correctly even when vignette impressions or revenue suddenly move. Google AdSense no longer uses the browser Back button as a vignette ad trigger, so a change in this format does not automatically point to broken code, a consent failure, or a traffic problem.

The practical question is narrower: how much of your vignette inventory depended on that navigation action, and are the remaining ad opportunities behaving normally? Answer that before you change placements, edit templates, or disable the format.

Key takeaways

The browser Back button no longer triggers an AdSense vignette ad. That does not mean the entire vignette format has been removed.
Treat an isolated decline in vignette impressions as a possible inventory change before treating it as an implementation failure.
Compare vignette impressions and revenue per session, not only revenue per pageview. A removed back-navigation opportunity may not correspond to a new pageview on your site.
Segment the change by browser, device, landing-page template, and traffic source. Sites with frequent land-and-return behavior may be more exposed.
Do not recreate the removed behavior by intercepting the browser Back button or trapping visitors. Improve useful internal navigation and evaluate the rest of your ad mix instead.

The change applies to a specific navigation action

Vignette ads are interstitial-style placements associated with navigation between pages. The important boundary here is the browser control itself: when a visitor presses Back in Chrome, Safari, Firefox, or another browser, that action is no longer a vignette trigger.

Do not translate that into the broader claim that vignette ads have stopped working. The change removes one trigger, not the format as a whole. It also does not establish that every link labeled Back will behave the same way. An on-page “Back to results” link is a site link, while the browser Back button operates through the visitor’s navigation history. Test those paths separately rather than grouping them by their visible label.

The behavior change alone is not evidence that you need to reinstall the AdSense tag, modify structured data, change a WordPress theme, or repair an SEO problem. Check those systems only if other evidence points to them. A decline across every ad format, for example, deserves a broader serving and traffic audit. A decline isolated to vignettes has a much narrower set of likely causes.

Why the revenue effect will vary between publishers

Three smartphones show different browsing paths, including frequent backtracking, mostly forward navigation, and a short exit route, with varying numbers of translucent ad panels.

Removing a trigger reduces the number of moments at which a vignette could be considered. It does not tell you how large the effect will be. That depends on how visitors move through your site.

A site can be more exposed when many visitors land on a page, consume what they need, and use the browser Back button to return to a search result, social feed, referring site, or previous page. A site with deeper internal journeys may rely less on that action. These are diagnostic hypotheses, not reasons to assume a loss before looking at your own data.

Page RPM can be a misleading first metric in this case. A vignette associated with an exit through browser history may have created an ad impression without creating another publisher pageview. If that opportunity disappears, pageviews can remain stable while vignette impressions and revenue fall. Revenue per session and vignette impressions per session provide a cleaner view of that mechanism.

Use these questions to determine whether the navigation change is a credible explanation:

Did vignette impressions per session fall while display and other ad formats stayed near their previous patterns?
Did the movement concentrate on landing pages that commonly end a visit?
Was it larger for search, social, or referral landings than for direct visitors who browse several internal pages?
Did one device or browser segment move more than the others?
Did sessions, pageviews, geography, consent rates, or the mix of page templates change at the same time?

The first four patterns make the removed trigger more plausible. A simultaneous change in traffic, consent, templates, or all ad formats means you have competing explanations and should not attribute the result to vignette behavior alone.

Audit the change without confusing correlation for cause

An analyst compares separate navigation, advertising, consent, traffic, and timing indicators across a laptop and smartphone using a central magnifying glass.

A useful audit separates format behavior from traffic behavior. You do not need a complicated attribution model, but you do need a comparison that preserves context.

Record possible confounders. Note any changes to consent management, AdSense settings, theme files, navigation, ad experiments, traffic acquisition, or page templates. If several things changed together, do not assign the full effect to one of them.
Find the first sustained movement in your own reporting. Compare equivalent periods on either side of that movement. Match the day-of-week mix and avoid using an unusually large campaign, outage, or seasonal spike as the baseline.
Isolate vignettes where your reporting permits it. Review vignette impressions and revenue separately from total advertising revenue. If you cannot separate the format, state that limitation instead of treating a sitewide result as proof.
Normalize for audience volume. Calculate vignette impressions per session and vignette revenue per session. Keep page RPM as supporting context, not the only decision metric.
Segment the affected traffic. Start with browser, device, traffic source, landing-page type, and new versus returning visitors. Stop adding segments when sample sizes become too thin to show a stable pattern.
Inspect navigation paths. Compare sessions that end on the landing page with sessions that continue through internal links. If available, examine flows from high-traffic landing pages to categories, related content, product pages, or site search.
Change one thing at a time. If you decide to adjust navigation or another placement, keep consent, templates, and other ad settings stable during the evaluation. Otherwise, the next comparison will be as ambiguous as the first.

A quick diagnosis matrix

What you observe	Most useful interpretation	What to do next
Vignette impressions per session decline while other ad formats remain stable	The removed trigger is a plausible cause	Monitor the new baseline before changing the implementation
All ad formats decline together	A broader traffic, consent, serving, or implementation issue is more likely	Audit sitewide changes and ad delivery
The decline is concentrated on high-exit landing pages	Visitor navigation patterns may explain the exposure	Review those pages’ internal paths and format-level metrics
Sessions or pageviews change materially at the same time	Raw revenue comparisons are confounded by audience volume or behavior	Normalize per session and compare stable traffic segments
Revenue changes but format-level impressions are unavailable	Causality remains uncertain	Avoid implementation changes based on the sitewide total alone

Respond by improving the journey, not recreating the trigger

If the audit shows a modest, isolated vignette decline and everything else is stable, the most defensible response may be to accept the new baseline. Fewer interruptions during browser Back navigation can change the balance between monetization and visitor control. There is no technical virtue in forcing the old interaction back into the experience.

If the effect is material, work on the parts of the journey you control:

Add a genuinely useful next step near the point where a reader has finished the current task, such as a related explanation, comparison, category page, or product detail.
Make internal links descriptive enough that visitors know what they will get before clicking.
Check whether intrusive elements, weak mobile navigation, slow pages, or dead-end templates are pushing visitors toward the browser Back button.
Evaluate other appropriate ad placements as part of the complete page experience, using both revenue per session and engagement signals.
Run controlled layout tests rather than changing navigation, ad density, consent behavior, and templates in the same release.

Do not hijack browser history, open unnecessary pages, or manufacture clicks to replace a lost ad opportunity. Those tactics work against visitor intent and make analytics harder to trust. The sustainable lever is a better internal path that a reader chooses because the next page is useful.

Set a new baseline before making an optimization decision

Your next action is simple: chart vignette impressions per session, vignette revenue per session, sessions, and total pageviews across the same comparison window. Then split the result by landing-page type and traffic source. If only vignette efficiency moved while other formats and traffic stayed stable, document the trigger change and establish a new baseline. If the decline reaches multiple formats or coincides with a site change, continue the broader audit before touching your ad strategy.

References

Search Engine Land — Google AdSense removes browser back button trigger for vignette ads

May 7, 2026

ChatGPT Ads Manager: A Practical Launch Plan for Marketers

You are probably not asking whether advertising in ChatGPT sounds interesting. You are asking whether it deserves a line in your media plan, which bidding model fits your goal, and how to test it without creating an expensive attribution problem.

The sensible answer is a bounded pilot. OpenAI’s self-serve ChatGPT Ads Manager removes the former $50,000 minimum for U.S. advertisers and adds CPC bidding alongside CPM. That lowers the barrier to testing, but it does not remove the need for a clear objective, validated measurement, and a hard spending limit.

The platform change is access, not proof of performance

Removing a minimum spend changes who can run an experiment. It does not tell you whether ChatGPT ads will work for your audience, what a conversion will cost, or how the channel should fit alongside search, social, display, and earned AI visibility.

Start by treating self-service access as permission to investigate, not as a reason to move budget immediately. The stated scope is U.S. advertisers. Do not assume that the same access, placements, policies, controls, or reporting apply in another country or account.

Before approving spend, open the account and answer these questions from the terms and controls actually shown to you:

Is your advertiser, billing entity, product category, and target geography eligible?
Where can the ad appear, how is it labeled, and can you preview its presentation?
What does the platform count as an impression and a click?
Which targeting, exclusion, frequency, placement, and brand-safety controls are available?
Which creative formats and landing-page destinations are accepted?
What conversion tracking, attribution windows, exports, or integrations can you use?
Which campaign, bid, budget, and account-level spending limits can you enforce?
How are invalid interactions, refunds, taxes, data use, and ad review handled?

These are verification questions, not assumptions about the product. Save the definitions and settings you use in the campaign brief. If an impression, click, or attribution rule changes later, you will need that record to interpret the trend correctly.

