Agentic commerce visibility is how often, and how accurately, AI assistants name your products when a shopper asks for a recommendation. Unlike classic SEO ranking, it depends on five separate data layers — merchant feeds, marketplace listings, review aggregates, live page reads, and checkout protocols — that each engine mixes differently.
That last part is what trips up most teams. A brand can be flawless on its own website and still vanish from ChatGPT's product cards because one feed field says out_of_stock. Another brand can have a thin site and dominate Perplexity because three reviewers wrote thorough comparisons.
This piece maps the whole stack, layer by layer, and shows which layer each engine actually leans on — using 621,000 shopping answers we captured over 90 days.
Why shopping answers use a different supply chain than B2B recommendations
Ask an assistant for "the best project management tool for a 12-person agency" and the model assembles an answer from editorial content: roundups, comparison posts, Reddit threads, docs. That is the familiar answer engine optimization surface.
Shopping is different. Ask for "a quiet espresso grinder under $300 that's in stock" and the model needs three things editorial content cannot supply: a live price, a live stock status, and a purchasable SKU.
So it reaches into a commerce-specific pipeline — merchant feeds, marketplace catalogs, structured review data — and only then layers editorial judgment on top. The retrieval path is different, the ranking inputs are different, and the failure modes are different.
That is why a B2B playbook for assistant-led buyer research does not transfer cleanly to a product catalog. The tactics that win a SaaS shortlist — thought leadership, comparison pages, analyst mentions — do almost nothing for a SKU missing a GTIN.
The five layers of the agentic commerce visibility stack
Every AI shopping recommendation is assembled from some combination of five layers. Think of them as suppliers: the engine is a general contractor, and each layer supplies a different material. Miss one and the build still happens — just without you in it.
| Layer | What it is | Who controls it | Fails when |
|---|---|---|---|
| 1. Merchant feed | Structured catalog file pushed to an engine | You, directly | Missing GTIN, stale availability, ineligible region |
| 2. Marketplace listing | Amazon/Walmart/Target catalog record | You, partially; resellers otherwise | Your listing outranks your own PDP |
| 3. Review aggregate | Star ratings plus review text | Customers, third parties | Rating below threshold; no attribute words |
| 4. Live page read | Real-time fetch of your PDP | You, technically | JS-only price, 403 to agent user-agents |
| 5. Checkout protocol | ACP / UCP transaction rail | You, via opt-in | Not enrolled; missing eligibility attributes |
Here is the stack in detail, from most structured to most interpretive.
Layer 1: Merchant feeds — the structured spine
A merchant feed is a machine-readable file of your catalog that you push directly to an engine, and it is the only layer where you control the exact words an assistant reads. Everything else is inferred, crawled, or aggregated.
OpenAI's product feed specification requires item_id, title, description, url, brand, image_url, price, availability, target_countries and store_country, plus explicit eligibility booleans — is_eligible_search and is_eligible_checkout. Recommended fields go further and include reviews (title, content, rating), q_and_a, variant_dict, return_policy and related_product_id.
Google's equivalent spine is the Shopping Graph, which Google says spans more than 50 billion product listings with billions refreshing every hour. At that scale feed accuracy is not hygiene — it is the ranking input.
Two practical consequences:
- The
descriptionfield is prose an engine will quote. Write it for a shopper, not for a keyword string. - The optional review and Q&A fields are the cheapest visibility win most catalogs have, because they let you supply trust signals inside the structured layer instead of hoping a crawler finds them elsewhere.
One caution from our own audits: engines read the feed you send that engine. A Merchant Center feed does not populate OpenAI's catalog, and neither populates Perplexity's merchant program. Treat each as a separate integration with its own required fields.
Layer 2: Marketplace listings — the borrowed catalog
Marketplace listings are the fallback catalog engines use when your own feed is missing, stale, or regionally ineligible. Amazon, Walmart, Target and category marketplaces already carry clean identifiers, price history, and review volume, so they are cheap for a model to trust.
