Methodology
Each leaderboard is a record of what AI assistants answered on one date. Here is exactly how it is made.
1. Questions
Every category has ten questions a buyer would really ask, from the broad ("what is a good…") to the specific ("which one works with Outlook…").
2. Engines
Each question goes to ChatGPT, Perplexity, Gemini, Claude, Google AI Overview and Google AI Mode, with web search on, through the Ranking API. Each question is asked 5 times per engine, so a full run is 10 × 5 × 6 = 300 answers.
3. Counting
For each answer we record which brands it names and in what order. Brands on the category's tracked list are matched by name, alias and domain, and by the brand list the Ranking API reports. Other brands are picked up only when they lead a list item in the answer and look like a proper name; this rule is strict on purpose, so it misses some brands rather than inventing any. A brand found this way is shown only if at least 2 answers (or 2% of answers) name it.
4. Numbers
- Mention rate: answers that name the brand ÷ all answers in the run.
- 95% interval: the Wilson score interval for that rate.
- Per engine: the same rate within one engine's answers.
- Average position: where the brand appears among the brands an answer names (1 = first), averaged over the answers that name it.
- Not mentioned: tracked brands that no answer named.
5. Publishing
A category is published only after a full run is complete. Until then its page says results are being collected and is kept out of search engines. Calls that fail after three tries are left out and the page says how many. Each category is re-run weekly; the page always shows the date of the answers.
What this is not
It is not a review, a test of the products, or a claim about which product is better. It is what AI assistants said.
Questions
Why ask each question several times?
AI answers change from one ask to the next. Asking 5 times per engine shows how consistently a brand is named instead of relying on one lucky or unlucky answer.
What is a Wilson interval?
A standard way to put an error bar on a proportion. If a brand was named in 30 of 300 answers, the interval shows the range its true rate plausibly falls in, given that sample size.
Do you change or summarise the answers?
No. Answers are stored and shown word for word. We only count which brands they name.
Can a brand pay to rank higher?
No. The numbers come only from the answers. The paid audit is advice for a brand; it does not change any leaderboard.