What does an agency do with a free AI visibility report on a client? One real check: 6 loads over 3 h 16 min, printed at minute 7, the client link copied at minute 156, 2 previews of the client view, 0 upgrade clicks

The question and the instrument

A free report tool aimed at agencies has two buttons an owner-operator has little use for: "Save as PDF" and "Copy client link". The client link opens the same report without the price, without the upgrade card, and with a "Powered by" line in place of our masthead. Whether an agency uses either button, and when, decides what the follow-up email should lead with. Until 2026-09-17 no real agency had run a check here, so there was nothing to read.

Two days ago we read every report load on the 6 real checks to that date: 12 loads, all inside 55 minutes of the check, and we wrote that a free report "is read in the session that produced it and not again". The seventh check does not fit that sentence. This post is the read of that one check, beside the six before it.

Counts are of real external persons as our instrument defines them; our own identities, our network egress, capture probes and two datacenter shapes are excluded. That the checker is an agency and the business is a client was read by hand from the check's own record: the report address is on a marketing agency's domain and the business checked is a different, local business. Neither is named here or in the artifact. "Visible seconds" is time the report tab was the foreground tab, reported by the page itself, per browser tab.

Before the check

The session began on the home page from a Google search, 23 minutes before the check. About a minute later the same browser loaded the agencies page with `utm_source=chatgpt.com`, which is what ChatGPT appends to a link in an answer; our AI-referral detector wrote one row for it. The reader moved between that page and the home page for about a minute. The first click into the form came 22 minutes after arrival; 5 fields and 55 seconds later the check ran.

The 6 report loads

6 loads in 3 sessions, 4 of the checker's view and 2 of the client view, 169 seconds on screen in total (152 across the loads plus 17 on the first tab when it was brought forward). 4 of the 6 loads came more than an hour after the check. At minute 196 all 4 open report tabs closed within 10 seconds of each other, which reads as a browser window being shut.

Two loads of the home page after the check (minutes 153.4 and 153.9) each showed the return bar, the "Your report" link for a browser that holds a saved check: shown 2, clicked 0. These are the first real "shown" rows since that sensor was repaired on 2026-09-16. The reader got back to the report without the bar both times.

The agency check beside the six before it

The print and copy sensors have recorded events since 2026-09-14. Every earlier row on either is one of our own capture probes; the two clicks above are the first real rows on both. The third $79-card row is the background tab from load 1 reporting the card 9203 seconds into a load that had been invisible until then, so it is counted as a row and not as a second look.

A row that looked like something else

At minute 154 the tape holds a pageview of the report URL carrying `utm_source=chatgpt.com`. Read alone, it looks like a reader pasting the report URL into ChatGPT and clicking the link it returned. It is not. The row's trigger is `session_resume`: it is the background tab from load 1 waking 2 h 34 min after the check, and the parameter had ridden along from the agencies-page landing through the form to the report URL. The tape holds 2 such resume rows on this check and neither is an arrival or a report load. Anyone counting AI referrals off a `utm_source` on pageviews will count kept tabs as new visits unless the count is keyed on the first pageview of a session and the resume trigger is excluded.

Reading

• One agency, one check: these are counts and not rates. The check contradicts a sentence we published two days ago. "Read in the session that produced it and not again" held for six owner-shaped checks and failed on the first agency-shaped one, where 4 of 6 loads came after the first hour.
• Both agency buttons were used: a PDF at minute 7, the client link 2 h 28 min later, then two loads of the client view by the agency itself. The tape shows no send: the link has 0 loads from any other browser up to this read, 17 hours after the copy.
• The $79 card was on screen twice and drew no click. The card sells a fix plan for the business checked. Whether an agency is the buyer of a plan for a client's business is not something one check can say, and we changed nothing on the report in response to one reader.
• The report's first tab opened in the background and stayed there for 2 h 34 min. The reading happened in a second tab opened 7 minutes later. That load carries no referrer, so the tape cannot say whether it came from the report email or a pasted URL.

The next read, pre-registered

1. 1On 2026-10-02: the client link from this check has been loaded by a second browser. Statement: it has not (0 loads by any person other than the checker). One or more and the copy was a send.
2. 2On the same date: the next real check whose report address is on an agency's domain and whose business is a different one. Statement: it produces a load after the first hour or a print or a copy, as this one did. A next agency check with none of the three means this one was a single reader's habit.
3. 3On the same date: upgrade clicks on this check. Statement: 0, including after the two follow-up emails.

A result outside its statement is published as a falsification with the same script and a dated artifact.

Artifact

`content/news/agency-report-cut-2026-09-18.json` holds the path to the form, each report load with minutes after the check, session, tab, view, seconds on screen and scroll depth, the actions with their minutes, the resumed-tab rows, the same counts on the six earlier checks, the lifetime rows of the print and copy sensors with probe rows named, and 17 self-checks, each an instrument count against a raw query. The script is `content/news/agency_report_cut.mjs`; this post is rendered from the artifact by `agency_report_post.mjs`. No business names, addresses or places are in either.