Choose CPC or CPM from the business objective

A marketer considers two paths, one showing individual interactions with blank cards and the other showing many viewed cards across an audience.

CPC and CPM do not merely offer two ways to pay the same bill. They place the immediate economic risk in different places.

Bid model	You pay for	Best starting objective	Main measurement trap
CPM	Impression delivery, priced per thousand impressions	Controlled exposure or message reach	Treating a served impression as attention, interest, or demand
CPC	Recorded clicks	Sending people to a page where a meaningful action can occur	Treating a click as a qualified visit, lead, sale, or customer

Choose CPM when exposure is the actual job. That may fit a campaign intended to introduce a category, establish a message, or reach an audience before a later action. You still need a way to judge whether exposure created useful movement. An impression count alone proves delivery, not attention or business impact.

Choose CPC when the landing page can carry the next part of the journey and you can measure what happens after the click. CPC transfers some delivery risk away from you because impressions without recorded clicks do not create click charges. It does not protect you from irrelevant clicks, weak landing pages, poor qualification, or broken conversion tracking.

Compare the models through a common business outcome rather than comparing their headline prices. Calculate effective CPC as spend divided by clicks, effective CPM as spend divided by impressions multiplied by 1,000, and cost per acquisition as spend divided by attributed acquisitions. Use the platform’s precise definitions for every input.

If your finance-approved allowable cost per acquisition is known and your landing-page conversion rate is reliable, a simple ceiling for CPC is:

Maximum CPC = allowable cost per acquisition x expected click-to-acquisition conversion rate.

This is a planning ceiling, not a bid recommendation. The conversion rate must come from a comparable audience and journey. If it comes from branded search, returning customers, or a different offer, it may overstate what unfamiliar ChatGPT traffic can support. If you have no reliable rate, describe the campaign honestly as a traffic-quality experiment rather than a test of profitable acquisition.

Build a pilot that can answer one decision

A marketer observes a blank advertising card moving through a small testing chamber bounded by a budget rail and a sealed container of tokens.

A useful pilot does not need to answer whether the entire platform works. It needs to answer one decision your team will make next: continue, stop, change the offer, change the audience hypothesis, or repair measurement before spending more.

Write one hypothesis. Use this form: For this audience and context, this message will produce this business action within our allowable outcome cost.
Select one primary business event. A qualified lead, completed purchase, activated account, or another value-bearing event is more useful than a page view. Define exactly when the event counts.
Validate the full measurement path before launch. Follow a test visit from the ad destination through the primary event, analytics, CRM or commerce system, and revenue record where applicable.
Match the advertisement to the landing page. Keep the promise, terminology, product scope, and expected next step consistent. A click bought with one promise and handed to a different page cannot diagnose channel quality cleanly.
Limit simultaneous variables. If you change the audience, bid model, message, offer, and page at once, a good or bad result will not tell you which change mattered.
Set financial guardrails. Record the total cap, any daily control available, the person allowed to approve an increase, and the condition that pauses spending. Paid experiments can consume budget before a delayed conversion report catches up, so the cap must exist before launch.
Write the decision rule in advance. State which primary metric, cost boundary, data-quality checks, and minimum evidence your team requires before it will scale, revise, or stop.

Do not use a cheap click as the decision rule unless a cheap click is genuinely the business outcome. Rank the metrics so that the platform metric remains subordinate to the business metric: delivery supports clicks, clicks support qualified actions, and qualified actions support revenue or another defined result.

Run an A/B test only when the campaign can produce enough observations for a defensible comparison. If volume is too low, do not declare a winner from a handful of outcomes. Treat the result as directional, retain the uncertainty, and use it to design the next test rather than to justify a broad rollout.

Keep paid performance separate from AI visibility

ChatGPT advertising and visibility inside unpaid AI answers belong in the same executive conversation, but not in the same measurement bucket. Paying for distribution does not, by itself, demonstrate that your brand will be mentioned, recommended, or cited in an unpaid response.

Maintain three distinct layers in your reporting:

Paid delivery: spend, impressions, clicks, effective CPC or CPM, and other delivery measures the account exposes.
On-site response: engaged visits, qualified events, conversion rate, cost per acquisition, revenue, and downstream lead quality where those measures apply.
Earned AI visibility: unpaid brand mentions, citations, answer inclusion, referral visits, and conversions from AI discovery measured through a consistent monitoring method.

Use consistent campaign parameters and retain platform, campaign, creative, and destination identifiers wherever the system supports them. Keep paid ChatGPT traffic out of organic AI referral reporting. Otherwise, an increase purchased through ads can be mistaken for progress in generative engine optimization.

Measure earned visibility with a stable prompt set, documented locale and account conditions, and timestamps. AI responses can vary, so a single favorable answer is not a trend. Compare repeated observations under the same method and label the result as monitored visibility, not guaranteed ranking.

The same separation applies to technical optimization. Clear entity information, useful content, and accurate structured data may support machine understanding, but JSON-LD is not an ad setting and does not guarantee an AI citation. Likewise, ad spend is not a substitute for the content and authority work required to earn unpaid visibility.

Automate reporting before you automate campaign control

Four OpenAI Ads nodes for Profound Agents can bring advertising data into agentic workflows. That creates useful options for recurring analysis, but the existence of four nodes does not tell you which data each one reads, which actions it can write, or which permissions it requires. Inspect those details before connecting a live account.

A safe first workflow should do the following:

Begin with read-only access if that permission is available.
Pull a defined account, campaign scope, date range, timezone, currency, and attribution setting.
Check for missing records, delayed conversions, duplicate rows, and inconsistent campaign identifiers before calculating performance.
Calculate derived metrics from the raw values and retain those values beside every conclusion.
Flag a breached budget, tracking anomaly, or performance threshold for review rather than silently changing the campaign.
Require human approval before an agent changes a bid, budget, audience, destination, creative, campaign status, or account permission.
Log the input data, generated recommendation, approver, resulting action, and rollback path.

If a connected node can write changes, give it the narrowest permission that supports the approved workflow. An agent asked to maximize click-through rate can rationally chase more clicks even when those clicks do not become customers. Every optimization instruction therefore needs a business constraint, a spending limit, and a metric that represents value after the click.

An automated report should also expose its boundaries. Include the reporting window, currency, attribution rule, conversion lag, excluded campaigns, missing fields, and the raw numerator and denominator behind each rate. A fluent narrative without those details is presentation, not a reliable decision system.

Key takeaways

Self-serve access and removal of the former $50,000 minimum make a smaller U.S. advertiser pilot feasible; they do not establish likely performance.
Use CPM when controlled exposure is the objective and CPC when a measurable post-click journey is the objective.
Judge both models against the same business outcome, not against impressions or clicks in isolation.
Launch one hypothesis with validated tracking, a hard spending cap, a pause condition, and a decision rule written before the first charge.
Report paid ChatGPT results separately from unpaid AI mentions, citations, referrals, and other GEO or AEO indicators.
Use agentic integrations for scoped data collection and anomaly detection first; keep spend-changing actions behind explicit human approval.

Your next step is a one-page test brief. Fill in the eligible account and geography, objective, bid basis, audience hypothesis, landing-page event, allowable outcome cost, attribution rule, budget cap, pause condition, and final decision rule. If any field is blank, the campaign is not ready to buy useful learning.

Once every field is defined, launch the smallest controlled test capable of answering the decision. At the first review, expand only when the business result and data quality support the rule you set in advance. Otherwise, repair the measurement, revise one variable, or stop.

References

Try Profound Blog — Unlocking ChatGPT Ads: A Seamless Integration for Marketers

May 6, 2026

ChatGPT Self-Serve Ads: A Practical Launch Framework

If you have been waiting for a practical way to test ChatGPT advertising without entering a large, managed pilot, self-serve buying changes the conversation. The important question is no longer whether the channel sounds interesting. It is whether you can run a controlled test without mistaking novelty, clicks, or platform-reported conversions for profitable growth.

You need a defined conversion, a defensible cost ceiling, a landing page that matches the ad, and tracking that reaches your order system or CRM. Put those pieces in place before you request access or allocate budget, and ChatGPT ads can be evaluated like a performance channel rather than treated as an open-ended experiment.

What self-serve buying changes, and what it does not

The announced rollout moves ChatGPT advertising beyond a tightly controlled pilot. Advertisers can pursue inventory through agency and technology partners or use a beta Ads Manager rolling out in the United States. The direct interface provides control over budgets, bids, creative uploads, and performance tracking.

That lowers the operational barrier for smaller businesses and teams that could not justify a high-touch engagement. It does not mean access is universal. The product remains in beta, so confirm that your account and market are eligible before you build a launch plan around it.

