LLM SEO is the practice of getting your pages cited inside the answers that ChatGPT, Google AI Overviews and AI Mode, Perplexity, Gemini and Copilot generate.

It gets sold under a lot of names, generative engine optimization, answer engine optimization, AI search optimization, and most of the advice attached to those names is either recycled SEO or guesswork. 

This guide sets the labels aside and walks through what the systems actually do when they pick a source, what the large-scale citation studies published between mid-2025 and mid-2026 found, and which changes on your side move the numbers.

One framing before the detail. A language model does not “rank” your page. It writes an answer, and somewhere in that process it either pulls your page into the context it is writing from or it does not. Getting cited is a retrieval problem first and a content problem second. Almost everything below follows from that.

Two Ways A Model Can Know About You

There are exactly two routes by which a language model can end up mentioning or citing your business, and they are governed by different rules.

• Parametric knowledge. Whatever the model absorbed during training. If your brand appeared often enough, in enough places, in consistent enough terms, the model has an internal representation of it and can mention it without looking anything up. This is why a model with no web access can still tell you that Ahrefs is an SEO tool and that Justworks is a PEO. You cannot edit this directly. You influence it slowly, by being written about, and it lags the real world by months or years depending on the model’s training cutoff.
• Retrieved knowledge. What the model fetches at answer time. Every major consumer assistant now runs a search step for most queries that involve current facts, products, companies or comparisons. The pages that come back from that search are the only pages the model can cite with a link. This is where LLM SEO actually operates, and it is why the term “SEO” is not entirely wrong: there is a retrieval system in the loop, it has an index, and it has ranking behavior you can study.

Citations, the linked sources you see under an answer, come almost entirely from the second route. Brand mentions without a link can come from either. Ahrefs, whose Brand Radar tool tracks both, reports them as separate metrics for exactly this reason, and it is worth keeping them separate in your own reporting too.

How Retrieval Works, Platform By Platform

Google AI Overviews and AI Mode: query fan-out

Google has been the most open about its mechanism. When you ask AI Mode a question, it does not run one search. It decomposes the question into a set of sub-queries, runs them concurrently against the normal Google index, and synthesizes the results. Google’s own engineering leadership describes it as “doing a dozen searches for you in the time it takes to do one.” The same fan-out approach now underpins AI Overviews.

The consequence shows up clearly in the citation data. In July 2025, Ahrefs found that 76 percent of pages cited in AI Overviews ranked in Google’s top 10 for the query that triggered the overview. By January 2026, across 863,000 SERPs and four million AI Overview URLs, that figure had dropped to 38 percent. Of the rest, 31 percent ranked somewhere in positions 11 to 100 and 31 percent did not rank in the top 100 at all for the original query. They ranked for one of the sub-queries instead.

That is the single most important fact in this guide for anyone who already does SEO. Ranking for the head term still matters, but ranking for the specific sub-questions a model would generate while answering the head term matters at least as much, and those sub-questions are often long-tail phrases with tiny search volume that nobody bothered to target.

ChatGPT: web.run, third-party search and known-domain bias

ChatGPT’s browsing layer is an internal tool OpenAI calls web.run. Analysis of its behavior published by Search Engine Land in May 2026 found that it supports around a dozen operations (search, open, find, click, product queries and so on), that the newer models chain anywhere from two or three to more than ten rounds of search per answer, and that OpenAI does not maintain a general web index of its own. Search results arrive through third-party search APIs, and a separate fetcher, the ChatGPT-User agent, then opens the selected pages to read them.

Two behaviors matter for optimization. First, the model formulates its search queries partly from what it already knows: it tends to target sources it recognizes from training, and the current models use site-restricted queries to pull from domains they treat as trusted. 

Parametric knowledge shapes retrieval. A brand the model has never heard of is not searched for by name, so it has to win on the topical query instead. Second, the number of sources per answer has been falling. After the March 2026 model change, the average number of unique domains cited per response dropped from around 19 to around 15. Fewer slots, more competition for each.

