We record what ChatGPT actually does before it answers - every search it runs and every page it reads - and publish what that says about getting recommended. Every number here comes from those recordings.
Each spoke is something ChatGPT put into its own searches before it answered. Longer means it came up more often.
Read off 158 searches across 72 recorded answers. Re-scanned every time we capture.
106 recorded answers on this build so far. The same panel of buyer questions is re-run every day, so the day ChatGPT changes build the old answers and the new ones are compared here on the same questions rather than the old figures being left standing.
We traced every recommended business back to whatever first put it there. Only one of these routes is open to a company that is not already famous.
ChatGPT writes its own searches before it answers a question. Every row below is a signal it put in those searches. Open one to see what it means for your site.
29.1% of searches carried a year, and some limited results to the last 30 days.
29.1% of searches asked for "best", "top" or a ranked list rather than a single named answer.
16.5% of searches asked about price, fees, a quote or an estimate.
In 12.0% of the searches it ran, ChatGPT added the word "official" - it was hunting for a company's own authoritative page.
9.5% of searches were restricted to a single website.
3.8% of searches carried award, accreditation, licensing or certification language.
3.2% of searches asked for reviews or ratings rather than price or location.
0.6% of searches named a specific directory, review platform or trade body.
0.6% of searches used proximity language.
71.8% of retrieved results were pages the named brands do not own. This is the condition under which placement can move an answer.
3 of 15 domain-scoped searches targeted a domain no named brand owns. Those are obtainable listings and should be treated as the priority placement targets.
Share of the 158 captured fan-out queries carrying each signal: year_qualified 29.1%, rankings 29.1%, pricing 16.5%, official_lookup 12.0%. A page that cannot satisfy the common ones fails the check silently.
is_claimed, popularity_score, provider, ranking_score, review_highlights, reviews — anything computed from these on an earlier build is no longer computable from the transport.
Across 22 repeat readings of the same question, 63.5% of the sources changed between one day and the next.
Asked twice in the same hour, the answer already moves. Across 14 same-day repeats, 60.4% of sources changed with nothing else different. The worst was “project management software for a small team” at 100%. A single check of whether AI mentions you is one sample of a moving thing, which is why we record on a schedule rather than once.
Between 2026-09-04 and 2026-09-06: tech publishers down 2.38 points.
Share of every citation in the sweep, counted by appearance rather than by distinct site, so a source cited five times in one answer weighs five times as much. Measured on Google's AI Overviews, 160 buyer questions a day. Every sweep is kept as an immutable snapshot at archive.json, raw domain counts included, so any of this can be recomputed rather than taken on trust.
The bar a change has to clear: 1.58 points. Run the identical sweep twice in one day and the answers still differ, so that is how far a class share drifts with nothing changed at all. Anything smaller is churn, and we do not report it as a trend. Measured across 2 same-day repeats.
We put the same 28 questions through a recorded ChatGPT session and through the two API routes that AI-visibility tools are built on. They agreed with the browser on 1.3% and 1.8% of the sources found. Both APIs searched on roughly half the questions the browser searched on, and neither returned a single local business record.
Same questions, same day. The gap is not a vendor problem: one is a third-party monitor, the other is OpenAI’s own API called directly. If a tool reports your AI visibility from an API, it is reporting on a different system from the one your customers use.
We record a real ChatGPT answer for the question your buyers actually ask, then walk you through the trace: the searches it wrote, the pages it opened, and who it recommended instead of you.
Get my ChatGPT visibility scanFree on the call. No pitch deck.We record the network traffic behind real ChatGPT answers - the streamed response and the stored conversation - rather than scraping the page or asking the model what it did. One question per fresh conversation, on a real account, under a fixed daily limit.
Measured effects we can see in our own data. We publish them because a figure you cannot challenge is not evidence.