LLM SEO: how language models decide what to cite

What is LLM SEO?

LLM SEO is the practice of making your content reachable, parseable, and quotable by large language model systems, so that you are named and cited when someone asks a question in your category. The goal is different from a blue link. You are not trying to occupy a ranking position, you are trying to become the source a generated answer leans on.

You will see this called AI search optimization, answer engine optimization, or generative engine optimization. The labels overlap more than the people selling them admit. I have written about how GEO, AEO, and SEO actually differ if you want that distinction drawn properly. What matters here is the shared underlying question: why does a model quote one page and ignore another?

How does a language model actually choose a source?

There is a pipeline, and nearly every practical decision follows from knowing where in that pipeline you are being filtered out.

1. Query interpretation. The system rewrites the user's question, often into several sub-queries. A single conversational question can fan out into half a dozen searches you never see.
2. Retrieval. It runs those queries against an index and pulls back a candidate set of documents. This is a search step, and it usually leans on a conventional search index.
3. Fetch and parse. It requests the candidate pages and turns the HTML into text. This is where a startling number of sites quietly drop out.
4. Chunking and ranking. The text is split into passages, and the passages most relevant to the question are selected. Not pages. Passages.
5. Synthesis and citation. The model writes an answer grounded in those selected passages and attributes the parts it used.

Stage three is the cruel one. You can write the best page on the internet, rank for the query, and still be invisible because the fetcher could not read you. That is not a content problem and no amount of content work fixes it.

Does it matter whether you are in the training data?

Far less than people assume, and you cannot influence it anyway. Training data is a frozen snapshot from the past. When someone asks ChatGPT or Perplexity a question about your market today, the answer is overwhelmingly assembled from pages fetched in that moment, not recalled from training.

This is the single most useful reframe available. It means AI visibility is a live, winnable, current-state problem rather than a historical one you missed the window on. It also means the lever is retrieval, so the work looks much more like technical SEO and content structure than like anything exotic.

The practical consequence: stop worrying about whether you were scraped for training, and start checking whether the retrieval crawlers can fetch you right now.

Why does passage quality matter more than page quality?

Because the system does not quote your page, it quotes a chunk of it. A 3,000 word guide that buries its answers inside long transitional paragraphs will lose to a shorter page where each section answers one question cleanly, even if the long guide is better written.

What this means in practice:

• Each section has to stand alone. Assume the reader of that passage has not read the preceding sections. A chunk that begins "as mentioned above" is a chunk that cannot be lifted.
• Answer in the first one or two sentences under a heading. Then elaborate. The classic build-up-then-reveal structure is actively harmful here.
• Use headings that match how people ask. Question-form headings give the retriever a clean pairing between a query and an answer.
• Prefer real structure over visual structure. An actual table, an actual ordered list, actual semantic HTML. A grid of styled divs that looks like a table parses like soup.
• Keep claims near their evidence. A number separated from its source by three paragraphs often gets quoted without the source, or not quoted at all.

Why does entity consistency decide whether you get named?

An AI system has to resolve who you are before it can recommend you. If your business name, location, service description, and founder attribution vary across your own site and your third-party profiles, you are a fuzzy entity, and fuzzy entities get described vaguely or skipped in favor of something the system can state confidently.

The fix is unglamorous and cheap. Pick one canonical description of what you do and who you serve. Use it consistently on your homepage, your about page, your service pages, and every external profile. Then make it machine-readable with structured data, giving your organization and your author a stable identifier that resolves to a real page.

This is also why an about page does disproportionate work in AI search. It is often the page a system uses to decide what kind of entity you are.

Does llms.txt actually do anything?

On the current evidence, not much, and not the thing most people are buying it for. The proposal is that you publish a Markdown file describing your site so language models can read it instead of parsing your HTML. It is a reasonable idea. The problem is adoption.

Google has said it has no plans to use llms.txt, according to testing published by Index Lab. Log analysis has repeatedly found the file is barely requested: one 2026 review of crawler behavior reported that GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, and Google-Extended overwhelmingly skip it and crawl HTML directly. A separate 2026 analysis concluded that having the file does not meaningfully improve visibility in AI search tools and that no major AI company has confirmed reading it as part of retrieval.

So treat it as optional housekeeping, not a strategy. It costs almost nothing to publish and it may serve as a content map for the handful of systems that do read it. Just do not let it displace the work that demonstrably matters, which is making your actual HTML fetchable and extractable.

Which crawlers do you actually need to allow?

This is where the decision gets real, because training crawlers and retrieval crawlers do different jobs and blocking them has different consequences. Blocking a training crawler keeps your content out of a future model. Blocking a retrieval crawler removes you from the answers being generated today.

