How Rempa measures AI visibility
AI answers change from one session, one country and one week to the next. A visibility number only means something if you know how it was collected. Here is exactly how Rempa asks the questions, which versions of the assistants it reads, what it extracts from each answer and how to read the results.
The questions
Rempa measures a brand on the questions its buyers ask, not on searches for its name. When a project is set up, Rempa reads the brand's website, proposes its competitors, its topics (what it sells, to whom) and, for each topic, prompts written the way a buyer would type them into an assistant: "which tool should a 50-person agency use for…", "what are the best… in the UK". You review and edit everything before the first update.
Most prompts don't name the brand: a buyer who already knows you doesn't need an assistant to find you. In the free check and in our studies, the questions never name the brand measured.
Where the questions are asked
Each AI model is read the way your buyers use it, with its web search on, rather than through a bare model API that answers from training data alone.
| Model | What Rempa reads | Located in your market |
|---|---|---|
| ChatGPT | The consumer ChatGPT, with web search, as on chatgpt.com | Yes, country |
| Gemini | The Gemini app (gemini.google.com), opened in a real browser session | Yes, country |
| Google AI Mode | The AI Mode answer on Google | Yes, country and language |
| Google AI Overviews | The AI Overview at the top of Google's results page, when Google shows one | Yes, country and language |
| Copilot | Microsoft Copilot answers | Yes, country |
| Perplexity | Perplexity's Sonar model with its web search | Yes, country |
| Claude | Claude with web search, as in claude.ai | Yes, country |
| DeepSeek, Qwen | The model through its API, without web search | No |
ChatGPT and Gemini can also be tracked through their API without web search, to see what the models know about you from training alone, separately from what they find online.
Clean sessions, set in your market
Every question is asked in a new session: no account, no conversation history, no memory, no custom instructions and no system prompt. The question is sent as written, with nothing added. The country and language are those of the project's market (52 markets are available), so a question tracked for the UK is answered as it would be for a new user in the UK.
This is a deliberate choice. A signed-in user's answers depend on their history and memory, which differ for every buyer and can't be reproduced. A clean session measures the baseline that every new buyer starts from, and it is the same at every update, so changes over time come from the models and the web, not from the person asking.
How often
At each update, every prompt is asked once to every model of the project. Updates run every 3 days on the Starter plan and every day on Pro and Enterprise; the free trial gets one update. The free AI visibility check and the studies on this site ask 12 questions once to ChatGPT, Gemini and Google AI Mode.
Asking each prompt once, over many prompts, gives more information than asking a few prompts many times: precision comes from the total number of answers behind a rate, and many prompts also cover more of what buyers ask.
What we read in each answer
- Mentions. Whether the brand and each competitor are named, by their name, their other names (short names, former names, spellings) or their website, as whole words. A name that is also a common word ("instantly", "notion") only counts when the answer is really about the brand: such matches are checked against the answer before they are kept.
- Position. The rank of the brand among the tracked brands the answer names: first, second, third…
- Sentiment. How positively the answer talks about each brand it names, on a scale from 0 (very negative) to 50 (neutral) to 100 (very positive). A language model scores it from the answer's text, writing a one-sentence reason before the score, which you can read next to every answer.
- Sources. The pages the model cites, as it returns them, with their site. The brand counts as cited when its own website is among them.
- Search demand. Each prompt is matched to one to three equivalent Google searches, with their monthly volume in the market. It is a proxy: no assistant publishes how often a question is asked.
The full text of every answer is kept, with the model version when the assistant states it, so any number in a report can be traced back to the answers behind it.
How to read the numbers
Visibility is a rate: the share of answers that name the brand. Every rate has a margin of error that shrinks with the number of answers behind it. With the Wilson interval at 95% confidence, a brand named in 18 of 50 answers (36%) is somewhere between about 24% and 50%; with 200 answers at the same rate, the margin falls to about 7 points either way.
In practice:
- Read rates at the level of a topic or a model over many prompts, not prompt by prompt. One prompt on one day is one draw.
- Read trends over several updates. A move from 36% to 42% on 50 answers is noise; on several hundred, it starts to mean something.
- Compare brands on the same prompts and models. Shares of voice are only comparable within one set of questions.
- When the assistant changes its model version, a step in the series can come from the model, not from you: the version is recorded with each answer.
Limits
- Personalised answers are not measured. What a signed-in user sees depends on their memory and history. Rempa measures the answer a new user gets.
- Location is by country. Answers are read for a country and a language, not a city: local questions ("near me") are not localised further.
- Not every answer has sources. Gemini, in particular, often answers without citing pages; citation rates are lower there for every brand.
- Sentiment is a model's reading. It is consistent across answers, but it can misread irony or a comparison. The reason given with each score lets you check it.
- Prompts are a sample. The results describe the questions you track. Adding the questions your sales team hears makes them closer to your market.
Questions about the method
Does Rempa use the AI models' APIs or the real assistants?
The real assistants, with their web search: the consumer ChatGPT, the Gemini app, Google AI Mode and AI Overviews, Copilot, Perplexity and Claude with web search. Only DeepSeek and Qwen are read through their API, without web search. ChatGPT and Gemini can also be tracked through their API without search, to compare what they know from training with what they find online.
Do personalised and location-based answers make AI visibility tracking inaccurate?
They make a single manual check misleading, which is why Rempa asks every question from a new session, without history or memory, in the country and language of your market. That measures the baseline every new buyer gets, the same way at each update. What it does not show is the answer of one particular signed-in user, which nobody can reproduce.
How reliable are the visibility numbers from one update to the next?
As reliable as the number of answers behind them. Each update asks every prompt once to every model; a rate over 50 answers has a margin of about ±13 points, over 200 answers about ±7. Read changes at the topic or model level and over several updates, not prompt by prompt.
How is sentiment measured, and how accurate is it?
A language model reads each answer and scores how positively it talks about each brand it names, from 0 to 100, after writing a one-sentence reason. The reason is shown next to the score, so you can check any surprising reading. It is a consistent measure across thousands of answers, not a human review of each one.
How many prompts do I need to track?
Enough to cover your buyers' main needs, and enough answers per topic for a stable rate: 30 to 50 prompts give a first picture across three models, 150 or more give rates you can follow topic by topic. Spread them across topics, buyer types and decision stages rather than writing variants of one question.
Which sources does Rempa count as citations?
The pages each assistant returns as sources for its answer, as the assistant gives them. Links an assistant gives to itself, such as a Google search page, are not counted as sources in our studies.