This creates the most common and least-noticed visibility leak: the assistant recommends your product but sends the buyer to someone else's checkout. You get the mention, a reseller gets the margin, and your analytics show nothing.
Amazon's own assistant matters here too. Rufus reads the full listing — title, bullets, A+ content text, backend attributes, Q&A and review bodies — which is why a rich Amazon listing so often beats a brand's own PDP inside AI answers. We cover that surface in detail in getting your product picked by Amazon Rufus and Perplexity Shopping.
The strategic question is not whether to be on marketplaces. It is whether your marketplace listing is better than your own product page — because if it is, engines will keep routing your buyer there.
Layer 3: Review and rating aggregates — the trust filter
Review data is not a ranking bonus in AI shopping answers; it is a gate. In our panel, products with an aggregate rating below 4.2 stars appeared in only 8% of shortlists even when they matched every stated constraint — price, category, feature, availability.
Engines use reviews two distinct ways, and teams routinely optimize for only the first:
- The number, used as a filter. Below a threshold, the product is quietly excluded before any text is read.
- The text, used as evidence. A model answering "quiet espresso grinder" needs someone to have written the word quiet in a review, a spec sheet, or an editorial test.
That second use is why review volume strategies underperform review coverage strategies. Fifty reviews saying "great product, fast shipping" give a model nothing to match against a constrained query. Eight reviews mentioning noise level, grind consistency, and hopper capacity give it three matchable attributes.
Practical test: list every attribute word that appears in your tracked prompts, then grep your review corpus for each one. Attributes with zero mentions are the exact queries you cannot win, no matter your star rating. This is the same attribute-matching logic behind feature-filtered AI shortlists.
Feed-level review fields, on-page Product and Review schema, and third-party aggregators all populate this layer.
Layer 4: Live page reads — the verification pass
Live page reads are when an assistant fetches your product page in real time to confirm what the other layers claim. This is the layer that grew fastest in our tracking window, driven by agentic browsing and deep research modes.
The pattern is consistent: the model builds a candidate list from feeds and reviews, then opens two or three finalist pages to check price, stock, shipping, and returns before committing. A page that fails that check gets dropped after it was already shortlisted — the most expensive kind of loss, because you did the hard work and lost at the last step.
Four failure causes we logged repeatedly:
- Price rendered only in JavaScript.
- Stock status hidden behind a variant selector.
- Shipping cost revealed only at checkout.
- Bot rules returning 403 to agent user-agents.
That last one is worth checking directly: OpenAI documents OAI-SearchBot and ChatGPT-User as separate crawlers with separate purposes, and blocking the user-triggered fetcher silently kills verification even when your feed is perfect. Server-rendered price and availability is no longer a nice-to-have. Multi-step agents lean on this layer hardest, which is why deep research modes change which brands get cited.
Layer 5: Checkout protocols — the transaction rail
Checkout protocols let an assistant complete the purchase without sending the shopper to your site, and participation is increasingly an eligibility signal rather than just a payments choice. Two open standards now dominate.
The Agentic Commerce Protocol (ACP), co-developed by Stripe and OpenAI, powers Instant Checkout inside ChatGPT. Google's Universal Commerce Protocol is the counterpart behind agentic checkout in AI Mode and Gemini, and Merchant Center gates participation on explicit eligibility attributes with a staged country rollout.
The visibility angle is subtle but real. OpenAI states that shopping results in ChatGPT are organic and unsponsored, ranked on relevance — you cannot buy a slot. But checkout eligibility changes which UI treatment a product qualifies for, and a buyable product card is a far bigger surface than a text mention inside a paragraph.
Decision rule from our panel: protocol work pays off only after Layers 1 and 3 are clean. Enrolling a catalog with 40% missing GTINs buys you a checkout button on products no engine shortlists.
Which layer does each engine actually lean on?
This is the part existing guides skip. They tell you all five layers matter. They do not tell you that Google's surfaces run on feeds while Claude runs almost entirely on editorial text — which means the same fix produces wildly different results depending on where you are losing.
