Tag: AI Discovery

AI Product Discovery Tracking: A Practical Measurement Plan

You can rank well in traditional search, maintain a complete product feed, and still have no clear answer to a basic question: when someone asks an AI assistant what to buy, does your product appear?

AI product discovery tracking closes that gap. It records how individual products appear in shopping-oriented answers, separates visibility from accuracy, and gives you evidence for deciding what to fix. The goal isn’t to collect screenshots of flattering mentions. It’s to understand which SKUs enter the recommendation set, under which buying conditions, and what happens next.

Track the buying decision, not a single brand mention

A brand-level visibility score is too blunt for ecommerce. An assistant can mention your company while recommending the wrong product, an unavailable variant, or an item that doesn’t satisfy the shopper’s constraints. That mention looks positive in a dashboard but does little for the buyer.

Use the SKU, or the most stable product identifier available, as the primary measurement unit. Connect each observation to the exact prompt, platform, market, date, product variant, cited page, merchant, and answer text. This lets you distinguish a product-level problem from a broad brand problem.

The relevant measurement surface is also wider than one chatbot. Commercial monitoring is now offered for SKU-level visibility across ChatGPT Shopping, Alexa for Shopping, Perplexity, and Google AI Mode. Keep results separate by platform. Combining them into one score too early can hide the fact that a product is consistently discoverable in one environment and absent in another.

For every observed answer, classify five different outcomes:

Presence: Did the brand, product family, or exact SKU appear?
Prominence: Was it a primary recommendation, a secondary option, or a passing reference?
Qualification: Did the answer connect the product to the shopper’s stated use case, budget, features, or constraints?
Representation: Were the name, variant, attributes, availability, and other offer details accurate?
Handoff: Did the answer provide a citation, merchant, product page, or another usable route toward purchase?

These outcomes answer different questions. Presence tells you whether the product entered the answer. Qualification tells you whether the system understood why it fits. Representation reveals whether the underlying product information is coherent. Handoff shows whether visibility can plausibly lead somewhere useful.

Build the measurement specification before choosing a tool

A tracker can automate collection, but it can’t decide what your business means by visibility. Write the measurement specification first. Otherwise, a vendor’s default prompts and scoring system will quietly become your strategy.

Create a product identity registry. Give every tracked item a canonical name and identifier. Add brand names, model names, common aliases, parent-child variants, canonical product URLs, and the merchants authorized to sell it. This prevents a shortened model name or alternate spelling from being counted as a different product.
Define the eligible product set for each prompt. A recommendation is only meaningful if the SKU could reasonably satisfy the request. If a prompt requires a feature the product doesn’t have, its absence isn’t a visibility failure.
Group prompts by buyer intent. Keep category discovery, feature-led discovery, problem-led questions, comparisons, branded validation, and purchase-ready requests in separate groups. A product that performs well on branded prompts but disappears from category discovery has an acquisition problem that a blended score will conceal.
Record the test environment. Store the platform, location or market setting, language, session state when controllable, device context when relevant, and collection time. If a condition can’t be controlled, label it unknown rather than assuming consistency.
Freeze a core prompt panel. Run the same core prompts repeatedly so changes are comparable. Maintain a separate exploratory panel for emerging language, new use cases, seasonal needs, and questions discovered in customer research.
Define what counts before collecting results. Decide how aliases, bundles, parent products, variants, repeated mentions, unordered lists, and cited merchant pages will be handled. Apply those rules to your brand and competitors alike.

Prompt wording needs particular care. “Best running shoe” and “running shoe for a wide forefoot on wet pavement” don’t represent the same decision. The second prompt supplies constraints that can change which products are eligible. Preserve those constraints in your reporting instead of collapsing everything into a generic keyword.

Don’t let exploratory prompts replace the fixed panel. New prompts improve coverage, but changing the entire prompt set between measurement periods destroys comparability. Use the fixed panel to detect movement and the exploratory panel to find new opportunities.

Use a scorecard that keeps visibility, accuracy, and outcomes separate

No single metric can represent the whole discovery journey. A useful scorecard shows where a product was eligible, whether it appeared, how it was described, and whether the answer created a usable path forward.

Metric	How to calculate it	What it helps you decide
Eligible prompt coverage	Eligible prompts containing the tracked SKU divided by all prompts for which that SKU was eligible	Whether the product enters relevant recommendation sets
Recommendation share	Recommendations of the tracked product divided by all product recommendations in the same prompt set	How often your product appears relative to alternatives
Primary recommendation rate	Answers treating the SKU as a leading option divided by answers mentioning it	Whether mentions are prominent or incidental
Qualification rate	Mentions that accurately connect the SKU to the prompt’s constraints divided by all SKU mentions	Whether the system understands the product’s relevant use cases
Attribute accuracy rate	Verified product claims divided by all checkable claims made about the SKU	Whether conflicting or incomplete product information needs attention
Handoff rate	SKU mentions with a usable citation, merchant, or product destination divided by all SKU mentions	Whether discovery can progress toward consideration or purchase
Competitor overlap	Eligible prompts where your SKU and a named competitor both appear divided by eligible prompts where either appears	Which products compete in the same answer contexts
Downstream engagement	Observed visits and commerce events attributed to an identifiable AI handoff	Whether measurable discovery activity contributes to business outcomes

The denominator matters. If you calculate coverage across prompts where a product couldn’t satisfy the stated need, you manufacture a weakness. If you count every brand mention as a product recommendation, you manufacture success. Keep the eligibility rules visible next to the score.

Preserve the underlying observations as well as the aggregate metrics. Store the returned product names, supporting language, cited URLs, merchants, competing products, and factual errors. When a score changes, you should be able to inspect the answers behind it.

Keep business outcomes in a separate layer. An AI mention isn’t a sale, and a sale that follows an AI interaction may not be fully attributable. Where a link, referral, or tagged destination is observable, connect it to product views, cart activity, and purchases. Where the handoff can’t be observed, report the outcome as unknown. Turning unknown activity into zero activity makes the dashboard look precise while reducing its usefulness.

Diagnose whether the failure is eligibility, selection, or representation

A three-stage product recommendation pipeline filters products, selects a smaller group, and displays them in translucent answer cards.

A missing product doesn’t tell you why it was omitted. The output gives you a symptom, not a causal explanation. Use it to form a testable hypothesis, then inspect the product information and competitive context that could support or contradict that hypothesis.

Eligibility failure: the product isn’t understood as a candidate

If the SKU is absent from non-branded prompts even though it genuinely meets their constraints, check whether its identity and qualifying attributes are expressed consistently. Review the visible product page, structured data, commerce feeds, variant records, category assignments, and merchant listings. Names, identifiers, sizes, colors, prices, availability, and feature claims shouldn’t contradict one another.

JSON-LD belongs in this audit, but don’t treat schema as a magic visibility switch. Its job is to express product information in a machine-readable form. It should match the visible page and the current offer data. If the markup describes a different variant or stale availability, adding more markup compounds the ambiguity.

Selection failure: the product is known but rarely recommended

A product may appear for branded validation prompts yet lose generic category, comparison, or problem-led prompts. That pattern suggests the system can identify the item but doesn’t consistently connect it to the buyer’s decision criteria.

Build a gap matrix from the actual answers. Put the prompt constraints in rows and the recommended products in columns. Record the reasons given for each recommendation. Then compare those reasons with claims your product can substantiate. If an important, verifiable attribute is missing from your product page or expressed only in an image, make it clear in the visible copy and structured product information. If your product doesn’t meet the criterion, don’t manufacture a claim to fit the prompt.

Representation failure: the product appears with incorrect details

Incorrect model names, mixed variants, stale offer details, or unsupported attributes are not positive visibility. Capture every checkable claim in the answer and compare it with the canonical record. Then locate conflicts across the pages, feeds, markup, and merchant data you control.

Correct the canonical product information before trying to increase mention volume. More exposure for a misrepresented SKU can send a shopper toward the wrong variant or create expectations the product can’t meet. Keep a record of the incorrect answer and the correction date so later observations can be evaluated against the change.

Turn tracking into a controlled optimization loop

An unbranded product sits at the center of a circular testing and optimization process with inspection, measurement, adjustment, and verification stations.

AI outputs can vary between runs, so a single before-and-after query is weak evidence. Treat optimization as repeated observation around a documented change.

Capture the baseline. Run the fixed prompt panel and preserve the complete responses, not just the calculated scores.
Choose one failure class. Decide whether you’re testing product identity, attribute completeness, use-case relevance, comparison content, offer consistency, or another specific hypothesis.
Change one information layer where practical. If you rewrite the page, replace the feed, alter structured data, and change merchant listings simultaneously, you may improve visibility without learning which correction mattered.
Log the deployment. Record the affected SKU, URLs, fields, platforms, markets, and publication time. Include rollbacks and feed errors in the same log.
Repeat the same core observations. Keep prompts, eligibility rules, and classification logic stable. Evaluate whether the direction of change persists across repeated collections.
Compare unaffected products. Similar movement across changed and unchanged SKUs may indicate broad output variation or a platform-level shift rather than the effect of your work.
Promote only durable findings. When an improvement continues to appear under the same measurement conditions, apply the lesson to other eligible products and keep monitoring for representation errors.

Report platform results independently and segment them by intent. A gain in branded prompts doesn’t prove stronger category discovery. A gain on one assistant doesn’t prove that another system changed. The useful reporting unit is the intersection of platform, market, intent group, and SKU—not an unsupported universal visibility score.

Competitor tracking should support diagnosis rather than imitation. Note which products recur, which buyer constraints they are associated with, what supporting pages are cited, and where their descriptions are inaccurate. This reveals the information standards operating within a prompt set. It doesn’t prove that copying a competitor’s wording, markup, or content structure will reproduce its visibility.

Key takeaways for a tracker you can trust

Measure exact products and variants, not brand mentions alone.
Define SKU eligibility for each prompt before treating an omission as a failure.
Separate presence, prominence, qualification, factual accuracy, handoff, and business outcomes.
Keep a stable core prompt panel for comparison and a separate exploratory panel for discovery.
Preserve raw answers and cited destinations so every aggregate score can be audited.
Use observed outputs to form hypotheses; don’t claim they reveal a ranking system’s hidden cause.
Audit visible content, structured data, feeds, and merchant records for consistency when product identity or attributes are wrong.
Evaluate changes through repeated observations and unaffected comparison products, not one favorable response.