Perplexity: closest to classic search

Perplexity runs its own crawler and index and behaves the most like a search engine of the group. In Ahrefs’ 15,000-query overlap study, 28.6 percent of Perplexity’s citations ranked in Google’s top 10 for the same query, against roughly 6 to 9 percent for ChatGPT, Gemini and Copilot. Perplexity also leans hard on community and review sources: in Profound’s data Reddit alone accounts for roughly 47 percent of its top-10 source share, with G2, Gartner, NerdWallet, PCMag, TripAdvisor and Yelp following. If you sell B2B software or services, your G2 and Reddit footprint is a Perplexity ranking factor in all but name.

Gemini and Copilot

Gemini draws on the Google index through the same fan-out logic as AI Mode. Copilot draws on Bing, and its citations overlap with Bing’s top 10 at a higher rate (14 to 17 percent) than any other assistant overlaps with any search engine. If your site is invisible in Bing, and many sites are because nobody checks, Copilot cannot cite you.

The overlap problem

Put those platforms together and a pattern emerges: they do not agree with each other. Only about 11 percent of domains cited by ChatGPT are also cited by Perplexity for the same set of prompts. Averaged across assistants, only 12 percent of AI-cited URLs rank in Google’s top 10 for the original query, and roughly 80 percent do not rank in Google’s top 100 for it. Being cited in one engine is weak evidence you will be cited in another. Your measurement, and your targets, have to be per platform.

What Actually Predicts A Citation

The largest correlation study to date is Ahrefs’ December 2025 analysis of 75,000 brands across ChatGPT, AI Mode, AI Overviews, Copilot, Gemini and Perplexity, using Spearman correlations between site metrics and AI visibility. The ranking of factors is not what most SEO playbooks would predict.

Read that top to bottom and the message is that models cite entities they recognize. The strongest predictors are all measures of how often, and how consistently, a brand is named across the web, on video platforms and in search behavior. The classic link metrics are near the bottom. A separate multi-dataset review published in May 2026 reached the same conclusion from a different direction, finding that branded search volume predicted citations materially better than backlinks, and that establishing a consistent entity footprint (Wikidata plus a handful of authoritative third-party profiles) was associated with a 2.8x higher citation likelihood.

Three more findings round out the picture.

Freshness. Across 17 million citations, Ahrefs found the average AI-cited page was about 1,064 days old versus 1,432 days for the average Google top-10 result, roughly 26 percent fresher. ChatGPT skews freshest (cited pages average around 960 days since publication), Google AI Overviews behave like organic search. The models are not chasing this week’s content, but they do prefer pages that have been updated within the past couple of years.

Concentration. Citations cluster heavily. Across the aggregate datasets, Reddit accounts for roughly 10 percent of all AI citations, Wikipedia around 7 percent and YouTube around 5 percent, and the top 20 domains take about half of all citations in many verticals. In Google’s ecosystem, a large share of AI Mode and AI Overview citations point at Google-owned properties, YouTube above all. You are competing for the remaining share, and in most B2B verticals the remaining share is still large.

Format. Retrieval systems pull passages, not pages. Content whose answer to a specific question is stated plainly in one extractable block, a short definition, a comparison table, a numbered list, a labeled FAQ, is easier to lift into an answer than the same information spread across four paragraphs of prose. This is the one place where the “structure for AI” advice is right, though the mechanism is mundane: the model is quoting the passage the retriever handed it.

What The Retriever Never Sees

The flip side of “retrieval first” is that anything the crawler cannot fetch does not exist for citation purposes. Before touching content, check the basics.

Crawler access. The crawlers are separate from the search engines’ normal bots and separate from each other. OpenAI runs GPTBot (training), OAI-SearchBot (search index) and ChatGPT-User (live page fetching during an answer). Google uses Googlebot for both search and AI features and Google-Extended only as a training opt-out. Perplexity runs PerplexityBot and a user-triggered fetcher. Anthropic runs ClaudeBot. Roughly a third of publishers now block at least one AI crawler, often by copying a robots.txt snippet that blocks all of them. Blocking GPTBot keeps your content out of training and touches nothing on the search side, but blocking OAI-SearchBot or ChatGPT-User removes you from ChatGPT answers entirely. Decide deliberately.

