Agents search and fetch live pages, which Ora says has a lot of impact on AEO

Ora, a guest on Vercel's show, shared research that tests two common beliefs. One is that agents read llms.txt, a file meant to summarize a site for AI. The other is that agents answer from memory when they cannot read a site. Ora said the findings have "a lot of impact on entire categories such as AEO". AEO is short for answer engine optimization, the effort to be recommended by AI tools. Ora's point: agents now search and fetch live pages, and they mostly find llms.txt only if you link to it.

What was said

Ora ran the talk as a true-or-false quiz built on the team's own research. On llms.txt, agents reached the file in 32% of cases. Of those, 86% happened because the site pointed to it somewhere. Ora's conclusion: an unlinked file has almost no chance of being read.

On memory, Ora posed a specific case: an agent cannot read your site. The question was whether it then answers from training data. Ora said the share fell to 7% in their research. The reason, as Ora told it: in 2023 models had no tools and hallucination was a constant worry. By 2026 they are trained to use tools and ground their answers.

So the common path is different from what many expect. An agent tries a site, gets blocked, searches elsewhere, and finds another source. Ora said: "we've actually seen real cases where they end up in your competitor's website."

Why it matters

Our reading: some AEO advice written for models that answer from memory may no longer fit, and the work shifts from writing files for models to making pages reachable and readable. Teams that spent years blocking bots may need to check what their rules now shut out. If answers appear only after scripts build the page, an agent may see very little. If you keep an llms.txt file, link it from the site. These checks are cheap compared with losing a recommendation to a rival.

The other side

Ora also said: "Bad bots still exist." The excerpt does not explain how to tell helpful agents from harmful ones, so that question stays open.

The figures come from Ora's own research. The excerpt gives no sample size or method, and nobody outside the team has checked them here. Ora spoke of AEO as a category and did not review any particular advice. Ora also noted that an agent can open a full browser, though Ora said that took 45 times more tokens. That is a cost, not a ban, so some agents may still do it.