Draft. This methodology will be finalized and dated before Edition 1 is published. Any change after that is recorded in the methodology changelog.
What is measured
The Census records two things in each AI answer, and reports them separately:
- Mention: a firm or company is named in the text of the answer.
- Citation: a web page is cited or linked as a source for the answer.
A firm can be mentioned without its website being cited, and a page can be cited without the firm behind it being named. Treating them as one number would hide both.
Engines
ChatGPT, Perplexity, Microsoft Copilot, Google Gemini, Google AI Overviews and Claude. Each engine keeps the same chart color in every edition.
Questions
Each sector has a fixed set of buyer questions, written the way a buyer would ask them, for example a question asking which firms to consider for a particular service in a particular place. The same questions are reused across editions so results can be compared. Every question is published in full with each edition.
Collection
Questions are asked on set dates from the market being measured. The date, location and engine version (where an engine shows it) are recorded with every answer. AI answers vary between runs, so each question is asked more than once and the variation is reported.
Limits
- Results are a snapshot of what the engines returned on the dates tested, not a measure of a firm’s quality.
- Answers can change when engines update their models or sources.
- A firm not named for one question may be named for a differently worded one.
Corrections
If a result looks wrong, the published question set and dates allow it to be checked. Confirmed errors are corrected and the correction is noted on the edition page.