What we did. Between 1 April and 30 June 2026 we ran 1,150 shopping-intent prompts daily across six engines — ChatGPT, Google AI Mode, Gemini, Perplexity, Copilot and Claude — spanning nine retail categories (running shoes, standing desks, espresso equipment, dog food, mattresses, laptops, skincare, cordless drills, car seats). That produced 621,000 answer captures. For every named product we logged the source the engine cited or linked, then classified that source into one of the five layers. Answers with no resolvable source were marked unattributed.
What we found. Source attribution by engine, as a percentage of named products:
| Engine | Merchant feed | Marketplace listing | Review / editorial | Live brand-site read | Unattributed |
|---|---|---|---|---|---|
| Google AI Mode | 61% | 8% | 18% | 9% | 4% |
| Gemini | 55% | 10% | 21% | 9% | 5% |
| ChatGPT | 34% | 21% | 29% | 12% | 4% |
| Copilot | 27% | 30% | 28% | 11% | 4% |
| Perplexity | 19% | 24% | 41% | 13% | 3% |
| Claude | 6% | 18% | 52% | 17% | 7% |
Three readings worth acting on:
Google's surfaces are feed-first. Six in ten named products traced back to Shopping Graph data. If you are invisible in AI Mode and Gemini, the diagnosis is almost always a Merchant Center problem, not a content problem.
Claude and Perplexity are editorial-first. Over half of Claude's product mentions traced to reviews, roundups and comparison articles. No feed work will fix a Claude gap. Getting into best-of listicles that AI quotes will.
Copilot is the marketplace outlier. At 30% it named marketplace listings more than any other engine — consistent with its Bing index dependency. Brands that skip Copilot in tracking miss the channel with the highest reseller-leak risk. It is one of the engines brands forget to track.
How fast does a broken layer cost you a shortlist slot?
We also timed the decay. When a tracked SKU flipped to out_of_stock in its feed, the median time to disappear from ChatGPT product cards was 31 hours. The same SKU kept appearing in Perplexity's prose recommendations for a median of 6 days, and in Claude's answers for over two weeks.
That asymmetry is useful. Feed-driven layers punish you fast and forgive you fast. Editorial layers reward you slowly and forget you slowly. A stockout is a short-term feed emergency; a bad review cycle is a long-term editorial liability.
It also sets your monitoring cadence: daily checks on feed-driven engines are worth the cost, weekly is enough for Claude.
A worked example: what happened when one brand fixed two layers
Panel numbers are directional. Here is a single brand's before-and-after, with its limits stated.
A US direct-to-consumer espresso-accessories brand in our panel — 62 SKUs, no marketplace presence — appeared in 11.3% of the 84 tracked prompts where its products genuinely matched the stated constraints. Its site was well built. Its feed was not.
In one week the team made four changes:
- Populated the optional
reviewsandq_and_afeed fields for all 62 SKUs. - Fixed 41 missing GTINs.
- Moved availability sync from once daily to every 15 minutes.
- Published a plain-HTML specification table on every product page, including decibel ratings and burr dimensions.
Six weeks later, tracked shortlist rate reached 28.6%. The gain concentrated exactly where the layer model predicts: ChatGPT went from 9% to 34%, Perplexity from 14% to 22%, and Claude barely moved — 6% to 9%.
The honest caveats. This was an observational before-and-after, not a controlled test. There was no holdout group. The window overlapped a seasonal demand rise in the category, and two competitors went out of stock during it. The engine-level split is the interesting result, not the headline number — Claude's flatness after a pure feed fix is the finding that would be hard to fake. Isolating cause properly is its own discipline, covered in proving causality in AI search.
How to audit your agentic commerce visibility in 10 steps
Work through these in order. The sequence matters — steps 1 to 4 are diagnostic, and doing them first stops you fixing the wrong layer.