The addition of cost-per-click bidding is the most consequential change for performance marketers. The initiative began with CPM-based buying, where cost is tied to impressions. CPC lets you bid around visits instead. That is useful because ChatGPT interactions can occur while people are exploring a problem, comparing approaches, or moving toward a decision.

A click is still an intermediate event. CPC is not CPA: paying for a click does not mean you are paying only when a sale, signup, or qualified lead occurs. You still own everything between the click and the business outcome, including page relevance, offer strength, conversion friction, follow-up, and measurement.

Use exploratory, comparative, and decision-ready intent as a creative planning lens:

Exploratory intent: Explain the problem and the practical outcome your offer supports. Avoid demanding a large commitment before the visitor understands the value.
Comparative intent: State the relevant difference, qualification, or tradeoff plainly. Give the visitor enough evidence to judge fit.
Decision-ready intent: Make the offer, next step, price condition, or eligibility requirement easy to find.

This is a messaging framework, not a claim that Ads Manager exposes individual prompts, conversation targeting, or query-level reports. OpenAI’s measurement model is aggregated, and advertisers do not receive access to individual conversations. Do not design targeting, attribution, or sales workflows that depend on identifying what a particular person told ChatGPT.

Direct access is not the only route. Agency and technology relationships include WPP, Publicis Groupe, Criteo, and Adobe. If you buy through a partner, ask who owns the account, which bidding controls you receive, how conversion data is implemented, what reporting can be exported, how frequently it is delivered, and which fees sit outside media spend. A familiar partner workflow is useful only if you can still audit the campaign’s economics.

Keep paid ChatGPT campaigns separate from organic AI visibility work. Ads buy exposure and traffic; AEO and GEO aim to improve how machines understand, retrieve, cite, and represent your content. Do not use paid click-through or conversion data as proof that organic ChatGPT visibility improved. Label the channels separately in analytics so paid traffic does not distort your AI-search reporting.

Decide whether your business is ready to test

Self-serve access makes launching easier, but it cannot supply the business logic that determines whether a campaign should run. Use the following readiness gate before committing spend:

You can name the primary conversion. Choose the event that represents value: a purchase, signup, or lead. If you optimize for a shallow action, such as a form start, keep the true business outcome visible in your reporting.
You know what that conversion is worth. Establish an acceptable acquisition cost from contribution margin, lead quality, close rate, retention assumptions, and fulfillment cost. Do not copy a target from another advertising channel without checking whether the traffic and sales process are comparable.
The destination can fulfill the ad’s promise. The landing page should repeat the core offer, explain who it is for, show relevant evidence, and provide the next step without forcing the visitor to reconstruct the argument.
You can connect ad activity to business records. Ads Manager reporting should be reconciled with web analytics and the system that records revenue or lead quality. Platform conversions alone cannot tell you whether a lead was qualified, duplicated, refunded, or closed.
You can afford an inconclusive test. A beta channel may not produce enough evidence to support a scaling decision. Treat the approved test budget as money at risk, not as revenue you expect the campaign to return on a fixed schedule.

For a performance campaign, calculate a planning ceiling before choosing a bid:

Maximum break-even CPC = acceptable cost per conversion multiplied by the expected landing-page conversion rate.

Use the conversion rate from genuinely comparable traffic when you have it. If you do not, model a conservative range rather than borrowing the best rate from branded search, email, or returning visitors. The result is a break-even boundary, not an automatic bid recommendation. Your actual bid still has to reflect available controls, delivery, competition, and the evidence generated by the campaign.

Lead-generation teams need an additional check. A campaign can appear efficient when it produces inexpensive forms but fail when sales rejects the leads. Define what makes a lead qualified, ensure the CRM records that status, and decide whether the beta’s Conversions API can receive the deeper outcome you want to optimize toward. If it cannot, use the deeper event for business evaluation even if campaign optimization must rely on an earlier event.

Wait to launch if nobody owns the landing page, conversion implementation, or lead follow-up. Buying traffic before those responsibilities are assigned creates a predictable dispute: the ad platform shows activity, analytics shows something different, and the sales team sees outcomes that neither report explains.

Build the first campaign around a falsifiable hypothesis

A tabletop testing setup splits one ad concept into two parallel audience and landing-page paths with a single visual variable changed.

Your first campaign should answer a narrow business question. Write the hypothesis before opening Ads Manager:

For people in a defined decision state, this offer and message will produce this conversion at or below this acquisition-cost ceiling.

That sentence prevents several common mistakes. It keeps brand awareness from being judged by last-click sales, stops a lead campaign from optimizing toward unqualified form fills, and gives you a reason to pause when the economics do not work.

Choose a single primary outcome. Purchases, signups, and leads require different pages, event definitions, and follow-up. Pick the event that matches the offer instead of mixing several goals into one test.
Define the decision state. Decide whether the message is helping someone understand a problem, compare alternatives, or act. Use that decision in your creative brief and landing-page structure. Apply only targeting options that are actually available in your beta account.
Write a specific promise. State the result, the relevant qualifier, and the next step. Avoid copy that merely announces your brand or repeats broad AI terminology. The visitor should know why the click is worth making.
Prepare controlled creative variants. Vary the claim, proof, or call to action separately so you can interpret the result. If every element changes at once, a winning variation does not tell you what to retain.
Build message continuity after the click. The landing page headline should resolve the promise made in the ad. Put the decision-critical facts, constraints, evidence, and action on the page rather than hiding them behind generic navigation.
Set stop and scale rules. Pause immediately if conversion tracking fails. Stop and diagnose when the approved test budget is exhausted without evidence that supports the hypothesis. Scale only when verified outcomes remain within the acquisition-cost ceiling.

Do not invent a universal testing threshold. The amount of evidence you need depends on conversion frequency, normal sales-cycle length, the cost of a false positive, and how much variation exists in lead or order value. Record the threshold you will use before seeing the result so a promising-looking dashboard does not move the goalposts.

Use a stable campaign naming and URL-tagging convention from the start. A workable UTM pattern is utm_source=chatgpt, utm_medium=paid_ai, a campaign value tied to the offer, and a content value tied to the creative variant. Record the exact values in the campaign brief. Consistency matters more than the label itself because it lets analytics, CRM, and finance records join the same test.

Your SEO and GEO work should support clarity on the destination page without being confused with ad configuration. Use visible, accurate facts and structured data that matches the page. JSON-LD can help machines interpret supported entities and attributes, but it is not a ChatGPT ad-targeting control, conversion tag, or substitute for persuasive page content.

Make measurement trustworthy before optimizing bids

An illuminated tracking path connects an ad interaction to a landing page, server, customer record, and verified order package.

ChatGPT advertising is adding pixel-based tracking and a Conversions API for actions such as purchases, signups, and leads. The pixel can capture supported browser-side events. A Conversions API can pass supported events from a server, commerce system, or CRM. Check the beta documentation available in your account before implementation because event fields and diagnostics may evolve.

If you use both methods, verify how duplicate events are handled before sending the same conversion through each path. Two tracking methods should improve resilience, not turn one order into multiple conversions. Test event names, identifiers, values, currency fields, timestamps, and final status against the platform’s current specification.

Build the measurement chain from the business outcome backward:

Business system: The order platform or CRM records revenue, qualification, cancellation, refund, or closed status.
Analytics: The session retains the expected campaign parameters and records the relevant onsite actions.
Conversion integration: The pixel or Conversions API sends the supported event with the correct value and status.
Ads Manager: The campaign reports clicks, spend, and attributed conversions using the attribution settings shown in the account.

Run a validation pass before meaningful spend begins. Confirm that the landing URL works through every redirect, UTM parameters survive navigation, consent behavior is understood, the intended event fires only when its real condition is met, and the backend stores the campaign identifiers you need. Save evidence of the test so later discrepancies can be compared with a known-good implementation.

Expect the systems to disagree at times. Attribution windows, consent choices, browser restrictions, server timing, duplicate handling, and later changes to an order or lead can all create differences. Reconcile the direction and magnitude of the data rather than forcing a false impression of perfect identity. The privacy model also means you should not expect a conversation-level customer trail: reporting is aggregated, and individual ChatGPT conversations are not exposed to advertisers.

