How the number is produced

Your questions are written for your category — category leaders, alternatives to incumbents, specific use cases, comparisons, budget qualifiers — phrased the way people actually type them, not the way marketers write them.

Your brand name never appears in a question. The whole measurement depends on whether an engine volunteers you unprompted. Every question is printed in your report.

Each question goes to every engine below, with web search enabled. This matters: a model answering from training data alone is months out of date and measures nothing useful.

ChatGPT OpenAIClaude AnthropicPerplexity PerplexityGemini GoogleGrok xAI

These are the vendors’ own search-grounded APIs. They are the closest faithful proxy available — a consumer account’s answers are shaped by that user’s history and settings, which can’t be reproduced for anyone else.

How many of these run depends on the plan — the entry plan covers 3, the others cover all 5. Your report names the engines it used, every time.

If an engine can’t be reached, your report says so on its face and scores only the engines that ran. We never fill a gap with an estimate.

These systems are non-deterministic: the same question asked twice can return different products in a different order. A report that asks once and prints a tick or a cross is telling you about a coin flip.

So every question is run several times on every engine — up to 600 graded answers for a single report — and results are reported as rates. When your report says “2 of 3 runs”, that is the finding, not a rounding of it.

• Mention rate — how often an engine named you, as a share of its runs.
• Average rank — where you land in the list when you do appear.
• Share of voice — how much of the answer space each competitor holds. Counted once per answer, so being named twice in one reply earns no extra credit.
• Overall score — the engines combined, weighted by position. Being named first counts for much more than being named eighth.

Issued for an engine only when it recommended you in at least 50% of its runs. Date-stamped, and linked to a public page showing the counts behind it.

recomention has no relationship with these vendors, and a badge is not an endorsement by them. It attests to one thing: what the engines said, on that date, in our test.

It’s a snapshot, not a status. Run it again next month and the numbers will move, because the web moved.

It’s not what one specific person sees. Consumer AI products personalise answers using history and settings we have no access to.

It’s a sample. Depending on the plan, that is 15 to 40 questions — enough to find a clear pattern, and never every question a buyer could ask. A wider plan widens the sample; it does not remove this limit.

Extraction is automated, and imperfect. Occasionally a product is miscounted or an unusual phrasing missed. Every verbatim answer is in your report so you can check our work rather than take it on faith.