LLM SEO: what changes, what does not, and what a site owner can actually control.

The short answer.

LLM SEO is search engine optimisation aimed at the answers large language models write, instead of the list of links a search engine shows. Most of it is the SEO you already know. What is new is where the model gets its information.

• 1There are two routes into an answer. What the model remembers from training, and what it reads from a live search while answering.
• 2The fundamentals do not change. Google says of its own AI features that “the best practices for SEO continue to be relevant.”
• 3Three things do change. What other websites say about you matters more, answers vary from one run to the next, and you measure how often you are named instead of where you rank.

Put most of your effort into live retrieval, where your actions show soonest, and treat training data as a slow by-product of doing the rest well.

What LLM SEO means.

LLM SEO means making your company easy for a large language model to find, understand and repeat correctly. You will also see it called LLM optimisation, LLMO, large language model optimisation or SEO for LLMs.

A large language model (LLM) is the software that writes the answer in ChatGPT, Gemini, Claude, Perplexity, Microsoft Copilot and Google’s AI Overviews. When a buyer asks “which firms should I consider for X?”, the model often names a handful of companies. LLM SEO is the work of being one of them, described accurately.

The label overlaps heavily with two others, generative engine optimisation and answer engine optimisation. We keep the three apart in our comparison of how AEO, GEO and SEO differ. What “LLM SEO” adds is a focus on the model itself: what does it know, where did it learn it, and what can it look up?

Google’s position is worth knowing before you buy anything with a new acronym on it. Its guide to generative AI search says: “optimizing for generative AI search is optimizing for the search experience, and thus still SEO.” A new name does not mean a new discipline. Our GEO and AEO services take the same view: one scope that covers Google search and AI answers together.

Two routes into an answer.

Everything a language model says about your company comes from one of two places: its training data, which is fixed at a date in the past, or pages it retrieves through a search while it writes the answer.

The distinction comes from how the systems are built. The 2020 research paper that named retrieval-augmented generation described language models as storing “factual knowledge in their parameters”, and proposed pairing that built-in memory with an external one the model can look things up in. For built-in memory alone, the authors noted, “providing provenance for their decisions and updating their world knowledge remain open research problems.”

The vendors’ own documentation shows both routes. Google describes search grounding for Gemini as a way to “provide more accurate answers and cite verifiable sources beyond its knowledge cutoff”: the model “analyzes the prompt and determines if a Google Search can improve the answer.” Anthropic’s documentation says Claude searches “when the request depends on information that is current, changing, or outside its training data”, and lists “information about specific organizations, people, or products that might have changed” as one of those cases. It answers without searching when the request “draws on stable knowledge”. Questions about which company to hire are the kind that tend to send a model to the web.

Route one: training data.

Training data is everything the model read before it was released. You cannot submit to it, edit it or check what it holds about you. You can only decide whether training crawlers may read your site, and keep what the public web says about you accurate.

What training memory is

AI companies collect web pages with crawlers and use them to train their models. OpenAI says its GPTBot crawler is “used to crawl content that may be used in training our generative AI foundation models.” Anthropic’s ClaudeBot does the same job for Claude. Google instead offers a robots.txt token, Google-Extended, to control whether content it already crawls “may be used for training future generations of Gemini models”.

Training stops at a cutoff date. Anthropic’s model documentation, for example, lists a “reliable knowledge cutoff” of June 2026 for its current models, and defines that as “the date through which the model’s knowledge is most extensive and reliable.” On the day this guide was published, that left a gap of about five months. A service launched or a rebrand made inside that gap is not in the model’s memory at all.

What you can control

• ✓Access. Whether GPTBot, ClaudeBot and the Google-Extended token are allowed in robots.txt.
• ✓Consistency. The same company name, description, location and service list on your site and on every profile you control.
• ✓Presence. Being written about, accurately, on sites other than your own.

What you cannot control

• ✕Whether a page that was crawled was actually used, or how much weight it carried.
• ✕A wrong fact already inside a released model. The crawler documentation cited here describes no way for a business to correct one.

