What Is LLM SEO? A Plain-English Guide to Getting Cited by ChatGPT, Claude, and Gemini

LLM SEO is the practice of shaping a website’s content and structure so that large language models like ChatGPT, Claude, and Gemini are more likely to name your brand, quote your page, or cite your domain when they answer a question in your category. It borrows the goal of traditional SEO, being found, but the mechanism underneath it is different: you are not optimizing for a ranking algorithm that returns a list of links, you are optimizing for a model that reads a question, writes an answer in its own words, and decides somewhere in that process whether your business is worth a mention inside it.

LLM SEO in one sentence

Traditional SEO earns a slot on a results page: ten blue links, a fixed number of positions, and a click that lands the visitor on your site. LLM SEO earns a mention inside a paragraph the model wrote itself, with no fixed number of slots. Some answers name one vendor. Others, in the same run, name close to twenty. There is no page two to rank up from, and no guarantee that being “first” in a ranking sense means being named at all: in the help desk teardown, two free, open source tools that were named less often than the category leader still averaged better positions when they did get named.

That is the part most people miss when they hear “LLM SEO” and assume it is search-engine SEO with an AI coat of paint. The target moved from a list a search engine assembles to a paragraph a model composes, and the composing model gets to decide, per question, whether you belong in it.

The phrasing of the question changes the outcome, too, in a way a fixed-keyword ranking never did. FoundCite’s help desk teardown asked eight differently phrased buyer questions in the same category, from “what’s the best help desk software for a support team of about 15 agents?” to “we want AI to draft replies to support tickets but never send them automatically, which platforms allow that?” A brand strong on general help desk positioning and weak on that second, narrower requirement can be named in one answer and skipped entirely in the other, inside the same category, on the same day. A page ranks for a keyword. A brand gets named, or does not, for a specific, oddly phrased real question, and the granularity of that question matters more than it ever did for a search ranking.

How a model actually decides what to cite

There are two broad mechanisms, and most assistants blend them. The first is training-time knowledge: whatever the model absorbed about your brand, your category, and your competitors during training, with no live lookup involved. The second is retrieval at answer time, often called RAG or, for a search-grounded assistant like Perplexity, just how it always works: the model (or a system wrapped around it) runs a real search, reads a handful of live pages, and writes an answer that cites what it just read.

You can see both mechanisms in the same run. FoundCite publishes a public teardown of one real measurement: eight buyer questions in the help desk software category, run fresh against three assistants (ChatGPT on openai/gpt-5.6-terra, Grok on x-ai/grok-4.3, and Perplexity on perplexity/sonar), for 24 answers total, measured July 29, 2026. In that file, ChatGPT’s answer to one question links out to a Zendesk support article with a utm_source=openai parameter still attached to the URL, which is a live citation, not something recalled from training. In the same run, answering the identical question, Grok’s answer carries its own set of numbered citation links to a different set of pages entirely. Same question, two assistants, two different reading lists behind the answer.

The practical takeaway is that “getting cited” is not one target. For a retrieval-grounded answer, the question is whether your page is the kind of page a live search would surface and a model would want to quote. For a training-time answer, the question is closer to traditional brand and content authority: has enough been written about you, by you and by others, that the association was there to absorb in the first place. You can read the full run, every question and every answer, in the raw teardown file.

LLM SEO vs SEO vs AEO vs GEO

These four labels overlap enough in current usage that most people use them interchangeably, and there is no governing body standardizing any of them yet. Roughly: SEO is the original discipline, optimizing for a search engine’s ranking. AEO, answer engine optimization, is usually used for optimizing content to be pulled into a direct answer, whether that is a featured snippet or a voice assistant’s spoken reply. GEO, generative engine optimization, traces back to an academic paper from late 2023 that tried to formalize which on-page tactics move the needle specifically inside a generative, LLM-written answer. LLM SEO is the plainest-English version of the same idea: optimizing so an LLM names you.

In practice, the distinctions matter less than the shared premise underneath all four: the surface where buyers first encounter your name has partly moved from a results page to a written answer, and the tactics that earn a spot in that answer are not identical to the tactics that earn a ranking. For more on the umbrella term people search for once they suspect this is happening to them, see what AI visibility means, and for how one specific number tries to summarize it, see what a GEO score is.

Two things LLM SEO is not

It is not keyword stuffing aimed at a bot instead of a crawler. Repeating a target phrase does nothing for a model that is trying to synthesize an answer in its own words; if anything, unnatural repetition makes a page look less like a source worth quoting and more like a page written for an algorithm, which is exactly the kind of writing a language model is comparatively good at spotting.

It is also not a single file you add once and forget. A page that ships a perfect llms.txt and a clean schema block, sitting on top of thin, vague prose that never actually states what the product does or who it is for, has fixed the metadata layer and skipped the layer that does most of the work: content specific enough for a model to want to lift a sentence from it. The metadata makes a good page easier to parse. It does not make a thin page worth citing.

The levers that actually move it

None of this is mysterious once you separate it from the ranking-algorithm mental model. The concrete things worth doing:

• Answer the question in the first sentence. A model looking to extract a citable fact rewards a page that states its point plainly instead of building up to it. This article’s own first paragraph is written that way on purpose.
• Name the entities explicitly. A paragraph that says “our platform” five times gives a model nothing to quote. A paragraph that names your product, your category, and your competitors by name gives it something concrete to lift.
• Add structured data. Schema markup does not force a citation, but it gives a model (and the crawler feeding it) an unambiguous, machine-readable version of the same facts your prose states in words. The schema markup generator builds the JSON-LD for that.
• Get named on the pages assistants already trust, not just your own. The citations inside the help desk teardown span more than sixty distinct domains: vendor docs and support pages, other vendors’ comparison posts, independent review sites like Capterra, and, for the compliance-focused question in the set, government pages like hhs.gov and marketplace.fedramp.gov. No single domain, including any one vendor’s own homepage, dominated the sourcing. Your homepage is one input among dozens a model might actually read.
• Watch llms.txt, but do not over-invest yet. It is an emerging, unofficial convention for a plain-text file that summarizes a site for a model. Some crawlers respect it today. It is worth having once your core content is in shape, not a substitute for that content.

How you’d actually measure whether any of this worked

Every tactic above is a guess until you check the answers themselves. The only way to know if you are named is to ask the assistants the questions your buyers actually ask, in a fresh session with no memory of you, and read what comes back. That is the entire method behind the teardown cited throughout this article: a fixed set of buyer questions, run cold, with every mention counted. Do it by hand for a handful of questions and you will already learn something. Do it every week across your full buyer question set, and you have a measurement instead of a guess.

One run also is not the whole picture. Models get updated, retrieval sources change, and a competitor publishing one strong comparison page can shift who gets named for a specific question within weeks. A single check tells you where you stand today. Repeating the same fixed question set on a schedule is what tells you whether any of the levers above actually moved the needle, or whether the answers simply drifted on their own.