Start with a narrow set of commercially important SKUs and the prompts for which they are genuinely eligible. Build the identity registry, freeze the core panel, and collect a baseline before editing anything. Your first useful result won’t be a universal visibility score. It will be a defensible answer to which product is missing, where it is missing, and what evidence you need to test next.

References

HiGoodie Blog — Agentic Commerce Suite: Track AI Product Discovery

October 9, 2026

How to Win Visibility in Agent-Driven Search

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

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

Search visibility now has four separate gates

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

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

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

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

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

Publish the facts agents need to qualify you

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

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

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

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

Use a suitability page pattern that answers the whole decision

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

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

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

Make proof machine-readable without hiding caveats

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

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

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

Remove the blockers between selection and completion

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

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

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

Audit the complete transaction, not just the landing page

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

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

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

Use audience preference where the platform supports it

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

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

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

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

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

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

Measure agent visibility as a decision path

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

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

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

Key takeaways

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

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

References

October 6, 2026

ChatGPT Virtual Try-On: An Ecommerce Optimization Playbook

You may be asking a deceptively simple question: what should your ecommerce team change now that a shopper can preview a product inside ChatGPT? The answer isn’t to add AI shopping phrases to every page. Virtual try-on moves part of product evaluation upstream, before the shopper reaches your store.

Your job is to make each product understandable during discovery, visually recognizable during evaluation, and easy to buy when the shopper finally reaches the product page. That requires coordinated work across imagery, catalog data, structured data, fit guidance, landing-page UX, and measurement.

Virtual try-on changes where product evaluation happens

On eligible clothing and accessory listings, ChatGPT can display a Try on button that lets a shopper take or upload a selfie. ChatGPT Images then generates a visualization of that person wearing the item. A product doesn’t have to originate in a ChatGPT recommendation: the shopper can also upload an image or screenshot of something found elsewhere and request a virtual try-on.

That creates a shopping path that may look like this:

The shopper describes the clothing or accessory they want.
ChatGPT surfaces products that appear relevant.
The shopper visualizes a candidate product on their own image.
They compare it with other possibilities.
They save promising items or visit a merchant to inspect the offer and buy.

Product discovery and visual evaluation can therefore happen within the same conversation. ChatGPT also lets shoppers save products to Favorites, organize them into Library folders, and return to them on mobile or the web. A recommendation is no longer necessarily followed by an immediate click. The shopper may build a shortlist first and arrive at your store later with a narrower set of questions.

For an ecommerce SEO or GEO team, that changes the optimization target. You need to support three decisions:

Recognition: Can the product be distinguished from superficially similar items?
Evaluation: Can the shopper understand its color, cut, pattern, material, and available variations?
Completion: Can your product page resolve size, price, availability, delivery, and return questions without introducing contradictions?

This doesn’t make the product page less important. It gives the page a more demanding role. The visitor may already like the apparent look; the merchant must now establish exactly what is being sold and reduce the remaining purchase risk.

The screenshot workflow matters just as much as native product discovery. A shopper may encounter your item in search, on a marketplace, in a social post, or on another page before bringing its image into ChatGPT. Your visual assets need to remain recognizable when separated from their original context.

Build a coherent product record, not an AI optimization gimmick

An unbranded sneaker is surrounded by connected product images, color swatches, size cells, packaging, and a product card.

There is no established Try on optimization formula, required image dimension, or special schema property that guarantees eligibility. Treat promises of guaranteed inclusion through a single field with skepticism. The practical goal is coherence across the product image, visible copy, variation selector, commerce feed, and structured data.

Use images that still make sense outside the product page

Start with the main image because it is the most likely visual shorthand for the product. It should make the item easy to identify without forcing a system or shopper to infer which object is for sale.

Show the complete garment or accessory clearly in at least one image.
Keep the product visually distinct from props, backgrounds, and neighboring items.
Use the correct image for each color or pattern variation.
Provide additional views when the front image hides important construction, shape, fastening, or pattern details.
Keep image treatment consistent enough that a shopper can recognize the same item across a listing, a screenshot, and the product page.
Avoid putting essential product facts only inside image text. Those facts also belong in visible HTML.
Write useful alternative text for accessibility and page comprehension, but don’t claim that alt text controls a virtual try-on rendering.

Run a simple crop test. View the product image without its title, price, or surrounding page. Ask whether a person could identify the item type, dominant color, pattern, and intended variation. If the answer depends on the missing copy, the image is doing too little. If several products compete for attention, it is doing too much.

Don’t replace accurate catalog photography with speculative AI composites merely to appear AI-ready. A visualization system needs a dependable representation of the product. Your controlled assets should establish ground truth, while the generated try-on remains a separate, personalized interpretation.

Make attributes explicit and consistent

Product copy should identify the attributes that distinguish the item. A poetic collection name may support branding, but it shouldn’t carry the entire descriptive burden. Pair it with plain product language that states what the shopper is looking at.

Use a stable product name, brand, and product type.
Name the actual color as well as any branded color name.
Describe the material or fabric without making unsupported performance claims.
State the silhouette, length, pattern, closure, and other decision-relevant features when they apply.
Map every displayed image to the correct selectable variation.
Keep price, currency, availability, and condition aligned wherever those fields appear.
Use valid product identifiers consistently. Never invent an SKU, GTIN, or other identifier to fill an empty field.
Provide measurements and size information in accessible page content rather than relying on an image alone.

Structured data should mirror that visible record. Product and Offer JSON-LD can express product and commercial facts in a machine-readable form, but markup is not a substitute for accurate page content and isn’t evidence of virtual try-on eligibility. If the page shows one price while the Offer markup publishes another, the problem isn’t a missing AI tactic; it is a conflicting product record.

Check variation handling closely. The selected color, image, SKU, availability, price, and structured data should refer to the same offer. If your implementation updates some of those fields dynamically, verify the rendered state rather than reviewing only the page template or source code. A technically valid block of JSON-LD can still describe the wrong variant.

Audit the complete product path

Use this sequence on representative clothing and accessory templates:

Open a live product and select each meaningful variation.
Compare the selected option with the main image, gallery, visible name, price, stock state, and product identifier.
Inspect the rendered Product and Offer data for the same variation.
Check the size guide, measurements, material details, delivery information, and return policy.
Capture the main product image as a shopper might encounter it elsewhere and verify that the product remains recognizable.
Resolve contradictions before adding more copy or markup. Consistency is the prerequisite, not the finishing touch.

This audit is useful beyond ChatGPT. It removes ambiguity from the catalog record that your own customers, feeds, analytics, search systems, and other shopping interfaces must interpret.

Separate appearance visualization from fit, then strengthen the handoff

The most important boundary is also the easiest one to blur: virtual try-on is a visualization, not a fitting room. The generated result may not represent the shopper or product exactly and doesn’t guarantee size or fit. Merchant measurements, product details, and return policies remain part of the buying decision.

Think of the preview and product page as answering different questions:

Shopper question	Best answer surface	What the answer must communicate
How might this style look on me?	Virtual try-on visualization	A directional visual impression, not a promise of exact appearance or fit
Which size should I order?	Merchant size guide and measurements	Clear measurement definitions, units, garment dimensions, and relevant sizing notes
What exactly am I buying?	Product page and variation selector	The selected color, material, construction, images, price, and availability
What happens if it isn’t right?	Delivery and return information	Applicable conditions, timing, process, and customer costs

Your size guidance needs enough context to be usable. Distinguish body measurements from garment measurements. Name the measurement points and units. Explain relevant stretch, cut, or layering considerations without pretending they can predict an individual’s fit. If sizing differs by product line or market, put the correct guide on the affected product rather than sending everyone to a generic chart.

The landing page should preserve continuity with what the shopper evaluated. The same variation should be easy to recognize, and the page should expose the remaining decision information without making the visitor hunt for it.

Keep the product name and selected variation visible near the main image.
Show current price and availability for that variation.
Place the size selector close to the relevant size guide.
Make material and care information easy to scan.
Present delivery and return terms before the shopper commits to checkout.
Explain unavailable variations honestly rather than silently switching the selection.
Keep mobile layouts usable because the shopping features are available on both mobile and web.

Favorites add another handoff consideration. A shopper may save an item, compare it with alternatives, and return after the original discovery session. Stable product URLs, persistent identifiers, current inventory, and clear replacement behavior matter more than a landing experience built only for an immediate click.

If you describe AI visualization on a page you control, keep the claim narrow. Plain language such as “The preview is a visual approximation; check the product measurements and return terms before ordering” sets the right expectation. Don’t call a generated image proof of fit, exact drape, precise color reproduction, or guaranteed appearance.

Measure discovery, merchant handoff, and post-purchase outcomes

Referral traffic alone will not describe the full effect. A shopper can upload a product screenshot found elsewhere, evaluate it in ChatGPT, save it, and return by another route. Some influence will therefore be invisible to your analytics or appear under a later source.

Observe visibility without treating one answer as a ranking report

Create a repeatable set of prompts based on real customer language. Include product type, material, color, occasion, style, and other attributes your catalog genuinely supports. Record whether your products appear, whether the correct variation is represented, whether the cited destination resolves correctly, and whether a Try on option is shown when relevant.

Use those checks diagnostically. They can expose ambiguous naming, weak imagery, broken destinations, and inconsistent variants. They do not establish universal market share, a permanent ranking, or the cause of a recommendation. Avoid turning a favorable answer from one session into a performance claim.

Instrument the merchant handoff

Preserve raw referrer information where your analytics and consent setup permit it, and group identifiable ChatGPT visits without overwriting the underlying source. Then evaluate the onsite sequence rather than counting sessions alone.

Which products receive identifiable AI referral visits?
Does the landing URL resolve to the intended product and variation?
Do those visitors use the gallery, variation selector, or size guide?
Where do they leave the product and checkout funnels?
Do they add the evaluated item to the cart, or switch to another variation or product?
Are analytics events firing consistently across mobile and desktop?