- Build a prompt set of 60–150 shopping queries your buyer would actually type, including constraints (price ceiling, feature, use case, region).
- Run them across all six engines, not just ChatGPT. Log every named product, not only yours.
- Classify each mention's source into one of the five layers. This gives you a layer-attribution profile per engine.
- Find your weakest layer per engine by comparing your profile against the category leader's.
- Validate your feed against each spec — required fields, eligibility booleans, and that every
urlreturns HTTP 200. - Test availability sync latency. Flip a test SKU and time how long the change takes to appear in each engine.
- Audit review coverage, not volume. List the attributes your prompts contain, then check whether any review text mentions each one.
- Fetch your own product page with a plain HTTP client, no JavaScript. If price and stock are missing, live-read verification will fail.
- Check marketplace parity. If your Amazon listing is richer than your product page, expect engines to route buyers there.
- Decide on protocol participation — ACP, UCP, or neither — based on where your buyers already are, and document the eligibility attributes each requires.
Steps 1 to 3 are the ones teams skip, and they are the ones that make the remaining seven cheap.
What to measure once the stack is fixed
Three metrics cover agentic commerce visibility properly, and most dashboards track only the first.
Shortlist rate is the share of qualifying prompts where your product appears in the named recommendation set. Qualifying matters — if a prompt asks for something under $200 and your cheapest SKU is $400, that absence is not a visibility failure.
Layer attribution rate is the split of your mentions across the five layers, per engine. This is the diagnostic metric. A brand at 70% marketplace attribution and 5% feed attribution has a very different problem from one at the reverse.
Destination share is where the recommendation sends the buyer: your site, a marketplace, or an in-assistant checkout. This is the metric that connects AI share of voice to revenue, and the one that exposes reseller leakage.
Alongside these, keep the distinction between being named and being cited clear — they answer different questions and move for different reasons, which we unpack in mention rate versus citation rate.
Retailers with physical locations have a sixth input on top of this stack — proximity and business-profile signals, which follow their own rules in local AI search visibility.
What we could not determine
Three things our data does not answer, stated plainly.
We cannot see inside the ranking functions. Layer attribution tells us which source an engine cited, not which source changed its mind. A model may read a feed and cite a review.
We could not cleanly separate personalization effects. Our captures ran from a consistent US-based, logged-out configuration; results for logged-in shoppers with purchase history will differ, and we cannot quantify by how much.
We ran English-language prompts against US-targeted catalogs. Layer weightings almost certainly shift by market, since feed coverage, marketplace dominance and review density all vary by country.
Frequently asked questions
What is agentic commerce visibility in one sentence?
It is how reliably AI assistants name, describe and link your products when a shopper asks for a recommendation — determined by five data layers rather than by a single ranking position.
Do I need a merchant feed if I already rank well in Google Shopping?
Your Merchant Center feed already covers Google's surfaces, which are feed-first. But ChatGPT reads a separate OpenAI feed, and Perplexity runs its own merchant program. One feed does not serve all engines; check each engine's spec independently.
Which layer should a small brand fix first?
Run the diagnostic before choosing. That said, for most small catalogs the highest-return first move is completing the optional review and Q&A fields in the feed — it upgrades two layers at once, costs nothing in media spend, and takes days rather than months.
Can I pay to get recommended by ChatGPT?
Not for the organic product results. OpenAI states these are unsponsored and ranked on relevance. Paid formats exist as separate, labelled surfaces; treat them as a distinct line item, not a substitute for fixing your data layers.
How often should I re-run tracking?
Daily for feed-driven engines, because stock and price changes decayed visibility within roughly 31 hours in our data. Weekly is enough for editorial-driven engines like Claude, where mentions persisted for weeks after the underlying content changed.
Does agentic commerce visibility replace product SEO?
No — it consumes it. Product schema, review markup and server-rendered pricing are inputs to Layers 3 and 4. The change is that these signals now get read by an assistant building a shortlist, not only by a ranking algorithm ordering ten blue links.