7 ChatGPT-referred sessions on one post: 5 had the call-to-action in view, 3 reached the form, 1 focused it, 0 ran a check. The latest reader stayed 35 minutes in two browsers and stopped at 54% of the page, above the form

Method, stated before the numbers

The page is /news/is-chatgpt-really-shifting-from-asking-to-doing-openai-signals-data, published 2026-08-08. The population is every PostHog session with at least one page view on that path, from the day it was published to the read at 09-09 16:44 UTC: 10 sessions. Sessions are classified with our visitor instrument's own lists, parsed from its source at run time: internal identities, our capture probe, our own IP, listed scanner IPs, a fake user-agent shape, AI-fetcher user agents, and two named non-human signatures. 1 ai fetcher and 1 capture probe were excluded, leaving 8 real sessions, 7 of them referred by ChatGPT (the referral detector fired with engine chatgpt, or the URL carried utm_source chatgpt.com, or the referrer was chatgpt.com) and 1 with no referrer.

Per session we read the seconds from the first page view to the first cta_viewed event (the call-to-action slot at the top of the body, inline_top, or the form itself, position form), the largest visible_seconds and max_scroll_pct beacons, whether a check_form_focused event fired, and whether a check_started event fired. A person is PostHog's person id; a network is the session's IP, printed as a letter and never as an address. A rolled session is a later session by the same person on the same path with no page view, which happens when a tab stays open past the session timeout; its beacon is folded into that person's dwell. An independent HogQL count of referral-detector sessions on the path, run without the classifier, returned 7 sessions, equal to the 7 classifier rows carrying the detector. The computation is the committed script `content/news/chatgpt_reader_form_cut.mjs`; it writes `content/news/chatgpt-reader-form-cut-2026-09-09.json`, and 13/13 self-checks pass.

7 sessions from ChatGPT, 7 person ids, 5 networks

5 of the 7 sessions had a call-to-action in view, the earliest one of them inside 6 seconds and the slowest at 34.1 seconds; the 2 that did not were 4 visible seconds at 7% and 4 visible seconds at 8%, both the second session on a network that had just opened the page (persons C and D; E and F, 197 seconds apart and 3 seconds apart). 3 sessions scrolled to the form. One focused it: person B, Gurugram, IN, desktop Chrome, 130 visible seconds and 73% scroll on 08-29 06:16 UTC, a click into the businessName field and nothing after it. The instrument's own lifetime list of real check_form_focused events on this path is that one row. 0 of the 7 ran a check. Together the 7 sessions were 438 visible seconds.

One reader, two browsers, 35 minutes

On 2026-09-08 a Stamford reader opened the post from a ChatGPT answer in Chrome at 18:23:44 UTC, with the call-to-action in view after 5.7 seconds, and read 27 visible seconds to 31%. 99 seconds later the same network opened the same URL in Microsoft Edge, with no referrer and no campaign tag, as a second person id: the call-to-action in view after 4.7 seconds, then 2094 visible seconds, 35 minutes, over a 2100-second session, reaching 54% of the page, with the same beacon sent once more at 19:37:23 UTC from a rolled session. The form sits below 54%; it was never in view, never focused. The instrument counts 2 persons here; the honest count is one reader who moved from ChatGPT's in-app view to a desktop browser and stayed. It is the longest read this page has recorded, and it ended without a touch.

The site's own line: 17 saw the form on /news in 30 days, 2 touched it, 1 ran a check

That is the instrument's 30-day see-to-touch line as it printed on 2026-09-09: a seer is a real external person with the form itself in view on that surface, a touch is a focused field among seers, a check is a submitted form among seers. The page above contributes 7 of its ChatGPT persons to the lifetime referral count of 25 net referral events (23 persons) and 1 of the 3 real form focuses the site has ever recorded. Held out of the 30-day line by the datacenter rule: 3 landings.

What changed on 2026-09-09: the form moved to the claim, 15% down this page

Until 2026-09-09 the form on this post sat at the bottom of the body, below the 54% the 35-minute reader reached. Engineering shipped the two-step check form at the claim on the two posts ChatGPT refers to most, this one and the does-chatgpt-read post, as deploy askedabout-d7onxs8vz; on this page its slot is the form at 15%, inside the first screen of body text. At the read, 09-09 16:44 UTC, the live HTML carried the claim anchor and the inline_claim position marker. The seven sessions above all predate it; none of their numbers changes. The slot is a new cta_viewed position, inline_claim, with its own focus placement, so the next read separates it from the old inline_top slot and from the form at the bottom.

What this means for a business that wants to be found by AI engines

• A ChatGPT referral is a reader, not a lead. On this page 7 referrals produced 438 visible seconds, 1 focus and 0 checks. If your plan for showing up in ChatGPT's recommendations ends at the referral, this is the shape of what arrives: people who read and leave.
• The reader who stays longest is the one your form never reaches. 2094 visible seconds and 54% of the page, with the form below 54%. Where the ask sits on the page is measurable in scroll depth against the form's position; on this site the answer was to move the form to the claim, and the 2026-09-15 read below says whether that changed anything.
• Your analytics will count this reader twice. Two browsers, two person ids, one network, 99 seconds apart, the second with no referrer. A referral read that stops at the referrer field misses the second, longer visit and files it as direct.

Limits

• One page, 7 sessions, 32 days. This is the complete record of ChatGPT referrals to one URL on a small site, reported because it is complete, not because it is large.
• The classifier reproduces the instrument's lists and read rules and not its user-agent burst detector or its audit-side joins. The independent detector count and the classifier's rows agree at 7, and the instrument's own focus rows on the path agree with the table.
• Persons are person ids. Two ids on one network minutes apart are printed as two persons and named as likely one reader each time; we do not merge them.
• visible_seconds is the page's own visibility beacon and stops counting when the tab is hidden; max_scroll_pct is the deepest point the viewport reached. Neither measures attention.
• The form's position on the page is the engineering log's figure for the new slot, 15%; the old form's position is not in the artifact and is stated only relative to the 54% the reader reached.

The next reading, pre-committed

2026-09-15: On this post and on the does-chatgpt-read post, the two pages that carry the form at the claim since 2026-09-09, we print from the instrument: real external sessions, sessions with cta_viewed at position inline_claim, check_form_focused with placement claim among them, and check_started among them, per page, against the baseline on this page of 7 ChatGPT-referred sessions, 5 with a call-to-action in view, 3 with the form in view, 1 focus, 0 checks. The instrument's 30-day see-touch line for /news is printed beside it (today 17 seers, 2 touches, 1 check). No decision is taken in-post; the counts are published either way.