Fifteen AI referrals in 76 days: one tap, zero form touches. The newest, a ChatGPT visit to our agencies page on August 27, lasted 0.9 seconds, sent four events and no leave beacon

Method, stated first

A visit counts as an AI referral when our detector, which runs in the page, sees a referrer or a `utm_source` naming an AI engine and writes an `ai_referral_detected` row. Lifetime the detector has written 17 such rows. Two of them, both at 23:50 UTC on 2026-08-23 with no page view anywhere in their session and a viewport height of zero, are held out under the same-instant referrer-spray rule we published on 2026-08-24. Fifteen remain. For each, the session is every PostHog event carrying the same `$session_id` as the detector row, and the span is the first to the last of those events, to a tenth of a second. Our own addresses, adjudicated scanners and internal identities are removed by the instrument before any of this runs, and the artifact records that it read the same instrument the analyst reads every morning.

Three sensors were born after the first referrals: the call-to-action visibility event on 2026-07-29, the engagement timer on 2026-08-02 and the form-focus event on 2026-08-04. Rows from before those dates carry a blank in those columns, not a zero. Two days ago we published the fourteen-row version of this table alongside the five people who ever reached the form; this cut re-read those fourteen sessions and found every span, tap and field count unchanged, which is one of the ten self-checks.

The fifteen, in order

Thirteen came from ChatGPT and two from Perplexity. Seven landed on an article, three on the homepage, three on the tools comparison and two on the agencies page. Spans: shortest 0.5 seconds, median 29.8 seconds, longest 30 minutes 27 seconds. Four sessions ended inside two seconds, six inside ten, five ran past a minute and four past five minutes. Nine of the fifteen never sent a `$pageleave`, and six never reached the engagement timer's first tick, which fires at five seconds of visible time.

The newest referral, event by event

That is the whole session: four events across 905 milliseconds, then nothing. No `$web_vitals`, no `page_engagement` tick, no `$pageleave`, no second page. We re-read the same browser identity at 16:42 UTC, seven hours and 38 minutes later: still four rows, none outside the session. Two readings fit the shape, a tab closed before the five-second tick or a client that stopped running our script after the first second, and the ledger cannot separate them. We are not choosing one here; the counts are the publication.

Where the short sessions land

All four sessions that ended inside two seconds landed on a product page: two on the tools comparison (0.7 seconds from San Jose, 0.5 seconds from Zwolle) and two on the agencies page (1.0 second from Gwangjin-gu, 0.9 seconds from Cheonan). None of the seven article landings ended inside six seconds; the shortest was 6.5 seconds and the median was 6 minutes 21 seconds. The two agencies-page referrals share more than a landing page: both came from South Korea, both were Chrome on Windows at 1280 by 680 pixels, both had the form on screen at zero seconds with the page scrolled 27 to 28%, and both were over at or under one second, six days apart. Two is two. We publish it as a count; Sunday's weekly review reconstructs all fifteen sessions against the page geometry and is the place a mechanism gets named, if one does.

The see-to-touch ledger, fifteen rows in

The three clicks are two outbound links, to Google's Search Analytics API reference and to OpenAI's Signals data download, and the Brisbane tap on the check button. The fifteenth referral moved the top three rows by one and the bottom four by nothing.

Why this matters if you care about AI visibility

• A referral is a count of arrivals, not of readers. On this ledger a quarter of the arrivals were over inside two seconds, and the visitor who stayed 30 minutes and the one who stayed half a second are the same row in any report that stops at "referrals from ChatGPT". If a dashboard gives you that number, ask for the session span next to it; the two columns tell different stories.
• The page the engine sends someone to decides what they do next, and for a small site that page is usually an article, because articles are what engines cite. Our article landings held people for a median of six minutes and produced zero form touches; our product-page landings produced every sub-two-second exit. Getting a business into ChatGPT's recommendations is a citation problem; what happens in the first second after the click is a landing-page problem, and the two need separate instruments.
• Measure the first second. A session that ends before the engagement timer's first tick is invisible to any dwell metric that starts at five seconds; six of our fifteen are. The events that captured the 0.9-second visit were the form-visibility fire at load, with the scroll percentage and the client time, and the detector row. Log those at load, not on a timer, or the shortest sessions will read as blanks.

The page both Korean sessions landed on is our page for agencies, where an agency evaluates the five-audit pack; the free check those fifteen visitors did not run is the same five-field form on the homepage. We know exactly how many AI-referred people have touched it, and the number is zero.

Limits

• Fifteen is fifteen. This is a complete list with timestamps, not a rate, and nothing in it 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 at all, so 0.5 seconds is the shortest session we can see, not the shortest that happened.
• Spans are client timestamps from one browser against one clock, so skew does not enter; below one second the figure is the order of four events, not a stopwatch.
• The detector names the engine from the referrer or from `utm_source`. The fifteenth arrived with referrer `$direct` and `utm_source=chatgpt.com`, so the engine label rests on the parameter ChatGPT appends to its links. A second basis that counts `utm` rows with no detector row reads 19 lifetime; it is reported and never scored.
• The two quarantined rows are held out by a rule published before this cut. Counting them would make the arrivals 17 and move none of the touch rows.

What we will publish next

The next AI referral gets its row within a day, with the same columns. Sunday's weekly review reconstructs all fifteen sessions against the page layout each one saw and has to name a mechanism before anything on a landing page changes. The cut script and artifact are `content/news/referral_dwell_cut.mjs` and `content/news/referral-dwell-cut-2026-08-27.json` in our repository; the ten self-checks in the artifact assert every number in this post, including that the fourteen earlier rows re-read unchanged.