Methodology summary
Questions
Fixed in advance and modelled on real buyer needs (category + spec + standard + region). Three types: non-brand, comparison and brand questions. These are our target questions; we do not claim to know their search volume.
Platforms and modes
- API tests of web-enabled ChatGPT, Perplexity and Gemini models: large, repeatable samples.
- Real-interface tests in web or app, with screenshots and login state recorded: our main client evidence.
- Google AI Overviews: no stable API, so manual screenshots in a stated region, reported separately.
Recorded for every answer
Date and time, platform and model where available, API or interface, login state, region, language and number of repeats.
Metrics
- Mention rate
- Answers naming the company ÷ valid answers
- Recommendation rate
- Answers listing the company as a recommended or candidate supplier ÷ valid answers
- Citation rate
- Answers linking to the company's own website ÷ valid answers
- Share of voice
- Company mentions ÷ all mentions of tested companies
- Accuracy
- Share of key facts in brand answers that match the company's own information
Limits
One answer is not long-term performance. Results change with time, region, account and wording. API answers can differ from web and app answers. Visibility is not the same as inquiries or orders, and no one can guarantee what a platform will answer.
Our own data
On 2026-10-15 we collected 151 API test answers about valve suppliers, from 25 buyer questions across 3 web-enabled engines. That test was API-only: no manual web/app re-check and no Google AI Overviews. Both are planned for our Day 0 test and paid diagnostics. Our sample report is based on it, and we test ourselves the same way on our public scorecard draft page.