Three steps turn a moving, per-query answer into a fixed, auditable record you can put on a product page. Here is exactly what happens — and what it does not claim.
We ask neutral buying questions for a real shopping need — many phrasings, no brand names — across ChatGPT, Perplexity, Gemini and Claude in their web-grounded modes. For each answer we record which products are named, and how often.
A count on its own proves nothing, so we test it. Frequency is compared to an honest per-intent null baseline (the chance level for that need) with a conservative bootstrap interval, and we check source diversity so a single affiliate can't manufacture a signal.
A record is Verified only when a product clears its baseline across engines, on fresh evidence, and holds over time. Full criteria (C1–C7) are on the methodology page.
When a product is consistently recommended for a real need, it earns a category record. You show it two ways, and both link back to the evidence:
Here is the badge as it lands in a store — quiet, monochrome, and always carrying its category, denominator, engines, date and evidence link:
"Frequently recommended by AI for office" — a measured frequency, on a date, across named engines.
It does not mean "the best", "AI's #1 pick", or that any provider endorses the product. It is not a promise of sales. If the signal drops below its baseline, the record is revoked and the badge stops showing. If its evidence simply ages past 90 days, the badge shows its date instead of a current claim — it never turns negative. We tell the merchant first, either way.