Does ChatGPT send visitors to pages Google buries? Three referrals in 92 minutes landed on a comparison page Google ranks 76th — one of them 3 minutes 11 seconds after ChatGPT's own fetcher read it

The night, timestamped

The page is `askedabout.com/compare/ai-visibility-tools`, a comparison of AI-visibility tools with disclosed prices. Google indexed it (last crawl 2026-08-11 23:17Z) and shows it at average position 75.8 — page eight — with 128 impressions and 0 clicks in the trailing 28 days; our whole `/compare` section reads 483 impressions, 0 clicks, average position 72.1 in the same pull. ChatGPT's traffic to it looked like this, every timestamp from our own instruments:

Method: a referral is counted by our client-side detector (`ai_referral_detected`, fired when the referrer or utm tag matches a known AI engine), read from real-visitor session rows with our own probe and scanner traffic excluded by tag and IP list. The crawler ledger records every bot-shaped user agent server-side. The Search Console numbers are the standard 28-day page pull. All three instruments are the same ones our weekly self-audit publishes from every week.

What cuts against the clean story

• No arrival lived three seconds. All three sessions — and the night's two other arrivals on other pages — recorded exactly one event: the landing pageview. No scroll, no web-vitals, no leave event. We verified the sensor itself was intact the same morning by replaying the served production build under interception and watching the full event sequence post within 9 seconds. Whether the three were humans who bounced immediately or automated clients that executed JavaScript and closed, the rows cannot say, and we are not deciding.
• One of the two tagged referrals was lost by our own batching. The 01:14Z visitor's detector event sat in posthog-js's client-side batch and the tab closed before it flushed; the session survives only as a pageview whose URL carries the utm tag. Counted on the raw utm basis our lifetime is 8, not 7. We fixed the defect the same day (the event now sends immediately) and we report both numbers; our scoring stays on the detector as written.
• A page can be fetched, referred to, and still convert nothing. Three referrals, zero engagement events, zero checks started. A referral onto a page is not a reader on the page.

Why this is worth a datum

Google and ChatGPT disagreed about the same URL on the same night. Google has shown the page 128 times in 28 days at position 75.8 and sent nobody; ChatGPT fetched it with its user-directed agent at 00:37Z and referred visitors to it minutes later — the fetch-then-referral gap of 3 minutes 11 seconds is the closest pairing our ledger has recorded between an AI fetcher and a human-shaped arrival on the same URL. It is one night, one page, three sessions, on one small site: too small to prove a channel, exactly the right size to timestamp. The pattern it suggests — AI engines routing users to pages classic search ranks nowhere — is the one our self-audit exists to measure, and the next weekly read (2026-08-24) will report whether the engines' citation sets moved with it. If you want to know whether AI engines send anyone to pages Google buries on your own site, the instrument is a referral detector plus a crawler ledger — or a free check that reads which engines currently cite you.

Limits

Three sessions in 92 minutes is an anecdote with timestamps, not a trend. The 00:41Z arrival carried no utm tag, so its attribution to ChatGPT rests on the 3-minute proximity to the ChatGPT-User fetch, which is a correlation. One of the three arrivals came from a colocation IP range, which is where automated browsers live — and also where VPN egress lives. The utm-basis count (8) and the detector count (7) will differ until the batching fix has been live long enough to prove no further loss. We publish every one of these numbers weekly and will say if the pattern repeats or dies.