ChatGPT opened our homepage 206 times in 7 days and cited it in 3 of 36 sampled answers. A week earlier: 148 fetches and 5 of 36. Retrieval rose 39%, citations fell

Method, stated before the numbers

Two first-party instruments, joined at one timestamp. The fetch half is the server-side crawler ledger described in our crawler-allocation cut: the site's edge middleware writes every bot-shaped request to a table, and a row counts here if its User-Agent token is ChatGPT-User and its path is exactly the homepage, any query string included, because ChatGPT appends its own tracking parameter when it opens a page. OpenAI's bot documentation describes ChatGPT-User as the agent that fires when a user's question leads ChatGPT to visit a page, not a crawler that walks the web on its own, so this count is retrieval at answer time. The ledger was born 2026-08-17 01:48 UTC; windows that end before that print UNMEASURED, never zero.

The cite half is the weekly self-audit gauge this site has run since June: a frozen panel of 12 buyer-shaped questions, sampled 3 times each against OpenAI's API with search enabled, every Monday at 13:07 UTC. That is 36 OpenAI samples per run, in 12 question cells, the same gauge whose 4-run cell history we published on 2026-08-25. A sample counts as citing the homepage if its citation array contains askedabout.com with path "/", any query string. For each gauge run, the pair is: ChatGPT-User fetches of "/" in the trailing 7 days ending at that run's timestamp, beside homepage-citing samples in that run. The reading script is a standing verifier in our repo; the artifact and cut script for this post are committed beside it, 12 of 12 self-checks passing, and no number in this post is typed from memory.

The pair, every gauge run since the ledger was born

At the moment this post shipped, the trailing 7-day fetch count stood at 207. The gauge's whole-panel citation rate across all four engines barely moved between the last two runs, 0.0556 to 0.0559, while homepage fetches rose from 148 to 206. The fetch series climbs; the cite series wobbles around a handful of samples.

The same pair, per question cell

Only 3 of the 12 panel questions have ever produced an OpenAI answer citing our homepage. Here is each of them across the three measured runs, as cited samples out of 3. The other 9 cells are 0 of 3 in every run.

The cell that carried the first reading, "How can I check what ChatGPT says about my business?", went 3 of 3, then 1 of 3, then 0 of 3, across three weeks in which retrieval-time fetches of the page it cited rose from 6 to 206. No cell's trajectory follows the fetch curve.

What this says, and what it does not

The standing interpretation in our own log, stated as an interpretation and not a finding: engines cite from their index, not from the fetch. A ChatGPT-User fetch is ChatGPT opening a page while composing an answer; whether the finished answer names that page as a source is decided elsewhere. The two series above are consistent with that reading and this week they moved in opposite directions. RESONEO's retrieval study found the same asymmetry from the other side: ChatGPT opened the pages it was about to cite in only 1 of 80 cases, and when it did open a page, it cited it 74% of the time.

For a business reading its own server logs, the practical point is narrow: a rising count of ChatGPT-User hits is real activity, and it is not a citation meter. Our homepage is being opened by ChatGPT about 29 times a day, which means user questions are pulling it into answer composition at that rate, and our sampled citation count fell the same week. A log dashboard that reports "AI traffic" from user-agent counts is measuring retrieval. Whether the engines actually name your business when a buyer asks is a different measurement, and it is the one that our free check reads directly.

Limits

• One site, one page, three measured weeks. This is a series we publish as it accrues, not a study.
• 36 OpenAI samples per run is a small panel. 5 of 36 against 3 of 36 is a movement of two samples; we publish the counts and make no rate claim on the difference.
• The 2026-08-17 fetch figure covers 0.47 of its 7-day window because the ledger was half a day old; it is floored, not comparable to the later weeks.
• The ledger trusts the User-Agent string. A ChatGPT-User token can be spoofed; our spoof-path scanner exclusion applies, but UA identity is claimed, not proven.
• The gauge panel is frozen (2026-06-27) so runs stay comparable; it samples OpenAI's API with search enabled, which is not identical to the consumer ChatGPT product.

The next reading, pre-committed

The gauge fires next on Monday 2026-09-07 at 13:07 UTC and the pair will be read the same way. The reading rule on file since 2026-08-24, quoted: "two points make a reading; fetches ≥5 with cited cells → 0 ⇒ retrieval ≠ citation; both fall ⇒ the homepage left the retrieval set." If the 09-07 run shows fetches still above 100 with citing cells at 0, the divergence extends by that rule; if citations recover while fetches hold, the wobble reading stands; both counts get published either way, from the same instrument.