Why AI assistants aren't recommending your store

Most merchants arrive at this question the same way. They typed a prompt into ChatGPT, read an answer that named three competitors and not them, and concluded they have a visibility problem. Sometimes that is right. Often the first thing to fix is the test, not the store.

Work through the causes below in order. They are arranged cheapest and most common first, so the early checks cost you minutes and rule out the false alarms before you spend a weekend rewriting descriptions. Each one has a way to tell whether it is your problem and a specific thing to do about it.

The diagnostic sequence

1. Re-test properly before you change anything

   Run the same prompt set several times across several days. One answer on one day is not evidence of anything.

2. Confirm the AI crawlers can reach you

   Read your live robots.txt and check for firewall or bot-protection rules. Cheap to check, catastrophic if wrong.

3. Read your own product data as a stranger would

   Take the constraints in the prompt and try to verify each one from the product record alone.

4. Question whether the prompt is winnable

   Some head prompts are settled. Check whether you are losing a fight you never had a route into.

5. Look for corroboration outside your own site

   Search your product name and see whether anything other than your store describes it.

6. Give the change time, then re-probe

   Different engines refresh on different cycles. Judge the trend over weeks, not the answer this afternoon.

7. Separate what is left into fixable and not

   Some of the variance is the channel behaving normally. Monitor it instead of chasing it.

Quick reference: symptom to cause

1. You are measuring it wrong

This is the most common cause and the cheapest to rule out, so start here. AI answers are non-deterministic. The same prompt sent twice can return a different set of brands with nothing changed on your end. A single absence is not a diagnosis.

How to tell: run the identical prompt several times across several days and write down the result each time. If your store appears in some runs and not others, you do not have an absence problem. You have normal variance, and the honest metric is how often you appear, not whether you appeared once.

There is a second measurement error worth naming. Asking an assistant do you know [brand] tells you almost nothing useful. That is a recall question, and recall is not how a shopping recommendation gets made. Test the prompt you actually want to win, phrased the way a customer would phrase it, with the constraints attached.

There is a third error: checking a single assistant and treating the result as your AI visibility. The distribution moves faster than most measurement habits do. Similarweb's tracking put ChatGPT's share of generative-AI website visits at roughly 53% by May 2026, down from about 76% in June 2025, while Gemini rose to around 27–28% and Claude to close to 9% (Similarweb, 29 July 2026). One assistant, one prompt, one day is three sampling errors stacked on top of each other.

What to do: fix the test first. Write a fixed set of realistic shopper prompts, keep the wording constant, run them on a schedule, and record mention rate over time. Our guide to measuring AI share of voice covers the method in detail, including which metrics are worth a dashboard and which are noise.

2. The AI crawlers are blocked

This is the failure that produces a total, permanent absence, and merchants hit it more often than they expect. A theme setting, an app that rewrote robots.txt, or a firewall rule someone added to cut bot traffic can exclude the named AI agents without anyone deciding to.

How to tell: open your live robots.txt in a browser and read it. Look for GPTBot, ClaudeBot, PerplexityBot, Google-Extended and any blanket disallow. Then check your CDN or bot-protection settings separately, because a rule there can block a crawler that robots.txt happily allows. If one engine never mentions you while others do, this is the first thing to suspect.

Google documents the controls that exist on its side, and it is worth reading them rather than guessing at them. Its guidance covers both AI Overviews and AI Mode, and sets out the robots.txt and nosnippet preview controls available to site owners (Google Search Central, updated December 2025). Use it as the reference for what you have actually switched off, because the setting that removed you is often one somebody applied without reading what it covered.

What to do: allow the agents whose answers you want to appear in, and remove the bot rules catching them. If you blocked AI crawlers deliberately to protect your content, that is a defensible position, but be clear it is a trade-off. You cannot be excluded from the crawl and cited in the answer at the same time.

3. Your product data cannot prove the match

An assistant answering a shopping question turns it into constraints and then needs candidates it can verify against every one of them. If your record cannot confirm a constraint, the safe move is to recommend a product that can. This is not a penalty. It is the absence of evidence.

How to tell: take the prompt you are losing and list its constraints. Then open your own product page and try to confirm each one using only what is written there, as if you had never seen the brand. Specs inside an image, sizing in a PDF, compatibility mentioned only in customer reviews: none of that counts. If you cannot verify a constraint yourself in under a minute, neither can a model.

What to do: fill the fields that turn claims into checkable facts, starting with product category, then attributes, then the description. The field-by-field product data checklist covers exactly what each field does and what a good value looks like. Doing that across a large catalog by hand is where most merchants stall, which is what Cuebase's Catalog Readiness Hub is for: it scores each product against an agent-readiness rubric, flags precisely which fields are missing, and generates suggested fixes you review before anything is applied to Shopify.

