AI Brand Mentions: Why a Mention Is Not a Visit
- A brand mention in an AI answer is the first of four steps, and it is the only one you cannot see from your own server.
- Pew Research Center found clicks in 8% of visits to a Google page with an AI summary, against 15% without one, and a click inside the summary in 1% of visits.
- A January 2026 SparkToro study found fewer than 1 in 100 runs returned the same list of brands, so one prompt sample tells you almost nothing.
- Tools that type prompts and count names are selling a number that moves on its own. Kymo does not track mentions at all.
- Kymo records what AI engines did on your site: which pages a crawler read and what traffic came back.
A mention in an AI answer is not a visit. It means a model wrote your name into a reply on someone else's screen. Nothing about that event touches your server, your analytics, or your revenue until a person decides to click.
The four steps from mention to money
Treat ai brand mentions as step one of a chain. Each step is a different event with a different owner and a different way to measure it.
Step 1: The mention
An assistant names you in an answer. This happens inside ChatGPT, Claude, Perplexity, or a Google AI summary. You have no log line for it. You may never know it happened unless someone screenshots it.
Step 2: The fetch
Either the assistant already had your page from an earlier crawl, or a live fetcher pulls it right now because one person asked. ChatGPT-User (OpenAI's live fetcher) is the clearest example of the second kind. This event does hit your server. It is a request in your logs.
The other kind of read is slower. GPTBot (OpenAI's training crawler) gathers training data in bulk with no question attached, while an indexer like OAI-SearchBot builds the store the answer gets drawn from later. If you want the full split, GPTBot vs OAI-SearchBot: One Trains, One Cites covers it. Anthropic's pair works the same way, and ClaudeBot vs Claude-SearchBot: What Each One Does on Your Site is worth reading because most directories get the two backwards.
Step 3: The click
The person follows a link to your site. This is a referral, and it looks like any other referral in your analytics. Pew Research Center studied this in July 2025 across 900 US adults and found that when an AI summary appeared, users clicked a result in 8% of visits, versus 15% without a summary. A click on a link inside the summary happened in 1% of visits. Compare that to the 15% baseline. The summary shrinks the click.
Step 4: The conversion
They sign up or buy. If your goal tracking works, you can trace this all the way back.
| Step | Where it happens | Visible in your server logs | Visible in Kymo |
|---|---|---|---|
| Mention | Inside the assistant's answer | No | No, Kymo does not track this |
| Fetch | Your server receives a request | Yes | Yes, by crawler category |
| Click | Browser navigates to your site | Sometimes | Yes, as a referral |
| Conversion | Your checkout or form | Yes, if tagged | Yes, goals and revenue |
Why the mention itself is the shakiest number
Even at step one, the thing you are counting moves under you.
The SparkToro study ran 12 prompts through ChatGPT, Claude and Google's AI surfaces 2,961 times with 600 volunteers. Fewer than 1 in 100 runs produced the same list of brands. Fewer than 1 in 1,000 produced the same list in the same order. Rand Fishkin described tools that claim to report a ranking position in AI as "full of baloney", and the data backs him.
This is not only sampling noise. Model outputs vary run to run for infrastructure reasons. Thinking Machines Lab showed in September 2025 that 1,000 completions of the same prompt at temperature 0 produced 80 unique outputs, driven by server load changing the batch size. And models get updated. Chen, Zaharia and Zou found GPT-4 scored 84% on a prime-number task in March 2023 and 51% in June 2023 with no change to the prompt. OpenAI rolled back a GPT-4o update on April 29, 2025 because it was too flattering.
One more wrinkle. Different AI surfaces do not even agree on what to cite. Ahrefs compared 540,000 query pairs in December 2025 and found Google's AI Mode and AI Overviews cited the same URLs only 13.7% of the time, with 86% semantic similarity. Same engine, different surface, different sources.
What a mention-percentage tool is really selling
Most AI visibility tools type prompts into assistants, read the answers, count the brand names, and sell the percentage as a score. That is the whole method. It measures one sample from an unstable system and reports it to the decimal.
The scoring itself is also a choice. Digital Applied took one brand on one dataset and got 20% on mention-based share of voice, 16.8% on position-weighted, and 31.4% on citation-based. Three numbers, one dataset, three scores. None of them is wrong. All of them depend on which method you picked.
None of this measures what happened on your site. A mention score does not tell you that Claude read your pricing page four times this week, or that ChatGPT-User fetched your comparison post and sent six visitors who signed up.
Kymo measures the steps you can verify
Kymo does not track mentions and does not report what any assistant said about you. It runs no prompt samples.
What it does is watch your own server. A server-side receiver classifies crawler user agents into three categories. ai_answer means a live fetch made for one person's question, like ChatGPT-User or Perplexity-User. indexing means a crawler building a search index for later, like OAI-SearchBot or Claude-SearchBot. training means bulk crawls, like GPTBot or ClaudeBot. You can read the full explanation in AI visibility, explained.
Alongside that, the same dashboard counts human visitors without cookies. Cookieless identity relies on a salted server-side hash that rotates at UTC midnight, so unique visitors are unique per day. No raw IPs are stored, and no human visitor's user-agent is stored. AI crawler user-agents are kept on the crawler's own event row so the classification stays auditable. A crawler is not a person, and Kymo treats them differently.
AI crawler events do not count against your event limit. Solo is $9 a month or $90 a year for up to 10,000 events a month. Studio is $29 a month or $290 a year for up to 100,000. Both plans include every feature and start with 14 days free, no card required. If you exceed the limit the dashboard pauses. Your data is not deleted.
You can see all of it without an account at the public demo dashboard, kymo.in/demo.
Robots.txt is a set of switches, and two of them are not crawlers
Disallowing a training crawler does not remove you from that operator's AI answers. The indexer and the live fetcher are separate tokens, and blocking one leaves the others working. That is why you see Google-Extended (a robots.txt control token, not a crawler) and Applebot-Extended (a robots.txt control token, not a crawler) discussed as if they were agents. They have no user-agent and never appear in a log. They control how content an operator already crawled may be used.
One more file people ask about. Publishing an llms.txt file costs a few minutes, and several documentation platforms and developer tools do read it. No major AI operator has publicly committed to honoring it. Publishing the file is not the same as a crawler fetching it, and only your own server sees which one happened. Our llms.txt best practices cover the cheap bet honestly.
The recommendation
Track the steps you can measure on your own site, and treat mention percentages as a mood, not a metric.
Do that if you are a solo builder or a small team who wants to know whether AI engines are actually reading your pages and sending people back.
Stop guessing. Watch.
Most AI visibility tools ask a chatbot and report a guess. Kymo watches your own site instead: which pages AI engines read, how many people their answers send you, and which of those people buy. Every number comes from your server, not from a prompt someone else chose.
Start free → 14-day free trial. No card required.