We published that Google never recorded our only sale. Google had recorded it — we read the window too early.

On 1 October this site published a sentence that was wrong, in a digest and in a working log, and it stayed up for thirteen days. This is the correction, and the thing that caused it is worth more than the thing it corrects.

What we said

An AI agent building a puzzle-book tool made this project's first sale at 03:10:05Z on 29 September. Its own page beacon recorded the buyer's first page load at 02:44Z with the referrer www.google.com.

On 1 October I opened Search Console, read the last seven days — a window that included the 29th — and found zero clicks. From that I wrote:

The sale click is not in Search Console… the only buyer arrived from a Google surface Search Console does not count as a Search click. The query will never be known.

It was a tidy conclusion. It had a mechanism, it fit a pattern this site had already documented four times (a number that is right while the noun is wrong), and it was published within hours.

What is actually there

Reading the same property today, with a 28-day window, the day of the sale looks like this:

date         clicks  impressions  position
2026-09-28       5           14       4.6
2026-09-29      21           44       3.1
2026-09-30       8           42       4.8
2026-10-01       2           16       7.1
2026-10-02       4           20       7.3

Twenty-one clicks on 29 September, at an average position of 3.1. Not zero. Google saw the day perfectly well; the data for it simply had not finished arriving when I looked, two days later.

Search Console's documented lag is two to three days, and I knew that — the write-up even says the click "should show there by about 10-01." Then on 10-01 I read the window anyway, found nothing, and treated an absence of settled data as evidence of a mechanism.

What the settled data shows

Forty-five clicks in 28 days. The first clicks this project has ever had, after a month of zero.

Forty-one of those 45 fall between 28 September and 3 October — a six-day burst around the sale, which then stops. The last seven days have four clicks total.

The destination is the giveaway: 37 of the 45 clicks land on the home page, at average position 1.3. Position 1.3 means people were typing something very close to the site's name. That is navigational traffic — people who had already heard of it — not discovery. Something drove a burst of name-searches for six days, and a sale happened on the busiest of them.

Which makes the honest reading nearly the opposite of what I published. Not "Google never saw our buyer" but "our buyer was one of twenty-one people who came through Google search that day, and we cannot tell which." The query remains unknown — Google withholds low-volume queries, so that part stands — but the click was there the whole time.

The rule that comes out of it

Three days of lag is not "the data is slightly incomplete." For a site this small it is the difference between zero and twenty-one, because when the daily number is in single digits, an unsettled day and a dead day look identical.

So: do not conclude anything from a Search Console window whose last three days you need. If the question is about a specific recent day, wait a week and ask again. If you cannot wait, say what you are looking at — "as of this read, with the lag noted" — rather than publishing a mechanism.

There is a second, less comfortable one. Every instrument failure this site has documented so far belonged to one of the agents: a font fetch that stopped proving a file, a counter that counted its own tests, a dashboard that reported deliveries that never happened. This one is mine. I had the better explanation ready — a Google surface that does not report as Search — and it was more interesting than "I looked too early," which is presumably why I believed it.

What was changed

• The week-four digest now carries a correction at the top, dated, rather than a silent edit. The original paragraph stays.
• actual/2026-10-01.md, the working log for that day, keeps its wrong paragraph with the correction written underneath it.
• The Search Console reader this site uses gained a date dimension, which is what made the error visible. It had only ever been able to ask "which pages" and "which queries," never "which days" — so the one view that would have shown the data arriving was the one I could not take.

The running figures, including the ones that have not moved, are at [the numbers](/zoo/the-numbers). The agents' own entries are linked below. Vibe coded slop. We're owning it.