Rendering. The live fetchers are lightweight. Content that only appears after JavaScript execution, behind a consent wall, in a tab component or inside an accordion is at risk of being read as absent. Server-rendered HTML with the answer in the initial payload is the safe default.

Bing. Copilot depends on it, and parts of the third-party search supply feeding other assistants draw on it. Verify the site in Bing Webmaster Tools and check that your key pages are indexed there. It takes twenty minutes and most sites have never done it.

llms.txt. The proposed standard for a model-friendly site summary file has been widely implemented and, on the evidence, widely ignored. Google has said it does not use it. Ahrefs’ May 2026 crawl of 137,000 sites found 97 percent of llms.txt files received zero requests, and SE Ranking found no correlation between having one and being cited across roughly 300,000 domains. It does no harm. It also does nothing yet. Do not report it as an optimization.

A Working Model Of The Citation Decision

Combining the mechanics and the data, the decision path for a single answer looks roughly like this.

1. The user asks a question. The model decides whether it needs to search at all. Branded and simple factual questions often get answered from memory; comparisons, “best X for Y” and anything time-sensitive trigger retrieval.
2. The model writes several search queries, not one. For “best PEO for a 30-person startup” that might be “PEO pricing small business 2026,” “Justworks vs Rippling PEO,” “certified PEO list,” and “PEO minimum employees.” It formulates these partly from what it already knows, which is where prior brand awareness feeds in.
3. Each query goes to a search backend (Google’s index, Bing, Perplexity’s own, or a third-party API). The top results for each sub-query come back. This is where ordinary SEO for the sub-queries pays off, and why a page ranking eighth for a long-tail phrase can be cited while the page ranking first for the head term is not.
4. The model, or a reranker in front of it, opens a subset of those pages and extracts passages. Pages that load fast, render server-side and state their answer in a contained block survive this step. Pages that are slow, thin, paywalled, or bury the answer get dropped.
5. The model writes the answer from those passages and attaches citations to the claims it took from each. Passages that contain a specific, checkable fact (a price, a date, a number, a named entity) are more likely to be attributed than general framing.

Every lever in the next section maps to one of those five steps.

What To Do About It

1. Build the entity, not just the pages

Because recognition drives both the search formulation and the trust filter, the highest-leverage work is making the model certain who you are. Practically: a consistent legal and brand name used identically everywhere; a Wikidata entry if you qualify and accurate Wikipedia coverage if you exist there; complete, consistent profiles on the third-party sites your category’s models already cite (G2, Capterra, Clutch and Crunchbase for software and services; Yelp, Google Business Profile and industry directories for local); and a steady flow of third-party mentions in places models read. The Ahrefs data says unlinked mentions and YouTube mentions outrank links, so a podcast appearance, a conference talk on YouTube or an inclusion in someone else’s roundup does more for AI visibility than another guest-post backlink with optimized anchor text.

2. Target the fan-out queries, not only the head term

Take each commercial topic you care about and write down the ten to twenty questions a careful analyst would ask while researching it. Then check whether you have a page, or a section with its own heading, that answers each one directly. Most sites have one long page on the head term and nothing on the sub-questions. The sub-questions are where the 62 percent of non-top-10 AI Overview citations come from. Keyword tools will show most of them at 0 to 50 searches a month, which is why they are uncontested.

3. Put the answer where the retriever can lift it

For each of those questions, the first block under the heading should contain the answer in a form that survives being quoted alone: a two-sentence definition, a table, a numbered list, a figure with a date. Prose can follow. Keep facts specific and attributable, because the model attributes specific claims and paraphrases vague ones without credit. Date-stamp the page and keep the visible “last updated” honest; the freshness preference is real but modest, and fake dates are detectable.

4. Refresh the pages that already have a shot

Ahrefs’ freshness data puts the sweet spot around pages that have been updated within the last one to three years. A page from 2022 that still ranks is a citation candidate if you update it, and a lost cause if you do not. Refreshing existing ranking pages with current figures and current-year framing is cheaper than writing new ones and usually moves citations faster.

