How we count the answers.

We ask four engines: ChatGPT (OpenAI), Gemini (Google), Perplexity and Claude (Anthropic). Every question goes through the vendor's official API with its web search turned on: OpenAI's web search tool, Gemini's grounding with Google Search, Perplexity's built-in search, and Claude's web search tool (up to 3 searches per answer).

Each call is a fresh conversation: no account, no chat history, no memory. A short instruction, like the one the consumer chat apps use, gives the model today's date and asks it to search before recommending apps. Each report lists the model it used for each engine.

A second model then reads every answer and lists the apps it names. An answer counts for your app when one of those apps matches your App Store ID or your app's name. An answer that names your app twice still counts once.

Questions are written from your App Store listing and grouped into intent clusters. A cluster is one need (“Outfit planning”, “Calorie counting”) asked several different ways, because real people phrase the same need very differently and assistants answer the need more than the wording.

• Up to 4 clusters of 3 to 5 questions, plus an optional “alternative to [rival]” cluster.
• Generic questions never name your app, your developer or any other app. They are written like someone who has never heard of you.
• Written in your market's language: English for the US and UK, Turkish for Türkiye.
• On the weekly plan you can read, edit and replace every question.

N is the number of answers we got back and read for one engine. X is how many of them named your app. “11 of 45 answers” means ChatGPT named your app in 11 of the 45 answers it gave that week. We show the count first because a percentage from a small sample looks more certain than it is.

A failed call is not a zero. If an engine times out or returns an error, that answer is left out of N. The report says so (“6 of 8 answered”), and an engine that didn't answer at all shows as not measured.

Ask the same question twice and you rarely get the same list. So each count comes with a 95% range: the rates that fit what we saw. We use the Wilson score interval, which behaves well with small samples and with counts of 0 or N. Zero mentions is still a range: 0 of 8 means “probably under 32%”, never “0%” or “invisible”.

Even these ranges are optimistic. Asking the same question 3 times gives answers that are more alike than 3 different questions would, so the real uncertainty is wider than the formula says. That is why the free card asks 8 different questions once instead of 4 questions twice, and why we don't read a trend into a single week.

We only call a change real when the ranges separate: the 4 weeks after you log a change against the 4 weeks before it. If the ranges overlap, the report says noise.

ChatGPT, Gemini, Perplexity and Claude use different models, different search backends and different sources. A single blended rate would hide which engine recommends you and would move whenever the mix of engines changed. Every number on AppCited belongs to one engine.

The free card asks the first 2 questions of each of 4 clusters. The weekly report asks your full set of 15 questions, 3 times each, with the same engines and settings, so the numbers don't jump when you sign up.

What the APIs answer can differ from what someone sees in the ChatGPT or Gemini app. The apps add accounts, memory, location and their own search settings; public comparisons of the two found little overlap in the brands they name. We use the APIs because they can be asked the same way every week, within the vendors' terms. Read our numbers as a consistent panel that shows direction, not as a copy of any one person's screen.

• Rank or position. Lists change order on nearly every run, so a position number would be noise.
• Sentiment. We count whether you were named, not whether the answer was kind.
• Google AI Overviews and AI Mode. They are not part of the panel, and we don't scrape consumer interfaces.
• Daily changes. We measure once a week; a single answer on a single day says very little.