How Llamat collects and counts AI answers

What a report measures, and what it does not

A report answers one question: when people ask ChatGPT and Gemini the questions your customers ask, or search for them on Google, do the AI answers bring up your brand? For every answer we record whether your brand is named, whether it is recommended, where it sits among the brands named, whether your website is cited as a source and which other brands are named instead.

From those records, your report shows your AI Visibility score, mention rate, recommendation rate, average position and citation rate, overall and for each engine. It adds a heatmap of prompts by engine, the competitors named in the same answers, how the engines describe your brand, which kinds of sources they cite and the opportunities to work on first. What these metrics mean in general is explained in our guide to AI visibility.

A report does not measure:

• Traffic, clicks or sales from AI assistants. It counts what the answers say, not what people do next.
• What signed-in users see. Answers shaped by an account, chat history, memory or a paid plan can differ from the logged-out answers we collect.
• Every question about your category. It covers the prompts in your report, up to 20, in one market.
• Whether the engines are right about you. The report shows how they describe your brand; it does not check those descriptions against the facts.
• A trend. One report is a snapshot of one day. A change shows only when you measure the same prompts again.
• Other AI surfaces. Google AI Mode, Perplexity, Claude and Microsoft Copilot are not part of the free report.

How the prompts are chosen

Prompts are the questions we put to the engines. We write them from real search phrases instead of making them up, because a phrase people actually search for is evidence of demand. The research runs after you confirm your e-mail address and usually takes two to four minutes:

• Reading your website. A language model we run on our own server reads the text of your website and writes a short profile: what the site offers, for whom, its category and its market, which is one of 20 language and country pairs. The website's text is passed to the model as data, never as instructions.
• Collecting real phrases. A research model looks up the suggestions a search engine shows as people type about that offer, in the market's language and country, and keeps searching from what it finds. It can only submit phrases the search actually returned; the code rejects any phrase the model wrote itself.
• Review by a critic. A separate conversation with the model scores each phrase from 1 to 10 against your website, and only phrases scored 7 or higher are kept. The critic's feedback goes back to the research model, which searches again: up to four review rounds, until there are 20 varied phrases or four minutes have passed. After that, it keeps what the critic has approved.
• Selection. The approved phrases are taken in order of score. Variety comes first: at most 5 from one starting search and no near-duplicates. Only if that leaves fewer than 20 is the list topped up from the rest.
• Rewriting into prompts. The model turns each phrase into a natural question of about 6 to 20 words in the market's language, the way a potential customer would type it into ChatGPT or Gemini. We ask for a mix: about 6 in 10 prompts ask the assistant to recommend providers or products, about 2 in 10 ask what something costs and whom to choose, and with 6 or more phrases at least one compares options and one asks how to choose. A prompt may not contain your brand or your domain, because the report measures whether the engines name you on their own; the code checks this. A prompt that fails a check stays as the original phrase.
• Straight to the engines. Nobody approves the prompts by hand: they go to the engines as soon as the research ends, and the report lists all of them. If your website has several language versions, we research one of them, usually the one at the address you entered. If the research cannot read your website or finds no prompts, we choose the prompts ourselves and the report usually follows within one working day.

Nothing goes to the engines until you confirm your e-mail address from the link we send, which is valid for 48 hours. Your results usually arrive within 30 minutes after you confirm: the research takes two to four minutes, and in a test on 4 October 2026, collecting and evaluating the 140 answers of a full report took 30.5 minutes after the research. ChatGPT and Gemini were done in about 11 minutes; most of the time went to the 20 Google searches for AI Overviews, which run one at a time, and Google blocked many first attempts, so those searches waited and were tried again.

How the answers are collected

Each prompt is asked three times on ChatGPT and three times on Gemini, and searched once on Google for AI Overviews: 7 answers per prompt, up to 140 for 20 prompts (60 from ChatGPT, 60 from Gemini and 20 from AI Overviews). Every ChatGPT and Gemini run opens a new logged-out session with no account, no chat history and no memory, so an earlier answer cannot steer the next one. The runs are interleaved: every prompt's first run goes out before any second run, so the three answers to one prompt are collected minutes apart, not back to back.

