LLM SEO (LLMO): How to Get Language Models to Find and Cite Your Pages

People now ask ChatGPT which accounting software suits a freelancer, ask Perplexity how to fix a leaking tap, and read Google's AI Overview instead of opening a result. Each of those answers is written by a large language model, and each one mentions some businesses and sources and leaves out the rest.

LLM SEO is the name for trying to be among the ones mentioned. The older name for the same idea is LLMO, large language model optimization. Neither is a formal discipline with agreed rules. What exists is a handful of things the model makers have documented, some research, and a lot of guesswork sold as certainty. This guide sticks to the first two.

Everything in LLM SEO follows from one distinction. A language model can use your content in two different ways, and they work on different clocks.

The first is memory. A model is trained on a very large body of text collected up to a cutoff date. If your business was written about widely before that date, the model may know it without looking anything up. You cannot edit that memory, and it only refreshes when a new model is trained.

The second is retrieval. Most assistants can now search the web while they answer, read a few pages and cite them. This is how ChatGPT search, Perplexity and Google's AI features work. Retrieval sees what is on your site today.

The practical upshot: if you changed your prices last month, memory is wrong and retrieval is right. A model answering from memory may also describe a competitor's product as yours. Retrieval is where a small site can compete, because the engine is choosing between pages, not between reputations.

Retrieval has three gates, and a page has to pass all of them.

Access comes first. The crawler has to be allowed to fetch the page. That means robots.txt, and it also means whatever sits in front of your site. Google's guidance on AI features says to make sure crawling is allowed by any CDN or hosting infrastructure, and OpenAI and Perplexity both ask sites to let their published IP ranges through. Bot protection that challenges unknown visitors is the usual culprit.

Indexing is second. Google's AI Overviews and AI Mode draw on the ordinary Google index, so a page must be indexed and eligible to show a snippet. ChatGPT search and Perplexity also send their own crawlers, OAI-SearchBot and PerplexityBot, to find pages. A page nobody links to, missing from your sitemap, may simply never be found.

The passage is third. The engine pulls a few paragraphs from each candidate page and keeps the ones that answer the question. Writing for that step is the part most sites skip, and it gets its own section below.

You cannot submit your site to a model's memory. What a model remembers about a business comes from everything written about it: its own pages, reviews, directory listings, forum threads, news and other people's articles. A business described the same way in twenty places is remembered more reliably than one described once.

Two things are worth doing. Keep the basic facts identical wherever you appear: the name, what you sell, who it is for, where you are. Contradictions are how a model ends up confidently wrong about you. And earn mentions in places a model is likely to have read, such as industry directories, comparison articles and community discussions where your customers ask for recommendations.

Then decide about training crawlers. Allowing them is the only direct way your own words get into the next round of training. Declining is a reasonable choice too, and it costs you nothing in live search.

Each company uses separate crawlers for separate jobs, and the names are documented. OpenAI says disallowing OAI-SearchBot means a site will not be shown in ChatGPT search answers, though it may still appear as a plain navigational link. Anthropic says blocking Claude-SearchBot may reduce a site's visibility in search results inside Claude. Perplexity says PerplexityBot is not used to train foundation models.

Google is the odd one out. Its AI features run on Googlebot's index, so there is no separate search crawler to allow. The Google-Extended token controls use for Gemini, and Google says it does not affect inclusion in Search.

The example below allows search and user-requested fetches and declines training. Change the second group to Allow if you want your writing in future models.

A retrieval system does not read your article the way a person does. It cuts the page into passages and scores each one against the question. A passage that opens with "It also helps with this" is useless out of context, because nobody knows what "it" or "this" refers to.

So write passages that survive being cut out. Name the subject in the first sentence of each section. State the fact, the number and where it comes from together. Use the same name for a thing every time; a model may treat "the Pro plan", "our premium tier" and "Pro" as three different things.

The GEO research paper measured which edits made pages more visible in generated answers. Adding citations, quotations and statistics were among the changes that helped most, with gains of up to 40% for the best methods, and the effect varied by subject. Keyword repetition did little.

Keep the important content in plain text. A price inside an image, a comparison that only appears after a click, or a specification loaded by a script may never reach the model. Google makes the same point for its own AI features.

Most products sold as LLM SEO tools are trackers. You give them a list of prompts, they ask several assistants on a schedule, and they report how often your brand is mentioned, which sources were cited and how you compare with competitors. Entry prices on the vendors' own pages run from $29 a month for Otterly.AI's Lite plan to $95 for Peec AI's Starter and $250 for Scrunch's Core plan. AthenaHQ has a free plan with 300 credits.

A tracker tells you where you stand. It does not change your pages. For a site that has not yet fixed access and openings, a monthly list of ten questions asked by hand gives the same signal for free. A tracker earns its fee once you publish enough that manual checks stop scaling. Our comparisons of GEO tools and AEO tools cover the options by price and by job.

For the page itself, the free AEO and GEO checker reads one URL and reports which AI crawlers your robots.txt allows and which answer-ready parts the page has. PixelClaw is our product, and the checker is free to use once a day.

Ask the same question twice and an assistant may cite different sources each time. Answers also change with the wording, the user's location and the model version. There is no position one to hold, only a share of answers you appear in.

No model maker publishes ranking factors for citations. Anyone promising a guaranteed mention in ChatGPT is selling something they do not control. What the documentation and the research support is narrower and less exciting: be reachable, be indexed, be specific, be sourced, and be described consistently across the web. Our ChatGPT SEO guide applies those five to one assistant in detail.