Read early results in a fixed order: tracking integrity, visitor behavior, conversion quality, and only then media efficiency. The pattern in the data tells you where to look first:

Observed pattern	First interpretation to test	Action
Ads Manager records clicks, but analytics sees few matching sessions	The click path, redirects, campaign parameters, consent handling, or analytics filters may be breaking attribution	Validate the final URL and session tracking before changing bids or creative
Analytics and the backend record completions, but Ads Manager records few conversions	The pixel or Conversions API event may be missing, malformed, delayed, or duplicated incorrectly	Repair and retest the conversion integration before judging campaign performance
Clicks arrive, but visitors do not reach meaningful onsite actions	The creative may be attracting curiosity, or the page may not continue the ad’s promise	Tighten the qualification in the message and remove landing-page mismatch
Platform conversions look efficient, but sales rejects the leads	The optimized event is too shallow to represent business value	Report qualified outcomes from the CRM and use a deeper supported event when possible
Verified conversions remain within the cost ceiling	The campaign is a candidate for controlled expansion	Increase exposure gradually and keep the offer, page, and measurement stable while evaluating the change
Delivery remains limited	Campaign settings, bid or budget constraints, access, or available inventory may be limiting the test	Check account diagnostics and settings before concluding that demand is absent

Do not respond to weak conversion economics by raising the bid first. Confirm that measurement works, inspect the promise-to-page transition, and check whether the recorded conversion represents real value. Increase bids or budgets only when account data indicates delivery is constrained and the verified acquisition economics can absorb more traffic.

Document every material change with its effective time, including bid, budget, creative, destination, event definition, and attribution setting. If several variables change together, the next reporting period may look different without telling you why.

Key takeaways

ChatGPT’s self-serve Ads Manager is a U.S. beta, so verify access and current account controls before planning a launch.
CPC bidding makes traffic easier to buy and evaluate, but a paid click is not a sale, qualified lead, or profitable customer.
Write the campaign hypothesis, conversion definition, cost ceiling, test budget, and stop rule before spend begins.
Use a matching landing page and consistent campaign parameters so Ads Manager, analytics, and backend outcomes can be reconciled.
Pixel and Conversions API tracking improve measurement, but data is aggregated and does not expose individual conversations.
Keep paid ChatGPT performance separate from organic AEO and GEO visibility. Neither should be used as proof that the other improved.

Your next move is to write the hypothesis and acquisition-cost ceiling, then trace the conversion from the landing page to the final business record. If either remains undefined, keep the budget closed. If both survive that check, you have the basis for a controlled beta test and a clear decision when the results arrive.

References

CrushPress.AI — Unlocking New Potential: ChatGPT’s Self-Serve Ads Revolution

May 5, 2026

How to Test Emerging Ad Platforms With Better Measurement

You have access to a promising new ad placement, the first click-through rates look excellent, and someone wants to know whether to increase the budget. That is exactly when measurement discipline tends to slip. A strong dashboard number feels like an answer even when it only describes the first step in the journey.

Your real task is to determine whether the platform creates valuable outcomes that would not otherwise happen, whether those outcomes remain economical as the test expands, and whether the available inventory can absorb more spend. This framework helps you answer those questions without expecting one attribution model to do every job.

Separate channel discovery from budget proof

An emerging platform can be interesting before it is investable. That distinction matters because discovery metrics and budget metrics answer different questions.

Click-through rate tells you whether people respond to a placement. It does not tell you whether the resulting customers are profitable, whether the ad caused those customers to act, or whether similar performance will survive broader distribution. This is especially important for conversational advertising, where early engagement has been strong but inventory and testing remain limited.

Run the test as a sequence of decisions. Each decision requires different evidence:

Decision	Evidence to inspect	What it does not prove
Does the placement attract attention?	Impressions, clicks, click-through rate, and engagement by query or audience segment	That the attention creates business value
Does the traffic produce the right outcome?	Purchases, qualified leads, subscriptions, revenue, lead quality, and downstream completion	That the advertising caused the outcome
Is the outcome incremental?	Holdout testing, geo experimentation, or another credible counterfactual	That the same return will persist at a larger spend level
Can the platform scale efficiently?	Available inventory, spend delivery, reach, frequency, conversion quality, and cost as exposure expands	That it improves the entire media portfolio
Should the portfolio budget change?	Experiment-calibrated media mix modeling alongside commercial constraints	That every individual conversion can be assigned to one touchpoint

This separation protects you from two common mistakes. The first is rejecting a potentially useful channel because it has not yet accumulated enough evidence for a permanent budget allocation. The second is scaling it because a high early click-through rate has been mistaken for incremental profit.

Label the stage of the evidence in every internal update. Use plain terms such as discovery signal, conversion signal, incremental evidence, and scale evidence. If the team only has a discovery signal, say so. That small piece of language prevents a preliminary result from hardening into a forecast.

Write the measurement contract before the first impression

Hands arrange matching campaign materials into separate test and control areas on a measurement planning table.

A measurement plan should be a decision contract, not a list of every metric the platform can export. Write it before launch so the team cannot redefine success after seeing the results.

Name one primary business outcome. Choose the event closest to value that the test can credibly observe: a completed purchase, a qualified opportunity, a subscription, or another commercially meaningful result. Keep clicks and engagement as diagnostics unless attention itself is the campaign objective.
State the causal question. Write what you are trying to learn in counterfactual terms: how many desired outcomes occurred because the ads ran, beyond what would have happened without them? This wording exposes the limit of ordinary attribution before anyone treats credited conversions as incremental conversions.
Define the test unit. Decide whether results will be examined by query theme, audience, geography, product, offer, creative, or another controlled unit. The unit must match the mechanism you expect to drive performance.
Set the comparison rules. Document the conversion definition, attribution window, revenue basis, treatment of returns or cancellations, and handling of duplicate records. Use the same definitions for the emerging platform and the benchmark channel.
Choose guardrails. Track conversion quality, acquisition cost, spend delivery, reach concentration, and any operational consequence such as low-quality leads. A channel that creates more form submissions but overwhelms sales with poor prospects is not passing the business test.
Predeclare the verdicts. Specify what evidence would justify scaling, continuing the test, pausing for an instrumentation repair, or stopping. Your thresholds should come from the economics of your own business rather than a generic platform benchmark.

The contract also needs a data lineage section. For every result, record where the event originates, how it is passed, which identifier joins it to campaign data, and which system is authoritative when two systems disagree. If a purchase appears in the ad platform but not in the commerce system, the team should already know which record governs the decision.

Do not postpone this work until reporting begins. Missing identifiers and inconsistent event definitions cannot always be repaired after exposure has occurred. If the primary outcome is not reliably captured, pause the test and fix the measurement path before buying more traffic. Otherwise, additional spend produces a larger dataset without producing a better answer.

Read early AI ad performance without fooling yourself

Conversational ads may appear beside a response at the moment a user is expressing a need. That context can make the placement feel more relevant than an interruptive format. It also creates several reasons for early results to look unusually strong.

Intent mix is the first reason. Prompts about Mother’s Day have been observed to trigger ads about three times more often than the overall average. A test concentrated in gift-seeking conversations is not representative of every prompt, product category, or stage of the buyer journey. Report results by intent class instead of averaging all conversations into one channel-wide figure.

Format novelty is the second reason. People may inspect a new placement because they have not seen it before. You cannot prove that novelty caused the clicks from an initial campaign, but you can watch for the pattern. Repeat the test across cohorts or campaign waves, keep the offer and conversion definition stable, and check whether engagement and downstream quality hold as the format becomes more familiar.

Inventory selection is the third reason. Limited supply can concentrate delivery in the prompts, advertisers, or use cases most likely to perform. Expansion may introduce weaker contexts, more competition, and different pricing. Track how much of the planned budget is actually delivered, where impressions cluster, whether new query categories enter the mix, and how acquisition cost changes as spend rises. A channel that cannot spend the approved amount is not yet a scalable acquisition engine, even if its small pool of impressions performs well.

The comparison channel matters too. Early conversational-ad click-through rates have exceeded display and podcast benchmarks, but that comparison describes engagement, not equivalent economics. Search, paid social, display, podcast advertising, and conversational placements differ in intent, buying method, inventory, and the role they play in a journey. Compare them on the same final outcome and accounting basis before moving budget.

At the review meeting, force the result into one of four decisions:

Scale: the primary business outcome meets the predeclared requirement, the evidence supports incrementality, data quality is intact, and the platform has enough inventory to test a higher spend level.
Continue testing: engagement and conversion quality are promising, but incrementality, pricing stability, or inventory depth remains uncertain. Name the next uncertainty and design the next test specifically around it.
Pause and repair: event loss, inconsistent definitions, broken joins, or missing downstream outcomes make the result unreliable. Fix the data path before resuming.
Stop: the test has enough reliable evidence to show that the business outcome does not meet your requirement, or repeated expansion causes economics or conversion quality to deteriorate beyond the accepted limit.

“Promising” is not a fifth verdict. It is a description that must be followed by a specific next decision.

Build an evidence ladder instead of trusting one model

An abstract ladder of measurement methods rises from raw signals to a verified outcome, with several evidence paths converging near the top.