In practice, a wrong statement from training memory is fixed through route two: you correct the source pages, and the model picks up the correction when it searches. Our guide to tracing and fixing wrong facts in ChatGPT goes through the process.

Blocking training crawlers keeps your content out of future models. A company that sells services usually wants the next model to know what it does. Choose on purpose and write it down.

Route two: live retrieval.

Live retrieval is the model running a web search while it answers, reading some of the results and writing from them. This is the route a site owner can influence.

How it works

The details differ by product, but the published descriptions agree on the shape. The model decides whether to search. OpenAI’s documentation says “the model can choose to search the web or not based on the content of the input prompt.” If it searches, it writes its own queries, often several. Google calls this “query fan-out” in AI Overviews and AI Mode: “issuing multiple related searches across subtopics and data sources.” It then reads what comes back and writes an answer with links.

The model reads more than it cites. OpenAI’s web search documentation says “the number of sources is often greater than the number of citations”, so a page can shape an answer without being linked. We cover the mechanics in how ChatGPT decides which companies to recommend.

Each product has its own gatekeeper. For ChatGPT search it is the OAI-SearchBot crawler: OpenAI says “sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links.” For Google’s AI features it is the normal Google index: a page must be “indexed and eligible to be shown in Google Search with a snippet.” Perplexity’s PerplexityBot “is designed to surface and link websites in search results on Perplexity”. Blocking a training crawler does not block these.

What you can control

Nearly everything on your own site: whether the search crawlers can reach your pages, whether the text is in the HTML, whether the page is indexed, whether it answers a buyer’s question plainly, and whether it is current. A change counts once the page is recrawled; for its own index, Google says “crawling can take anywhere from a few days to a few weeks.” The technical half of that list is in our guide to AI crawlers, robots.txt and llms.txt. To see which bots actually arrive, read our guide to AI crawler log file analysis.

What you cannot control

The queries the model writes, which results it opens, which passages it keeps, and whether it names you in the final answer. Nor can you control what third-party pages say, although you can influence it.

What you can and cannot control.

You control access and your own pages, you can influence what other sites say, and you cannot control the model’s choices.

Note the last row. Accurate, independent pages that mention you are the one kind of work that pays into both routes at once.

What does not change.

The technical and content fundamentals of SEO carry over unchanged, because retrieval runs on search indexes and web pages.

Google is explicit. Its AI features documentation says “there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”, and that “all existing SEO fundamentals continue to be worthwhile.” Its generative AI guide adds: “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search.”

On rendering, Vercel’s December 2024 analysis of crawler traffic on its network found that “none of the major AI crawlers currently render JavaScript”, with Google’s Gemini (which uses Googlebot’s infrastructure) and Applebot as exceptions, and advised that “any important content should be server-rendered.” The study is nearly two years old, but key text in the HTML helps every crawler.

What does change.

Three things are different enough to change how you plan: the weight of brand mentions on other sites, the variability of answers, and the shape of the passages that get used.

1. Mentions matter more than links

In May 2025 Ahrefs studied 75,000 brands and compared how often each appeared in Google’s AI Overviews with a range of familiar SEO measures. The strongest relationship was with branded web mentions (a correlation of 0.664); the number of backlinks showed a much weaker one (0.218). Ahrefs notes that correlation is not causation, and the study covers AI Overviews only.

Treat this as a direction, not a formula. In our November 2026 study of where AI answers get their sources, we recorded which kinds of sites were cited for 12 B2B buying questions, and our guide to the off-site sources AI engines rely on ranks those sources by effort and value.

2. Answers are not stable

Research by SparkToro and Gumshoe.ai, reported by Search Engine Journal on 30 January 2026, repeated 12 brand-recommendation prompts 60 to 100 times per platform on ChatGPT, Claude and Google’s AI search features. ChatGPT and Google’s AI each returned the same list of brands less than 1% of the time. The study was run with an AI tracking company and is not peer-reviewed, but the practical lesson holds: a single answer proves nothing.

3. The passage is the unit

A search engine ranks pages. A model writes from pieces of pages. A paragraph that answers one question completely and names its subject can be lifted cleanly; one that depends on the three before it cannot. Each section should make sense when read alone. Our guide to writing passages AI engines can quote has the patterns and a checker.