A high click count with frequent variant switching may indicate that the upstream image or product description set the wrong expectation. Strong product-page engagement with weak size selection may point to incomplete fit guidance. Treat these as diagnostic signals to investigate, not automatic proof of causation.

Connect the experiment to business outcomes

Virtual try-on is intended to help a shopper evaluate a product, so the useful outcomes sit deeper than impressions. Track completed purchases, cancellations, exchanges, returns, and available reason codes for the affected products. A generated preview that increases curiosity but creates a mismatch at delivery is not an unqualified success.

Use a controlled improvement cycle:

Save a baseline for the selected product group, including its images, visible attributes, structured data, funnel behavior, and return outcomes.
Fix one interpretable layer, such as variation-image mapping or measurement content.
Repeat the same visibility checks and review the same onsite events.
Annotate concurrent changes in price, promotion, inventory, seasonality, and delivery terms.
Read the result as directional unless the design actually isolates the changed variable.

Don’t label every post-change sale as AI-driven revenue. Report what you can observe directly, separate identifiable referrals from inferred influence, and name the blind spots. Favorites activity inside ChatGPT and screenshot-based exploration are not merchant-side analytics events.

Key takeaways

ChatGPT virtual try-on can combine product discovery, selfie-based visualization, comparison, and shortlisting before a merchant visit.
A shopper can upload a product image found elsewhere, so clear and recognizable assets matter beyond native ChatGPT listings.
There is no basis for promising eligibility from one schema field, keyword, image treatment, or feed attribute.
Product images, visible content, variations, commerce feeds, and Product and Offer structured data should describe the same item and offer.
Virtual try-on visualizes a possible look; merchant measurements, size guidance, product facts, and return terms must handle fit and purchase risk.
Measure visibility, onsite behavior, purchases, and returns while acknowledging that screenshot and Favorites activity may leave no direct referral trail.

Start with a representative clothing or accessory template and follow one product from its standalone image through variant selection, JSON-LD, size guidance, return information, and analytics events. Fix every contradiction you find before scaling the audit across the catalog. That gives you a durable commerce foundation whether the next shopper discovers the product through ChatGPT, another AI interface, a conventional search result, or a saved screenshot.

References

Search Engine Land — ChatGPT adds virtual try-on for clothes and accessories

October 2, 2026

How to Hire Senior SEO Talent for Judgment, Not Tasks

You are not hiring a human backlog. You are hiring someone to decide which search problem is real, which evidence deserves trust, and which work should win scarce support.

A polished candidate can discuss crawling, content, links, reporting, and AI visibility. The harder test begins when those signals disagree. Organic clicks can fall while conversion and quality indicators improve. Visibility can grow in a market the business cannot serve. An impressive AI score can have no demonstrated relationship to revenue. Your hiring process needs to reveal who can navigate those conflicts without retreating into a generic checklist.

Start with the decision this person must improve

Before writing the job description, finish this sentence: We need this person to help us decide and execute…

The words that follow should describe a business problem, not an SEO department. You may need more qualified demand in a particular industry, better conversion from existing traffic, clearer priorities for a neglected backlog, or someone who can move discovery work through engineering, content, product, PR, and legal. You may need to learn whether visibility in ChatGPT and other AI experiences produces valuable customer behavior. Those are different mandates.

Build a short role charter before listing responsibilities. It should define:

The business problem: What is currently underperforming, uncertain, or blocked?
The outcome: What should improve for customers or the business if the hire succeeds?
The constraints: Which budgets, markets, technical limits, compliance requirements, or capacity limits are real?
The dependencies: Which teams must approve, build, publish, measure, or support the work?
The decision rights: What can this person prioritize directly, and where must they persuade others?
The non-goals: Which adjacent responsibilities belong to other people?

The non-goals matter. A description that combines technical SEO, content strategy, AI discovery, analytics, conversion optimization, link acquisition, reporting, and project management may conceal several jobs inside one salary. It also makes evaluation incoherent: one interviewer rewards technical depth, another expects an editorial strategist, and a third wants a cross-functional program leader.

Decide whether you primarily need a specialist who will complete defined work or a leader who will determine what the work should be. A senior discovery leader may not personally execute every migration ticket or content brief. They should be able to diagnose the system, select a defensible sequence, obtain support, and keep the work connected to a business outcome.

Build a scorecard that rewards judgment

Five symbolic assessment objects form a balanced structure while an unnecessary metallic piece is set aside.

Technical competence remains a threshold requirement. A senior SEO leader must recognize technically plausible explanations, interrogate the right systems, and understand the consequences of a recommendation. But technical fluency should not consume the entire scorecard. One useful hiring model treats SEO- and AI-specific knowledge as roughly 25% of what makes a senior discovery hire effective, with the rest carried by critical thinking, communication, persuasion, prioritization, and business judgment. That percentage is not a universal formula. It is a practical guardrail against hiring the candidate with the largest vocabulary.

Use evidence-based dimensions instead of adjectives such as strategic, data-driven, or collaborative. Those words are too easy to claim and too difficult to score consistently.

Dimension	What to ask the candidate to do	Strong evidence	Risk signal
Problem framing	Interpret a situation in which search and business metrics disagree	Separates the observed symptom from the decision the business must make	Accepts the prompt’s framing and immediately recommends familiar tactics
Measurement judgment	Identify what must be validated before comparing performance	Questions tracking, consent, definitions, time comparisons, and the relationship between proxy and outcome metrics	Treats every dashboard value as equally reliable and meaningful
Prioritization	Choose work under a real resource constraint	Names what will be deferred, explains the opportunity cost, and states what could change the order	Labels most of the backlog urgent or critical
Business connection	Map search demand to capacity, conversion, and revenue	Distinguishes available demand from demand the business can profitably serve	Treats rankings, traffic, or AI mentions as the final objective
Influence	Explain the same recommendation to technical and commercial stakeholders	Changes the language and level of detail while preserving the reasoning	Uses channel jargon in place of a business case
Technical and AI literacy	Develop and test plausible causes across conventional and AI-mediated discovery	Knows what evidence would support or falsify each explanation	Repeats platform announcements or best practices without connecting them to the case

Listen for causal reasoning. A candidate should be able to say: this observation could have several causes; this is the evidence that would separate them; this decision is safe while we investigate; and this is the point at which we would change course. Memorized recommendations rarely contain that structure.

Do not penalize a candidate for challenging the premise. Senior judgment often appears as "I need more information." The phrase becomes useful only when the candidate identifies the missing information, explains why it changes the decision, and offers a provisional path instead of stopping the conversation.

Use an ambiguous work sample instead of a trivia test

A candidate sorts ambiguous evidence cards and selects one resource token while two interviewers observe.

A realistic exercise should contain enough evidence for a recommendation and enough ambiguity to make a checklist inadequate. Keep it close to your operating environment, but fictionalize sensitive data so every candidate receives the same case.

A useful brief could contain these conditions:

Organic clicks are down, while conversion and customer-quality indicators are up.
Keyword trends and an AI visibility score are available, but neither has been connected conclusively to the business outcome.
Some markets have unused service capacity, while others cannot absorb much more demand.
An analytics or cookie-consent change may have affected the year-over-year comparison.
Engineering can contribute only 40 hours during the quarter.

Ask the candidate to make a recommendation, not produce an audit. The deliverable should require them to:

Define the decision the business actually needs to make.
Identify the assumptions and measurement questions that could materially change that decision.
Offer a working recommendation while those questions are being resolved.
Allocate the constrained engineering capacity and state what will not be done.
Choose outcome measures that distinguish commercial progress from visibility alone.
Explain what new evidence would cause the plan to change.

A weaker response usually expands the scope. It proposes a technical audit, content refresh, cleanup program, link initiative, and AI visibility project at the same time. Every tactic may be legitimate in isolation, but the candidate has not shown why any of them deserves priority in this situation.

A stronger response first tests whether the apparent decline is a problem. If conversions and customer quality are improving, the lost clicks may include less valuable demand, the measurement may have changed, or another part of the journey may be performing better. The candidate should not assume which explanation is correct. They should specify how to tell them apart.

Market capacity creates another revealing choice. Improving visibility where the business cannot serve more customers may produce attractive charts and operational frustration. A candidate with business judgment will examine where additional demand can become a completed sale, appointment, subscription, or other real outcome. They may prioritize a market with unused capacity even when its search opportunity looks less glamorous.

Treat the AI visibility metric the same way. It is a hypothesis-generating signal until the candidate can show a credible relationship to customer discovery and business results. The right next step may be a bounded test, better attribution, or closer analysis of the queries and citations involved. It is not automatically a mandate to maximize the score.

Use a short panel discussion after the exercise. Grade the candidate’s reasoning, questions, tradeoffs, and communication – not whether the final recommendation matches an answer your team decided in advance. If there is only one answer you will accept, you are testing compliance rather than judgment.

Interview for tradeoffs, influence, and restraint

The best interview questions make the candidate choose. Broad prompts such as "How would you improve our SEO?" reward confident improvisation. Constrained prompts reveal whether the person can protect the business from low-value work.

Questions that reveal diagnosis

Organic traffic has declined while qualified conversions have improved. Under what conditions is that good news, bad news, or a measurement problem?
Which data would you validate before comparing this period with the previous one, and why?
What finding would make you decide not to run a broad technical audit?
One market has a visibility gap but no service capacity. Another has spare capacity but lower apparent search demand. How would you choose where to work?
Our AI visibility score increased. What would you need to see before treating that increase as business progress?
Which recommendation would you make now, and which decision would you deliberately postpone?

Do not judge the candidate by the number of questions asked. Judge whether each question can change the decision. Asking about a consent implementation that may invalidate a trend is valuable. Asking for every report the company owns may simply delay commitment.

Questions that reveal leadership

Engineering gives you 40 hours this quarter, while the proposed work would take six months. What ships, what waits, and what do you tell the executive team?
Explain your recommendation first to a CFO and then to a CTO. What changes in the explanation, and what remains constant?
A technically sound recommendation is blocked by product or legal. How do you determine whether to modify it, build a stronger case, or stop pursuing it?
When can conversion, inventory, follow-up, reputation, or product preference be a more important discovery constraint than crawlability?
Tell us about a recommendation you would reject even if it increased rankings or visibility. What makes the tradeoff unattractive?

A senior leader should be able to operate outside the SEO silo. Search performance connects to product experience, customer support, paid landing pages, brand reputation, conversion paths, operational capacity, and revenue. That does not mean the SEO leader owns every function. It means they can recognize when the limiting factor sits elsewhere and bring the right owner into the decision.

Restraint is part of the job. If the candidate describes six months of work as critical despite a narrow engineering allowance, they have not prioritized. They have reformatted the backlog. Look for explicit deferrals, sequencing logic, reversible first moves, and thresholds that would justify further investment.

During the debrief, record evidence before discussing overall impressions. Ask what assumption the candidate challenged, what they chose not to do, how they connected discovery to business capacity, and whether a non-specialist could follow the logic. A charismatic presentation should not compensate for an undefined problem or an unbounded plan.

Key takeaways and your next move

Define the business decision before defining the SEO role.
Treat technical and AI fluency as essential foundations, not the whole senior-level scorecard.
Use conflicting metrics and real constraints to expose how a candidate thinks.
Reward requests for more information when they identify decision-changing evidence and still produce a provisional recommendation.
Make candidates connect search and AI visibility to capacity, conversion, customer quality, and revenue.
Grade tradeoffs, communication, and restraint rather than agreement with a predetermined answer.

Before you publish the role, replace its opening list of channel responsibilities with the decision this person must improve. Then replace the generic take-home audit with an ambiguous case drawn from that decision. The candidate who clarifies the problem, makes a choice, and earns support for it is showing the judgment you are actually hiring.

References

Search Engine Land – The best SEO hires bring judgment, not just expertise

October 1, 2026

Google Discover’s “Dive Deeper” Test: A Publisher Playbook

If Google Discover sends you meaningful traffic, the new “Dive deeper” experiment deserves a measurement plan, not a panic rewrite. The AI-powered card can occupy a feed position that might otherwise show publisher content, then answer part of the user’s need before offering links to the web.

Your immediate job is to separate a plausible traffic risk from an observed traffic loss. Establish a Discover baseline, isolate the content most exposed to the test, and make the value of clicking unmistakable. You can do all of that without guessing at an undisclosed ranking factor or inventing a new schema strategy.

What the test changes in the Discover journey

A normal publisher card offers a relatively direct choice: open the content or continue scrolling. “Dive deeper” introduces another route. A person can enter an AI-generated topic overview and then decide whether one of its linked stories, community reactions, or pieces of original reporting deserves another click.

Google describes those destination links as prominent, but prominence doesn’t remove the added decision point. The overview itself may satisfy a casual reader. A publisher also has to win selection among several related resources rather than win the initial feed interaction alone.

That creates three distinct risks for publishers:

Displacement: the topic card may use feed space that could have carried a publisher’s individual item.
Intermediation: the user reaches an overview before reaching a publisher, adding another choice between discovery and the site visit.
Substitution: the generated overview may provide enough context that some people no longer need the underlying coverage.

Those are mechanisms, not measured outcomes. Google is starting the experiment with videos and trying multiple designs. That makes it premature to treat the interface as a completed rollout, assume every Discover user can see it, or attribute every traffic decline to it.

Measure the test without mistaking correlation for cause

Two streams of content tiles pass through separate test pathways while a magnifying lens and measuring vessels represent controlled analysis.

A total traffic chart won’t tell you whether “Dive deeper” affected your site. Publishing volume, subject mix, headline quality, seasonality, and changing audience interest can all alter the same line. You need a Discover-specific view and enough page-level detail to identify the shape of the change.

Preserve your baseline. Export Discover clicks, impressions, click-through rate, and landing-page performance from Google Search Console. Use a period long enough to show your site’s normal range rather than selecting only a convenient high point.
Record editorial context. Annotate major changes in publishing frequency, topic selection, video output, headlines, and distribution. Otherwise, a newsroom decision can look like a platform effect.
Separate video-led content. Because the experiment begins with videos, compare pages built around video with the rest of your Discover inventory. Keep the classification consistent; don’t move a page between groups merely because its performance changed.
Inspect pages before aggregates. Identify which landing pages lost impressions, which retained exposure but lost clicks, and which continued to convert after the visit. A sitewide average can conceal all three patterns.
Connect visits to outcomes. Pair Discover traffic with the action that matters on your site, such as engaged reading, registration, subscription, or revenue. Fewer visits would still be harmful at scale, but a publisher should know whether the remaining visits became more or less valuable.

Use the pattern below as a diagnostic guide, not as proof of exposure to the experiment.

Pattern in your data	What it may indicate	What to check next
Impressions fall while CTR stays near its normal range	Reduced feed exposure or weaker topic relevance	Compare publishing volume, subject mix, and video-led versus non-video pages
Impressions hold while CTR falls	A more competitive or more satisfying interface, or weaker packaging	Review the affected headlines, media, and the distinctive value promised by each page
Clicks fall while value per visit holds	A volume problem rather than a visitor-quality problem	Model the total subscription or revenue impact and reduce channel concentration
Only a small group of pages declines	A page, format, or topic issue rather than a sitewide platform effect	Compare those pages with stable content before changing the whole editorial plan

If you cannot identify which users encountered “Dive deeper,” describe any relationship as an association. A decline that begins during a platform experiment is worth investigating, but timing alone doesn’t establish causation.

Give readers a reason to continue beyond the overview

A reader moves from a small translucent summary card into a series of deeper chambers filled with visual research and practical resources.

The wrong response is to make content longer or more mysterious. An overview competes most easily with generic coverage that repeats known facts. Your stronger position is content whose useful part cannot be reproduced by a short topic summary.

Google says the expanded experience can link to related stories, community reactions, and original reporting. Treat those labels as clues about the types of destination that can complete a reader’s journey, not as confirmed ranking factors.

Make the unique asset visible in the headline. Name the interview, analysis, data, demonstration, timeline, local detail, or expert interpretation the reader will receive. A broad topic label gives the overview little reason to send the user onward.
Put original evidence near the top. If the page contains reporting, show what was learned and how. Don’t bury the differentiating material beneath a generic explanation that an overview can already provide.
Define the unanswered question. A useful headline and opening should reveal what the short overview cannot settle: why an event happened, what changed, who is affected, how competing claims differ, or what the viewer can verify in the full video.
Match the promise to the page. A headline that implies original reporting must lead to original reporting. Artificial curiosity may win an occasional click, but it creates a poor destination and weakens the value of being selected.
Build a useful next step on your own site. Connect the landing page to genuinely related analysis, primary material, or an update path. If Discover supplies a more fragmented entry point, your internal journey has to restore context quickly.

For video-led pages, audit the complete package: title, thumbnail, opening text, video, transcript or summary, and supporting evidence. The page should make clear what the video contributes beyond the surrounding topic overview. Don’t assume that embedding a video makes otherwise generic coverage distinctive.

Do not invent a schema fix for a user-interface test

This is where technical teams can lose time. Google’s disclosed description of “Dive deeper” does not specify a new structured-data type, an opt-in setting, or a publisher control for the feature. There is therefore no responsible basis for promising that a markup change will secure placement or prevent summarization.

Keep existing Article or VideoObject markup accurate when those types properly describe the page. Make sure visible titles, dates, authorship, media, and structured properties agree. That is sound technical hygiene, but it shouldn’t be presented internally as a “Dive deeper optimization.”

Use this decision rule before approving Discover-related technical work:

If the change repairs inaccurate or inconsistent markup, make it.
If the change improves how people understand and navigate the page, evaluate it on that merit.
If the change depends on an undocumented “Dive deeper” signal, hold it until Google provides supporting guidance or your own controlled evidence justifies the work.

Also keep the product distinction clear in reports. “Dive deeper” is an experiment inside Discover; it is not evidence that every Discover card is being replaced, and it should not be casually relabeled as the search results feature commonly called AI Overviews. Blurring those surfaces makes your measurements and recommendations less reliable.

Reduce the business risk before the interface settles

You don’t need to predict the final design to manage the exposure. Start with channel concentration. Calculate how much traffic, engagement, subscription activity, and revenue comes from Discover, then identify the pages and formats responsible for most of that contribution.

Build scenarios from your own historical range rather than borrowing an arbitrary industry percentage. Your baseline scenario can reflect normal variation. A lower-range scenario can show what happens when Discover underperforms without disappearing. A stress scenario can show which editorial products become uneconomic if referral volume contracts materially.

Assign an action to each scenario before traffic moves. That action might be protecting distinctive reporting, changing the volume of generic video coverage, improving conversion on the visits you retain, or accelerating channels you control more directly. Email subscriptions, direct visits, feeds, memberships, and durable search demand won’t reproduce Discover’s feed distribution exactly, but they can reduce the damage caused by dependence on any single interface.

Avoid across-the-board cuts based on one weak reporting period. A narrow decline in commodity coverage calls for a different response from a broad loss of impressions across original work. The first may be a content-positioning problem; the second may justify a larger distribution and revenue review.

Key takeaways

“Dive deeper” inserts an AI-generated topic overview between parts of the Discover experience and publisher destinations, creating a credible risk of click compression.
The experiment begins with videos and may use multiple designs, so its current form should not be treated as a settled, universal rollout.
Track Discover impressions, clicks, CTR, landing pages, content format, and downstream value separately; aggregate traffic alone cannot diagnose the cause.
Make original evidence and the reason to continue beyond a summary explicit in the headline, opening, and page experience.
Do not promise a structured-data solution when Google has not identified special markup or a publisher control for the test.
Model Discover dependency now so your response is based on business impact rather than fear generated by an unfamiliar interface.

Start with a clean export of your current Discover performance. Classify the leading pages as video-led or non-video, note the distinctive value each one offers, and record the editorial conditions behind the baseline. If the interface begins affecting your audience, you will have evidence for a targeted decision instead of a reason to overhaul everything at once.

