Methodology
When you run a check we read your website, work out what you do and who your plausible competitors are, then generate 8 to 12 realistic buyer-intent questions: category recommendations, problem-phrased questions with no brand named, direct questions about your company, comparisons and alternatives.
We send every question to the real APIs of four providers: OpenAI (ChatGPT), Anthropic (Claude), Perplexity and Google (Gemini). Every answer in your report is a response we actually received and stored, verbatim. We never ask one model to imagine another model's answer.
The 0 to 100 score is weighted as follows:
API answers differ from the consumer chat apps: different grounding, no user history, no memory of you. Treat the report as a controlled, repeatable measurement of how the underlying models answer, not a screenshot of any one user's chat. Answers also vary between runs; that variability is a property of the systems being measured.
Get Recommended runs a full cycle once a month, not once a week. That is deliberate. Week-to-week movement in these answers is mostly the run-to-run variability above, not anything that changed about you or the market. Sending a fresh score every week would mostly be reporting noise and dressing it up as insight. A month is roughly how long real movement takes to show: the sites AI search cites in a category turn over substantially from one month to the next, and the pages and files a fix produces need time to be crawled and to actually change an answer. Between cycles, we only alert you when something unambiguous happens: a competitor newly recommended where you were, you dropping out of an answer you held, or the model measuring you changing.
Every report states the date it was generated and the model identifiers used, so you can compare like with like against the same domain's earlier reports or cycles.