Beta GEONALY is in open beta. Found a rough edge? Tell us.
Blog · How to
There is no report inside ChatGPT that tells you when it names your brand. To see it, you ask the assistant the questions your buyers ask, record what it says, and repeat on a schedule. This post gives four ways to do that, from a free afternoon to a tool, and the fields worth recording every time.
26 of 80 answers name you. Every run goes in the log, not only the good ones.
Key facts
Open ChatGPT, ask it to recommend a provider in your category, and see whether you are named. That takes a minute and it is where everyone starts. It is also close to worthless on its own, for three reasons.
The answer varies. Models sample, so the same prompt produces different shortlists on different runs. The web results behind the answer change day to day, and the sources the assistants lean on shift quickly: in Semrush's 230,000-prompt study, Reddit's share of ChatGPT's top 25 cited domains fell from about 60% to about 10% within six weeks3. And the answer depends on who is asking. A logged-in account with memory switched on, a chat with history in it, or a different language all change what comes back.
The rule
The thing to build is not a check. It is a series: the same prompts, the same conditions, on a schedule, with the answers kept.
Whatever method you pick, the record is the same. Capture these fields each time, one row per prompt per assistant per run.
| Field | Why it matters |
|---|---|
| Prompt, market, language | A changed prompt is a new series, not a trend. |
| Assistant and model, date | Model updates move answers. You want to know which day the line jumped. |
| Mentioned: yes or no | The basis of mention rate. |
| Position in the shortlist | Named first and named sixth are different outcomes. |
| Cited: yes or no, and the URL | Whether the assistant linked your page, or learned about you from somebody else's. |
| Competitors named | The list of who is being recommended instead of you, by count. The most useful column. |
| Sentiment or caveat | "Good but expensive" is a mention with a cost attached. |
| The full answer text | So you can re-read it when the number moves and see what changed. |
Write ten to twenty buying questions in the buyer's language. In each assistant, start a fresh chat with no history and with memory off where the product allows it, ask each question, and paste the full answer into a document with the date. Repeat monthly. It costs an afternoon a month for one market and one run, which is why it gets skipped by the third month. But it is real data, and it is free.
The same routine, with the fields above as columns. Now mention rate and share of voice are a formula away, and the competitor column can be counted. This is the point at which most teams first see that the four assistants disagree with each other, and that the disagreement is the interesting part.
OpenAI, Anthropic, Google and Perplexity all sell API access, so the run can be automated and repeated as often as you like. Two cautions. API answers can differ from what the consumer apps show, because web search, defaults and models differ between the two. And parsing "was the brand named, and in what position" out of free text needs care, or the counts drift. Keep the raw answers alongside the parsed fields so drift can be found.
A tool earns its price when it does four things: stores every answer word for word, runs on a fixed cadence you did not have to remember, keeps a separate prompt set and competitor list per market and language, and records citations as well as mentions. GEONALY's AI Visibility module is built to do exactly that across the four assistants; other tools cover some of the same ground. Whichever you choose, ask to see the raw answer behind a number before you trust the number.
| Method | What it costs | What it gives you | Where it breaks |
|---|---|---|---|
| Manual check | An afternoon a month, per market and run | Real answers, dated, free | It gets skipped by the third month |
| Spreadsheet log | The same afternoon, plus set-up | Mention rate and share of voice as formulas, competitor counts | Hand entry drifts, and one run a month is a thin sample |
| Scripted API run | Developer time and API fees | Any cadence, any number of prompts, raw answers stored | API answers can differ from the apps, and parsing needs care |
| Monitoring tool | A subscription | Every answer stored, a fixed cadence, one prompt set per market, citations recorded | Trust only a number you can trace to a raw answer |
Monthly is the floor. It matches how share of search is reported and it is enough to catch a model update or a competitor's rise. Weekly is worth it while a campaign runs or after a large change to your own site. Daily produces mostly noise.
| Cadence | When it is the right one |
|---|---|
| Monthly | The floor. It matches how share of search is reported and catches a model update or a competitor's rise. |
| Weekly | While a campaign runs, or after a large change to your own site. |
| Daily | Mostly noise. |
Ten to twenty prompts per market is enough to start, and the count matters less than the consistency. In our share of voice write-up, 80 answers gave 114 brand mentions, so each mention is 0.9% of the total and one extra mention of the brand moves its share by about 0.7 points. Below a hundred mentions, treat swings of a few points as sampling. The fix is more prompts, or more runs of the same prompts, not more decimals.
Three outcomes cover almost every row in the log.
| Outcome | What it means | What to do |
|---|---|---|
| Not mentioned | The assistant learned the category from other pages. | Look at who is named and at the pages cited for them. Your claim needs to exist in a place like that, in the same language. |
| Mentioned, not cited | The assistant knows your name from somebody else's page, usually a review site or a forum. | Write the page it could have cited instead: a plain claim, a number, a source, kept indexed. |
| Cited from a page you do not control | Good news with a risk attached. The page can change or vanish. | Put the same claim on your own site so the assistant has a second source. |
The reasoning behind that list, and the click data that makes the mention matter more than the link, is in GEO vs SEO. The arithmetic for turning the log into a share of voice is in the share of voice post.
Published 21 September 2026. If a model or a figure above changes, this line says what changed and when.
Short answers. The detail is in the sections above.
No. There is no report or setting inside ChatGPT, Claude, Gemini or Perplexity that shows how often a brand is named. The only way to know is to ask the questions your buyers ask, record the answers, and repeat on a schedule.
No. AI Overviews and AI Mode are Google products with their own retrieval, and Google reports more than 2.5 billion monthly users of AI Overviews alone7. Track them as a separate series, with the same prompts.
Models sample, so the same prompt gives different wording and sometimes a different shortlist. The web results behind the answer change too, and the assistants' preferred sources move week to week, as the Semrush study showed. That is why a series beats a check.
Yes. Pew found people clicked a link inside a Google AI summary in 1% of visits, so for most readers the mention is the entire exposure. And the visits that do arrive convert better: Semrush measured AI search visitors as 4.4 times as valuable as organic search visitors across 500 topics.
The AI Visibility trial runs on your real brand for 7 days, with no card and no call. The first answers are in it when the trial opens.