The record of what AI actually recommends.

The answer moves, and no one can see behind it.

Ask an assistant what to buy and it collapses millions of signals — documentation, reviews, forum threads, comparisons, tutorials — into a single spoken recommendation. That answer shifts by phrasing, by engine, by day. The shopper sees only the answer, never how it was reached or how stable it is. The brand can’t see it at all — and can’t prove it.

We measure the output, not the training.

We don’t retrain models or claim to read their internals. We ask real buying questions — many phrasings, no brand names — across ChatGPT, Perplexity and Gemini in their web-grounded modes, record which products are named and how often, and compress that into one honest number per brand × category × market.

• A reproducible summary of what the AI shelf says — not a review we wrote.
• We count frequency, not ranking: how often a product recurs is the signal.
• Every observation keeps its sources and a timestamp.

Measured, tested, dated, independent.

• Measured, not asserted — a frequency across engines, not an editorial pick.
• Tested — compared to an honest per-intent null baseline with a conservative interval, so a single affiliate can’t manufacture a signal.
• Dated & attributed — every record names the engines and the date it was measured.
• Reproducible — the sources are kept; the measurement can be re-run.
• Independent — not affiliated with, endorsed by or paid by the AI providers, and the record can’t be bought (see pricing).

So you can buy well-informed.

On a product page, a shopper sees at a glance whether AI assistants actually recommend that product for their need — with a link to the evidence behind it. Honest by construction: when there’s no signal — a new product, or one the assistants simply don’t name — we show nothing, never a negative mark. The full method lives on the methodology page.

Aggregated across a whole category, this same measurement becomes a competitive metric we call AI Market Share — coming soon.