References

Search Engine Land – Google testing “Dive deeper” for Google Discover – it’s not good for publishers

September 21, 2026

Agentic Ecommerce: A Playbook for Discovery and Advertising

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

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

The shopping funnel now contains a conversation

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

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

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

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

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

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

Build a product record an agent can safely choose

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

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

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

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

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

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

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

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

Design conversational ads around the next unanswered question

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

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

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

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

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

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

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

Measure the path from question to profitable order

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

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

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

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

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

Key takeaways: use this launch sequence

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

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

References

September 16, 2026

How to Build Brand Visibility in Personalized AI Discovery

You search for your brand in an AI-assisted experience, see a reasonable answer, and assume visibility is handled. That check is too narrow once a discovery surface can remember what someone wants, favor publications they have chosen, or recommend different options under different contexts.

Your job is no longer to chase a single universal position. You need to make the brand eligible for the right discovery moment, easy for the audience to prefer, and difficult for an AI system to misrepresent. Here is a practical way to work on all three without pretending that every platform uses the same signals.

Personalization turns a ranking check into a context check

Google Discover is introducing conversational controls that let a person use their own words to request more or less of particular topics or links. The feed can then adjust in response and remember those requests. A generic check of whether your content appears cannot capture that kind of audience-specific filtering.

Google Preferred Sources adds a different type of personalization. A searcher can star a publication in the Top Stories section, giving Google an explicit signal to show more stories from that selected outlet. One mechanism expresses topical interest; the other names a preferred publisher.

Do not combine these features into a supposed universal AI ranking factor. They are platform-specific controls, and neither proves that a preference passes into every chatbot, answer engine, or language model. What they do reveal is the operating model you now need: discovery can depend on both the subject a person wants and the entities that person already trusts.

Separate brand visibility into three questions:

Eligibility: Do you have content that directly satisfies the person’s stated topic, task, and constraints?
Preference: Has the person been given a clear reason and a supported mechanism to choose your publication or brand again?
Representation: When an AI system includes the brand, are its claims accurate, current, and relevant to the recommendation?

This distinction prevents a common measurement error. A brand can be eligible but not preferred, visible but inaccurately described, or mentioned without being recommended. Those are different failures, so they require different fixes.

Make explicit preference an audience action, not a ranking theory

Explicit preference is valuable because the audience is choosing the relationship. Google has said people have selected more than 600,000 unique Preferred Sources and are twice as likely to click. That makes the feature worth considering for a qualifying publication, but its documented scope is Google Top Stories. It is not evidence that the same choice improves your standing everywhere else.

The newer embedded flow reduces interruption: a reader can select the Preferred Source button, confirm the addition, and then return to the page they were already reading. If your site is eligible, place the platform-provided control where the reader has just received enough value to understand why they might want more.

Use this implementation checklist:

Put the control on pages that demonstrate your editorial specialty, not only on a generic home page.
Place it after a complete answer or useful analysis, where preference is a natural next action rather than an interruption.
Explain the platform-specific benefit plainly: selecting the publication can result in more of its coverage appearing in Top Stories.
Keep the explanation beside the control. Do not imply that selection affects unrelated AI products.
Test the full confirmation and return path on the devices your audience uses.
If your analytics setup permits it, distinguish an initial button interaction from a completed addition. Otherwise, you may mistake interest for a successful preference action.

If Preferred Sources does not apply to your business, keep the strategic principle and discard the unsupported ranking claim. Give satisfied visitors a clear way to subscribe, follow, save a resource, join a relevant community, or return to a named recurring feature. These actions create a direct audience relationship. Treat that relationship as an asset in its own right, not as a secret way to manipulate an unrelated model.

Build content around the language people use to shape feeds

Conversational personalization makes vague topical relevance less useful. A person does not have to choose from your internal taxonomy. They can describe the exact material they want to see. Your content architecture should therefore reflect recognizable needs, not just broad keyword categories.

For each important content lane, define four elements before choosing a title:

Situation: Who is making the decision, and what is already true for them?
Subject: Which product, platform, entity, or problem must be unmistakably present?
Task: What is the person trying to decide, fix, compare, or implement?
Constraint: What condition would make a generic answer inadequate?

For example, WordPress schema tips names a broad subject but leaves the task and constraint unclear. How to remove duplicate Organization schema in WordPress when an SEO plugin already outputs it describes a recognizable situation. Someone asking a feed for more technical WordPress schema debugging has a much clearer reason to match with the second page.

Run a preference-fit test before publishing:

Write the natural-language request a qualified reader might use, such as a request for more implementation guidance, fewer introductory explainers, or deeper coverage of a narrow platform issue.
Identify the page in your library that should satisfy that request. If several pages seem interchangeable, the content lane is probably not distinct enough.
Check whether the title and opening paragraph make the situation, subject, and task explicit without requiring the reader to infer them.
Use headings to answer the component questions that follow from the main task. Remove sections that belong to a different intent.
Connect the page to a stable hub that names the broader specialty, then link to adjacent pages only when they solve a genuine next problem.
State boundaries and limitations. A page becomes more trustworthy when readers can tell who should not follow its advice.

This is also where entity consistency matters. Use the same brand name, product labels, authorship information, and core factual descriptions across your pages. Structured data can reinforce that consistency for machines, but it cannot rescue an editorial premise that is unclear to a person.

Avoid producing near-duplicate pages for every imagined wording of a preference. The goal is not to manufacture endless variants. It is to create a distinct, complete answer for each materially different situation. If changing the audience phrase does not change the appropriate advice, it probably does not justify a separate page.

Audit what AI says, who it recommends, and under which context

Traditional monitoring often stops at whether the brand was mentioned. That misses the two outcomes that matter most: whether the description was accurate and whether the brand was selected for the user’s actual need.

Goodie markets Brand Command as a reputation-management layer designed to detect false AI claims and identify which brand receives the recommendation. Treat that as a vendor capability claim to evaluate, not proof that any monitoring product can inspect every model, explain every recommendation, or repair an answer automatically.

Build a context matrix before choosing a tool

Start with the decisions that matter to your audience. For each decision, record the contexts that could legitimately change the best answer: the person’s role, use case, experience level, constraints, location when relevant, and buying posture. Do not invent persona variations that would not alter the recommendation.

For every check, preserve these fields:

The platform and model or experience name shown to the user.
The exact prompt, conversational history, and declared preference context.
Whether the account or session had known personalization that you could observe or control.
The answer as displayed, including citations or linked destinations.
Whether the brand was absent, mentioned, accurately represented, or recommended.
Which alternative was recommended and which criteria were used to justify that choice.
The date of the observation and the page or evidence that supports your accuracy assessment.

Generative answers may vary between runs, so do not turn a single observation into a trend. Keep the prompt and conditions consistent when comparing results, and preserve meaningful audience differences instead of averaging them away.

Route each visibility failure to the right action

Observed pattern	Question to ask	Next action
Brand is absent across relevant contexts	Do you have a clear, authoritative page that answers this exact decision?	Create or improve the canonical answer. Make the brand’s relationship to the problem explicit and connect the page to the appropriate content hub.
Brand appears for one audience context but not another	Does your content genuinely address the missing audience’s constraints?	Preserve the split in reporting. Build content for the missing context only when the offering and evidence actually fit it.
Brand is mentioned, but another option is recommended	Which suitability criterion drove the recommendation?	Publish verifiable facts about fit, limits, requirements, and differentiators. Do not answer with unsupported superlatives.
The answer contains a false or outdated brand claim	Is the correct fact explicit, consistent, and easy to locate in your owned materials?	Correct conflicting owned information, strengthen the canonical factual page, and document the answer before and after the change.
The brand is accurately described, but the linked page does not produce a useful next step	Does the destination complete the job implied by the answer?	Align the page with that intent and provide a clear next action without hiding the promised information behind it.

Keep reach, representation, preference, and actionability as separate reporting dimensions. A blended visibility score can hide the most damaging case: the brand appears frequently but is described incorrectly. It can also make a legitimate audience split look like a general performance decline.

When you correct a factual problem, do not promise an immediate model update. You can control the clarity and consistency of your public evidence; you cannot control when or whether a particular system incorporates it. Continue monitoring the same context, retain the previous output, and treat a changed answer as an observation rather than proof of causation.

Key takeaways

Personalized discovery makes visibility context-dependent. Record the audience, preferences, session conditions, and prompt behind every result.
Explicit source preference is a valuable platform feature and audience relationship, not evidence of a universal AI ranking signal.
Build content lanes around a person’s situation, subject, task, and constraint so conversational preference filters can find a recognizable fit.
Measure inclusion, factual accuracy, recommendation outcome, and next-step usefulness separately.
Fix the observed failure: improve eligibility when absent, clarify fit when passed over, and strengthen canonical facts when misrepresented.

Start with the highest-value decision your audience brings to AI discovery. Map its meaningful contexts, identify the page that should answer each one, add an appropriate preference action, and record how the brand is represented. That focused loop will tell you more than another broad visibility score, and it gives your team a concrete change to make next.

References

August 20, 2026

How to Build SEO Across Social Search and AI Discovery

Your website can rank, your social posts can earn views, and your brand can still disappear when someone asks an AI assistant what to buy. The problem is usually not one missing keyword. It is a broken discovery chain: the answer exists, but the proof is fragmented across surfaces that never reinforce one another.

You fix that by planning website SEO, social search, third-party distribution and AI visibility as one operating system. The goal is not to publish the same content everywhere. It is to give each surface a clear job while keeping the underlying facts, expertise and evidence consistent.

Optimize a discovery chain, not an isolated page

Start by keeping the SEO foundation intact. Your important website content still needs sound indexability, crawlability, internal linking, semantics, taxonomy, layout and consistency. Those elements help machines retrieve a page, understand its subject and connect it to the rest of your site.

But a technically strong page cannot do the whole job. A buyer may first encounter your expertise in a short video, hear your company discussed on a podcast, see a creator demonstrate your product, compare reviews and only then search your name. An AI assistant may draw on several of those touchpoints before it decides whether your brand is relevant enough to mention.