No single measurement method can tell you whether an ad was served correctly, influenced an individual journey, created incremental demand, and deserves a larger share of the portfolio. Use a ladder in which each layer answers a narrower question and checks the layers below it.

Layer 1: instrumentation and platform diagnostics

Start with clean event collection. Connect ad delivery, site or app behavior, commerce results, and CRM outcomes. Preserve campaign identifiers where possible, deduplicate events, and reconcile totals against the system that records the actual transaction or qualified lead.

The direction of Google’s tooling shows how central this plumbing has become. Data Manager is being expanded with a map-based view of connections involving systems such as BigQuery, HubSpot, and Shopify, while Google tag changes are intended to extend existing setups without requiring additional code. The useful principle is broader than any vendor: make the flow of data visible enough that a marketer can locate a missing connection before it distorts a campaign decision.

Platform reports remain useful at this layer. They help you diagnose delivery, creative response, query mix, and conversion paths. Treat attributed conversions as claims that need reconciliation, not as automatic proof of causality.

Layer 2: controlled experiments

An experiment estimates the counterfactual that ordinary attribution cannot observe. A holdout keeps an eligible group from receiving the treatment. A geo experiment varies advertising across comparable regions and evaluates the difference in business outcomes. Neither method is a decorative validation step. It is the evidence used to decide how much of the platform-reported performance is genuinely incremental.

Google’s Meridian GeoX reflects this shift toward causal validation. It is built on an open-source framework and connects geo experimentation with the broader Meridian media mix modeling system. For your team, the practical lesson is to plan experimentation and portfolio modeling together. Experimental results can challenge an attribution narrative and provide a firmer basis for calibrating broader budget models.

Choose an experimental design only when the platform and your market provide a defensible control. If exposure leaks heavily between groups, the regions behave differently for unrelated reasons, or the outcome volume is too sparse to distinguish change from noise, do not dress the result up as causal proof. Document the limitation and continue at the lower rung of the evidence ladder.

Layer 3: media mix modeling

Media mix modeling examines aggregated changes in spend and outcomes across channels and time. It is suited to portfolio questions: how channels work together, how budget shifts may affect total results, and where marginal investment may be more productive. It does not need to identify a single ad as the exclusive cause of a single purchase.

An emerging channel may initially be too small or too stable in spend for a portfolio model to isolate reliably. That is not a reason to invent precision. Use controlled testing to establish an initial incremental read, create meaningful and documented variation when expanding the channel, and add it to the model when the underlying data can support the distinction.

Google is also working to reduce the operational burden of this layer through Meridian Studio, a Google Cloud-powered environment for building, customizing, and scaling media mix models. Easier tooling does not remove the need for sound inputs, transparent assumptions, or experimental checks. A faster model built on inconsistent revenue, incomplete spend, or unexplained tracking changes is still an unreliable model.

Keep a measurement change log alongside the model. Record tag updates, consent changes, platform launches, campaign restructures, pricing changes, promotions, and breaks in source data. When performance moves, this log helps you distinguish a market effect from a measurement artifact.

Key takeaways for your next platform test

High click-through rate is a discovery signal. It is not evidence of incremental revenue, efficient scaling, or portfolio impact.
Define the business outcome, counterfactual, comparison rules, guardrails, and decision thresholds before the campaign begins.
Segment conversational-ad results by intent and query class. A concentration of high-intent prompts can make the channel average look more transferable than it is.
Evaluate scale separately from efficiency. Limited inventory can produce good economics while preventing meaningful budget deployment.
Use platform reporting for diagnostics, experiments for causal lift, and media mix modeling for portfolio allocation.
Pause when instrumentation is broken. More spend cannot repair missing identifiers, inconsistent events, or an unreliable outcome definition.

Before accepting the next emerging-platform test, write the measurement contract on one page and identify the weakest rung in your evidence ladder. Fund the test that resolves that uncertainty. Increase the budget only when the business outcome, incremental effect, data quality, and available inventory all support the same decision.

References

May 5, 2026

ChatGPT Advertising Insights: A Practical Pilot Playbook

If you are deciding whether ChatGPT advertising deserves budget, do not start by asking whether it resembles paid search. Start with the moment the ad enters: the user has already described a need, added constraints, and moved partway toward a decision.

A ChatGPT ad can appear inline within that conversation, marked as Sponsored and presented with a headline, short body, and destination. Your job is not to interrupt the journey. It is to offer a credible next step that fits the journey already underway. That difference should shape your creative, measurement, landing pages, and relationship between paid advertising and organic AI visibility.

Use the early data as a format signal, not an ROI benchmark

The first useful insight is about the strength and limits of the evidence. The early U.S. trial launched on February 9 for Free and Go users, while Adthena tracked more than 50,000 daily placements from over 600 advertisers across B2B software, ecommerce, fintech, and consumer categories.

That is enough activity to reveal recurring creative conventions. It is not enough to establish a universal cost per acquisition, return on ad spend, or incrementality benchmark. The observations come from a vendor-tracked index during a trial, span materially different verticals, and do not provide one standardized performance baseline for every advertiser.

Use the data to answer questions such as how much copy the format can carry, which information tends to appear first, and how closely creative reflects the conversation. Do not use it to forecast your return before you have campaign-level evidence from your own offer, audience, and destination.

Before assigning meaningful budget, make sure your pilot can answer a defined question:

Can you identify a narrow group of commercial topics where the user is likely to be comparing options or preparing to act?
Do you have a specific, verifiable benefit that can be understood without several lines of explanation?
Does the destination continue the exact promise made in the ad?
Can you separate ChatGPT placements from your other paid traffic when evaluating outcomes?
Have you defined what would justify expanding, revising, or stopping the test before spend begins?

Rollout status is time-sensitive, so confirm actual inventory and account eligibility before committing budget or launch dates. A projected geographic expansion is not the same thing as inventory you can buy.

Write an answer fragment, not a compressed search ad

A distinct sponsored module fits into a flowing sequence of text-free conversation cards while a separate banner sits outside the flow.

A traditional search ad often has several components competing for attention: multiple headlines, descriptions, sitelinks, extensions, and other assets. The early ChatGPT format is more restrained. That makes every word carry more of the decision.

The strongest working model is an answer fragment. It should make sense beside the assistant’s response, acknowledge the user’s decision criteria, and introduce a next step without pretending to be the neutral answer.

The tracked placements show several compact patterns. Headlines averaged about 30 characters and peaked at 36, body copy averaged roughly 19 words, and many ads used two short sentences. These are observed conventions, not confirmed platform character limits.

Creative element	Early pattern	What to do with it
Headline	About 30 characters on average, with a peak at 36	Lead with the decision-driving benefit. Do not spend the available space on a generic slogan.
Headline opening	Most begin with the brand name	Test a Brand: Benefit construction when recognition and accountability matter.
Body	About 19 words, commonly split into two sentences	Use the first sentence for proof and the second for a low-friction action.
Relevance	Stronger creative mirrors the user’s context	Reflect the category, constraint, or desired outcome instead of repeating a loose keyword.
Offer detail	Dollar signs, rates, and concrete figures were associated with stronger conversion performance	Prioritize a specificity test, but treat the pattern as a hypothesis to validate in your own campaign.

Build each variation from three prompt components

When a user asks for accounting software for a small team, for example, accounting software is only the category. Small team is the constraint. The unstated decision criterion might be fast setup, predictable cost, or limited administrative work. Creative that reflects only the category will feel generic even if it contains the right keyword.

Extract the category: what kind of product, service, or action does the user want?
Extract the constraint: what price, use case, location, feature, risk, or timing narrows the choice?
Choose one decision criterion your offer can substantiate.
Write the headline as Brand: Verified Benefit.
Use the body for one proof point and one proportionate call to action.
Remove any claim that the landing page cannot immediately confirm.

A useful template is: Brand: [specific outcome]. [Proof tied to the user’s constraint]. [Simple next action]. The brackets are not an invitation to stuff several benefits into one placement. Choose one reason to continue.

Specificity needs controls. If you advertise a price, rate, discount, delivery window, or availability claim, it must be current, approved, and visible at the destination. A concrete figure can improve clarity, but an outdated figure creates both conversion friction and potential compliance exposure. When the value changes frequently, build a review process before testing it in ad copy.

Test in an order that explains the result

Changing the headline, proof, call to action, and landing page at the same time may produce a winner, but it will not tell you why it won. Start with the variables most closely tied to conversational relevance:

Specific offer versus general benefit.
Query-matched benefit versus broad category language.
Quantified proof versus qualitative proof.
Low-commitment call to action versus immediate purchase or signup language.
General landing page versus a page that continues the same constraint and benefit.

Hold the other elements steady during each comparison. The point is not merely to improve the ad. It is to learn which part of the conversation your audience needs resolved before moving forward.

