Seventeen AI referrals in 78 days: two check-form touches 13 hours apart, both from AI engines onto news posts, one check, one 0/100 report read twice, one pricing visit, zero upgrades

Method, stated first

The population is the same ledger as the fifteen-row post: every `ai_referral_detected` row our detector has fired since 2026-06-27, after the instrument removed owner, internal and scanner traffic and quarantined two same-instant rows under a rule published on 2026-08-24. That leaves 17 eligible referrals of 19 raw. On 2026-08-29 at 16:44Z we pulled each visitor's event stream from PostHog, keyed by person rather than by browser id, and cut it to the session carrying the detector row. The person key is new. A submitted check re-identifies the browser, so a visitor's check and report rows sit under a second id; the earlier cuts keyed by browser id and would have missed them. Each row is then scored on a ladder: a check call-to-action in view, the form in view, any click, the check button tapped, a field focused, a check started, the report viewed, the pricing section visited, upgrade clicked, checkout created. Later sessions of the same person count for the check-and-after steps; there were none. The rule for this page, set before the pull: engine, page, device, timestamps and counts only. No business, person, town or address appears here or in the artifact.

The series: 14, 15, 17

The first fifteen rows re-read unchanged today (timestamps, landing pages, spans, taps and input clicks match the 2026-08-27 artifact; a self-check asserts it). The two earlier posts counted "fields" as autocapture clicks inside an input, which read 0 on the fifteen. Today's count uses the `check_form_focused` sensor, live since 2026-08-04, and on the two new rows both bases read 2. The 08-25 and 08-27 "form in view" cells are today's re-read of those rows.

The seventeen, in order

Fourteen came from ChatGPT, two from Perplexity, one from claude.ai, and the claude.ai row is the only Claude referral the detector has ever recorded. Nine landed on /news posts, three on the homepage, three on the tools-comparison page, two on the agencies page. Both field touches and the one check sit on the /news nine. The comparison page does not carry the form, so its three never had it in view. The two agencies-page sessions had the form in view at 0 s and lasted a second or less.

The ladder, seventeen rows in

Down the first column the shape is 17, 7, 3, 2, 1, 1, 1, 0. Every rung to the pay step has a count on it for the first time. The rung that reads zero is the one after the report.

The first touch: claude.ai to a news post to a check, 380 seconds

Row 16 is the first check ever started from an AI referral on this site and the only row of the 17 with a `report_viewed`. The session is 104 events over 7 minutes 51 seconds. The rows that matter, with time since the first event:

The report scored the business 0/100. The score is read from the audit record, and the report email our system sent at 17:01Z carries the same figure in its subject. Twenty-nine seconds into that report the upgrade block was on screen and the reader did not click it. They went to the homepage, clicked "Pricing", spent 13.7 seconds on the pricing section, used the link there to re-open their own report, were back on the pricing section a second later and left from it 18 seconds after that. Zero `upgrade_clicked`, zero checkout sessions, nothing from that browser since.

The second touch: ChatGPT to a news post, one field, nine minutes

Row 17 arrived 13.3 hours after row 16's field focus, from chatgpt.com onto a different /news post, and reached the form through the button written for agencies ("Check a client's AI visibility →"). It is 26 events over 9 minutes 25 seconds:

After the one click into the business-name field at 28.6 s there is no further click, no second field, no submit. Engagement ticks keep arriving for nine more minutes with the scroll position frozen at 73%, which is a tab left open with the form on screen. Autocapture does not record whether characters went into the field. It records that the form was never sent.

What one session cannot separate

• Price from object. A reader who scores 0/100, sees the $79 upgrade block at 29 s, goes to the pricing section unprompted and re-reads the free report instead is one data point on what the report offered. One session cannot tell "$79 is too much" from "the report gave me no reason to pay" from "I wanted another look before deciding and got interrupted".
• Back button from second visit. The pageview on /#pricing one second after the second report view fits a browser back and fits a deliberate return. The rows do not say which.
• Coincidence from pattern. Both touches are AI engine, page-one /news post, check button, business-name field, 13 hours apart. But nine of the 17 landed on /news, so /news holding both touches is 2 of 9 against 0 of 8, and two events are two events.
• Sensor age from behaviour. Rows 1 to 3 predate the field-focus sensor, so their cell is unknown, not zero. Rows 4 to 15 had it live and read 0.

Why this matters if you care about AI visibility

• AI-referred readers of an informational page will use a tool embedded in it. Row 16 read for five minutes, left the post twice and came back, then filled the form on the post itself rather than on the homepage it had already seen. The form was on the page the engine sent them to, and that is the one that got used.
• The step that failed was report to pay, on a reader who went looking for the price unprompted. That is a different failure from "nobody reaches the report", and it is the one we are working on now. Our 2026-08-31 decision on what the free report shows and offers takes this session as its primary input.
• Row 17 came in through the agencies button, from a ChatGPT answer, onto a post about ChatGPT's own product direction. If you audit AI visibility for clients, that is the path we are watching: an engine answer, a specific post, then the client check on the agencies page. One tap is not a rate. It is the first row on that path.

Limits

• Seventeen is seventeen. A complete list with timestamps, not a rate, and nothing here generalizes past this site.
• Every row depends on the visitor's browser running our script. A tracker-blocked visitor is invisible, and a tab closed before the script loaded never enters the detector. The population is "AI referrals we could see".
• Keying by person is new in this cut. It changes nothing on rows 1 to 15 (asserted by self-check) and adds the check and report rows to row 16, which a browser-id cut would have shown as a field touch with no check.
• The engine label comes from the referrer or from the `utm_source` an engine appends; row 16's referrer was claude.ai itself. A second basis that also counts `utm` rows with no detector row reads 19 lifetime. It is reported beside the detector and never scored.
• "Read twice" is two `report_viewed` rows 54 seconds apart from one person; the second fired 0.4 s after the page loaded because the report was already computed.
• The two quarantined rows are held out by a rule published on 2026-08-24. Counting them would make the arrivals 19 and move none of the touch rows.

What we will publish next

The next AI referral that touches the form, starts a check or opens the report gets its row within a day, same columns. The 2026-08-31 decision on the report and offer will be published with the session it was argued from. The cut script and artifact are `content/news/referral_see_touch_cut.mjs` and `content/news/referral-see-touch-cut-2026-08-29.json` in our repository. The fifteen self-checks in the artifact assert every count in this post, including that the earlier fifteen rows re-read unchanged, that the two touches match the field-focus instrument to the minute, and that the artifact carries no identifier.