That changes the planning question. Instead of asking only, “How do we rank this page?” map the full route from a person’s question to a defensible answer:

Demand: What complete question is the person asking, including qualifiers such as location, use case, budget, eligibility or timing?
Answer: What direct conclusion would resolve that question?
Evidence: Which product facts, demonstrations, customer experiences, expert opinions or original findings support the conclusion?
Format: Does the person need a detailed page, a visual demonstration, a short answer, a comparison or location-specific information?
Reinforcement: Where could the claim be independently discussed, reviewed or cited?
Action: What should the person be able to do next – compare options, verify availability, book, buy or continue learning?

Turn those fields into a discovery brief before commissioning anything. If the team cannot identify the evidence or the next action, changing a title tag will not solve the underlying problem.

This also exposes the difference between a keyword and a conversation. A keyword may describe a topic. A conversation contains the follow-up questions, objections, constraints and proof a person needs before making a decision. Website pages, social formats and external mentions should cover different parts of that conversation without contradicting one another.

Give every discovery surface a distinct job

Cross-channel SEO becomes wasteful when every team receives the same instruction: promote the new page. A link and a shortened caption rarely make a useful social asset, while a social clip rarely contains the depth, navigation or conversion path expected from a durable website resource.

Use the website as the durable evidence layer

Your site should hold the complete version of important factual and commercial answers. It is where you can explain conditions, show supporting material, connect related entities, maintain current policies and offer a controlled next step.

That does not mean every query deserves a new page. Create one when the person needs more depth, stronger verification or a better conversion path than an existing search result can provide. If another owned asset already satisfies the intent, a duplicate page may merely split attention between two weak destinations.

Treat social content as a searchable answer

A social post is no longer just a promotional route back to the site. Social and video content can surface directly in Google, which means a short-form answer may become the first result a prospective customer sees.

Suppose a video starts earning clicks for variations of “how to lace running shoes for wide feet” while the website has no useful answer. That pattern reveals search demand, the language people use and a format that already attracts attention. If those searchers need product guidance or a purchase path that the video cannot supply, build a detailed site resource, embed the useful demonstration and connect it to the appropriate products.

Run the logic in reverse as well. If the social result answers the question and leads people to the right action, do not clone it into a thin page just to add another URL. Strengthen the result you already have with a clearer caption, an accurate profile, a relevant destination and a planned follow-up.

The transferable unit is not identical copy. It is a stable claim supported by the same evidence. The website can provide depth, a short video can demonstrate the method, a static post can isolate the decision criteria and a profile can establish who is speaking. Each expression should feel native to its surface.

Use creators and independent coverage to fill trust gaps

Your own search data can reveal conversations where the brand has no presence. Use those gaps to brief creators by query territory and audience need, not follower count alone. A useful brief identifies the question to address, the evidence available, the claim boundaries, the preferred format and the action the audience should be able to take.

Format evidence belongs in the brief too. If your short-form content repeatedly gains search visibility while long-form video does not, that is a production signal rather than a matter of taste. Creators can then be selected for their ability to explain the right subject in the right format.

Independent coverage serves another purpose: corroboration. Your website is the appropriate authority for your hours, specifications, policies and availability. It is not an independent judge of whether you are the best or most convenient option. Reviews, publications, communities and creators can supply the external experience that a self-authored claim cannot.

Build evidence an AI system can connect and verify

AI discovery raises the cost of ambiguity. An assistant trying to recommend a business has to connect an entity to the right products, audience, locations and claims. Contradictory profiles, generic location pages and unsupported superlatives make that connection harder.

Create a controlled fact sheet for the claims that must remain stable across your digital presence. It should cover:

The official brand and location names you use publicly.
A plain description of what the business does and whom it serves.
Product, service and category relationships.
Locations, service areas, hours and available contact paths.
Eligibility, fees, policies, availability and appointment conditions where relevant.
The original evidence that supports distinctive claims.

Use that sheet to audit the About page, location pages, social profiles, speaker biographies, event descriptions and other copy you control. The wording can adapt to each setting. The facts should not drift.

Structured data supports this work when it describes the same information people can see on the page. JSON-LD can clarify relationships among a business, its locations, services and content, but markup cannot reconcile conflicting opening hours or turn an unproven claim into authority. Publish the complete, current fact in visible content first; represent it accurately in structured data second.

Specific context matters most when the question contains several constraints. Someone may ask for a nearby bank with free small-business checking and Saturday hours rather than typing “banks near me.” Answering that request requires fees, eligibility, proximity and branch hours to be available and verifiable together.

Part of the question	Evidence the machine needs	Strongest place to maintain it
“Near me”	Accurate location and service-area information	Location pages and maintained business listings
“Free small-business checking”	Current fees, conditions and eligibility	Official product and policy content
“Open on Saturdays”	Current hours for the specific branch	Branch-level pages, listings and operational data
“Recommended” or “most convenient”	Independent experience and reputation evidence	Reviews, publishers and other third-party platforms

For a multi-location company, do not treat this as one brand-level record. Each location needs its own accurate context. A service offered in one branch, an appointment policy used in one region or weekend hours at one address should not silently become a claim about every location.

Go beyond operational facts by creating material that cannot be replaced with a generic rewrite. Proprietary data, internal experiments, customer stories, product insights, industry findings, expert opinions and examples from real work give other people something concrete to cite and discuss.

Package each evidence asset so it can travel. Give it a stable page, a direct conclusion, enough method or context to evaluate it and clear limits on what it proves. Then adapt the finding into social explanations, creator conversations, presentations or interviews without changing the underlying claim.

Turn social search data into publishing decisions

Guesswork becomes less defensible when first-party query data is available. Google Search Console Platform properties can connect a verified social or video account to performance data from Search, Discover and News. The available reporting includes clicks, impressions, click-through rate, average position and the queries associated with the account’s content.

If the property type is available for an account you control, verify it promptly. Collection starts after verification and does not backfill earlier performance. Waiting does not preserve an option; it permanently leaves a gap in the query history.

Use the data in a repeatable workflow:

Record the verification point. This prevents the team from treating an incomplete early reporting window as a performance decline.
Check the 24-hour view after publishing. If a new asset begins gaining search demand quickly, cross-promote it while the subject is active or prepare the follow-up people are likely to need.
Review query groups. Separate leading, rising and declining themes. Use the language of genuine searches to refine captions, future topics and the questions covered on your site.
Compare like with like. Use URL-based filters to compare short-form and long-form video, or video and static posts, instead of letting total account performance hide a format difference.
Connect discovery to the next action. A query and click show that content was found. They do not show that the visitor reached a useful destination, understood the offer or completed a business action.

The report should end in a publishing decision, not a slide of metrics. Use these rules:

Observed signal	Likely issue or opportunity	Decision to consider
Social content earns relevant queries, but the site has no complete answer	Demand is proven, while the conversion or depth layer is missing	Create a useful site resource and connect the successful media to it
A social result already satisfies the intent	A second page may add duplication rather than value	Preserve the winning result, improve its destination and publish a logical follow-up
One format repeatedly earns more search visibility	The audience or result surface favors that mode of explanation	Change the production brief and test more topics in the stronger format
A topic rises in the 24-hour view	There may be a short window for related demand	Cross-promote it or release the next answer while interest is active
Impressions increase but useful actions do not	Visibility may be attracting the wrong intent or leading to a weak destination	Inspect the query, promise, landing path and action before scaling output

Keep the limits visible. Platform properties contain first-party information for accounts you can verify. They do not provide a competitor view, category benchmark or share-of-voice report. Native platform analytics and website conversion data still have different jobs.

AI visibility is less deterministic still. Responses can vary with context, location and prior activity, while current visibility tracking is better suited to directional patterns than exact attribution. Measure whether important facts and citations appear more consistently across a controlled set of relevant prompts, but do not present that sample as a complete market view.

Install one operating loop across SEO, social and AI

The final obstacle is usually organizational. SEO manages pages, social manages feeds, public relations manages mentions and local teams manage operational facts. Each group can hit its own target while the overall discovery experience remains inconsistent.

Organize the recurring review around conversations rather than channels:

Select a query territory. Start with a question that matters to the audience and has a plausible next action.
Classify the evidence requirement. Decide whether the answer depends on an official fact, a demonstration, independent experience, original analysis or several of them together.
Choose the primary asset. Name the website page, social result, video or location record that should carry the complete answer. Do not assume it must always be a new page.
Close factual gaps. Correct conflicting profiles, incomplete location data and unsupported claims before increasing distribution.
Create native adaptations. Preserve the conclusion and evidence while changing the length, format and framing for each surface.
Earn reinforcement. Put useful findings and demonstrations in front of the communities, creators and publications the audience already trusts.
Read the combined signals. Use query demand, format performance, external references, destination behavior and directional AI visibility to choose the next update.

Assign ownership at each handoff. Someone must be accountable for canonical facts, someone for platform-native production, someone for third-party distribution, someone for location accuracy and someone for business outcomes. Job titles can vary. Unowned handoffs are where contradictions and dead-end traffic accumulate.

Key takeaways

Keep technical SEO strong, but plan discovery around a person’s complete question rather than one page or keyword.
Use the website for durable depth, social content for searchable explanations and third parties for independent validation.
Make important brand and location facts consistent in visible content before representing them in JSON-LD.
Verify eligible Google Search Console Platform properties early because performance data is not backfilled.
Convert query and format signals into explicit publishing decisions instead of reporting visibility as an end in itself.
Treat AI visibility measurements as directional and improve the evidence available across the surfaces an assistant may consult.

Begin with one query cluster where your social traction, website coverage and business destination do not line up. Decide which asset should answer it, repair the supporting evidence and distribute the answer in formats suited to each surface. That single completed loop will teach your team more than another disconnected content calendar.

References

August 18, 2026

Commercial Product Discovery in ChatGPT: An Action Plan

Your product can rank well in conventional search and still disappear when a buyer asks ChatGPT what to purchase. The useful question is not simply, “How do we rank in ChatGPT?” It is, “What would ChatGPT need to understand, verify, and distinguish before placing this product on a relevant shortlist?”