ChatGPT. We use the logged-out chatgpt.com web app, the version any anonymous visitor gets, with web search requested; ChatGPT still decides whether to search and which sources to cite. ChatGPT has no country setting, and our sessions run through our own proxy farm with Czech IP addresses. The answer language follows the prompt, so a prompt in English gets an answer in English.

That matters for every other market, the US included. OpenAI's help page says “ChatGPT may use your IP address to estimate your general location, such as your country, state, or city” and that ChatGPT “may share that general location with third-party search providers to improve the accuracy of your results” (OpenAI help, read 4 October 2026). For a question with a local side, such as services in a city, shops, prices or availability, ChatGPT may answer as if the person asking were in the Czech Republic, and a visitor in the US could get a different answer. We have not yet measured how much this changes answers to questions without a local side, such as which software to choose.

Gemini. We use the logged-out gemini.google.com web app through the same proxy farm, with the interface in the market's language. We also pass the market's country, but it does not move the exit: in a test on 4 October 2026 with the country set to the United States, Gemini placed the visitor in Prague, Czechia. So Gemini answers, like ChatGPT answers, come from Czech IP addresses. Gemini sessions accept the cookie notice, because with cookies refused, Gemini cut answers off with an error.

Google AI Overviews. We search for each prompt once on Google through the same proxy farm, so from Czech IP addresses too, with the search set to your market's country and language, and record the AI Overview Google shows for it, with the links in it. Google shows an AI Overview for some searches and not for others, and the same search can get one and, minutes later, none. A search where Google shows no AI Overview is still a successful check, because it is what a searcher saw: it counts as an answer that does not name, recommend or cite you.

Each prompt is searched once, not three times, because Google limits automated searches: in our test on 4 October 2026, 19 of 20 searches sent three at a time were blocked, and searches sent one after another were mostly blocked unless they were about a minute apart. The searches therefore run one at a time, alongside the ChatGPT and Gemini answers. A search that Google blocks before the results load is not an observation of your brand; it counts as a failed answer and is tried again.

The logged-out ChatGPT web app does not say which model wrote an answer, and we do not record a model version for Gemini or AI Overviews either. A report therefore describes what an anonymous visitor got on that day, not a specific model, and the vendors can change the model behind their apps at any time.

If an engine does not answer, we try again through other exit addresses. An answer that still fails is left out of the counts. A report goes out only when each engine returned at least 60% of its answers; otherwise the order goes back into the queue and is tried again later. The one exception is AI Overviews: if Google still blocks so many searches that fewer than 60% get through on the order's last attempt, the report goes out without AI Overviews, covering ChatGPT and Gemini only, rather than not at all. Every answer is stored with its sources as soon as it arrives.

ChatGPT, Gemini, AI Overviews and the paid add-ons

The free report covers ChatGPT, Gemini and Google AI Overviews. Google AI Mode, Perplexity, Claude and Microsoft Copilot are available as paid add-ons on request. This page describes the free report only.

Google AI Overviews joined the free report on 4 October 2026; until then it was one of the paid add-ons.

Perplexity was listed in the free report until 2 October 2026. It did not answer logged-out visitors in our tests in September 2026, and in our own measurement on 1 October 2026 it returned 0 of 12 answers and showed a sign-in wall instead. It is now one of the paid add-ons.

How each answer is read

Each answer is read twice: by code that matches text and by a language model we run on our own server. They have separate jobs, and the model never decides whether your brand was mentioned.

