An AI agent that measures whether an engine mentions your company every day is cheap to build and easy to build wrong. The version that holds up asks the same frozen buyer questions every morning from a session that can see nothing but the question, keeps the engine's own search results beside each answer, counts two control brands you do no work for in the same run, and refuses to read a one-day change as a result. Ours has done that since October 15 on GPT-6 Sol, the model behind the search a buyer uses, and on Claude: 20 questions, one answer each per engine, 40 answers a morning, 18 readouts so far. On November 1 our own three counts were 0, 0 and 0. This post is what the five numbers are, what moved in the week, and what one morning can and cannot tell you.
What GPT-6 Sol's search returns for this question
On the morning of November 1 we put the question through GPT-6 Sol's web search and kept the list: 26 results, none of them ours. Five were Reddit threads, among them "Anyone using tools to track whether your brand shows up in ChatGPT/AI answers?" and a founder's report of 0 mentions in 11 category questions. Twelve were tracker pages: Lettertrace, Teun.ai ("Is your business mentioned in ChatGPT? Check for free"), Noma, Shareof.ai, Answer Insight ("Know When ChatGPT Mentions Your Brand"), Genwolf, Nuwtonic, Aelo and others, and five were TechRadar reviews, of HubSpot's AEO tool, Surfer's tracker and Ahrefs Brand Radar among them. We read the titles only. The titles sell the yes or no. What none of them can say in a title is how the yes or no was taken, and that is the whole difference between a measurement and a mood.
What our agent does every morning
| Step | What happens | Why it is there |
|---|---|---|
| Probe | Each engine is asked what its session can see: tools, folders, files, connectors. Anything beyond the question stops the run | A session with memory or project files is answering about you, not about what a stranger would get |
| Ask | The 20 frozen buyer questions, none naming a brand, once each to GPT-6 Sol and to Claude Opus 5, from an empty folder with a meaningless name, web search on | The questions are the ones a buyer types before knowing who exists; changing them changes the series |
| Keep | Each answer is kept with its sources and the search results the engine saw, on the machine, never published | Retrieval is the earliest number to move, and only the stored results show it |
| Score | String matches for the brand's spellings and domain, the vendor dictionary, and two control brands; every hit on our name is read by a person before it counts | Another company shares our name, so an unread hit is a false positive |
| Summarise | Five numbers beside yesterday's and the seven-day pool of 280 answers, with the run record on top | A day is 40 answers; the pool is the unit a change is read in |
The five numbers are retrieved (our domain was in the results the engine searched, cited or not), mentioned (the answer's text names us), cited (the answer's sources include our domain), agency named (the answer names a done-for-you agency other than us, the slot we compete for), and the two controls, answers naming Profound and answers naming Peec. Counts are added across days and engines, never averaged into rates, and an answer the engine failed to give leaves the denominator rather than counting as a zero.
An AI agent that measures whether ChatGPT mentions your company every day: what it read this week
| November 1 (40 answers) | October 31 (40) | Pool Oct 26 to Nov 1 (280) | Pool Oct 19 to 25 (280) | |
|---|---|---|---|---|
| Retrieved (ours) | 0 | 1 | 3 | 2 |
| Mentioned (ours) | 0 | 0 | 0 | 0 |
| Cited (ours) | 0 | 0 | 2 | 1 |
| Agency named | 10 | 8 | 70 | 73 |
| Profound named (control) | 10 | 12 | 89 | 92 |
| Peec named (control) | 9 | 13 | 86 | 85 |
Read the day first, the way a buyer would. Peec fell by four answers between October 31 and November 1, Profound by two, and our retrieval went from one to zero. At one answer a question a day, a control brand moving by three is the engine's own variation, so the Peec move is one over the band and nothing else on the row is a reading at all. Our one retrieval on October 30 and 31 was Claude's search returning one post of ours on one question, on choosing an AEO agency without an in-house team; on November 1 the same search did not return it. A single search does that from one morning to the next, and the series is built so that nobody has to decide on November 1 whether it meant anything.
Then the pool, which is the unit.
Over seven days, 280 answers, our domain was in the search results three times and among the cited sources twice, against two and one in the seven days before; both control brands stayed inside a band of ten answers (89 against 92, 86 against 85). The engines searched: GPT-6 Sol on all 20 questions on November 1, Claude on 19, answering one from memory, and a question answered from memory cannot reach a page written last week. The honest sentence for the week is that our pages reach an engine's search results on one or two questions and its answers on none, and that both controls say the engines did not move. We have published that sentence, or a flatter one, every week since the series began.
Why one day is noise and a pool is not
The rule we wrote down before the first readout is that the only daily reading that is never noise is our own count leaving zero, and even that is read in the answer's text before it is published, because the name is shared with an arcade. Everything else is read in the 280-answer pool against the pool before it, and a move in our numbers only counts when the controls did not move the same way. The daily series exists to catch the day a page first appears in the results, which is the earliest signal there is, and to spread the asking so no one day's engine mood decides the week. A tracker that reports "mentioned today, not mentioned yesterday" without controls and without the stored search results is reporting the coin, and a founder reading it will change pages for reasons that were never there.
What a buyer should ask of any daily tracker
Whether the questions are frozen and who wrote them. Whether the asking session is clean, and how that is checked each morning. Whether the search results are stored, not only the answer, since retrieved moves weeks before cited does. Which brands are measured that the tracker's owner does no work for. And what the tracker said in a week when nothing moved: ours said so on October 12 after a 200-answer round and has said so in most of the 18 daily records since, each on the AI visibility agent's log with the run's steps and the pool's dates. A tool that has never published a flat week has either not had one or not published it, and a buyer should ask which.
How we checked
The daily figures are from the AI visibility readout of November 1 (40 of 40 answers, both probes clean, Claude Opus 5 and GPT-6 Sol) and its seven-day pool of October 26 to November 1, beside the pool of October 19 to 25; the first daily readout was October 15, and the eighteenth is November 1. The definitions of the five numbers, the noise band and the review of flagged answers are from the agent's definitions table. The search results for the buyer question are from our GPT visibility round of November 1, one search with the list kept; we did not open any tracker's page in it, so no vendor's price or method is quoted. The raw answers stay on the machine and are not published, because each carries the account it was asked from.
Disclosure
ClawWorld runs AI growth agents for early-stage startups and small businesses: they carry out the growth routine a team already planned (posting, replying in communities, outreach, launches and measurement) on a schedule, with every run on a public log, and AEO (answer engine optimization) is one of the services. The measuring agent in this post is one of them, free to copy from its page or priced per agent to have us run it, so we have an interest in this question. The numbers above are our own, and the zeros are left in.