Measure prompt coverage and response duplication before calling it reach

An overhead arrangement of varied prompt tokens connects to response cards, including a magnified cluster of visibly duplicated cards.

Clicks and conversions still matter, but they do not tell you whether your brand is present across the conversations that matter. Conversational inventory needs an observation layer organized around topics, prompts, and individual responses.

That becomes especially important because one brand has been observed appearing twice within the same ChatGPT response. This double-parked behavior creates more placements, but it does not automatically create more unique reach. Counting each placement as a separate conversation would overstate coverage.

For every observed placement, record the topic, prompt or prompt class, response identifier, timestamp, position, advertiser, headline, body, and destination. Add post-click outcomes when your analytics can connect them. That record supports several more useful measurements:

Observed prompt coverage: the portion of your monitored commercial prompts in which your brand appeared.
Observed response presence: responses containing your brand divided by eligible responses you actually monitored.
Duplication rate: brand-present responses containing more than one placement for the same brand.
Competitor overlap: responses where your brand and a named competitor appeared together.
Creative-context match: whether the ad reflects the category, constraint, and decision criterion in the prompt.
Post-click continuity: whether the destination preserves the offer and language that earned the click.
Business outcome: qualified lead, sale, signup, or another result defined before the pilot.

Call these observed rates, not platform-wide impression share. A monitoring sample cannot tell you the total number of eligible conversations unless the platform provides that denominator. This naming discipline prevents a directional visibility metric from turning into a false market-share claim.

Review duplication separately from performance. Two appearances might reinforce recall, or they might add no incremental value. The placement pattern alone cannot settle that question. Compare duplicated and single-placement responses only when you have enough campaign data to evaluate their downstream outcomes.

Your landing-page review should be just as specific. Check whether the advertised benefit appears without searching, whether the price or rate matches, whether the next action is obvious, and whether the page answers the constraint expressed in the originating conversation. A relevant ad that lands on a general homepage throws away the context that made the placement useful.

Coordinate ChatGPT ads with AEO and GEO without merging the KPIs

Paid presence and organic AI visibility can occur in the same conversational environment, but they are not the same achievement. A sponsored placement buys labeled exposure. An organic citation, recommendation, or brand mention depends on how the system constructs its answer. Early placement observations do not establish that buying ads improves organic answer inclusion.

Keep the two lanes separate in reporting. If you combine them into one AI visibility number, you will not know whether a change came from media spend, content improvements, brand demand, or answer-engine behavior.

Use one shared topic map. Organize paid monitoring and organic visibility work around the same commercial questions, constraints, entities, and decision criteria.
Give paid media its own outcomes. Track observed presence, duplication, clicks, qualified actions, and campaign economics.
Give AEO and GEO their own outcomes. Track whether the brand is mentioned, cited, represented accurately, and connected to the intended category across monitored answers.
Align the factual layer. Prices, rates, features, availability, and offer terms should agree across ad copy, visible page content, and applicable structured data.
Investigate cross-channel clues. A commercial prompt with competitor ads but weak organic answers may expose a content opportunity. Strong organic visibility with no paid presence may identify a conversation worth testing, but neither observation guarantees demand or return.

JSON-LD can clarify entities, products, offers, and other machine-readable facts when it accurately represents visible content. It does not purchase inventory, guarantee inclusion in an AI response, or repair a weak offer. Use structured data to reduce ambiguity, then use advertising to test whether a clear commercial promise earns action.

This coordinated model also gives you a cleaner competitive view. You can distinguish a competitor that is buying exposure from one that is repeatedly earning non-sponsored visibility. The response is different: one may call for a media test, while the other may require better content, stronger entity signals, clearer proof, or a more competitive offer.

Key takeaways for your first ChatGPT ad pilot

Treat early placement data as evidence about format and creative conventions, not as a guaranteed ROI benchmark.
Write for a user who has already supplied context: lead with the brand, one verified benefit, one proof point, and one next action.
Use the observed 30-character headline and 19-word body patterns as editing discipline, not as assumed platform limits.
Test concrete figures before vague claims when your offer supports them, but keep every price, rate, and term synchronized with the destination.
Measure prompts and unique responses as well as placements, because two appearances in one response do not equal two reached conversations.
Coordinate paid, AEO, GEO, landing-page content, and structured data around one topic map while reporting paid and organic outcomes separately.

Your next move is a narrow pilot, not a platform-wide commitment. Choose a small set of high-intent topics, document the user’s constraints, create controlled variations, and establish an organic visibility baseline before ads run. You will then be able to decide from your own evidence whether conversational advertising adds qualified demand, merely adds placements, or reveals a larger content opportunity.

References

CrushPress.AI – Unlocking Insights from ChatGPT Ads: What You Need to Know

May 1, 2026

How to Prepare for ChatGPT’s Advertising Expansion
If you’re deciding whether ChatGPT belongs in your paid media plan, don’t treat its advertising expansion as a cue to move budget immediately. Treat it as a cue to become test-ready. The opportunity may be meaningful, but availability, targeting, reporting, and campaign economics still need to be proved.

Your advantage won’t come from being first at any cost. It will come from knowing exactly what you want to learn, what evidence would justify more investment, and how paid placement fits beside your existing SEO, AEO, and generative engine optimization work.

The expansion addresses inventory, not the whole advertising case

Early observations indicate that ads are appearing within conversations for some logged-out users, although OpenAI had not formally announced the expansion. That uncertainty matters. A visible rollout can establish that inventory is growing without establishing who can buy it, which users are eligible, how delivery is priced, or whether the experience is stable enough for forecasting.

The immediate pressure appears to be supply. Pilot advertisers have reportedly struggled to spend their intended budgets because inventory was limited, even after the financial hurdle fell from $200,000 to $50,000. Opening more conversations to ads is a logical way to create additional opportunities for delivery.

That doesn’t automatically make ChatGPT a scalable performance channel. More inventory can help campaigns spend, but it doesn’t prove that the added impressions will produce qualified traffic, incremental customers, or acceptable acquisition costs. Logged-out reach could also differ from logged-in reach in ways that affect relevance and measurement. Until the buying interface or your agreement provides the details, don’t assume the platform can recognize, target, exclude, or report on these two audiences in the same way.

Keep ChatGPT out of your dependable base forecast for now. Put it in an experimental budget with its own success criteria and loss limit. That protects the budget you already rely on while giving you room to learn if access becomes available.

Key takeaways
- Wider logged-out reach may relieve an inventory constraint, but it doesn’t yet establish stable campaign economics.
- Conversational placement deserves its own creative and landing-page strategy; repurposing a display banner is unlikely to answer the user’s immediate need.
- Require definitions for delivery, targeting, attribution, and logged-in versus logged-out reporting before committing meaningful budget.
- Measure paid placement separately from organic AI visibility. Buying an ad doesn’t demonstrate that ChatGPT knows, cites, or recommends your brand.
- Prepare a controlled pilot now, but release money only after the platform can support the decisions you need to make.
Build the pilot around one commercial decision

Novelty is not a campaign objective. A useful pilot answers a decision such as: Should we add this channel to our acquisition mix? Can it reach buyers earlier than search ads? Does it create qualified demand we wouldn’t otherwise capture? Choose one question. A pilot designed to prove awareness, traffic quality, lead generation, and revenue at once usually produces an ambiguous answer to all four.
1. Choose one demand state. Define the situation in which your offer helps, such as comparing approaches, narrowing a shortlist, solving an urgent problem, or selecting a provider. Don’t assume the platform lets you bid on exact prompts. Ask what targeting controls actually exist, then translate your demand state into the controls available.
2. Name one primary business outcome. Use a completed purchase, qualified lead, activated account, booked consultation, or another event connected to value. A click can diagnose delivery, but it shouldn’t become the business case merely because it is easy to count.
3. Set a quality guardrail. For lead generation, that could be lead acceptance or sales qualification. For commerce, it could be cancellation, return, or contribution margin. A campaign can report an attractive acquisition cost while sending customers who never become profitable.
4. Create a landing page for the conversational handoff. Restate the promise plainly, answer the next likely question, provide evidence for important claims, and make the next step obvious. If the advertisement answers one question but the page opens with a generic corporate message, you lose the contextual advantage of the placement.
5. Prepare multiple message angles. Ads have been observed fitting into the conversation rather than behaving like conventional banners. Write concise copy around the user’s task: a direct answer or benefit, a relevant qualification, and a proportionate next step. Keep every claim defensible when read outside the surrounding conversation.
6. Write the expansion rule before launch. Define the acquisition cost, conversion quality, and measurement confidence needed for more investment. Also define the conditions that stop the test. Historical economics from your own business are more useful here than an arbitrary industry benchmark.
Your test charter should also identify the comparison that matters. If ChatGPT merely receives budget that would have converted through paid search, platform-reported conversions may look encouraging without adding much business value. Compare the pilot with your normal channel mix, not with doing nothing in an imaginary market.

