A page with one Google impression in 28 days took three referrer-less arrivals and two ChatGPT-User fetches in 21 hours: one iPhone reader for 6 min 28 s with the form on screen three times, 0 focus, and 0 checks

The page and its search record

The page is a data post on where AI engines look before they recommend a local business, published in June. Search Console has one row for it in the 28 days to 2026-09-05: 1 impression on 2026-08-15 at position 8, 0 clicks, no query disclosed. Since 2026-06-01 the page has 1 impression and 0 clicks in total. URL Inspection at the read: "Submitted and indexed", verdict PASS, last crawled 2026-07-16. Before 2026-09-04 our analytics had no real session on it at all in the 90-day window; the only rows were our own capture probes (0 probe sessions in the window, excluded by tag).

Three arrivals in 10.7 hours, none with a referrer

Every session on the path is listed. Sessions are numbered and persons lettered; no identifiers are published. The ruling column is our arrivals instrument's, not a judgment made for this post: a session is a read when it carries at least one engagement, scroll, form-view or leave event after its first pageview, and a two-event direct hit from a listed hosting block on Linux at 1920×1080 with 0 engagement is an arrival that is not a read.

Person A opened the page at 14:52 UTC on 2026-09-04, on an iPhone at 390×641, and stayed 3 min 58 s visible: the form was on screen 16 s in at 10% scroll, the scroll milestones ran to 100% inside 80 s, and the session ended with a heartbeat at 238 s and a page-leave. The same person came back at 17:20 UTC, 2.5 hours later: 3 pageviews in one session (a reload at 5 s and a return after a leave), 150 s visible, 83% max scroll, 7 dead swipes, the form on screen 2 more times. Across both sessions: 6 min 28 s visible, 3 form views, 0 form focus, 0 checks started.

Person B is the third row: 01:34 UTC on 09-05, Chrome on Linux at 1920×1080, 2 events (one pageview and one form-view 21 ms apart on capture time), 0 engagement events, from an address inside a listed hosting block. The instrument's rule for that shape was added the same morning and names this row; it stays an arrival in the day's count and is not a read.

The fetch ledger around the visits

Our server records every request whose user agent names a crawler, agent or unknown bot, with the path and the time. On this path the 30 days to the read hold 26 such rows from 11 tokens. The table is the by-token count; the rows within 24 hours of the first visit follow.

Two rows are ChatGPT-User, the agent OpenAI sends when a person's ChatGPT session opens a page: 08:45 UTC and 12:08 UTC on 09-05, 17.9 and 21.3 hours after person A's first visit and 7.2 and 10.6 hours after person B's. They are the first ChatGPT-User rows on this path in the ledger's 30 days; the earlier OpenAI rows are OAI-SearchBot, the index crawler, on 09-02 and 09-03. Meta's crawler fetched the page 3 times on 09-05, the first at 01:05 UTC, 10.2 hours after the first visit. Before the first visit the path has 0 user-class fetch and no row from a messaging or mail link-preview agent.

What a referrer-less iOS arrival can and cannot be attributed to

The arrivals carry nothing that names a source: no referrer, no utm parameter, no engine marker in the URL. On this site a visit that starts from a ChatGPT answer has carried `utm_source=chatgpt.com` in 21 of 21 ChatGPT-origin sessions in the last 30 days (0 arrived with a chatgpt.com referrer and no tag), so these three did not come through the link shape ChatGPT's web and app answers have used here. Sitewide, sessions with no referrer are 68 of 153 (44.4%) in 30 days, against 46 from Google and 21 from ChatGPT; a referrer-less row is the largest class and the least attributable.

• Consistent with the rows: a link opened from a native app that strips the referrer (a mail client, a messaging app, a notes app, an AI app's in-app browser), a pasted or typed URL, or a bookmark. The two ChatGPT-User fetches are consistent with someone pasting or citing the page inside ChatGPT after the visits, or with ChatGPT retrieving it for an unrelated answer; either way they came 17.9 hours later and cannot have caused a visit that preceded them.
• Not supported by the rows: a Google click (Search Console records 0 clicks and one impression, on 2026-08-15); a ChatGPT link-out in the shape seen on every other ChatGPT arrival here (no utm, no referrer); a shared link in a messaging app that fetches previews (no preview-agent row before the visit; Meta's crawler came 10.2 hours after).
• Unknowable from first-party data: which app or person put the link in front of the iPhone reader. The honest attribution for all three rows is "no referrer"; the page's only external signal in the same day is that it was fetched by ChatGPT-User twice and by Meta three times after the visits.

The day's counts on the instrument

For scale: the arrivals/reads instrument, which excludes our probes, listed scanners and AI fetchers, reads the three days around the visits as below. Person A's two sessions are two of the 4 reads on 09-04; person B's row is the one arrival and zero reads on 09-05 at the time of the read. The trailing 7 days hold 29 arrivals and 22 reads on the whole site.

What this means if you care about AI visibility

• A page can be invisible in Google (one impression in 28 days, indexed, last crawled 2026-07-16) and still be the page a real reader spends 6 min 28 s on. Search Console does not see any of this traffic and never will; the referrer-less class was 44.4% of this site's sessions in 30 days.
• ChatGPT-User rows in a server log are a person's ChatGPT session opening your page, and they are timestamped. Read them against your own analytics before crediting ChatGPT with a visit: here both fetches came hours after the visits, so they are evidence that the page was in a ChatGPT conversation, not that ChatGPT sent the reader.
• The most engaged reader on the page saw the free check form three times and did not touch it. A see-to-touch leak on the site's best-read arrival is the finding that matters more than the attribution; the form is on this site's data pages at 10% scroll and the reader passed it three times.

Limits

• Three sessions and two persons. This is a description of one page's day, not a rate; nothing here is scored against a bar.
• The ledger's user-agent detail is read from a 5,000-row site-wide pull whose floor at the read was 2026-08-21 16:40 UTC; the 4 older "other" rows on the path are counted from the path index without their user agent. All rows near the visits are inside the detail window.
• Visible seconds are our own heartbeat counter (15-second cadence plus milestone events); the counter restarts on each pageview and the per-session figure is the sum of the segment maxima. Scroll depth is the highest milestone reached.
• Search Console's last two days are marked fresh by Google and can still move; the 28-day figures are stated as read at 2026-09-05 16:44 UTC.
• The hosting-block ruling on person B is a rule of our instrument, added the same day from this row, and is stated as such. The row is still counted as an arrival.

Trailing indexation, read the same run

The daily read of the 14 posts our rebuilt /news index links, day 9 of a 15-day read that ends 2026-09-11 with a bar of 7 of 14 indexed: 1 of 14 indexed, 10 "Discovered, not indexed", 3 "unknown to Google", 2 Googlebot rows on the 14 in 7 days, both on the one indexed post. The 8 posts published 2026-09-02 to 09-04 were inspected in the same run: 7 "Discovered, not indexed", 1 "unknown to Google", none with a lastCrawlTime.

The next reading, pre-committed

2026-09-12: the same three pulls on the same path (analytics sessions, ledger rows, Search Console). Published either way, with three lines that can be wrong: whether a fourth real session lands on the page and carries a referrer or a utm tag (a `utm_source=chatgpt.com` row would name ChatGPT as the source of at least one later visit; none would leave the three as "no referrer"); whether person A returns and touches the form; and whether Search Console records a second impression or a first click. The 14-post index read closes 2026-09-11 at its bar of 7 of 14.