Methods
How we measure AI visibility
Every number is a count of real answers. Click any number in the app to read them.
How we ask
- We ask each prompt the way a buyer types it. We never add your brand name.
- Six engines: ChatGPT, Google AI Overviews, Google AI Mode, Claude, Perplexity and Gemini.
- Logged out, with no memory and no chat history.
- Each prompt is asked from its own country, so a prompt for Germany gets the answer a buyer in Germany sees. A prompt written in German is asked in German. Gemini is the exception: its API has no way to set a country, so it answers as it would for anyone.
- ChatGPT, AI Overviews and AI Mode are read as a buyer sees them. Claude, Perplexity and Gemini are asked through their APIs, with web search on.
- We keep every answer, with the pages it cites, from 2026-11-20 on. Older numbers have no answers to open.
How often
- Every day: ChatGPT, Google AI Overviews and Google AI Mode.
- Every month: Claude, Perplexity and Gemini.
- Each prompt is asked once per engine in each check. The same question gets different answers on different days, so a share is counted across many prompts and days, never from one answer.
The 95% range
- A share is answers that name you, out of answers asked.
- Example: 441 of 1,050 answers is 42%. The 95% range is 39–45%.
- The true share is inside that range 95 times in 100.
- Fewer answers make a wider range. 4 of 12 answers is 33%, with a range of 14–61% (± 24 points).
- We use the Wilson score interval. It stays sound for small samples and at 0% and 100%.
The 95% test for fix results
- Each fix names the prompts it should win.
- We compare your share on those prompts in the 42 days before it shipped with your share from 7 days after. The 7 days let engines pick the change up.
- Only engines asked in both windows count. Turning an engine on or off is never a change.
- We use a two-proportion z-test at 95%. A fix is marked up or down only when the move is bigger than the noise.
- With fewer than 8 answers on either side, we say it is too early to tell.
Why one answer is not the truth
- Assistants sample each answer. Ask twice and you can get two different lists.
- Your own ChatGPT knows you: your memory, past chats, location and account. It may name you more often, or less.
- A screenshot is one draw. The share across many answers is the signal.