AI shopping visibility is whether your product gets named when an AI answers a buying question. It is not a ranking. There is no position to track, because the answer is generated fresh each time and differs by phrasing, by engine and by session. Five surfaces decide it today: Alexa for Shopping (formerly Rufus) and CoSMo on Amazon, Sparky on Walmart, and ChatGPT, Gemini and Perplexity off marketplace. Each reads a different source of truth. You measure visibility by asking a fixed set of buyer questions on a fixed schedule and counting how often you are named.
Why this matters now
Adobe Analytics reported that AI traffic to US retail sites rose 393% year on year in the first quarter of 2026, from over 1 trillion visits analyzed. In March 2026 that traffic converted 42% better than traffic from other sources. Adobe's own conclusion, in its AI traffic report, is that retail sites are not machine readable enough for the shoppers now arriving through AI.
The buyers are already there. The question is whether your product is in the answer they get.
What is AI shopping visibility, and what is it not?
Three things it is not.
→ Not a ranking. A search results page is a list in a fixed order. An AI answer is a short set of products, written on demand. Nothing sits at position four.
→ Not a keyword match. The shopper asks "which insulated tumbler fits a car cup holder and does not leak". The assistant looks for facts that answer that sentence. A listing that carries the right keywords and none of those facts has nothing to offer.
→ Not only a website problem. Most advice on this topic is written for brands that own their site, their schema and their product feed. Sellers on a marketplace do not control any of that. They control one thing: the listing. For a marketplace seller, the listing is the unit of visibility.
That last point is where this guide differs from the generic explainers. If you sell on Amazon or Walmart, the page the AI reads is the one inside the marketplace, written into fixed fields, under character limits you did not set.
The five surfaces that pick products today
An AI "surface" is any place where an assistant chooses which products to name. Five matter now.
1. Alexa for Shopping on Amazon. Amazon's AI shopping assistant, renamed from Rufus on 13 May 2026. It answers buyer questions inside the search bar and on product pages, using the listing as its material. The field-level detail is its own guide: Optimizing Amazon Listings for Alexa for Shopping.
2. CoSMo on Amazon. Amazon's commonsense knowledge graph. Amazon researchers described it at SIGMOD 2024 as a system that learns why people buy things (who a product is for, where it is used, what it is used with) from search and purchase behavior. It maps 15 relation types. It is a separate system from Alexa for Shopping, with its own backend. See What Is Amazon CoSMo?.
3. Sparky on Walmart. Walmart's AI shopping assistant. It answers from Walmart's own catalog, so the same rule applies: the listing is the source.
4. ChatGPT, and 5. Gemini and Perplexity. The general assistants are not tied to one store. They answer from whatever they can retrieve and cite: retailer pages, brand sites, product feeds, reviews, editorial roundups. You have far less direct control here, and far more routes in.
The distinction that matters is between closed and open surfaces. On a closed surface (Amazon, Walmart) the assistant reads one catalog, and your listing is most of what it knows about you. On an open surface the assistant assembles an answer from the web, and your listing is one source among many.
What does each surface read, side by side?
The table below is a map, not a ranking of importance. Amazon has not published how it weights its sources, and neither have the other platforms, so no column here claims a weighting.
| Surface | Main source of truth | What you control | What you do not |
|---|---|---|---|
| Alexa for Shopping | Listing fields, structured attributes, reviews and customer questions | Title, item highlights, bullets, A+ text, attributes | Reviews, how Amazon weights sources |
| CoSMo | Listing content mapped to relations such as audience, use and occasion | Whether the listing states those relations in plain words | The graph itself |
| Sparky (Walmart) | The Walmart item page and catalog data | Item content and attributes you submit | How Walmart's assistant selects and phrases |
| ChatGPT | Whatever it can retrieve and cite from the web, plus feeds where available | Your own pages, schema, third-party coverage | Which sources it retrieves on a given day |
| Gemini and Perplexity | Retrieved web sources and shopping data | Your own pages, product data feeds, third-party coverage | Retrieval choices, citation choices |
One thing runs through every row. Each surface answers from facts it can find. A fact that exists in your product data but not in a retrievable form does not exist to the assistant.