Demand measurement answers before you demand scale

Conversational advertising can create a less familiar path than keyword, feed, or social advertising. A person may ask several questions, see a commercial placement, leave, research the brand elsewhere, and convert later. That makes a clean platform dashboard especially tempting. It also makes unexamined platform attribution especially risky.

Before launch, get written answers to the questions that can change your interpretation of performance:
- What event counts as an impression, and can one conversation generate more than one?
- What counts as a click or other engagement?
- Which click-through or view-through attribution windows are used?
- Can you change those windows or compare them with your analytics standard?
- Can results be segmented by logged-in status, placement type, geography, device, creative, and audience method?
- What contextual, behavioral, demographic, or account-level signals can influence delivery?
- Which exclusion, frequency, suitability, and sensitive-topic controls are available?
- How are duplicate conversions, invalid interactions, refunds, cancellations, and offline outcomes handled?
- Can you export event-level or sufficiently granular campaign data for independent reconciliation?
A missing answer is information. If you can’t distinguish the new logged-out inventory from the rest of delivery, you won’t know whether the expansion improved reach, reduced quality, or simply changed the mix. If you can’t align attribution windows, you won’t be able to compare ChatGPT with another channel fairly.

Build reporting in four layers. Delivery tells you whether the campaign can spend. Response tells you whether people engage. Business quality tells you whether those interactions become valuable outcomes. Incrementality asks whether the outcomes would have happened without the campaign. Keep these layers separate so a strong click rate cannot disguise weak economics.

Use a controlled comparison if one is available and proportionate. A randomized holdout is the clearest option when the platform supports it. Otherwise, use a carefully chosen geographic or time-based comparison and document its limitations. Seasonality, promotions, sales activity, and changes in other media can all create false lift. Don’t call a before-and-after difference incremental merely because the dates line up.

Preserve campaign and creative identifiers in your analytics, connect conversions to revenue or lead quality where consent and applicable rules allow, and deduplicate outcomes across platforms. Compare the platform’s totals with your own analytics before increasing spend. A disagreement doesn’t automatically mean one system is wrong; attribution systems can assign the same conversion differently. It does mean you need to understand the difference.

Keep paid ChatGPT reach separate from organic AI visibility

ChatGPT advertising and generative engine optimization address different problems. An ad buys an opportunity to appear under specified campaign conditions. Organic visibility depends on whether a system can discover, interpret, trust, and use information about your brand or subject. Paid delivery is not evidence of organic inclusion, and an organic mention is not evidence that advertising caused it.

This distinction should shape both your dashboard and your content plan. Report paid impressions, engagements, conversions, acquisition cost, and incrementality as campaign metrics. Track organic citations, brand mentions, referred visits, answer accuracy, and visibility across relevant prompts as a separate program. You can examine relationships between them, but don’t combine them into one score that hides which mechanism changed.

The landing pages used for conversational ads should still meet the same evidence standard as your organic content:
- Answer the visitor’s central question before forcing them through a broad brand narrative.
- Use descriptive headings that make each section understandable on its own.
- Identify products, services, organizations, and authors consistently across the page and site.
- Support material claims with evidence a reader can inspect.
- Keep prices, availability, policies, and other changeable facts current wherever you publish them.
- Use schema types and properties that accurately represent visible content. JSON-LD can clarify entities and relationships, but it cannot guarantee inclusion in an AI answer or eligibility for an advertisement.
- Make ownership, contact details, and the path to a real next step easy to verify.
Use paid learning to improve content only when the data supports the connection. If a message angle attracts qualified visitors, examine the underlying need and build a fuller answer around it. Don’t manufacture near-duplicate pages for every phrasing variation, and don’t turn an advertising result into an unsupported claim about what all ChatGPT users want.

The reverse is useful too. Organic visibility analysis can reveal questions where your brand is absent, misunderstood, or poorly supported. Those gaps can inform a paid hypothesis while you improve the underlying content. The advertisement may create immediate reach; the content fixes the durable information problem.

Use a readiness gate before committing budget

You don’t need to choose between rushing in and ignoring the channel. Use three readiness states.
- Prepare now if ChatGPT is relevant to how your buyers research or compare solutions. Create the test charter, conversion definitions, landing page, creative hypotheses, suitability rules, and reporting requirements without assuming access.
- Test when available if you can isolate a meaningful business outcome, cap the downside, reconcile conversion data, and learn something that affects a real channel decision. Learning value matters, but it should be named rather than used as an excuse for unlimited spending.
- Delay investment if access requires a commitment your experiment cannot justify, essential targeting or safety controls are missing, results cannot be independently reconciled, or your landing experience is not ready. Scarcity of access is not proof of value.
The reported reduction from $200,000 to $50,000 still represents material exposure for many organizations. Don’t commit merely to reserve a place in a pilot. Confirm the contract terms, cancellation rights, measurement access, inventory expectations, and responsibility for unsuitable placement before funds become difficult to recover.

Start with a one-page test charter. Write down the user need, primary outcome, quality guardrail, maximum acceptable downside, required platform answers, and expansion rule. When broader access arrives, that page will let you evaluate the opportunity on business evidence instead of launch momentum.

References
- Search Engine Land – ChatGPT ads expand to logged-out users
April 23, 2026

How to Audit Campaign Controls Before You Optimize Spend

Your campaign can look more efficient while becoming harder to control. Spend may be compressed into fewer active days, conversion signals may be incomplete, and a polished dashboard may show activity without giving you the controls needed to explain or stop it.

If performance changes without a clear bid, audience, or creative change, audit the control layer first. You need to know what the platform is allowed to do, what data its optimizer can see, and whether your reports describe the same system you configured.

Key takeaways

Budget, schedule, consent, optimization, and reporting are separate controls. Changing or validating one does not validate the others.
A restricted ad schedule may concentrate spending rather than reduce the campaign’s monthly spending limit.
Consent diagnostics should help you locate missing or inconsistent signals. A consent rate is not a target to maximize at the expense of genuine user choice.
A dashboard is not a mature control system unless you can inspect state, enforce changes, verify their effects, and reconstruct who changed what.
Paid placement in an AI interface and earned visibility in a generated answer require separate attribution and reporting.

Audit the whole control chain before touching bids

Campaign optimization is usually treated as a bidding problem. In practice, bidding is only one link in a chain. The platform first determines whether an ad is eligible, then how much it may spend, which signals it can use, what decision automation should make, and what evidence you get afterward.

A weakness anywhere in that chain can produce a misleading result. A schedule can alter the concentration of spend. A consent implementation can reduce observable conversions. A reporting delay can make a stable campaign appear volatile. Raising or lowering a bid before resolving those conditions adds another variable without answering the original question.

Control layer	Question to answer	Evidence to record
Business constraint	What outcome, total cost, or operational load can you accept?	Approved spending ceiling, capacity limit, and stop condition
Eligibility	When is the campaign allowed to enter auctions?	Active days and hours, plus the business reason for each restriction
Delivery	How may the platform allocate spend while the campaign is eligible?	Budget values, bidding mode, spending caps, and documented pacing behavior
Signal	Which conversions and consent states can the optimizer observe?	Conversion definitions, consent diagnostics, and coverage by relevant dimension
Observation	Can you explain what happened after delivery?	Reporting latency, available breakdowns, exports, attribution settings, and change history

Run the audit in that order. Starting with reports is tempting, but a report cannot tell you whether the configured business constraint was correct. Starting with bidding is worse because the optimizer may be responding rationally to a budget, schedule, or signal state you did not intend.

Write down the campaign’s intended result and its hard constraint. Separate a performance target from a limit the platform must not cross.
Capture the current schedule, budget, bidding mode, conversion actions, consent state, targeting, and exclusions. Use actual settings, not what the launch plan says should be configured.
Translate settings into effective exposure. For example, calculate the monthly spending ceiling and inspect how much delivery could be compressed into eligible periods.
Check whether the optimizer receives the signals you expect across apps, platforms, regions, and traffic sources. Treat gaps as unresolved until you have distinguished user choice from an implementation problem.
Verify that important controls are enforceable. A pause button, budget edit, or exclusion is useful only if you can confirm its scope, timing, and effect.
Record each change with the old value, new value, timestamp, reason, expected effect, evaluation window, and stop condition. Where practical, avoid changing another layer before the first change can be evaluated.

This gives you a baseline that optimization can build on. Without it, every performance movement invites a new theory, and several contradictory theories may fit the same aggregate chart.