The prioritised checklist.

Work in this order: access first, then your own pages, then what other sites say, then measurement.

The checklist in detail

• ✓Check robots.txt and your firewall for OAI-SearchBot, Claude-SearchBot and PerplexityBot.
• ✓View the page source of your main service pages and search for a sentence from the middle.
• ✓Write one plain description of the company (what you do, for whom, where, since when) and make your site, LinkedIn page and directory profiles match it.
• ✓List the ten questions buyers ask before they shortlist. Make sure each has a page, or a section, that answers it in the first two sentences.
• ✓Find the third-party pages AI answers cite for your category and work out which ones you could credibly appear on.
• ✓Set the baseline before you change anything, so you can tell what the work did.

If your company is absent from every answer, we set out the diagnosis in why a brand goes missing from AI shortlists.

Measuring LLM SEO.

Measure how often you are named across a fixed set of buyer questions, repeated over time, and then whether named visibility turns into visits and enquiries. Do not report a “rank”; position in an AI answer does not hold still.

A workable measurement set has three parts.

1. A question panel. Ten to twenty questions real buyers ask, run through each engine several times a month. Record whether you were named and who else was. Our guide on how to check your AI search visibility gives the manual method, and for a first reading without building the panel, the free AI visibility report covers a sample of your buyers’ questions.
2. The engines’ own reports. Bing Webmaster Tools launched AI Performance in public preview on 10 February 2026. It reports how often your pages are cited across “Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations”; Bing says the data “does not indicate ranking, authority, or the role of any page within an individual answer.” Google announced Search Generative AI performance reports for Search Console on 3 June 2026, showing impressions in its AI features in Search and Discover.
3. Referrals and enquiries. Visits from AI tools in your analytics, and “how did you hear about us?” on your forms. Our guide to tracking AI referral traffic through to pipeline covers the set-up.

Keep the two routes apart when you read the results. If answers given without search still describe the old you, that is training memory, and it will not move until a new model is released.

The LLM SEO self-check.

Tick each statement that is true today. You get a score for each route and the fixes to make first. Nothing you tick is sent anywhere.

Retrieval fixes are listed first because they show a result soonest. The scores show where to start; they do not measure your visibility.

Questions owners ask.

What is LLM SEO?

LLM SEO is the work of getting large language models, such as those behind ChatGPT, Gemini, Claude and Perplexity, to mention, describe and cite your company accurately. It combines SEO fundamentals with attention to the two sources a model draws on: training data and live search.

Is LLM SEO different from SEO?

Mostly not. Google says optimising for generative AI search “is optimizing for the search experience, and thus still SEO.” What changes is the weight of brand mentions on other sites, the variability of answers and how results are measured.

Can I get my company into an LLM’s training data?

Not directly. You can allow training crawlers such as GPTBot and ClaudeBot in robots.txt and keep accurate, consistent information about your company on the public web, but you cannot see or control what a model retains.

How long does LLM optimisation take to show results?

It depends on the route. Changes to pages can appear in search-backed answers once the pages are recrawled. Google says crawling “can take anywhere from a few days to a few weeks.” Changes to training memory appear only when a new model is released.

Our scope covers access, pages, the sources AI answers draw on, and measurement. See how we help companies get recommended in AI answers.

Sources: Google Search Central, generative AI search guide; Google Search Central, AI features and your website; Google, common crawlers (Google-Extended); Google, ask Google to recrawl; Google, Grounding with Google Search; OpenAI, crawlers; OpenAI, web search tool; Anthropic, crawlers; Anthropic, web search tool; Anthropic, models overview; Perplexity, crawlers; Lewis et al., Retrieval-Augmented Generation, arXiv, 22 May 2020; Vercel, 17 Dec 2024; Ahrefs, 26 May 2025; Search Engine Journal, 30 Jan 2026; Bing Webmaster Blog, 10 Feb 2026; Google Search Central Blog, 3 Jun 2026. All accessed 12 November 2026.