Being famous isn't the same as being recommended by AI

There's a comfortable assumption in a lot of CPG boardrooms right now. It goes like this: we spent decades and a lot of money making our brand famous, so when shoppers ask an AI assistant what to buy, the assistant will think of us. Fame was the moat on the shelf. Surely it's the moat in the chat window too.

A study published this month is the clearest evidence yet that this assumption is wrong, or at least badly incomplete. Being known to an AI assistant and being recommended by it turn out to be two different measurements, and a brand can score perfectly on one and close to zero on the other.

The study, briefly

Envoyra, a company that works on AI visibility, asked four assistants (ChatGPT with web search, Claude, Gemini and Perplexity) the same 170 buyer-style questions every week from September 13 to October 4, 2026. That produced 2,720 answers across five software categories: meeting assistants, customer support AI, CRM for small business, marketing automation and AI visibility tools. Then, in a separate test from October 7 to 9, it asked each assistant about 30 of those brands by name, 24 times per brand, and labeled whether the answer correctly described what the brand does.

Search Engine Land's write-up covers the headline numbers. Envoyra's own report has the detail, and it's worth reading the method section before quoting any of it, which we'll come back to.

Known, but not picked

Here's the result that should bother established brands. Every assistant knew Keap, a CRM for small businesses: all 24 name tests described it correctly. Yet when buyers asked about CRM for small business without naming anyone, Keap was recommended in 3 of 400 answers. Close, another CRM, was recognized in 24 of 24 and recommended in 16 of 352.

And the reverse: some brands were recommended often but recognized poorly when asked about directly.

Recognized by nameRecommended in category

Show the data

Brand	Recognized by name	Recommended in category
Keap	100%	0.8%
Close	100%	4.5%
Profound	21%	35%
Fathom	33%	36%
Otterly.AI	54%	32%
Peec AI	46%	29%

Source: Envoyra AI Recommendation Report, Fall 2026 (name tests Oct 7-9; category answers Sep 13 to Oct 4, 2026). Percentages are our calculation from the published counts, for example 3 of 400 is 0.8% and 136 of 384 is 35%. The name test used 24 answers per brand.

Profound was recommended in 136 of 384 category answers and recognized in only 5 of 24 name tests. Fathom, 134 of 368 against 8 of 24. Envoyra traced many of those misses to name collisions: asked about Fathom, several answers described a financial reporting company also called Fathom; asked about Profound, some described a developer tool from a company called Profound Logic. The assistants weren't inventing things. They picked the wrong entity.

Put the two halves together and you get a picture that doesn't match how most marketers think about brand. Recognition is a fact the model holds about you. Recommendation is a choice the model makes when it builds an answer to someone else's question. Different inputs feed each one.

The shelf is different in every aisle

The second finding makes the first one harder to manage. The four assistants named the same top brand in only one of five categories, customer support AI, where Zendesk led everywhere. In the other four, the leader depended on which assistant you asked.

Category	ChatGPT	Claude	Gemini	Perplexity
Customer support AI	Zendesk	Zendesk	Zendesk	Zendesk
CRM for small business	HubSpot	HubSpot	Zoho CRM	HubSpot
Meeting assistants	Otter.ai	Fireflies.ai	Zoom	Fireflies.ai
Marketing automation	HubSpot	Klaviyo	ActiveCampaign	ActiveCampaign
AI visibility tools	Semrush	Profound	Semrush	Profound

The starkest gap was in Envoyra's own category: Claude named Profound in 89 of 96 answers, and Gemini named it in none. Envoyra reran the Gemini questions on October 9 to check, and still got zero.

Week to week, the top of each list was steady. HubSpot was recommended in 305 of 368 small business CRM answers, and its weekly counts were 77, 76, 75 and 77. Below the top, things moved: only 144 of 680 question-and-assistant pairs gave the same brand list all four weeks. That's the same pattern SparkToro and Gumshoe found in January, when their research saw the same list repeat fewer than once in 100 runs, while a handful of leaders kept turning up. We wrote about what that means for tracking in Stop tracking your AI rank.

Show the data

Weekly run	Answers naming HubSpot
Sep 13	77
Sep 20	76
Sep 27	75
Oct 4	77

Source: Envoyra AI Recommendation Report, Fall 2026. 305 of 368 answers across the four runs. Run dates are Envoyra's four Sunday runs.