Scheduled campaigns need a spend-concentration audit

A hand adjusts a scheduling gate above a blank calendar grid where glowing budget tokens are concentrated into only a few active tiles.

A budget limits spending; a schedule limits eligibility. Those settings may feel interchangeable when a campaign runs only on selected days or hours, but they answer different questions.

Under Google’s scheduled-campaign pacing model, a campaign can pace toward its full monthly spending limit even when its ads are not eligible every day. Disabled days remain disabled, but the system has more reason to capture available demand during the periods that remain open.

The stated limits make the exposure calculable: the monthly spending cap remains 30.4 times the average daily budget, while spending on an individual day can reach up to twice that daily budget. These are ceilings, not promises about what the campaign will spend.

The practical correction is simple: do not assume that fewer eligible days will produce a proportionally smaller monthly bill. If you intend to reduce total exposure, set the budget to reflect that intention. Keep the schedule focused on when the business can serve demand or when traffic is valuable.

Find every non-continuous schedule. Include campaigns limited to particular weekdays as well as those restricted to certain hours.
Write down why the restriction exists. A schedule tied to staffing, inventory, response time, or lead quality is an operational guardrail. Do not remove it merely to smooth a spending chart.
Calculate the monthly ceiling. Multiply the average daily budget by 30.4, then compare that amount with the total monthly exposure you actually approved.
Check the active-day boundary. Ask whether spending up to twice the average daily budget on an eligible day would create a cash-flow, inventory, or service-capacity problem.
Review eligible periods directly. Monthly averages can hide concentrated delivery. Inspect spend, conversions, and downstream quality during the windows when ads were allowed to run.
Change the correct control. Lower the budget when the total amount is too high. Narrow or widen the schedule only when eligibility itself is wrong.

This distinction also improves diagnosis. Faster spending during active periods does not automatically mean bidding has become more aggressive or demand has improved. It may be the predictable result of the pacing system trying to use the same monthly allowance within fewer opportunities.

Consent diagnostics tell you whether the optimizer can learn

An analyst examines anonymous data signals passing through transparent consent gates toward an unbranded optimization engine, with some signals blocked or fading.

An optimizer cannot act on a conversion it cannot observe. That makes consent signal quality part of campaign operations, not a separate technical housekeeping task.

Google Ads’ App Consent Insights exposes consent diagnostics across apps, platforms, regions, and traffic sources. The view includes an overall rating of Excellent, Good, or Poor, a live count of apps sending consented data, and conversion consent rates with EEA and non-EEA differences.

Use those dimensions to localize a gap. Do not interpret the account-level rating as a complete diagnosis. A lower rate could reflect genuine user choices, traffic composition, a deployment inconsistency, or missing signal transmission. Those possibilities need different responses.

List the apps and platforms that should be sending consent information. Compare that inventory with the live count shown in the diagnostic.
Locate the narrowest break. Determine whether the difference belongs to one app, one platform, one region, one traffic source, or a wider implementation.
Compare EEA and non-EEA results without assuming geography is the cause. Review the regional consent implementation and the underlying traffic mix separately.
Validate the technical path from the consent choice to the advertising platform. Confirm that the relevant state is collected, transmitted, and associated with the intended conversion setup.
Annotate the release or configuration change that corrected a gap. Keep unrelated budget and bidding edits out of the same evaluation window where possible.
Reassess campaign performance only after the corrected signal flow has had an appropriate observation period for your normal conversion lag.

The overall rating is a diagnostic indicator, not an optimization objective. Do not make a consent experience more coercive just to lift a platform metric. Changes to consent language or interaction design should remain under the appropriate privacy and legal review. The campaign team’s job is to make sure a valid choice is transmitted accurately and that missing instrumentation is not mistaken for user behavior.

This protects decision quality in both directions. You avoid blaming creative when measurement is incomplete, and you avoid treating every consent-rate difference as a tagging failure. Once signal coverage is understood, bidding and conversion reports become easier to interpret.

Prove an AI ads manager can control delivery before scaling it

New advertising interfaces can improve access long before their control systems become mature. OpenAI is testing a ChatGPT Ads Manager that moves beyond weekly CSV reporting toward real-time campaign management, monitoring, and optimization. That is meaningful progress, but testing an interface is not evidence that every targeting, reporting, governance, or automation capability is complete or broadly available.

Evaluate an emerging ad manager by what you can verify, not by how familiar its dashboard looks. For every requirement, distinguish between a control that is promised, a control visible in the interface, and a control whose effect you have confirmed.

Authority: Can the authorized operator pause delivery, edit budgets, and reverse a change at the required account or campaign scope?
Budget semantics: Is the budget daily, monthly, lifetime, or another form? How is pacing described, and what prevents an unexpected concentration of spend?
Eligibility and exclusions: Which scheduling, targeting, placement, brand-safety, and exclusion controls actually exist? Do not assume parity with Google Ads or Meta because the navigation feels familiar.
Measurement: Which event counts as a conversion, what attribution rules apply, how quickly do results appear, and can reported totals be reconciled with your analytics?
Diagnostic depth: Can you break performance down far enough to separate delivery, audience, creative, placement, and signal problems?
Auditability: Is there a change history showing who changed a setting, when it changed, and what the previous value was?
Portability: Can you export campaign, delivery, and conversion data in a form your reporting system can retain and compare?
Governance: Can access be limited by role, and can a second operator review high-impact changes before they affect delivery?

If a required control is missing or unverified, limit the test to exposure your organization can tolerate and define a manual stop path before launch. A report that arrives quickly is helpful, but speed does not replace enforcement, audit history, or the ability to reconcile results.

Keep paid AI advertising separate from GEO and earned AI visibility as well. An ad impression purchased inside an AI experience is not proof that the brand was selected, cited, or recommended organically by a model. Give paid campaigns their own attribution labels, landing-page tracking, and reporting view so an increase in paid traffic cannot be presented as improved generative visibility.

Before your next optimization cycle, open one consequential campaign and record its monthly spending ceiling, the reason for its schedule, its maximum active-day exposure, its consent-signal coverage, the controls that can stop delivery, and the delay in its reporting. Resolve any unknown that could change the meaning of the results. Once those controls are observable and enforceable, bid and creative changes can produce evidence you can actually use.

References

April 22, 2026

Category: Advertising

Decide what AI may decide before you activate it

Fix the input layer before generating ads or answers

Give Google and Microsoft different operating roles

Measure decisions, not just automated output

Key takeaways

References

Key takeaways for your platform decision

Separate decision support from automated execution

Audit the data contract before evaluating the AI

Use conversational dashboards as an investigation layer

Run a bounded pilot before expanding authority

References

Give the AI an optimization contract before giving it data

Build a business outcome layer across five data domains

Reconcile the systems without forcing their numbers to match

Expand the agent’s permissions only after the data proves reliable

Stage 1: Observe in read-only mode

Stage 2: Produce structured recommendations

Stage 3: Allow bounded execution

Key takeaways: your AI advertising readiness check

References

Key takeaways

The change applies to a specific navigation action

Why the revenue effect will vary between publishers

Audit the change without confusing correlation for cause

A quick diagnosis matrix

Respond by improving the journey, not recreating the trigger

Set a new baseline before making an optimization decision

References

The platform change is access, not proof of performance

Choose CPC or CPM from the business objective

Build a pilot that can answer one decision

Keep paid performance separate from AI visibility

Automate reporting before you automate campaign control

Key takeaways

References

What self-serve buying changes, and what it does not

Decide whether your business is ready to test

Build the first campaign around a falsifiable hypothesis

Make measurement trustworthy before optimizing bids

Key takeaways

References

Separate channel discovery from budget proof

Write the measurement contract before the first impression

Read early AI ad performance without fooling yourself

Build an evidence ladder instead of trusting one model

Layer 1: instrumentation and platform diagnostics

Layer 2: controlled experiments

Layer 3: media mix modeling

Key takeaways for your next platform test

References

Use the early data as a format signal, not an ROI benchmark

Write an answer fragment, not a compressed search ad

Build each variation from three prompt components

Test in an order that explains the result

Measure prompt coverage and response duplication before calling it reach

Coordinate ChatGPT ads with AEO and GEO without merging the KPIs

Key takeaways for your first ChatGPT ad pilot

References

The expansion addresses inventory, not the whole advertising case

Key takeaways

Build the pilot around one commercial decision

Demand measurement answers before you demand scale

Keep paid ChatGPT reach separate from organic AI visibility

Use a readiness gate before committing budget

References

Key takeaways

Audit the whole control chain before touching bids

Scheduled campaigns need a spend-concentration audit

Consent diagnostics tell you whether the optimizer can learn

Prove an AI ads manager can control delivery before scaling it

References