Which AI engines send visitors to a small site, and what do those visitors do? 7 net referrals in 8 days from ChatGPT, Gemini and Claude cleared our pre-stated bar of 3; median 38 visible seconds, 2 form touches, 1 check

What was pre-stated, and what the instrument reads

On 2026-08-19 we wrote down four counts to be scored on 2026-09-23, before the window opened. One of them, leg D, is the number of AI-engine referrals detected inside 2026-08-26 00:00 to 2026-09-22 23:59 UTC, bar at least 3. On 2026-08-30 the count was tightened in the stricter direction only: it scores net of two named scanner shapes, with the gross count printed beside it and never scored. The sensor is a client-side event that fires when a page loads with an AI engine as referrer or as a tagged UTM source; the instrument then drops the owner, internal identities, adjudicated scanner addresses and the two shapes. This post is the day-8 read of that instrument, at 2026-09-02 16:50 UTC.

Which engines, in the window and lifetime

Four engines have sent at least one net referral since the sensor was born. Three of them appear inside the window. ChatGPT is 15 of 19 lifetime and 5 of 7 in the window; all 4 held-out rows in the window also carry a ChatGPT referrer, which is why the gross ChatGPT count is 9.

What the seven did, one row each

Each row is the session that carries the detector event, keyed by session id so that a check submit (which swaps the visitor's identity) cannot split it. Visible seconds are the page's own engagement counter, which ticks only while the tab is visible; scroll is the deepest point reached. "CTA in view" counts any call-to-action element that entered the viewport; "form in view" is the inline check form specifically.

Row 1 is the session we wrote up on 2026-08-29 in the see-to-touch ladder: a Claude referral onto a /news post, 104 events, a field focus, a check, a report read twice, a pricing visit, no upgrade. It remains the only check ever started from an AI referral, 1 of 19 lifetime. Row 2 focused a field and typed nothing. Rows 5 to 7 are one phone that opened our agencies page three times in 41 minutes for 1, 2 and 6 visible seconds, scrolled 8% each time, and never had a call to action in view; two of the three sessions ended before the page's copy swap fired. For scale, the one Google Search click onto a /news post on 2026-09-01 read 138 visible seconds, 61% scroll and 3 CTA views, with no field focus.

The two shapes held out, as shapes

Both quarantine classes are defined by what the session contains, not by where it comes from, and each carries a dated decision. The fixed-window shape was described in full on 2026-08-30, in the 1280×680 post, when it was 7 of 7 such sessions in 30 days and 0 of the 14 other referral sessions. It has since added 2 rows, one of them today at 10:13 UTC; it is 6 lifetime and 4 in the window, all with a ChatGPT referrer, landing on /for, /compare and /guide pages, never on /news.

Lifetime against the window, as counts

Landing classes, lifetime net: /news 11, / 3, /for 3, /compare 2. In the window: /news 4, /for 3. Both field touches and the one check landed on /news posts. The engagement and form sensors were born on 2026-08-02 and 2026-07-29; the three lifetime rows without a tick are older than the sensor and carry nulls, not zeros.

What this means if you care about AI visibility

• AI referrals can be counted per engine with one client-side sensor that reads the referrer and the utm_source ChatGPT appends, and the count is small enough on a small site that every row can be read by hand. Report it net of scanner shapes and print the gross beside it, or a fixed-window scanner will double your ChatGPT number, as it did ours (9 gross to 5 net this window).
• The visits are short. Median 38 visible seconds in the window, and the newest three under 7 seconds each. A page that an engine sends a buyer to has one screen to answer the question the buyer asked; the one check we have ever earned from a referral came on a 235-second read that reached 100% scroll.
• Being sent a reader and being recommended are different counts. Nineteen referrals in 67 days is traffic; whether ChatGPT, Perplexity or Gemini names your business when a buyer asks for a recommendation is what our free 60-second check reads, per engine, for your market.

Limits

• The sensor reads referrer and UTM. An engine that opens a page in an in-app browser with no referrer is invisible to it; the count is a floor.
• The quarantines remove named shapes; they do not prove the remaining rows are humans. A patient scanner with a consistent window and a scroll passes.
• "Visible seconds" is capped by the tab being visible and the engagement sensor ticking; a reader who prints the page or reads it in reader mode under-counts.
• The Google Search comparison row is a raw referrer read, not the instrument, and is one click.
• Identity counts come from the analytics vendor's person model; the three-rows-one-phone reading is stated from the sessions' shared device dimensions and timing, not proven.

The next reading, pre-committed

The instrument freezes on 2026-09-16 and the window closes on 2026-09-22. On 2026-09-23 the same script runs on the closed window and both counts are published, gross and net, per engine, with this table repeated on the full 28 days. Two lines are falsifiable now: if net falls below 3 the checkpoint leg fails as written; if the window closes with 3 or more checks started from referral sessions, the line "one check per 19 referrals" is dropped from this series, otherwise it is restated at its new denominator.