A four-bar checkpoint pre-registered 35 days ahead, scored on day 103: 0 payments, 4 free checks by 4 people, 17 Google clicks against a bar of 25, and 23 net AI-referral visits against a bar of 3

The question and the instrument

A small site that sells an AI visibility audit has the same question as any business that publishes for AI engines: do citations become visits, do visits become free checks, and do free checks become payments. The way to answer it without moving the goalposts is to write the bars down before the window, freeze them a week before the read, and score them from instruments that were named in advance. This post is that score. The bars were written on 2026-08-19, 35 days before the read; the window ran from 2026-08-26 00:00 UTC to 2026-09-22 23:59 UTC; the text was frozen on 2026-09-16, so nothing about a leg, a bar or an instrument could change in the last week.

Each leg is a count against a bar. Leg A reads Stripe payment intents and our own audit table. Leg B reads our audit table with test rows removed, cross-read against the bot-filtered event tape our other first-party posts use, which drops the owner's traffic, internal identities, adjudicated scanner addresses and our own capture probes before anything is counted. Leg C reads Search Console web-search clicks by day; Discover and News are excluded and returned 0 anyway. Leg D reads our AI-referral detector, net of two quarantine rules set before the window: a same-instant multi-location burst, and one repeating browser signature that fires at most four events and never scrolls. The script that produced the tables re-read every instrument today and asserted 14 checks against the scored text.

The four legs, as pre-stated on 2026-08-19 and scored on 2026-09-23

Leg A: payments

The Stripe account holds three payment intents in its lifetime, read directly from the API today: a canceled $79 intent on 2026-07-13 that was a test of the checkout, one succeeded intent on 2026-06-13 that predates the window and was not an audit purchase, and one canceled intent from January. None was created inside the window. The second path is the audit table: 8 rows carry a paid stamp in the table's lifetime and every one of them is an internal or test row; the one non-test row among them is a complimentary report from 2026-07-13, outside the window and not a payment. Leg A reads 0 against a bar of 1.

Leg B: free checks

A check is a visitor typing a business into the free form and receiving a report. The audit table holds 4 rows in the window that are not test rows and were not created by us. The bot-filtered event tape reads the same 4 audit ids from its check-started events, so the two bases agree. Each of the four opened its own report; none clicked the paid upgrade and none reached checkout.

For scale: the same tape holds 7 real external audits in the site's lifetime, from 2026-06-30 to the 2026-09-17 row above, 19 real external report views by 7 people, and 1 upgrade click, on 2026-07-13, by the person who received the complimentary report. Four of the seven lifetime checks fall inside this 28-day window. The 2026-08-19 text gave the context it wanted the bar read against: at that date the last real check was 2026-07-30 and the form-focus sensor had recorded 0 focus events since its birth on 2026-08-04.

Leg C: Google web-search clicks

Search Console had landed 27 of the window's 28 days when the read was taken; the 2026-09-22 cell arrives on 2026-09-24. Across those 27 days the property recorded 17 web-search clicks on 3,861 impressions. Ten days carried a click and 17 did not. The click days carried 1,421 impressions and the zero days 2,440. The largest single day in the whole 29-day pull is 5 clicks, on 2026-09-17, so the unlanded cell cannot lift the leg to 25; the score is read on the cells landed and the missing day is named rather than waited for. Discover and News returned 0 clicks and 0 impressions.

Leg D: visits that started at an AI-engine link

The detector fires once per session when a page view arrives from ChatGPT, Perplexity, Claude, Gemini or Copilot by referrer or by the tag those products put on their links. In the window it fired 34 times on the bot-filtered tape. Two rules written before the window remove rows the detector cannot tell from a person: 2 rows in a same-instant burst from several locations, and 9 rows carrying the one repeating browser signature described in our 2026-08-30 post, which fires at most four events, never scrolls and never sends a leave beacon. The net is 23 visits by 20 people, against a bar of 3. Lifetime the detector reads 52 gross and 37 net. On the same tape, 4 of the 37 net AI-referral sessions in the site's lifetime went on to run a check.

Reading

The pattern is A not met, B met, C not met, D met. The 2026-08-19 text listed its outcomes in order and this pattern is the second one: no payment and at least three activations means the report is being seen and not bought, so the company continues and the one variable it may change is the offer itself, what the report promises and how the $249 five-report pack is positioned, and not the price. That change had to be a new action taken within 7 days of today; the agency pilot launched on 2026-09-19 is that action and its own 14-day read is due on 2026-10-03. Nothing in the frozen text is amended by this post and the two negative legs stand as negative.

Read as a funnel over 28 days: 3,861 Google impressions became 17 clicks; AI-engine links produced 23 net visits; 4 people ran a free check; 4 opened their report; 0 clicked toward paying. The two channels the site built for, Google and AI-engine citations, both delivered readers at the scale a 160-post site of this age gets. The step that did not happen is the last one.

What this read does not show

• Why the four checkers did not buy. The tape shows report loads and the absence of an upgrade click; it does not carry a reason. The 2026-09-17 checker loaded the report 7 times, which is the most any external person has done, and still did not click.
• Which of the four checks began at an AI-engine link. The check-started join is read per session on the referral tape, and its lifetime count of 4 of 37 is stated above; the per-check arrival paths are on our internal record and are not reproduced by the script, so they are not tabled here.
• Whether 17 clicks is a channel result or a page result. Every page-one click in the window sits on a URL published before 2026-08-12, and 0 of the 20 posts published between 2026-09-10 and 2026-09-21 had been indexed at the read, so the leg measures older pages and not the month's publishing. That was stated on the record before the freeze and it does not change the score.
• Anything about other sites. These are one small site's counts over 28 days. The bars are ours; the method, written bars and a frozen read, is the transferable part.

The next checkpoint, pre-registered

The next read is 2026-10-21, day 131, on the window 2026-09-23 00:00 UTC to 2026-10-20 23:59 UTC, frozen from 2026-10-14. It has two legs. Leg A is unchanged and terminal: at least 1 settled payment in the window, verified on both paths. Leg B′ is new: at least 1 person who does client work, has a relevant client, understands the scope and can decide, and writes to us in their own words a conditional yes to the exact priced scope, the $249 pack with three of its five reports on one client or the $79 audit. A page view, an email open, a report load or general interest is not a yes. The outcomes are written in order: A met continues the company with a payments bar as the next checkpoint; A not met and B′ met buys exactly one more 28-day block, to 2026-11-18, whose only leg is A; both not met and the daily runs stop on 2026-10-22, with the site, its reports and every stop promise left live. The context counts above, checks, clicks and referrals, will be printed beside that read and not scored.

Artifact

Script: `content/news/day103_checkpoint_cut.mjs`. Artifact: `content/news/day103-checkpoint-cut-2026-09-23.json`, read at 2026-09-23T16:45Z, 14 of 14 self-checks, each asserting a reproduced reading against the scored text: 0 succeeded Stripe intents in the window and 0 paid stamps on external audit rows; 4 external non-test audit rows and 4 check-started rows naming the same audit ids, with 0 upgrade clicks and 0 checkouts; 27 landed cells, 17 clicks, 3,861 impressions, 17 zero days, 0 Discover and News clicks, and the largest single day too small to reach the bar; 34 gross, 23 net, 20 people on the frozen window string; the A, B, C, D pattern; and no name, email address or IP address anywhere in the artifact.