Brand mentions in AI mean an assistant names your business; recommendations suggest your product for a shopper’s buying question, with a reason to consider it. First place is a separate placement measure. In Metrisque’s stored case, a store passed Brand Recall yet received no first-place recommendations across 11 saved reads.
Brand mentions in AI answer a different question
When a marketing lead says “ChatGPT knows us”, they usually mean the model gives a sensible answer when asked about the brand by name, which is recall. When a shopper types “what’s the best magnesium supplement for sleep”, the model is choosing, and recall only gets you considered. The shopper never typed your name, so the model has to decide on its own that your product fits the need better than the others it knows, which means a useful report must preserve the original question and the reason given for the choice before an editor can decide whether the answer exposes a page problem or simply reflects a different shopping need.
Reports blur the two. A report that counts both events as a mention hides that distinction. A brand named in reply to “tell me about Brand X” and a brand named first in reply to an unbranded buying question add the same single point to share of voice. For reporting a trend, that simplification is tolerable. For deciding what a store should change on its pages, it hides the one distinction that matters most: whether the model picks you when the shopper has not decided yet.
A stored case known by name and never first
Stored runs make this possible. Metrisque keeps every read in stored runs, which lets us look at one store’s history in detail without naming it. The models recalled this store when asked directly, and its Brand Recall verdict, which covers only the exact name asked, came back positive. That positive verdict establishes recall for the exact name tested, rather than visibility across every possible branded question.
Question Finder told a different story. In 11 saved reads, which asked the same shopper questions of the models’ memory and again with web search turned on, the store received no first-place mentions. Rivals took the top position each time. We call the distance between how well a model knows a brand and how often it recommends that brand first the recall gap, and this store had the widest version of it: full recall, zero first-place picks.
The case shows that recognition alone did not secure first place on these questions, while leaving the underlying cause unresolved. Diagnosis outcome: Not measured by Metrisque. Product pages, category fit and cited sources remain possible checks, rather than established explanations for this store’s results.
Answer types and their effect on buyers
Answers vary in kind. A buying-question answer can include a brand in four ways, and each one lands differently with a shopper. Reading answers by type, instead of counting names, shows which ones are worth chasing.
| Answer type | What it looks like | What it means for a buyer |
|---|---|---|
| Named in a list | “Other options include Brand A, Brand B and Brand C.” | The buyer sees an option without a specific reason to choose it |
| Named with a caveat | “Brand B is popular, though some reviewers mention sizing problems.” | The answer gives the buyer a reason to look elsewhere |
| Recommended first | “For a small apartment, Brand C’s compact model is the best fit because...” | The buyer receives a suggestion and a reason to consider it |
| Cited as a source | A link to the store’s product or guide page beneath the answer | The answer cites the page; it may or may not recommend the product |
Table 1. These answer types provide different signals; a mention count alone cannot establish which answers lead shoppers toward purchase.
The caveat row deserves attention. A caveat often comes from a review site or forum thread the model read through web search, and Metrisque’s Competitors and Citations view shows which sources the assistants lean on for each question. If the same complaint keeps appearing, the fix may be a product page that answers the concern directly, or a reply on the source itself.
The case on the recall and recommendation chart
The clearest way to see the recall gap is to cross the two measures. Recall goes on one axis (does the model know the name or not) and recommendation on the other (is the product recommended first for buying questions or not). Every brand lands in one of four quadrants, each with its own next step.
The top-left quadrant surprises people. A model can recommend a product for a buying question even when the model gives a weak answer about the brand itself, usually because web search found a clear product page or a strong review. That store’s pages are doing the work, and the risk is a page change that breaks what is working.
Known brands lose on product evidence
The cause sits lower down. When recall is fine and recommendation is not, it sits below the brand. Category filing is a starting point: Metrisque’s Category Fit marks a product Outside when its Gemini similarity to the expected category falls under 0.40, and the latest read for 53 stores found 13 of 923 products Outside. A category mismatch suggests a question to investigate, without proving why the model chose another product for a shopper.
Metrisque’s Buyer Match then tests whether the product page answers the buyer’s question. Description clarity follows: in a synthetic test, models classified clear descriptions correctly 150 of 150 times and vague ones 116 of 150 times. Sources come last. Citations naming rivals can suggest where to investigate the evidence available to the assistant, without establishing what caused its choice.
One caution keeps this honest. In Metrisque’s synthetic tests, similarity scores did not predict what a model recommends. A product that passes Category Fit can still lose to a rival with stronger third-party coverage, so each check rules out a cause and the final proof is a recommendation read after the fix.
Memory and web search close the gap at different speeds
The 11 reads in the stored case covered both of the places an assistant answers from, and the split matters for what a store does next. Model memory holds what the model learned in training. Page edits do not directly update its training, so changes to your site may leave memory answers unchanged for now. Web search pulls live pages when the shopper asks, including your product pages, retailer listings and reviews, so an edit to your own page can reach those answers as soon as the page is recrawled.
A store with a recall gap on web search answers can start by checking its own pages and cited sources. A rewrite of a vague description, a short section that answers the buyer’s most common question, or a fix to a breadcrumb that filed the product under the wrong category can all show up in the next round of reads. A gap that appears only in memory answers is slower, since it depends on the wider web and on the next model update.
That difference is why Metrisque records the two separately in Question Finder. Pooling them into one number would hide whether a change affected search answers while memory answers stayed the same, making the next editorial decision harder to justify from the saved results. Keeping a stored baseline for each, and re-reading both after a change, shows which answers moved and which still need wider coverage.
A recall gap test for your own store
You can measure the gap with two lists of questions. The first list names your brand (“what does Brand X sell?”, “is Brand X any good?”) and tests recall. The second list names only the need, the way a shopper would (“best compact sauna for an apartment”), and tests recommendation. Ask each question of each model with memory only and with web search, at least three times, and record where your product lands in each answer.
Compare the two lists. If recall is high and first-place picks are rare, you have a recall gap, and the work sits in pages, filing and sources. If both are low, the model needs to learn who you are, which is slower and depends on wider coverage. Our guide to measuring AI search visibility explains how recall, recommendation, citation and placement fit together, and how often to re-read each one.
Why does ChatGPT mention my brand but not recommend it?
Recognition can coexist with weak product evidence for a particular buying question, so test recall and recommendation separately before choosing changes.
What counts as a brand mention in AI? A brand mention is any appearance of your brand name in an AI answer. It includes a place in a list of options, a mention with a caveat, a first-place recommendation or a citation, which a simple mention total can group together.
How are AI product recommendations different from mentions? A recommendation is the model suggesting your product as the answer to a buying question, usually in first place and with a reason. A mention can be a passing reference that a shopper will not act on.
How do I check ChatGPT brand visibility? Ask ChatGPT about your brand by name to test recall, then ask unbranded buying questions to test recommendation. Repeat each question three times with and without web search and record your position.
The stored case does not establish which explanation applies to your store, so treat these checks as hypotheses for testing.
Sources
- Metrisque stored runs: Brand Recall and Question Finder case (one unnamed store, 11 reads, memory and web search). Category Fit latest read (13 of 923 Outside, 53 stores). description clarity study (300 classifications). synthetic similarity tests. Read on 16 November 2026.
Start with a free Metrisque check to see whether assistants know your brand and choose your products for buying questions.