How is this different from SEO?
SEO and AI shopping visibility share a goal, which is being found. They differ in almost every mechanism.
| Search engine optimization | AI shopping visibility | |
|---|---|---|
| What the shopper gets | A ranked list of links or listings | A written answer naming a few products |
| What you compete for | A position | A mention |
| How the query works | Short keywords | Full sentences with constraints |
| What gets rewarded | Keyword relevance, links, sales velocity | Facts the assistant can quote |
| How you check it | A rank tracker | Repeated test questions |
| Is the result stable? | Mostly, within a day | No. It varies by phrasing, engine and session |
Keywords do not stop mattering. They stop being sufficient. For the longer comparison of the three disciplines, see AEO vs GEO vs SEO for ecommerce.
Why is there no ranking to track?
Because the thing being measured does not exist in a fixed form.
A search results page is built once and shown to many people. An AI answer is generated each time it is asked. Ask the same question twice and you may get two different sets of products. Change "waterproof speaker" to "speaker I can take in the pool" and the answer changes again. Switch engines and it changes again.
So three habits from SEO stop working:
→ Checking one query once. One answer is one sample. It tells you nothing about the rate.
→ Treating a screenshot as a result. A screenshot of you being named is an anecdote. It proves the product can be named, not that it usually is.
→ Chasing a position. There is no position. There is only the share of answers in which you appear.
The metric that replaces rank is mention rate: out of a fixed set of buyer questions, how often does the assistant name your product.
It is a crude measure and it is honest about being crude. The method for running it properly, and what to do about the variance, is covered in How to measure AI shopping visibility with no ranking to track.
Why is product content now the deciding factor?
When the shopper's question is a sentence, the product that wins is the one whose content contains the matching fact. That moves product content from a conversion asset to the entry ticket. The argument is made in full in Why product content is the new ranking factor, and the mechanics of how an assistant reads a page are in How AI assistants actually read a product page.
The short version has three parts.
Facts beat adjectives. "Keeps drinks cold for hours" gives an assistant nothing to quote. A number, a material, a dimension or a compatibility does.
Consistency beats volume. When a title says 30 oz, an attribute field says 32 and a bullet says "about a litre", the safest thing an assistant can do is not answer, or answer from someone else's listing. This is what we call answer debt: the cost of contradictory product data. It has its own article: Answer debt.
Absence is invisible. A human skimming your page does not notice a missing fact. An assistant asked about that fact simply has nothing to say, so it names a product that does.
Here is what that looks like on a real public listing. ASIN B0G2MZ7JHC, a JBL Grip portable speaker, carries a title of 194 characters. Amazon capped titles at 75 characters including spaces on 27 July 2026, per its Seller Central announcement of 10 June 2026, in every category except media.
→ Before (194 characters, 119 over the cap): "JBL Grip Portable Bluetooth Speaker, Waterproof, Ambient Light - Teal | Up to 14 Hr Playtime, AI Sound Boost, IP68 Waterproof And Dustproof, Drop-Proof, App Control, Ideal For Outdoor And Travel"
Cut at the cap, that title ends mid-phrase at "Up". The playtime, the IP68 rating, drop-proof and app control all fall off the end. A shopper who asks for a speaker "that survives the pool and lasts a day" is asking about exactly the facts that were truncated away.
→ After (74 characters): "JBL Grip Portable Bluetooth Speaker, Waterproof IP68, 14 Hr Playtime, Teal"
The overflow, "AI Sound Boost, drop-proof, app control, ambient light. Ideal for outdoor and travel.", moves into the 85-character item highlights field, which Amazon caps at 125 characters. Nothing was deleted. The facts a buyer asks about are now in a field the assistant can read. The full walkthrough is in the Alexa for Shopping guide.
What can you find out with a simple test?
You do not need a tool to see where you stand. You need ten questions and an hour.
A useful set mixes three kinds of question:
→ Category questions: "best insulated tumbler for a car cup holder". Does the assistant name you at all?
→ Constraint questions: "leakproof tumbler under a given size that is dishwasher safe". Does it find a fact on your page?