Why we think this carries over to groceries and personal care

Envoyra studied business software, and we want to be careful here. Nobody has published the same split for snacks, supplements or shampoo, and this is our reading, not a finding.

But the mechanism isn't specific to software. A model's knowledge of a brand comes from everything it was trained on: decades of press, packaging copy, retailer listings, Wikipedia. A model's recommendation for a category question comes from how it frames that question, which sources it pulls (for assistants that search), and which brands those sources put forward for that need. A 60-year-old cereal brand will be in the first pile. Whether it's in the second depends on whether anyone writing "best high-fiber cereal for kids" in the last two years put it there.

There's a reason to think the gap could be wider in CPG than in software. Software buyers ask fairly standard questions. Grocery shoppers ask oddly specific ones: a snack for a toddler with a sesame allergy, a sunscreen that won't sting a swimmer's eyes, a protein bar under 200 calories that isn't chalky. Those questions reward brands that have published something precise about that exact need. Fame doesn't answer them; facts do.

And shoppers are asking. Salesforce's 2026 holiday predictions report that half of shoppers say they've used an AI assistant at some point while shopping, and 74% say they trust recommendations from AI chat. That doesn't tell us how many CPG purchases start in a chat window. It does tell us the answers are being read.

The two traps

We see two ways a CPG team gets this wrong.

The incumbent's trap. A legacy brand asks ChatGPT "what is [our brand]?", gets an accurate, flattering paragraph, and concludes its AI visibility is fine. That's the Keap test, and Keap passed it 24 times out of 24. The question that matters is the one where your name isn't in the prompt: "what's a good dairy-free creamer that froths?" If you're not in that answer, your awareness isn't doing the work you think it is.

The challenger's trap. A newer brand sees itself recommended for a category question and assumes the assistant knows who it is. Then a shopper asks a follow-up about the brand by name, and the assistant describes a different company with a similar name. Envoyra's Fathom and Profound misses are what that looks like. For a CPG brand with a common-word name, it's a real risk, which is why our tips on unambiguous names, entity facts and Wikidata exist.

Both traps come from treating one number as if it were the other.

What to measure instead

The fix is simple to describe and a bit tedious to do. Keep two scores, per assistant.

The first is recognition. Ask each assistant you care about, by name, what your brand is and what it sells, a few times over a few days. Note whether it gets the company right, the category right and the facts right. This is ten minutes of work and our tip on checking ChatGPT walks through it. If the answers are wrong or mixed up with someone else, that's an entity problem, and it's fixable with consistent facts on your site and the places models learn from.

The second is recommendation. Ask the category questions your shoppers ask, without your name, and count how often you appear across several runs and several weeks. Don't read one answer as the truth; Envoyra and SparkToro both show a single check tells you very little. Tools that ask assistants shopper-style questions and keep the answers, such as SearchPanel from XENA Intelligence, measure this second number at the product level. You can also build a rough version in a spreadsheet.

Then look at the two side by side, assistant by assistant. Known but rarely picked means you need more third-party and on-site content that ties your product to specific needs. Picked but misidentified means you need cleaner entity information. Both low means start with the basics. And because the assistants disagree, a strong showing in one of them says little about the others.

Awareness is a fact a model holds about you. A recommendation is a choice it makes for someone else.

The caveats, stated plainly

Envoyra competes in the AI visibility category, which is one of the five it studied, and it says so in the report. It doesn't publish its questions. It queried the assistants through their APIs, and two of the four (Claude Sonnet 4.6 and Gemini 2.5 Flash) ran without web search, which a separate analysis by GeoParrot flagged as a confound: models without search are answering from training data, which may lag behind the market. Consumer apps can behave differently from APIs. The recognition test was small, six answers per brand per assistant. None of this undoes the core finding, but it's a reason to treat the exact numbers as one snapshot.

Our own claim is narrower than "fame doesn't matter." Fame clearly helps: the steadiest names in Envoyra's data, HubSpot and Zendesk, are famous. The point is that fame isn't sufficient, and for most brands below the very top of a category, it isn't even the main thing. In an earlier piece we argued that small AI referral numbers hide where decisions are being made. This is the same argument from another angle: the scoreboard you already have, awareness, doesn't tell you whether you're on the shortlist.