Because in-chat recommendations can compress the route from discovery to decision, you have less room to repair a vague product description later in the journey. Your product information must connect a specific buyer situation to a defensible recommendation, while your reporting must keep generated answers and paid placements separate.

Map the decision ChatGPT is being asked to make

A commercial prompt is rarely just a category keyword. A buyer may describe the job they need to complete, who will use the product, a limiting requirement, an unacceptable tradeoff, and the alternatives they are considering. Follow-up questions can narrow the decision further.

Treat the prompt as a compact purchasing brief. Before changing pages or adding schema, build a commercial question map for each important product:

Buyer: Who is the product designed for, and who is likely to find it unsuitable?
Job: What concrete problem or task is the buyer trying to handle?
Constraints: Which requirements can rule the product in or out, such as compatibility, location, budget structure, capacity, or implementation effort?
Comparison criteria: Which differences matter when the buyer compares this product with another option?
Evidence: Which product page, specification, policy, or help page substantiates each claim?
Transaction details: What must the buyer know about price conditions, availability, delivery, returns, warranties, or the next purchasing step?

Use real questions from sales conversations, customer support, site search, product reviews, and search-query data where you have access to them. Then remove any wording that your public evidence cannot support. The map is not a keyword list. It is an inventory of the decisions your content must help someone make.

A simple test exposes the gaps: can a buyer find a short, factual passage on your site that answers each mapped question without combining clues from several pages? If not, ChatGPT may also have to infer too much. Add the missing decision fact to the appropriate product, comparison, policy, or support page.

Make product evidence recommendation-ready

An unbranded modular device is inspected on a workbench alongside its components, material samples, accessories, and household use-case objects.

Your primary product page should do more than announce benefits. It should make product identity, suitability, limitations, and buying conditions explicit. A persuasive claim can attract attention, but a precise fact is easier to use in a recommendation.

Audit the evidence layer in this order:

Establish one identity. Use the same product name, brand, category, model, and variant labels across product pages, documentation, feeds, comparison content, and structured data.
State fit in plain language. Name the audience, use case, prerequisites, and meaningful limitations. A clear not-for statement can be more useful than another broad benefit.
Expose decision criteria. Publish compatibility, included capabilities, implementation requirements, commercial conditions, and tradeoffs in text that can stand on its own.
Support comparisons. Organize comparison pages around buyer-relevant dimensions. Explain where each option fits instead of declaring your product the universal winner.
Connect claims to proof. Link feature claims to specifications or documentation and policy claims to the applicable policy page. Remove unsupported superlatives.
Show update state. Display when time-sensitive specifications, prices, or policies were last reviewed, and assign someone to keep them current.

Where it accurately describes the page, Product and Offer structured data can provide a machine-readable version of facts such as the product name, brand, identifiers, offer URL, price, currency, and availability. Use only identifiers and commercial details that actually apply. Do not invent a product code to fill a field, and do not leave an old price in JSON-LD after changing the visible page.

Structured data is not a guaranteed entry ticket to a ChatGPT recommendation. Treat it as a precise mirror of visible, maintained product information. If the markup, product page, shopping feed, and support documentation disagree, fix the underlying fact before adding more optimization.

Treat generated recommendations and ads as separate channels

A split scene shows an unbranded product on a neutral comparison table on one side and on a brightly spotlighted display on the other.

Commercial discovery in ChatGPT can contain two distinct surfaces: the generated answer and a sponsored placement. Combining them in one visibility number produces false confidence.

In an analysis of more than 50,000 commercial prompts across 20 niches, sponsored placements appeared on 25.94% of the sampled prompts. Every observed ad appeared below the generated response, and each placement contained one sponsored offer rather than a group of competing advertisers. Your campaign may not reproduce that delivery rate because prompt context and category can change what appears.

The overlap between paid placement and generated visibility was small. Only 3.63% of advertisers also received a citation in the answer above the ad. The advertised URL appeared in citations in 0.09% of cases, while advertiser brands were mentioned in 4.44% of responses. On this evidence, buying an ad does not appear to make the brand materially more likely to enter the generated recommendation.

Surface	What success means	Primary optimization work	What to record
Generated answer	The product is correctly included, described, and supported for a relevant buyer situation.	Clear product facts, suitability criteria, comparisons, documentation, and consistent structured data.	Product mention, recommendation rationale, cited URL, factual accuracy, and competitor inclusion.
Sponsored placement	The offer appears in a relevant commercial conversation and sends qualified prospects to an appropriate destination.	Precise context hints, focused keyword-style phrases, suitable creative, and a landing page aligned with the conversation.	Placement data, landing-page engagement, lead quality, purchases, and other business outcomes available to you.

Keep separate dashboards, targets, and budgets. A paid impression is not earned answer visibility. A citation is not an advertising conversion. You need both measurements before you can tell whether ChatGPT is influencing discovery, traffic, or revenue.

Run a controlled discovery program instead of chasing screenshots

Benchmark the generated answer

A screenshot proves that one response occurred. It does not tell you whether the product appears consistently, whether the recommendation is accurate, or which missing fact is preventing inclusion elsewhere. Use a fixed prompt set and a repeatable record.

Create prompts from the commercial question map. Include category discovery, use-case fit, constraint-led selection, direct comparison, and branded validation questions. Keep each prompt focused enough that you can identify why an answer changed.
Record the conditions. Capture the prompt, date, whether the test began in a new conversation, the answer, citations, sponsored placement, and any follow-up question used.
Grade the response. Mark whether the product was mentioned, recommended for the right reason, linked or cited, and described accurately. Record unsupported claims and omitted limitations as failures, even when the brand appears.
Trace each weakness to a page. For every missing or incorrect fact, identify the public URL that should resolve it. If no appropriate URL exists, you have found a content gap rather than a prompting problem.
Change one evidence cluster at a time. Update the relevant product, comparison, or support content and its structured-data mirror together. Retest the same prompt set on a regular cadence, but do not declare success or failure from one response.

Constrain paid targeting with conversational detail

ChatGPT ad matching uses natural-language context hints alongside keyword-style phrases. Those hints guide matching rather than operating as strict keyword rules, and advertisers did not have visibility into the individual queries or conversations that triggered their placements. That makes precision in the context description and measurement after the click especially important.

Draft each context hint internally with this structure: buyer type evaluating product category for a defined job, under a named constraint, with a stated decision criterion. The structure forces you to describe a conversation in which the offer genuinely belongs. A broad category label does not.

Separate materially different audiences and use cases instead of blending them into one targeting theme.
Send each context cluster to a distinct, tracked landing-page destination aligned with that buyer, job, and criterion.
Repeat the relevant suitability facts and limitations on the destination so the visitor can confirm the fit immediately.
Use your own analytics and customer records to judge qualified engagement, lead quality, and purchases because the underlying triggering conversation may be unavailable.
Rewrite or pause a broad context when it produces irrelevant visits. Do not try to repair weak relevance by adding more generic phrases.

This control matters because 14.35% of the observed ChatGPT ads were semantically unrelated to the prompt beside them. That rate describes the sampled placements, not every campaign, but it is large enough to make relevance auditing a launch requirement rather than an optional cleanup task.

Key takeaways

Optimize for a buyer decision, not a single category keyword. Map the buyer, job, constraints, comparison criteria, evidence, and transaction details.
Publish explicit suitability, limitation, tradeoff, and commercial facts. Keep visible content, documentation, feeds, and JSON-LD consistent.
Measure generated recommendations and sponsored placements as separate channels. Paid placement does not imply inclusion in the answer.
Use a fixed prompt benchmark to track mentions, citations, reasoning, accuracy, competitors, and ads under recorded conditions.
Make ad context hints narrow enough to describe the right conversation, then use distinct landing destinations and your own outcome data to expose mismatches.

Start with the product that matters most commercially. Build its decision map, audit the public evidence against every question, and capture a generated-answer baseline before expanding content or buying placement. That sequence gives you something more useful than visibility for its own sake: a clear view of where the commercial discovery path is breaking and what to fix next.

References

August 10, 2026

How to Optimize Product Feeds for AI Shopping Discovery

If your products have strong pages and good reviews but rarely appear in AI shopping carousels, writing more copy may not solve the problem. The missing layer may be the product data that helps an AI system decide which items deserve consideration in the first place.

Your product feed now has to do more than support Shopping ads. It must identify each item, keep commercial facts current, distinguish variants, and answer the kinds of questions people ask conversational shopping tools. The practical goal is not to choose between feed optimization and product-page SEO. It is to give each surface a clear job and keep both synchronized.

Treat the feed as the consideration layer

ChatGPT can use shopping-oriented query fan-outs that are separate from the searches used to compose its written answer. In one observational sample of more than 43,000 products from March 2026, 83% of the matches appeared within Google’s top 40 organic Shopping results. Only 11% matched Bing results, and almost all of that smaller group also appeared on Google.

Position mattered within that sample. Sixty percent of strong matches came from Google’s top 10 Shopping results, and the order of products in a ChatGPT carousel tended to follow their Google Shopping order. A shopping fan-out often drew one results page to build an eight-product carousel. This is observational evidence, not a guarantee that every carousel comes from Google, but it gives you a useful diagnostic: a product that is missing or poorly ranked in organic Shopping may struggle before its product page gets a chance to persuade anyone.

A separate vendor dataset covering more than one million ChatGPT shopping offers in June 2026 showed why feeds can be attractive to a retrieval system. When ChatGPT cited a merchant feed directly, about 99.9% of those citations appeared on the top product offer. The share of feed-sourced retrievals rose from 4.3% to about 20% over six weeks.

Within that same dataset, feed-sourced offers populated the brand, image, and merchant fields 100% of the time, compared with 0% for page-scraped offers. They also carried the "best price" label 100% of the time, compared with 21% for scraped offers. Those percentages should not be treated as universal benchmarks. They do show the operational advantage of structured fields: the system can read an explicit value instead of inferring it from a page.

OpenAI describes ChatGPT product results as organic and unsponsored, with relevance influenced by availability, price, quality, and whether the merchant is the primary seller. You cannot control every signal, but you can stop forcing the system to guess about facts that belong in your catalog.

Surface	Primary job	What failure looks like
Product feed and catalog	Make the item eligible, understandable, current, and competitive for shopping retrieval	The product is excluded, misclassified, ranked poorly, or shown with incomplete information
Product detail page	Confirm the offer, answer deeper questions, support retrieval, and persuade the shopper	The item is considered but the offer is inconsistent, unconvincing, or difficult to verify

Fix the fields that can exclude or misclassify a product

An isometric comparison shows a hiking shoe with complete, organized product attributes entering a discovery path while a shoe with missing and mismatched data is diverted.

Begin with the feed’s factual core. Enhancements cannot compensate for an invalid identifier, stale availability, or a price that disagrees with the live page. Approval is the floor; accurate, discriminating data is what gives the product a chance to match the right request.

Confirm that the intended products are actually present and eligible. Check the items you expect to sell, not only the catalog total. A missing variant, rejected item, or unintended destination setting can make an otherwise excellent product page irrelevant to shopping retrieval.
Validate product identity. Supply the correct brand and a valid GTIN where the product has one. Do not invent an identifier to fill an empty field. A false identifier creates a worse entity match than a properly represented product without one.
Make the title identify the actual item. A title should distinguish the product and its meaningful variant without turning into a string of repeated keywords. Use attributes that are true, commercially important, and necessary to tell this item from neighboring products.
Match price and availability to the live offer. Compare the submitted values with what a shopper sees on the corresponding product page. If a sale begins or inventory changes, the feed and page should change as one commercial system.
Inspect the primary image. It should render cleanly and represent the exact product or variant attached to the record. A technically valid image is not useful if it depicts a different color, pack size, or configuration.
Use the correct category. Preserve both the most accurate Google taxonomy assignment and a useful internal product type. Broad or incorrect classification weakens the system’s ability to place the item in the right comparison set.
Check the destination page for consistency. The URL should resolve to the same product, variant, identity, price, and availability described by the feed. Treat any disagreement as a data-quality defect, not a copywriting opportunity.

The fastest audit is a line-by-line comparison between the source catalog, the submitted feed, the processed Merchant Center record, and the live page. That sequence tells you where a defect entered the pipeline. If the source catalog is wrong, fix it there and regenerate downstream data. Repeated manual corrections in Merchant Center create a second source of truth that is easy to forget during the next inventory, price, or platform update.

Defect	Likely interpretation problem	Corrective action
Wrong GTIN or brand	The item can be associated with the wrong product entity	Correct the identifier in the catalog system and resubmit it
Feed price differs from page price	The offer appears stale or unreliable	Fix the update path or timing before changing promotional copy
Generic title across several variants	The system cannot confidently distinguish the requested option	Add the truthful attributes that separate the records
Broad or incorrect category	The product enters an unsuitable comparison set	Choose the most specific accurate taxonomy value and retain your product type
Image shows another variant	The visual evidence conflicts with the structured record	Map each record to the image for that exact option

Prioritize defects in this order: eligibility problems, factual mismatches, missing identity or category data, weak differentiation, and then optional enhancements. This keeps the team from polishing fields on products that cannot yet enter the selection set.

Add conversational attributes around real buying decisions

Google introduced optional conversational attributes for Merchant Center at Google Marketing Live 2026. They are intended to support experiences such as AI Mode and Gemini. These fields do not determine product approval, so treat them as a second layer: first make the core record correct, then make it more useful to an agent handling a specific buying task.

Choose the field that matches the shopper’s question

Question and answer: Store concise answers to recurring pre-purchase questions about compatibility, fit, intended use, care, installation, or constraints. Answer the question directly and avoid unsupported claims.
Related product: Express relationships such as often_bought_with, required_part, accessory, and substitute. This can help an agent assemble a workable solution instead of recommending one isolated item.
Document link: Connect the product to a relevant manual, specification sheet, or sizing guide. Use the document that resolves a buying question rather than linking every PDF associated with the SKU.
Item group title and variant option: Tie records to a recognizable product family and expose the available options. These fields matter when the request includes a constraint such as a particular color and size.
Popularity rank: Represent how a product performs relative to the rest of your catalog. This can support questions about your best-selling or most popular option, but only if the score has a stable definition and remains current.

For a travel bag, for example, a question-and-answer pair could address the product’s documented dimensions, a document link could point to the sizing sheet, variant fields could connect capacities and colors, and a related-product relationship could identify a compatible accessory. That is more useful than repeating "ideal for travel" across several fields. One approach supplies evidence and relationships; the other supplies a slogan.

Build enhancements from evidence you can maintain

Collect recurring decision questions. Use internal search terms, customer-support questions, return reasons, reviews, and merchandising knowledge to find the uncertainties that prevent a confident purchase.
Map each question to a structured field. Use a Q&A pair for a direct factual answer, a relationship for compatibility or substitution, a document for detailed evidence, and variant fields for product-family navigation.
Identify the owner of the underlying fact. Dimensions may come from product operations, compatibility from technical documentation, and popularity from commerce data. The feed should distribute an authoritative value rather than create one.
Check the claim against the page and supporting material. If the feed promises compatibility that the manual or page cannot confirm, the extra field increases inconsistency instead of reducing it.
Retire stale enhancements. Remove or update relationships, documents, answers, and popularity signals when the catalog changes. Optional data is still product data and needs an operating owner.

Do not measure this work by field coverage alone. A catalog full of generic Q&A pairs can be complete and still fail to resolve a single buying decision. The better test is whether each enhancement helps an agent answer a question that the core title, category, price, image, and availability fields cannot answer on their own.

Run the feed and product page as one discovery system

A rain jacket is connected to contextual buying attributes, an organized product feed, and a product page through one luminous discovery pathway.

A feed-first strategy does not make the product detail page secondary in every sense. Across the June 2026 shopping-offer sample, about 88% of ChatGPT offers still came from product pages rather than feeds. Even among merchants that used feeds, roughly 76% of offers were sourced from the page.

The apparent contradiction disappears when you separate selection from presentation. Feed data can help the system identify and rank a candidate while the final merchant offer still points to, or is extracted from, the product page. A visible PDP citation therefore does not prove that the feed played no role. Citation source and selection input are not necessarily the same thing.

Your product page should repeat the feed’s core facts without ambiguity, explain benefits and constraints the feed cannot hold, expose the correct variants, and provide credible supporting material. Reviews and independent coverage can influence how an AI system characterizes the product or brand. They do not guarantee selection.

That distinction matters when evaluating content-led tactics. In one examination of brands that ranked themselves first in their own listicles, about 69% were cited without being recommended; a larger competitor mentioned on the same page often received the recommendation. The brand supplied retrievable content, but the system selected someone else. Treat that outcome as a selection problem to diagnose, not as proof that another self-authored ranking page is needed.

Site architecture is not a substitute for product-data work either. A June 2026 review of 11,400 shopping answers across ChatGPT, Perplexity, and Gemini did not find category structure affecting whether a brand was recommended on those platforms. That does not mean category pages are useless for shoppers or conventional search. It means you should not assume that reorganizing them will repair an AI shopping eligibility, identity, or ranking problem.

Merchant listing structured data can help keep the page machine-readable, but it belongs in the same fact system as the feed. In July 2026, Google added support for product-category information covering both its taxonomy and merchant product types, along with sale-duration fields. If the page markup, visible offer, and submitted catalog describe different categories or sale windows, adding more schema only formalizes the disagreement.

Use a diagnostic loop instead of a one-time feed cleanup

Choose representative shopping requests. Include category searches, attribute-led requests, use cases, comparisons, compatibility questions, and requests for a popular or lower-priced option.
Establish the Shopping baseline. Record whether each relevant product is eligible, whether it appears in organic Google Shopping, and where it sits relative to competing offers.
Record the visible AI outcome. Note the exact request, selected products, carousel order, merchant, displayed price, cited URL, and whether the requested variant or constraint was respected.
Classify the failure before editing anything. Missing or rejected products point to eligibility. Eligible products with weak Shopping visibility point toward feed relevance or competitiveness. Wrong prices or variants point to synchronization. Selection without engagement points toward the offer or PDP. A citation that recommends a competitor is a selection problem, not automatically a markup problem.
Change the responsible layer and retest. Correct catalog facts upstream, improve only the attributes involved in the request, and preserve a record of the before-and-after result. Do not rewrite the PDP, feed title, taxonomy, and schema simultaneously or you will not know what fixed the defect.

Discovery can move quickly, but there is no dependable instant-indexing promise. One documented merchant appeared in Google Shopping the day after its feed was connected and then surfaced in ChatGPT. Use that as evidence that the pipeline can respond, not as a service-level expectation. Recheck after material catalog changes and keep the observation date with every test because rankings, availability, prices, and retrieval behavior can all move.

This work needs shared ownership. Commerce operations controls inventory and price, merchandising controls categorization and relationships, SEO controls page discoverability and structured consistency, and analytics observes selection and downstream behavior. A feed defect should not wait in a paid-media queue simply because Merchant Center was originally configured for ads.

Key takeaways

Use organic Google Shopping visibility as an early diagnostic for AI shopping discovery, while recognizing that the observed overlap is not a universal retrieval guarantee.
Fix eligibility, GTIN, brand, title, price, availability, image, category, and page consistency before adding conversational enhancements.
Treat the feed as a consideration and ranking layer, and the product page as the offer-verification, explanation, and conversion layer.
Add Q&A, related-product, document, variant, and popularity data only when it answers a real buying question and has a maintainable source of truth.
Do not infer the selection path from the visible citation alone; a page-sourced offer may still have benefited from structured catalog data.
Measure eligibility, Shopping position, AI selection, offer accuracy, and shopper response as separate stages so the team fixes the layer that actually failed.

Start with one commercially important product family. Compare its source catalog, processed Merchant Center record, live page, and structured data line by line. Fix every disagreement, add one enhancement tied to a real customer question, record its Shopping and AI visibility, and then extend the process to the next family. That turns feed optimization from a setup task into a repeatable discovery system.

References

Search Engine Land — AI shopping starts with your product feed, not your product page

July 31, 2026