LLM SEO

Where the term comes from

Nowhere in particular, and that is worth knowing before you buy any of it.

GEO has a paper behind it, with authors, a date and a benchmark you can argue with. LLM SEO has no equivalent. It appeared in agency marketing, and it competes with AEO, AI SEO, LLMO and GEO for roughly the same territory. The names differ; the pitch rarely does.

That absence is the first useful fact about the term. When a field has five names for one thing and no shared definition of it, the labels are being generated by sellers rather than by the thing being sold.

What the model providers actually document

Every major provider publishes documentation for site owners. It is short, and it is about one thing.

OpenAI

OpenAI separates its crawlers by purpose. OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features", and the documentation is direct about the consequence of blocking it: "Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links." GPTBot is a different matter, used "to crawl content that may be used in training our generative AI foundation models". ChatGPT-User covers actions a person initiates, and because of that, "robots.txt rules may not apply."

The only recommendation on the page is this: "To help ensure your site appears in search results, we recommend allowing OAI-SearchBot in your site's robots.txt file."

Anthropic

The same split. ClaudeBot collects content that "could potentially contribute to their training". Claude-SearchBot "navigates the web to improve search result quality for users". Blocking the search crawler "may reduce your site's visibility and accuracy in user search results".

Perplexity

PerplexityBot is "designed to surface and link websites in search results on Perplexity. It is not used to crawl content for AI foundation models." The recommendation adds one item beyond robots.txt: allow the published IP ranges too, which the documentation calls the source of truth for WAF configuration. A firewall can block a crawler that robots.txt welcomes.

Google

Google is the only one that addresses optimisation directly, and it rules it out: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary."

What they have in common

Four providers, four documentation pages, one lever between them: whether the crawler can reach the page. None of the four names a content technique, a format, a markup type or a ranking factor. Three do not raise structured data at all. The fourth, Google, raises it only to say that none is needed.

That is a measured absence rather than a claim about what the systems do internally. What the vendors document is access. Everything past that point is somebody's inference.

The llms.txt confusion

The file llms.txt is proposed as a machine-readable index a site publishes for language models. It is the most visible artefact sold under the LLM SEO label, and the evidence around it is regularly read backwards.

OpenAI's crawler documentation does mention llms.txt: "For the complete documentation index, see llms.txt." Perplexity's does too, under the heading "Documentation Index": "Fetch the complete documentation index at: /llms.txt." A second line on the same page names the audience: "For AI agents: see the complete llms.txt documentation index."

Both of those are the provider publishing its own file so that agents can read its documentation. Neither is a statement that the provider reads yours. Anthropic's crawler page does not mention the file at all.

Google, meanwhile, says the opposite of the pitch in as many words: "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add."

So the position, stated precisely: one provider says the file is unnecessary for its surfaces, two publish such a file for their own documentation, and none of the four documents reading one from your site. A vendor who cites OpenAI's llms.txt as proof that OpenAI reads llms.txt has confused a publisher with a consumer.

What is left that is real

The unglamorous part survives, and it is worth doing.

Check that the search crawlers can reach you. They are separate user agents from the training crawlers, and a site can block one while intending to block the other. A robots.txt written in 2024 to keep content out of training data may also be keeping the site out of ChatGPT search results, which was probably not the intention.

Check the firewall as well as robots.txt. Perplexity documents published IP ranges for exactly this reason. A WAF rule can quietly undo a robots.txt permission, and nothing in Search Console will tell you.

Decide separately about training and about retrieval. Blocking GPTBot and allowing OAI-SearchBot is a coherent position: no to the training corpus, yes to being findable. Most robots.txt files treat these as one decision because they were written before the split existed.

Measure before you optimise. Asking assistants what they say about your category and recording the answers is straightforward and yields real data. Changing content to influence those answers is a claim, and the strongest published claim in this field comes with an upper bound, a domain caveat and a benchmark from 2024.

Common mistakes

Treating LLM SEO as a discipline with established methods. It has no standard, no shared definition and five competing names. That does not make every technique worthless. It does mean nobody can appeal to a consensus, because there is not one.

Blocking training crawlers and expecting to stay visible in AI search. Different bots, different consequences. OpenAI, Anthropic and Perplexity all document the distinction.

Reading a provider's own llms.txt as evidence it reads yours. The file name is shared. The direction is not.

Buying AI-specific markup for Google's AI surfaces. Google's documentation addresses this and says it is not needed. Whether it helps elsewhere is a separate question, and no vendor documentation currently answers it.

Assuming silence means confirmation. None of the four providers documents a content technique. That is an absence of published guidance, not proof that content does not matter. Both overclaiming and dismissing go beyond what is on the record.

Related terms

• Generative Engine Optimization (GEO): the same territory, with a paper behind the name
• AI Overview: the Google surface, and the one with the clearest vendor statement
• llms.txt: the proposed file, and the most frequently misread evidence about it
• Answer engine: the broader category these systems belong to