A second reading of the agency index
The index measured what an AI crawler can read on 300 agency homepages. This measures the other half: in 36 markets we asked the three questions a buyer asks, with live web search on, and read which agencies the answer names.
240 of the 300 agencies are not named once. 60 are named in at least one of the three answers for their market. In several markets the answers named none of the agencies we had measured.
The pages those answers cite are mostly not agency websites. They are directories, listicles and roundups, which is the part worth knowing: the channel returns a handful of names, and the names come from somewhere other than the sites themselves.
An answer names three to five agencies. A market with sixty of them cannot have sixty named, so most of any market is absent by arithmetic, not by fault. That is the finding rather than the flaw: the channel a buyer now uses returns a handful of names, and a high score in the index does not put an agency among them.
Being readable is what lets a model use a page once it has found it. It is not what makes a model pick the page. Both have to be true, and they are measured separately. Anyone who sells you one as the other is selling something.
Readable here means a score of 80 or more in the index. The four counts below are the whole sample split both ways. Read them as a floor rather than a lever: a site a crawler cannot use is out of the running, and a site it can use is merely in it.
Median index score of the agencies an answer names: 85. Median score of the agencies no answer names: 82.
Every answer stored the pages it cited. These are the pages that decide which names an answer carries, counted across every market we asked. A firm absent from all of them is hard to name, however well its own site reads.
Every agency in a market was checked against the same three answers. The named ones are listed; the rest of the market is the remainder.
No agency of this market was named in any of the three answers.
No agency of this market was named in any of the three answers.
No agency of this market was named in any of the three answers.
No agency of this market was named in any of the three answers.
No agency of this market was named in any of the three answers.
No agency of this market was named in any of the three answers.
No agency of this market was named in any of the three answers.
No agency of this market was named in any of the three answers.
No agency of this market was named in any of the three answers.
No agency of this market was named in any of the three answers.
No agency of this market was named in any of the three answers.
Three questions per market, identical for every agency in it, because a question invented per agency would make the readings incomparable. Each answer comes from gpt-4o with live web search enabled, not from what the model remembers: a model answering from its own weights cannot be moved by anything on a page, so measuring that would say nothing about findability.
An agency counts as named when the answer contains its domain or the distinctive words of its name. Words every agency shares never count on their own. A question the engine could not be asked is stored as nothing at all, never as an absence.
Answers change between runs, and between engines. This is one reading at one moment, and the date is on each market page.
Your own site, read the same way: