Mentioned is not recommended

There is a moment in most AI visibility demos where someone points at a number going up and says “we’re at 62% visibility.” It sounds like progress. Then you ask what the buyer actually saw, and it turns out that in a good share of those answers the brand appeared in a sentence like “cheaper alternatives include…” or “some teams also evaluate…”.

That is a mention. It is not a recommendation. And the gap between them is where most of the money is.

Why the distinction suddenly matters

For twenty years the unit of search visibility was a position on a results page. Ten blue links, a rank, a click-through rate you could model. The buyer did the comparing.

That is no longer reliably true. OpenAI said in February 2026 that ChatGPT had passed 900 million weekly active users.¹ Pew Research Center’s 2026 survey of 5,119 US adults found that about half of American adults now use AI chatbots, up from a third in 2024, and that around four in ten use them for information searching. Six in ten say they read AI search summaries when they appear.²

When the buyer reads a summary instead of a list, the comparing happens before they ever see you. The model has already decided who to name, in what order, and with what framing.

That last number is the one people underestimate. Pew tracked 68,879 Google queries from 900 US adults in March 2025; roughly 18% of those searches returned an AI Overview. Clicks on the citations inside the summary happened on 1% of visits.³ The citation is not primarily a traffic channel. It is a credential the model shows the reader while telling them what to think.

What the two metrics actually measure

A mention rate answers: across the questions we track, how often does our brand name appear anywhere in the answer?

A recommendation rate answers something narrower and harder: how often is our brand the thing the answer tells the reader to do?

Both columns contain the brand name. A tool that counts string matches scores them identically. A buyer reading them does not.

This is why a visibility number can climb through a quarter of work while pipeline does nothing. You got named more often — as the comparison, the cheaper option, the one people used to use.

The traffic argument, and its honest limits

There is a second reason to care, and it is worth handling carefully because the numbers here are widely overstated.

Similarweb’s own analysis found that visitors arriving from ChatGPT behaved differently to search visitors: 15 minutes on site against 8, 12 pages against 9, and a 7% conversion rate against 5% on transactional sites.⁴ Ahrefs published a case study of their own site showing AI search visitors converting at 23 times the rate of organic visitors — but Ahrefs were explicit that this is one company’s 30-day window, not an industry benchmark, and said a larger multi-site study was still in progress.⁵

Treat the direction as real and the magnitude as unsettled. You will see “AI traffic converts 4–5x better” quoted widely; we could not trace that figure to any primary study, and neither should you.

Why mentions and recommendations move independently

Three mechanics push them apart.

1. Comparison questions name everyone A question like "X vs Y vs Z" will name all three regardless of merit. Track enough comparison prompts and your mention rate rises mechanically, with no change in how you are regarded.
2. Answers are shaped by sources you do not own Models assemble answers from third-party material — reviews, forum threads, roundups, documentation. Your own site can be immaculate while the sources the model leans on describe you as the budget option.
3. Recommendation is conditional Answers increasingly qualify: for a small team, for regulated industries, for high volume. You can be the recommendation in one condition and absent from another. A single blended percentage hides which.

That third point is the one that changes what you do on Monday. “We are at 62% visibility” is not actionable. “We are recommended on integration questions and absent on compliance questions, on Claude and Gemini but not ChatGPT” tells you what to write next.

How to measure it without fooling yourself

1. Separate the two counts from the start Mention rate and recommendation rate as distinct metrics on the same set of questions. If your tooling only gives you one, you cannot tell these stories apart.
2. Keep the questions fixed Visibility is measured over the prompts you track, so adding or removing prompts moves the number without changing what the models say. Change the question set when the business changes, not when the chart looks bad.
3. Report per model, not blended Models disagree, and they update at different times. A blended score is an average of things that should be read separately.
4. Sample repeatedly The same prompt returns different answers on different runs. One check is an anecdote. A trend needs volume.
5. Keep the underlying answers A number you cannot open is a number you cannot argue with. Store the text so anyone can read the sentence your brand appeared in.

Questions we get asked

Where this leaves you

The useful question is not whether AI knows you exist. For most established brands it does. The question is whether, at the moment a buyer describes their situation and asks what to do, your name is the answer or the footnote.

Those are different numbers. Measure them that way.