Last month closed with a deferral. The site had been restructured between runs, no effect was shown or expected, and I wrote that indexing lag meant any effect belonged to September or October. This is that sentence being answered, at the early end of its own window.
For anyone new to the series: every month the same 22 questions go to three AI providers, twice each, through their APIs with web search available. The method is in the July report. The August run is here. API behaviour differs from what you see in the consumer apps, so read every finding with that in mind.
Run 3 fired on September 2. Panel v2, 22 questions, 132 responses, none failed, none truncated.
The instrument found the site
On the question "Who should I hire to check if my website is readable by AI?", Perplexity cited yellowhousedigital.com/ai-readability-audit. In both passes.
That page did not exist before July. It was built in the site restructure, for that question, and for nothing else. Three runs in, this is the first time the brand domain has appeared anywhere in the instrument's citation data.
I want to state the causal claim carefully, because it is the kind of claim that gets inflated. The page was built for the question. The page is now cited on the question. That happened one month after indexing began. All of that is consistent with the intervention working. It is also one provider, one question, one month. It is not proof of a repeatable mechanism. The finding sits in the ledger with a confirm-or-lapse status, and run 4 decides which. Both outcomes publish.
Cited is not named
Here is the part that matters more than the citation itself.
Zero tracked commercial entities were named anywhere in this run. Not one answer, across 132 responses, said Yellow House Digital or Jordi Buskermolen.
So the current state, precisely. The site is read: AI fetchers hit it daily. The site is now cited: Perplexity, twice, on one question. The site is still not named: no answer recommends it.
Three rungs. Three different numbers. A crawler dashboard shows you rung one, and it looks like progress because the line goes up. A citation is rung two: the model pulled the page and pointed at it. The recommendation is rung three, and it is the only rung anyone is actually buying visibility for. Nobody wants to be a footnote. They want to be the answer.
The gap between the rungs is where the work lives. Being read did not produce a citation for two months. Being cited has not produced a name. Whatever moves a page from rung two to rung three, this run did not measure it, and I do not know what it is yet. That is the honest position, and it is a better position than a month ago.
The stable exception got company
In July, one company survived both passes on the highest-churn question, the one about who builds AI tools for agencies. Buildberg. In September it appeared again, in an OpenAI pass. Three consecutive runs. That is the longest stable presence the instrument has measured, on the question where names otherwise come and go the fastest.
What changed is the company it keeps. Buildberg is cited via a dedicated page for marketing agencies. Of the eight vendors OpenAI cited on that question this month, five are cited the same way: a made-for-this-buyer page, an industries-slash-marketing-agencies style URL. What looked in July like one company's trick now reads as a pattern the ecosystem is converging on. The dedicated intent page is becoming the entry ticket.
Same breath, same caveat. Five of eight is a pattern, not a rule. One month of it is one month.
The machinery stayed boring, with one wobble
OpenAI's set of questions answered purely from memory, never searching in either pass, is identical for the third straight month. The same six questions, to the question. Anthropic's equivalent set churned again. The stability is a property of OpenAI, not of the questions.
Overall search rates dipped this month. OpenAI went from 66% to 59%, Anthropic from 77% to 73%, Perplexity held at 100%. That is one interval. I am reporting it without a story attached. Run 4 says whether it means anything.
The churn continued. Month-over-month entity turnover on the buyer questions stayed heavy, same character as August. July's argument still stands: the machinery is consistent, the output is a lottery.
What this does not mean
August's temptation was despair. September's is the opposite, and it is the more dangerous one.
One citation from one provider on one question is a rung, not an arrival. It does not mean the site update worked. It does not mean "build intent pages and get cited" is advice anyone should take from this report. The Buildberg convergence is a pattern observation about eight vendors on one question. The YHD citation is a single measured event with its confirmation criterion stated. Those are different kinds of claims, and I am keeping them apart on purpose.
The wrong answers, for the record, also changed this month. On the questions that exist to test for confusion, the confidently wrong answers of August were replaced by entirely different confidently wrong answers, delivered with the same confidence. The models did not get better. They got differently wrong.
Cost
The computed cost of the run was $11.88 for 132 responses. The meter computes from token usage; the provider bills are the final truth.
What happens next
The panel stays frozen at v2, third month. Run 4 fires October 1.
The citation either re-confirms or it lapses. The report publishes either way. That is the same sentence I wrote in July. It is still true. The difference is that it now has a track record behind it.