The distinction people get wrong most often is within OpenAI's own fleet. GPTBot is the crawler associated with training data collection. OAI-SearchBot is the one that builds the search index behind ChatGPT's browsing, and ChatGPT-User fetches a page when a user's question requires it. Block GPTBot and you have made a defensible copyright decision. Block OAI-SearchBot and you have removed yourself from ChatGPT's results. Those are very different outcomes and a single blanket rule cannot express both. Search Engine Journal maintains a verified crawler reference if you need the current user-agent strings.

My recommendation for most businesses: allow the retrieval and user-fetch crawlers, decide the training crawlers deliberately based on how you feel about your content being used that way, write the reasoning down in a comment in your robots.txt so the next person does not undo it, and then verify in server logs that reality matches your intent. The verification step is not optional. Robots rules fail silently.

So what actually moves LLM visibility?

Ranked roughly by how much difference they make relative to effort:

1. Be fetchable without JavaScript. Several retrieval systems read raw HTML and do not execute scripts. If your content only appears after hydration, you are invisible to them. View source and look for your own headline.
2. Allow the retrieval crawlers. Deliberately, per bot, verified in logs.
3. Restructure for passage extraction. Question headings, answers first, real lists and tables, sections that stand alone.
4. Fix entity consistency. One description of your business everywhere, backed by structured data with stable identifiers.
5. Earn third-party corroboration. Models weight claims that appear in more than one place. Being described consistently on sites you do not control is the strongest form of this, which is the uncomfortable point where AI visibility work turns back into ordinary reputation and link work.
6. Keep URLs stable. A citation points at a URL. Move the page and you throw away the accumulated evidence that it is the right answer.

Notice what is absent from that list: keyword density, word count targets, and publishing frequency. None of them meaningfully influence a retrieval system's decision about which passage answers a question best.

How is LLM SEO different from traditional SEO?

Less than the marketing suggests, and more than the skeptics allow. The foundations are shared almost completely, because retrieval depends on crawling, rendering, and indexing exactly as ranking does. What changes is the unit of competition and the definition of success.

Dimension	Traditional SEO	LLM SEO
Unit optimized	The page	The passage
Goal	Ranking position	Being cited or named
Query shape	Short keywords	Long conversational questions
Winner count	Ten positions	Often three to five sources
Measurement	Rankings and clicks	Mentions, citations, crawler hits
Content shape	Comprehensive depth	Extractable answers

The last row is the tension worth naming. Comprehensiveness and extractability pull against each other, and the resolution is not to pick one. It is to write comprehensive pages assembled out of individually extractable sections.

The honest summary: if your technical foundations are broken, LLM SEO is technical SEO. If they are solid, it becomes a content structure and entity clarity problem. Almost nobody needs a separate discipline, and the practitioners selling one usually cannot tell you which stage of the retrieval pipeline they are influencing.

For the tactical layer of this, see how to rank in ChatGPT and Perplexity. For the foundations, see technical SEO for AI search. And once you have made changes, you need to know whether they worked, which is AI visibility tracking.

LLM SEO FAQ

What is LLM SEO?

LLM SEO is the practice of making content reachable, parseable, and quotable by large language model systems so your business is cited when someone asks a question in your category. It focuses on the retrieval pipeline that selects sources, rather than on ranking positions in a list of links.

Is LLM SEO different from regular SEO?

The foundations are nearly identical, because retrieval depends on crawling, rendering, and indexing just as ranking does. What changes is that the unit of competition becomes the passage rather than the page, and success means being cited rather than ranked. If your technical foundations are broken, LLM SEO is technical SEO.

Do I need to be in the training data to appear in ChatGPT?

No. Modern AI search assembles answers from pages fetched at the moment of the query rather than from training memory. That makes AI visibility a current-state problem you can influence now, and it means the lever is retrieval access rather than historical inclusion in a dataset.

Does llms.txt improve AI search visibility?

Current evidence says no, or at least not meaningfully. Google has said it has no plans to use it, and log analyses in 2026 found that the major AI crawlers overwhelmingly skip the file and parse HTML directly instead. It is cheap optional housekeeping, not a strategy.

Should I block AI crawlers in robots.txt?

Decide bot by bot, because training crawlers and retrieval crawlers have different consequences. Blocking a training crawler like GPTBot keeps your content out of future models. Blocking a retrieval crawler like OAI-SearchBot removes you from the AI answers being generated today. Then verify in server logs that your rules match reality.

Which matters more for LLM SEO, content or technical work?

Whichever one is currently blocking you. If retrieval systems cannot fetch or render your page, no content improvement will help. Once fetching works, the constraint shifts to whether your content can be extracted as a self-contained passage that answers a specific question.