→ Comparison questions: "how does your product differ from the market leader". Does it describe you accurately?
Put each to Alexa for Shopping, ChatGPT and Gemini. For every answer, record three things: whether you were named, whether what it said about you was right, and where the answer seems to have come from. Where you are not named, the cause is usually one of two things: the fact was not on your listing, or it was on your listing and contradicted somewhere else.
A broader surface list, and a way to decide which of them deserves your time first, is in The AI shopping surface map.
How do you measure it without a rank tracker?
Fix the inputs, vary nothing, and repeat.
- Fix the question set. Write ten buyer questions in shopper language, not in your own product vocabulary. Keep the exact wording.
- Fix the surfaces. Pick the engines your buyers use. For most marketplace sellers that is Alexa for Shopping first, then ChatGPT and Gemini.
- Fix the schedule. Run the set on the same day each month.
- Run each question more than once. Because answers vary, a single run is a sample, not a reading.
- Record named, accurate, and source. Named or not. Accurate or not. Which field or page the answer appears to draw on.
- Compare between runs. The change from month to month is the measurement. One run tells you nothing.
There are limits you should know about. The test measures your own questions, not the market's. Variance means small movements are noise. And no engine tells you how it chose. Treat the result as a trend line, not a score.
What do you control on a marketplace, and what do you not?
This is the part generic guides skip, because they are written for sites you own.
You control:
→ The title and item highlights, within the caps.
→ The bullet points and description.
→ A+ content, as text rather than as images. Amazon has not said whether the assistant reads text inside an image, so do not bet on it.
→ The structured attribute fields: material, dimensions, capacity, compatibility, intended use. These are the cheapest gap to close and the one most sellers leave half empty.
→ Consistency between all of the above.
You do not control:
→ Reviews and customer questions. When your listing does not answer a question, the assistant can answer from reviews instead, in your customers' words.
→ How a platform weights its sources. Amazon has not published weights, and anyone quoting exact ones is guessing.
→ Which products the assistant chooses to compare you against.
→ How fast a change shows up in answers. Amazon has not said, and seller estimates are observation, not documentation.
→ Whether and how sponsored placements sit beside organic answers.
The practical consequence: spend effort on the first list and measure the second. Do not build a plan on a weighting nobody has published.
Do you need to be on the open web if you only sell on Amazon?
Not to be found by Alexa for Shopping. That surface reads your Amazon listing. But shoppers also ask ChatGPT, Gemini and Perplexity, and those draw on the web and on retailer pages. A product with a strong Amazon listing and nothing else retrievable can be named on one surface and absent on the rest. How the open-web surfaces differ, and what each reads, is the subject of the pillar 3 hub: The Multi-Surface Product Visibility Guide.
This guide maps the territory. It does not teach you to optimize each surface. That depth lives in the guides above.
What makes a product invisible to AI?
Most invisibility is not mysterious. It comes from a small set of repeatable gaps:
→ Facts present in images but absent in text.
→ Attribute fields left empty.
→ Titles over the cap, so the useful part is cut off.
→ Contradictions between title, bullets, attributes and description.
→ Benefits written as adjectives instead of measurements.
→ Audience, use case and occasion left unstated because they feel obvious.
The longer catalog, with the gaps ranked, is in The 12 content gaps that make products invisible to AI.
What to do next
- Write down the ten buyer questions a shopper would ask before buying your product, in their words.
- Put all ten to Alexa for Shopping, ChatGPT and Gemini. Record whether you are named, whether it is accurate, and where the answer came from.
- For every question where you are not named, find the fact in your listing that would have answered it. Usually it is absent.
- Fill the structured attribute fields first. They are the cheapest gap to close.
- Repeat the same ten questions monthly. The trend is the metric.
Where to start
The quickest way to see what an AI cannot answer about your product is to look at the gaps directly. A free Cosmy account at cosmy.ai runs an Alexa for Shopping analysis on one of your ASINs and returns a gap report: the buyer questions your listing does not answer, and what is missing. It covers the Amazon side. For the surfaces beyond it, the test in this guide is the place to begin.