• Mention, matched in code. Your brand name is taken from your website, or from your domain name if the website does not give one. The code looks for it in the answer as a whole word, ignoring capital letters and accents, in a few variants: the name as written, the name without spaces or punctuation, a name like PragueCoding split into words, and your domain in full and, if the part before the ending has at least four letters, without the ending. Because no model is involved, a mention cannot be invented. In an AI Overview, the source labels and source cards Google shows around the text are removed first, so your website's name on a source label does not count as a mention.
• Citation, matched in code. An answer cites you when one of the sources the engine lists for it is on your domain or one of its subdomains. An answer that lists no sources counts as not citing you. For AI Overviews, the sources are the links shown in the Overview; a link that still points to a Google redirect is ignored.
• Brands named as options, by our model. The model lists up to 10 brands, companies or products that the answer names as options for the reader, leaving out the publishers of cited sources and generic categories. For each one it marks whether it offers the same kind of thing you do, so that a reader could choose it instead of you. Names the code cannot find in the answer text are dropped, and the list is put in order of where each brand first appears.
• Recommended, by our model. The model decides whether the answer presents your brand as a suggested choice. Every item in a list of suggested options counts, as do a top pick, “best for …” and “I recommend”. A mention in passing, as a point of comparison, as something to avoid or as a source does not. An answer that does not name you never counts as recommending you.
• Position, from the order. Your position is your place among the brands the answer names as options, in order of first appearance. Named first is 1; with two brands before you, it is 3.
• Descriptions and source types, by our model. For each answer that names you, the model writes how the answer presents your brand in at most 15 words, and a second step sums up the descriptions that recur. These are our model's paraphrases, not quotes; the answer excerpts in the report are quoted from the answers, shortened where they are long. The model also sorts the cited domains into types such as review sites, forums or retailers, and a simpler rule-based grouping takes over if it fails.

The model is Qwen3.8-Flash-Next, an open-weight model that we run on our own server; the same model runs the prompt research. The answers are not sent to another AI provider to be read. If the model fails on an answer, the mention and the citation still count. That answer then has no competitor list, does not count as a recommendation, and your brand is treated as the only brand named in it, at position 1.

How the numbers are calculated

Every number in a report is computed by code from the stored answers; the language model never writes a number. The sentences around the numbers are fixed templates filled in with them, so the report does not claim anything the measurement does not show.

Three rules apply throughout. Rates are shares of the successful answers on one engine, so a failed answer neither helps nor hurts you. Overall values are the average of ChatGPT, Gemini and AI Overviews with equal weight, however many answers each returned, so AI Overviews, with one search per prompt, counts as much as ChatGPT; the one exception is average position, which pools every answer that names you. And the majority of runs decides only the ChatGPT and Gemini heatmap tiles, while an AI Overviews tile shows its one search: in every rate, each answer counts on its own. Percentages are rounded to whole numbers and average position to one decimal place. The table below gives every formula.

Weekly re-measurement and tone in the free dashboard

Weekly re-measurement. When your report is ready, we set up your free dashboard at app.llamat.com for your e-mail address, and the e-mail with your results links to it and to the full report. From then on, we measure your prompts again once a week, on the same weekday as your first report, with the same method as the report: each prompt three times on ChatGPT and three times on Gemini, every time in a new logged-out session, and once as a Google search for AI Overviews, with the numbers computed by the formulas on this page. The re-measurement is automatic and is never paused; it runs for as long as your account exists. You can change the prompts in your dashboard, up to 20, and a new prompt is first measured in the next weekly run. The results appear in your dashboard, where you can see how the numbers change from week to week; after each run we also e-mail you a short summary of what changed, which you can turn off. The full report stays in your dashboard.

Tone of the answers. For every answer that names your brand, our model also notes the tone towards it: positive when the answer presents the brand favourably (strengths, praise, a recommended pick without major caveats), negative when it is mainly about drawbacks, complaints, warnings, or advises against it, and neutral for a factual, passing or balanced mention (mixed = neutral). Answers that don't name your brand have no tone. Your dashboard shows the share of each tone among the answers that name you, on each AI platform and over time; the overall share is the average, with equal weight, of the platforms whose answers name your brand (a platform that doesn't name it has no tone and is left out). Quotes on the Sentiment page are copied word for word from the answers; the themes summarise how the answers describe your brand.