LLM SEO: what actually works, according to research
LLM SEO isn't a new trick, it's a different question: not how you rank higher, but how you get cited. What research and practice actually show, without guarantees.
The short answer: LLM SEO is not a new flavour of classic SEO, it is a different question. On Google you optimise to rank high in a list. With a language model there is no list, only a generated answer with a handful of sources folded into it. So the question shifts from "how do I rank higher" to "how do I get in at all".
We already covered what GEO is and where it came from and concretely getting recommended by ChatGPT. This article zooms out: what is LLM SEO (also called SEO for AI) exactly, what does research actually show works, and how does it relate to the SEO you already know?
What is LLM SEO exactly?
LLM SEO, sometimes called SEO for AI, is optimising your content so language models (the engine behind ChatGPT, Gemini, Claude and Perplexity) use you as a source more often in a generated answer. Not a position, not a ranking, but a mention or citation inside running text.
The term is younger than you'd think. Academically this is called "generative engine optimization" (GEO), a name first coined in a paper by Pranjal Aggarwal and colleagues in 2023 (the underlying idea, that AI answers can be influenced, existed informally before that). Outside academia, "LLM SEO" or "SEO for AI" is the more common label for the same approach. We use the terms interchangeably in this article.
Why classic SEO only partly applies here
Some of what you know from SEO still matters. Crawlability, for one: a language model cannot cite what it cannot read. Clear structure too: headings that answer a question help both a search engine and a language model match your content to a user's query.
But a few core classic SEO factors lose weight:
| Classic SEO factor | Role in LLM SEO |
|---|---|
| Keyword density | Generally not much of a direct factor; models match on meaning, though clear wording still helps in finding the right passage |
| Backlink count | Not a direct factor at the model level; indirectly relevant when an assistant leans on an underlying search index |
| Page speed | No known direct ranking role; still matters so crawling and live fetching don't fail |
| Factual density and quotability | Becomes more important: models cite concrete, checkable statements |
| Unambiguous entity (who you are, for whom) | Becomes more important: the model needs to be able to match you to the question |
In short: you don't start from zero if you already do SEO, but the dials you turn shift.
What research shows actually works
Most of the evidence here comes from the same GEO paper that introduced the term generative engine optimization. The researchers tested nine content adjustments on a benchmark of real search queries and measured the effect on how often and how prominently a source was used in the generated answer. Two adjustments stood out:
- Adding statistics (concrete figures backing up the text) worked particularly well for factual and opinion-type questions.
- Adding quotations and citing sources (direct quotes, references to authorities) worked particularly well for explanatory and society-related topics.
Combining the two, together with readable, fluent writing, gave the strongest effect. What did not help: keyword stuffing and stylistic padding without substance. Take the reported effect (the authors claim up to 40% more visibility in their test setup) for what it is: a result from an academic benchmark, not a guaranteed outcome for your content in practice. Effects also varied strongly by topic; what works for a legal question does not automatically work for a product comparison.
Five things that improve your odds
Translated into something you can do today:
- Answer the question directly, ideally in the first paragraph. A model is more likely to pick a clear, self-contained answer than a text that only gets to the point after a long introduction.
- Back claims with concrete, dated figures. "170 searches per month in the Netherlands, measured in September 2026" is quotable. "Many people search for this" is not.
- Add quotations and cite sources where relevant, instead of only stating your own opinion.
- Make your entity unambiguous: who you are, for whom, what sets you apart. A model needs to be able to match you to the question before it can cite you.
- Make sure you're crawlable for the relevant AI crawlers. This works slightly differently per assistant; for the ChatGPT-specific steps (which bots, which robots.txt rules) see our article on getting recommended by ChatGPT.
One example: the same sentence, written twice
An indicative, made-up example (no real client or dataset) of what this looks like in practice.
| Before (classic SEO copy) | After (LLM SEO) | |
|---|---|---|
| Claim | "We are the leading specialist in invoicing software." | "Our software sends invoices in three Dutch accounting formats and processes VAT filings automatically." |
| Backing | None, just the claim itself. | "According to our own customer survey (n=140, August 2026), this saves an average of 2 hours per month." |
| Source | None. | Reference to the research report or an external industry source. |
The right-hand column isn't necessarily more persuasive to a human, but it is easier for a model to pick up as fact: specific, dated and backed up. If you want a number like that to actually be citable, publish the methodology behind it too, not just the outcome.
SEO for AI is not the same as AI for SEO
A common mix-up, visible in the search volumes too: "AI for SEO" and "SEO for AI" sound almost the same but are opposite things. AI for SEO means using AI tools to speed up your classic Google optimisation (generating content, doing research, running technical audits). SEO for AI, or LLM SEO, means optimising your content itself so an AI assistant cites you. The first is about your tooling, the second about your target. This article is about the second.
What LLM SEO is not
A few things worth keeping sharp:
- Not a guarantee. Anyone promising a guaranteed spot in an AI answer is selling something that doesn't exist. Generated answers shift by question, by model and by version.
- Not a hack. There is no trick that tricks a language model into mentioning you. The techniques that work (factual clarity, quotability, a clear entity) are just good writing, aimed at a different target.
- Not a replacement for classic SEO, more of an addition. As long as people still search on Google, classic SEO stays relevant; LLM SEO sits alongside it, not instead of it.
How you measure whether it's working
Because there is no fixed position to track, measuring here looks different from classic SEO. You can't check your ranking every morning; you need to periodically ask the same buyer questions to multiple assistants and track whether and how you get mentioned. Watch for:
- Whether you get mentioned, and for which question phrasings you do and don't.
- With which assistant (ChatGPT, Gemini, Claude and Perplexity regularly give different answers to the same question).
- How stable that is across repeated measurements; a single test tells you little.
That's exactly what a GEO audit is for: not a one-off sample, but a repeatable measurement across multiple assistants and question variants, so you can see whether a change actually moves the needle instead of guessing. We measure and advise; making the actual changes to your content stays your work.
Frequently asked questions
What is LLM SEO? LLM SEO, sometimes called SEO for AI, is optimising content so language models like ChatGPT, Gemini, Claude and Perplexity cite or mention you more often in their answers. It is not about a position in a results list, it is about the odds of being one of the sources the model actually uses.
Is LLM SEO the same as GEO? Pretty much. LLM SEO is the more common, informal name for what researchers called generative engine optimization (GEO). Both terms point at the same thing: writing so a language model can more easily fold your content into a generated answer.
Do classic SEO techniques still work for language models? Partly. Crawlability, clear structure and an unambiguous entity help with both. But keyword density and backlink volume, two classic ranking factors, matter far less to language models than factual clarity and quotability.
Can I guarantee a language model will cite me? No. A generated answer is not a fixed ranking and shifts by question, by model, and by version. You can improve your odds with the right content properties, but no one can guarantee a citation.
Is using AI for SEO the same as SEO for AI? No, they are opposite things. AI for SEO means using AI tools to speed up your classic search engine optimisation. SEO for AI, or LLM SEO, means optimising your content itself so an AI assistant cites you.