Five external visitors in 74 days reached our check form. The three who typed all got a report; the two who only tapped the button left without touching a field. Fourteen AI referrals produced one tap and zero typed fields

What we counted, and what we did not

A visitor "reached the form" if PostHog recorded either a click on a call-to-action whose link target is the check form (`href="#check"`, the hero button on the homepage and the inline ask on every article) or a click inside one of the form's fields. We read every such click since autocapture began on 2026-06-12, then removed our own rows (two from the owner's IP, one from a smoke-test identity) and four clicks from one AWS address in Ashburn wearing an iPhone user agent that clicked four elements in 2.0 seconds, including a rage-click, and cycled three pages in 8 seconds. That address is not on our standing scanner list; we are excluding it for this cut and saying so. Three real call-to-action taps survive. Adding the people who went straight to a field without tapping anything gives five external persons, lifetime.

Two sensor limits shape the read. Our dedicated `check_form_focused` event was born on 2026-08-04, so for the first three people "typed" means autocapture saw a click on an `<input>`; for the last two it means both sensors agree. And our scroll and visible-time sensors were born on 2026-07-29 and 2026-08-02, so the June and mid-July sessions carry no scroll depth.

The five people, in order

One correction to our own morning note. The analyst's summary called these "five hero-CTA tappers". The record says two of the five never tapped a button. The Colorado visitor's first click on the site, 10.9 seconds after landing, was inside the business-name field, and the England visitor's first click on the homepage was a field too, 5 seconds after arriving there from `/for/dentists`. Both submitted inside a minute. The button was touched by three people, and only one of the three typed afterward.

The two who tapped and left

The London reader found the inline ask 34 seconds in, at 11% scroll, tapped it 10 seconds later, and the anchor jump put the form on screen at 83% of the page. On a 2560 by 958 viewport that is the entire five-field form and the button with room to spare. The reader stayed another 88 seconds, reached 89% scroll, and left on a `$pageleave`. No field click, no focus.

The Brisbane visitor is the one that made us write this. They arrived from a ChatGPT answer (the `utm_source=chatgpt.com` tag; iOS strips the referrer) at 03:06:47.082 UTC, and 2.18 seconds later tapped "Run my free check". That is the first tap on that button by any of the 14 AI-referred visitors. The form fired its in-view event 16 milliseconds after the tap. Our engineer re-measured the deployed homepage today in an iPhone-shaped browser at the real layout viewport this session reported, 390 by 648, and after the jump every one of the five fields and the submit button sit inside those 648 pixels. The first field was 133 pixels from the top and 296 by 65 pixels to hit. The session's last event, a page-engagement heartbeat reporting 11 visible seconds and 17% scroll, came at +10.1 seconds. No `$pageleave` arrived, which on iOS Safari usually means the tab was backgrounded or closed from the app switcher.

So the wall for that visitor was not a hidden form. It was a fully visible five-field form on a phone, looked at for about eight seconds, and declined. We do not know why. The candidates our engineer listed are the size of the ask (four required fields before any result), a 15-pixel input font that triggers Safari's zoom on first focus (which no one triggered), and the plain possibility that a ChatGPT-referred reader wanted to read, not to fill in a form. We are not changing the homepage on one session; it stays frozen until Sunday's review of all fourteen sessions.

Fourteen AI referrals, one tap

These are the 14 sessions our referral detector counts as eligible (16 raw; two rows from the 2026-08-23 23:50 referrer spray are held out by the quarantine rule we published last week). "Form in view" means the check form itself entered the viewport; "inline ask only" means the article's mid-page call-to-action was seen but the form was not.

Totals: form in view for 4 of 14; some call-to-action seen in 9; any click at all in 3 (two of them outbound links to OpenAI's data pages, one the Brisbane tap); field clicks 0; checks started 0. 3 sessions ended inside 2 seconds and 4 ran past five minutes, including two that read a single article for about 30 minutes each. The long readers are the part of this we would not have guessed: a ChatGPT referral that stays 30 minutes and still does not touch the form is not a bounce, it is a reader who wanted the article.

Why this matters if you care about AI visibility

• Getting cited is the top of a funnel that has more steps than most AI-visibility dashboards show. On this site the chain is answer → click → article read → call-to-action seen → button tapped → field touched → submitted, and the 14 AI referrals fall off at the button. If your reporting stops at "we were cited" or "we got a referral", it stops four steps before money.
• The people who converted here did not need persuading by the page. All three typers were in a field inside 11 to 77 seconds and submitted inside 95 seconds. They arrived with the intent (a launch-day Product Hunt browser, an agency owner answering a cold email, a reader returning three hours later). The two who tapped and left arrived through a search or a chat answer. That is a small sample and a real pattern, and it argues for treating a referral as the start of a relationship, not a lead.
• Measure the tap, not just the form. Our funnel events start at `check_started`; the button tap lives in autocapture, which is why the analyst's count and ours differed by two people. If you run a form behind a call-to-action, log the tap as a first-class event with the viewport height at the moment the form appeared, or you will be measuring the wrong step for months, as we were.

If you want to see what the engines say about your own business, the free 60-second check is the same five-field form described above. We know exactly how many people have filled it in, and we would rather you knew that too.

Limits

• Five people is five people. Nothing here is a rate; it is a complete list of what has happened, with timestamps.
• "Typed" is inferred from a click inside an input before 2026-08-04 and from the focus sensor after. A visitor who tabbed into a field with the keyboard and never clicked would be invisible to the first method; none of the five sessions shows a `check_started` without a preceding field click, so no such case exists in this set.
• Autocapture depends on the client running our script. A visitor with the tracker blocked reaches the form without appearing here. Our referral count has the same blind spot.
• The Ashburn exclusion is our adjudication for this cut, printed in the artifact with the reason. If it was a human, the count of people who tapped and left is three, not two, and the conclusion does not move.
• The 2.2-second figure is PostHog client capture time on the visitor's phone against the same clock's `$pageview`, so clock skew does not enter.

What we will publish next

The next external person to tap the button or touch a field gets the same row in a follow-up within a day: source, viewport, seconds to first field, seconds to submit, outcome. Sunday's weekly review reconstructs all 14 AI-referred sessions against the page geometry and has to name a mechanism before anything on the homepage changes. The cut script and artifact are `content/news/cta_tap_cut.mjs` and `content/news/cta-tap-cut-2026-08-25.json` in our repository; the self-checks in the artifact assert every number in this post.