Guide · Inspeccia
Ecommerce AI visibility: what actually gets measured
The conversation about AI visibility in ecommerce carries a vocabulary problem that costs money. People talk about "appearing in ChatGPT" as though it were a single thing you win or lose, when it is in fact three separate competitions, with different rules, different material and different teams inside the company. Winning one does not give you the others. And worse: most stores measure just one of the three and draw conclusions about all of them.
The three races, one line each.
1. The category conversation. Getting named when someone asks about your sector without mentioning you. Won with reputation and content. Rivals: other brands.
2. The slot on the product card. Your store showing up when the assistant lists where to buy something specific. Won with catalog data. Rivals: everyone selling the same item, marketplaces included.
3. The informational answer. Getting cited in "how do I choose" answers. Won with well-structured classic SEO.
Why separating them changes the decisions
A scene that repeats. A store discovers that when it asks the assistant about its flagship product, three marketplaces come up and it does not. The team's immediate conclusion: "there is nothing to be done, the big players win this." The topic is dropped.
What actually happened is that they measured race number two, the product-card slot, which is precisely the one that looks worst for a small store, and decided for all three. Had they asked in category form — "I'm looking for X for Y, which brands would you recommend?" — they would have measured race number one, where the rival is not Amazon but the other four brands in the sector, and where a store with good content and good reputation competes on level terms.
The separation also clarifies who fixes what. Race one belongs to marketing and content. Race two belongs to catalog and systems: if the price being shown is wrong, no blog post corrects it. Race three belongs to SEO. When everything shares one name, the problem lands on the wrong desk.
Race 1: getting named in the category
This is the one that most resembles old-fashioned reputation, and the one you control least directly. Someone describes a need without naming any brand and the assistant answers with a list. The question is whether you are on it.
What moves it is not your product page: it is what is written about you across the web, yours and other people's. Reviews, comparisons, mentions in trade press, forums, your own pages explaining what you do and who for. When a model searches live to answer, it opens sources; when it answers from memory, it repeats what it learned. In both cases, the material is text that exists somewhere.
How to measure it. By asking about the category and never about your brand. If you ask "what do you know about [your store]?", the model will talk about you: that confirms you exist, not that you are visible. The useful test is the one a customer would run without knowing you. And it has to be repeated: the same question gives different answers in different sessions, so one attempt is not a measurement. The figure that matters is how many times out of how many you appear.
Race 2: the slot on the product card
Here the terrain changes completely. This is no longer about content but about structured catalog data, and about it being current.
OpenAI's public product feed specification gives the measure of how operational this layer is: price and availability are required fields, and there is a field that explicitly controls whether a product can be surfaced in ChatGPT search results — with the wrinkle that this permission has to be on before the product can also enter the checkout flow. There is also an expiration date to retire an item after a given day.
Translated into practice: this race is lost on plumbing. A feed that does not refresh, a short promotion your catalog publishes and the assistant has not caught up with, a badly filled field. None of that is fixed by writing better. It is also why the price you see in an assistant often does not match the one on your site: it did not come from there. We cover it in wrong product data in ChatGPT, and the way in, in how to get your products into ChatGPT Shopping.
How to measure it. Product by product, starting with the few that matter. Take the five with the highest margin, ask about each one the way a buyer would, and compare four fields against your site: price, availability, who is listed as the seller, and whether the link goes to your listing or someone else's. It is tedious and it is the only method that yields an actionable number.
Race 3: the informational answers
The third is the most familiar to anyone coming from SEO, because to a large extent it is SEO. These are answers to questions that are not immediate purchases: how to choose, what the difference is, which one suits a given use. There the assistant cites pages, and those pages can be yours.
Google's documentation on its AI features is explicit about the part most often misread: you don't need to create new machine readable files, AI text files, or markup to appear in these features, and there is no special structured data to add. What you do need is the usual: to be eligible as a supporting link, a page must be indexed and eligible to be shown with a snippet. There are no additional technical requirements.
For a store this reads concretely, and a little uncomfortably: the content section many ecommerce teams treat as filler — buying guides, comparisons, genuine frequently asked questions — is exactly the material of this race. And it is the only one of the three where the work you already did for Google transfers almost intact.
The three honest caveats
Anyone promising you a clean number here is simplifying something. Three things worth accepting before building a dashboard.
- The same question does not give the same answer. Repeating the query in two sessions can return different lists. That is why an isolated measurement is not a measurement: the minimum useful unit is a repeated batch, and the metric is a frequency, not a yes or no.
- Traffic undercounts, by design. Many mentions carry no link. Of those that do, some lose the referrer and get logged as direct, and some turn into a later branded search that appears as organic. Measuring visibility with traffic analytics is measuring the shadow of the object.
- There is no official dashboard. Google documents that pages appearing in its AI features are included within general Search Console traffic, with no breakdown of their own. There is no report telling you how many times an assistant named you. Auditing, here, still means asking the question again.
The one-afternoon diagnostic. Pick three category questions (without your brand) and your five highest-margin products. Ask each of the three questions five times in separate sessions and note how often you appear and who appears when you do not. Then review the five products field by field: price, availability, seller, link.
At the end you have two numbers that actually mean something — your category mention rate and your catalog errors — and you know which of the three races you are losing. Which was the only question that needed answering.
Frequently asked questions
What is AI visibility for an online store?
It is the sum of three things the industry usually collapses into one word. First: getting named when someone asks an assistant about your category without mentioning you. Second: your listing winning the slot when the assistant shows where to buy a specific product, with its price and availability. Third: your content being cited in informational answers, the how-do-I-choose kind. They are three separate races because they are won with different material: the first with reputation and content, the second with catalog data, the third with well-structured classic SEO.
Does ranking well on Google give me AI visibility?
It helps a lot in one of the three races and fairly little in the other two. For Google's informational answers the relationship is direct: the official documentation says that to be eligible as a supporting link a page must be indexed and eligible to be shown with a snippet, and that no special files or markup are needed. But the slot on a product card is decided by catalog data that does not live on your site, and getting named in a category conversation also depends on what other sources say about you. That is why a store can rank first and still not appear where the purchase is decided.
How do I measure whether my store appears in ChatGPT?
By asking about the category, never about your brand, and by repeating. If you ask about your own name, the model will talk about you almost every time: that does not measure visibility, it measures that you exist. What you want to know is whether you show up when someone describes the need without knowing you. And you have to repeat the same question across separate sessions, because the answer varies from one to the next. The useful metric is not did-I-appear, it is how many times out of how many.
Is AI traffic a good way to measure this?
It is a poor measure and a misleading one, especially for a store. Many mentions carry no link, so they generate no visit at all. Of those that do, a good share loses the referrer along the way and ends up counted as direct, and another share turns into a later branded search that shows up as organic. The result is that analytics undercounts by design. Traffic measures a partial consequence; to measure visibility you have to look at the answer itself.
Which catalog data matters for product cards?
The basics, and above all that they are current. In OpenAI's product feed specification, price and availability are required fields, and there is a field that explicitly controls whether a product can be surfaced in ChatGPT search results. There is also an expiration date to retire a product after a given day. None of this is editorial content: it is catalog plumbing, and it is the reason the marketing team sometimes cannot fix the problem on its own.
Is competing in AI the same as competing with marketplaces?
It depends on the race. In the category conversation your rivals are usually other brands, not Amazon: the question is who the model names when someone asks for recommendations. In the product-card slot you do compete head-on with everyone selling the same item, marketplaces included. Mixing the two leads to a very common false conclusion, that there is nothing to be done because marketplaces always win. We cover it separately in the marketplaces-versus-your-store guide.
How often should I measure it?
Often enough to see the variation, not to get a number. A single measurement of a single question says almost nothing, because the same query can return different answers in different sessions. What does inform you is a batch of category questions repeated at intervals, watched as a trend: whether the share of times you appear goes up or down, and who takes your place when you are absent.
Sources cited
- Google Search Central — "AI features and your website" (no new machine readable files, AI text files or markup needed; no special structured data; to appear as a supporting link a page must be indexed and eligible to be shown with a snippet; no additional technical requirements; pages appearing in AI features are included in general Search Console traffic). Official documentation.
- OpenAI Developers — "Product Feed Reference" (
priceandavailabilityrequired;is_eligible_searchcontrols whether the product can be surfaced in ChatGPT search results and must be true to enableis_eligible_checkout;expiration_dateto retire a product). Specification.
Which of the three races are you losing?
Every analysis asks about your category live and captures what the AI answers: whether your store appears, which brands take your place, and which sources get cited. It is the starting point for knowing where to put the effort.