LLM SEO: optimizing for models, not just crawlers

The name is unfortunate. "LLM SEO" suggests you're optimizing a model's weights — writing content that somehow flatters the model itself. You aren't. Almost every AI answer about a current topic is assembled at query time from pages a retrieval system fetched seconds earlier. The model didn't remember you. It read you.

That distinction decides everything else here. If the mechanism is retrieval, the job is making a page easy to find, easy to parse, and easy to quote without distortion.

Why LLM visibility became a separate concern

For twenty years the assumption held that ranking produced traffic. That link is loosening. Pew Research Center tracked 68,879 Google searches from 900 US adults in March 2025 and found that when an AI summary appeared, users clicked a traditional search result on 8% of visits — against 15% when no summary appeared. Clicks on the sources cited inside the summary happened on about 1% of those visits.

That cuts both ways. Being cited in an AI answer sends very little traffic. Being absent sends none, and increasingly the answer is where the decision gets made. You're optimizing for consideration now, not only for sessions.

The supply side moved too. Cloudflare's July 2025 analysis of crawl-to-refer ratios found AI platforms requesting pages vastly more often than they send visitors back. Cloudflare itself notes the ratios likely overstate the gap, since some AI apps send referrals without a Referer: header — but nobody disputes the shape.

What Google actually says you need

This is where most LLM SEO advice diverges from the documentation. Google published explicit guidance on optimizing for generative AI features, and it is blunter than the industry around it:

We sell SEO and AI search visibility and we still think you should read that table at face value. Structured data isn't a magic AI ingredient; it earns its place through rich results and entity disambiguation. llms.txt is cheap and probably a small edge — the honest version is in llms.txt explained. Neither is the mechanism that gets you quoted. The mechanism is that your page was retrievable, and the passage the system pulled was good enough to use.

What the research says works

The most-cited empirical work here is the GEO study by Aggarwal, Murahari and colleagues, presented at KDD 2024. The researchers built GEO-bench — 10,000 queries across domains — and tested content edits to see which increased a source's visibility inside generative answers.

The headline finding: targeted edits boosted visibility in generative engine responses by up to 40%. The edits that did best were the ones you'd hope for. Adding citations. Adding relevant quotations. Adding statistics. Improving fluency and making the writing read as authoritative.

The finding worth pinning to a wall is the inverse one: keyword stuffing was the weakest strategy tested. Keyword density was a lever in classic search for two decades. In generative retrieval it does nothing useful and can cost you — these systems select passages that read as credible evidence, and stuffed text reads as the opposite.

Two caveats keep this honest: the study measured visibility on its own benchmark, not revenue, and the authors note effectiveness varies by domain, so no single tactic list transfers cleanly everywhere.

The work, in the order that matters

1. Be retrievable. Crawlable, indexed, fast, one H1, real HTML rather than text painted by JavaScript. Nothing downstream matters if this fails, and a surprising share of "we're invisible to AI" diagnoses end here.

2. Answer the question in the first sixty words. A model extracting an answer needs a self-contained sentence that resolves the query without the surrounding paragraphs. If your best sentence is in paragraph nine, it's functionally invisible.

3. Attribute your evidence. Named source, named publisher, date, link. The GEO study found this measurably raised visibility, and it forces you to check whether your numbers are real. Many circulating statistics aren't.

4. Structure so a machine can lift cleanly. Real headings, real lists, real tables. Comparison tables get extracted more than almost anything else, because they compress a decision into rows.

5. Build depth on one subject. One good page rarely establishes credibility on a topic. A set of them does. This is the slow part, and the part almost everyone abandons.

6. Measure citations, not just positions. Ask the engines your buyers' questions on a schedule and record who gets named. AI visibility tracking covers the mechanics.

For the sequenced version with diagnostics attached, see AI search optimization.

Where LLM SEO is the wrong investment

If your pipeline is empty this month, this is not the fix. Retrieval-based visibility compounds over quarters; paid search and outbound buy demand now. Build this underneath them, not instead of them. It's also a poor fit if nothing publishes without a six-week review cycle, or if nobody can edit the site's markup — the structural work is most of the value, and it requires touching templates.

Frequently asked questions

Is LLM SEO different from regular SEO?

Partly. The technical foundation is identical — a page must be crawlable, indexed and fast before any system can retrieve it, and Google's own guidance states that optimizing for generative AI search is still SEO. What genuinely differs is content structure and measurement. Traditional SEO optimizes a page to win a ranking position a human clicks. LLM SEO optimizes a passage to be extracted and quoted accurately, and measures success in citations and brand mentions. Anyone selling it as a wholly separate discipline with a separate invoice is mostly repackaging work you're already paying for.

Do large language models remember my website?

Generally no, and this is the most common misunderstanding in the field. When an AI assistant answers a question about something current, it typically runs a search, fetches live pages, and composes an answer from what it just read. Your content is being retrieved at query time, not recalled from training. That's good news: it means changes you publish today can influence answers within days or weeks, rather than waiting for a model retraining cycle you have no visibility into or control over.

Does keyword stuffing help with AI search?

No — the evidence points the other way. The GEO study presented at KDD 2024 tested a range of content edits across a 10,000-query benchmark and found keyword stuffing the weakest strategy examined, while citations, quotations and added statistics performed best. This inverts twenty years of classic search intuition. Generative systems select passages that read as credible, self-contained evidence. Text engineered for density reads as low-quality to exactly the mechanism you are trying to influence, so the tactic costs you rather than helping.

Do I need an llms.txt file to be cited by AI?

No. Google's published guidance on generative AI features lists no special file as a requirement, and no major engine currently guarantees it will honor one. llms.txt is an emerging convention rather than a ratified standard. It costs almost nothing to ship, makes a site easier for an agent to navigate, and carries no real downside, so implementing it is reasonable — as long as you understand it is a small edge rather than the mechanism that earns citations. Retrievability, structure and credible evidence do the heavy lifting.

How long does LLM SEO take to show results?

Technical fixes can move within weeks: indexation problems, page speed and heading structure often produce measurable change in the first month or two. Citation gains are less predictable than rankings, because retrieval varies between engines, users and sessions. A well-structured page answering a specific question can get quoted before it ever ranks in the top ten, which sometimes makes AI visibility faster than traditional rankings. Topical authority, the part that compounds, generally takes several months. Anyone promising citations in 30 days is describing something else.

Should I block AI crawlers to protect my content?

That's a real business decision, not an obvious one. Cloudflare's analysis found AI platforms crawl far more than they refer, so the traffic-for-content trade is worse than it was in classic search. Blocking protects your work from being summarized away. It also removes you from answers your buyers are reading, and from the consideration set entirely. Publishers with subscription revenue often block. Businesses that need to be found by people who don't know them yet usually should not. Decide deliberately rather than by default.