It is worth naming what this is not. If you are hunting for a file or a markup tag that unlocks the answer, Google's own guidance says there isn't one: "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search", and "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add" (Google Search Central, updated July 2026). That is Google describing its own surfaces and not a statement about every engine, so it does not settle the question everywhere. But it does mean the missing ingredient is far more likely to be a fact you never wrote down than a file you never created.

4. You are targeting prompts you cannot win

Some prompts are effectively settled. Broad head questions of the form best running shoes tend to return the same large, heavily reviewed, heavily discussed brands regardless of what any individual store does. Losing those is not a data problem and no amount of catalog work will change it.

How to tell: rephrase the prompt several ways and note who gets named. If the same handful of large brands appears every time no matter how you ask, that prompt is not a realistic target for you right now. Then test a narrower version with an attribute, a use case and a constraint attached, and see whether the answer opens up.

What to do: move your measurement set toward prompts where your catalog genuinely is the best answer. A prompt combining an attribute, a use case and a constraint has a much smaller candidate pool, and a store with excellent data on exactly that product can be the most confident match in it. Winning ten narrow prompts that describe real purchase intent beats losing one famous one. The positive playbook goes deeper on choosing and writing for those.

5. Nothing outside your site backs you up

Assistants lean on independent sources when they answer comparison, ranking and opinion questions. Reviews, category roundups, marketplace listings, forum threads and editorial mentions all give a model something to corroborate. If the only place on the internet that says your product exists is your own store, there is very little to corroborate, and a model asked to compare options has nothing to compare you with.

How to tell: search your product name and your brand name and look at what comes back that you did not publish. If the results are entirely your own domain and your own social accounts, this is likely part of your answer. It also explains the frustrating case where your data looks as good as a competitor's and they still get named.

What to do: be realistic. This is the slowest cause on the list and the one you control least directly. You can ask for reviews properly, get listed where your category is aggregated, and give the people who write roundups accurate specs to work from. What you cannot do is manufacture independent coverage on a schedule. Treat this as a long effort running underneath the faster fixes, not a task with a due date.

6. Your fix has not propagated yet

Changes to your catalog do not arrive in assistant answers immediately. Some engines read live feeds and reflect price or stock changes quickly. Others depend on crawled pages and cached indexes that refresh on their own cycles, and those cycles are not published.

How to tell: compare the date you shipped the change against the dates of your probes. If you rewrote descriptions last week and have been checking daily since, you are looking at data from before the fix. It is also normal for one engine to reflect a change well before another, which reads like a broken engine and usually is not.

What to do: set the expectation in weeks and hold your prompt set constant so the history stays comparable. Probing on a fixed cadence gives you evenly spaced points on a trend line, which is what makes a real movement readable against the noise. This is the part Cuebase's Share of Voice module automates: the same shopper prompts sent to ChatGPT, Perplexity, Gemini, Claude and Google AI on a schedule, with mention rate, position and citation rate tracked per query and per engine.

7. Some of it is genuinely not yours to fix

This section is the honest one, and it does not have a fix at the end.

A share of your visibility is decided by things no merchant controls. Models are updated without notice and behaviour changes with them. Engines differ in what they retrieve, how they weight sources and how willing they are to name a small brand at all. Answers vary run to run by design, so a prompt you won consistently last month can go quiet for reasons that have nothing to do with your store. There is no setting, no tool and no agency that removes this. Any product promising you a guaranteed mention or a fixed position in an assistant's recommendation is describing something that does not exist.

The reasonable posture is to separate the two categories and treat them differently. Crawler access, product data quality and prompt selection are yours, and they are worth real effort. Model behaviour and answer variance are not, and the only sane response is to measure them consistently enough that you can tell a genuine decline from a bad week. Chasing the second category costs you time you should be spending on the first.

The controllable half is not wasted either way. Complete categories, accurate attributes and specific descriptions improve your on-site filtering, your product feeds and your conversion rate regardless of which assistant reads them.

FAQ

Sources

Every figure quoted above is linked to its original publisher with the date attached, so you can check it and see how current it is. Where the underlying measurements disagree with each other, we have said so rather than picking the most flattering number.

1. Google Search Central. AI features and your website. Updated December 2025.
2. Google Search Central. Optimizing your website for generative AI features on Google Search. Updated July 2026.
3. BrightEdge. Rank overlap after 16 months of AIO. 18 September 2025.
4. Similarweb. Generative AI usage statistics. 29 July 2026.