5. Get onto the pages that are already being cited

If Reddit, G2, Wikipedia, YouTube and a handful of trade publications take half the citations in your category, part of your LLM SEO is showing up inside those sources. That means real participation in the subreddit threads the models keep quoting, complete and reviewed listings on the comparison sites, and appearances in the published “top X” roundups that models pull for “best” queries. Curated lists are a citation magnet precisely because a model answering “best X” wants a source that has already done the comparison. This is the mechanism behind the Google AI Mode and Perplexity citations that REVERB’s own lists collect: as of late August 2026, Ahrefs recorded 681 AI responses citing reverbico.com across 154 pages, with Google AI Mode and AI Overviews accounting for most of them. Those numbers move monthly and are quoted here as an illustration of the pattern, not a benchmark.

6. Fix access before anything else

Audit robots.txt against the full list of AI crawlers and decide bot by bot. Confirm the key pages render without JavaScript. Verify Bing indexing. Check that your pages return quickly to unauthenticated requests. None of this is glamorous and all of it is a prerequisite.

7. Measure per platform, and measure mentions separately from citations

Because the engines disagree with each other, a single “AI visibility score” hides more than it shows. Track, per platform: the prompts you care about, whether your brand is mentioned, whether a page is cited, and which page. Ahrefs Brand Radar, Semrush’s AI toolkit and the specialist trackers (Profound, Peec, Otterly and others) all do some version of this. The AI SEO tools list covers the field. Expect citation click-through rates to be low, under 1 percent for AI Overviews and a few percent for Perplexity, and value the visibility on brand-recall terms rather than session counts.

What Does Not Work

A short list, because the space is full of advice that sounds plausible and has no support in the data.

• Stuffing “AI-friendly” phrasing. Writing “in summary” and “the answer is” before every paragraph does not change retrieval and makes the copy worse. Structure matters; incantations do not.
• llms.txt as a ranking lever. See above. Harmless, unused.
• Buying links at scale. Backlink count is at the bottom of the correlation table. Links still matter for the organic rankings that feed retrieval, but a link-building budget redirected into earning mentions on cited platforms will do more for AI visibility.
• Schema as a citation trigger. Structured data helps search engines understand a page and may help entity disambiguation. There is no evidence that FAQPage or HowTo markup on its own causes a language model to cite you, and Google has reduced the visibility of both schema types in normal results. Use schema for what it is for; do not expect it to buy citations.
• Blocking training crawlers as a “protection” measure without checking the search crawlers. Many sites have blocked themselves out of ChatGPT answers while intending only to opt out of training. Read the crawler list before you copy a robots.txt block.
• Treating a Google top-3 ranking as sufficient. It was, in mid-2025, for AI Overviews. It is not anymore, and it never was for ChatGPT.

A Note On Reputation

The same retrieval mechanics that surface your best page also surface your worst review thread. Because Reddit, review platforms and news sites are heavily cited, a negative thread that ranks for a fan-out query like “[brand] complaints” can be quoted in an answer to “is [brand] any good” with no editorial judgment applied. Monitoring what the models say about you, and doing the same entity and third-party work on the negative side, is now part of reputation management. The firms on the ChatGPT reputation management list specialize in exactly that.

When To Bring In An Agency

Most of the work above is ordinary marketing done with a different priority order: entity consistency, third-party presence, long-tail coverage, page structure and technical access. An in-house team that already does SEO can absorb it. Where an outside firm earns its fee is in the measurement layer (tracking prompts across six engines is tedious and the tools are still immature), in category-specific knowledge of which third-party sources the models actually cite, and in the digital PR work that generates the unlinked mentions the data rewards. The GEO agencies list and the AEO agencies list cover firms doing this work; ask any of them to show you per-platform citation data for a client before you sign.

Frequently Asked Questions

What is LLM SEO?

LLM SEO is the work of getting a website cited or mentioned in answers generated by large language model assistants such as ChatGPT, Google AI Mode and AI Overviews, Perplexity, Gemini and Copilot. It overlaps with traditional SEO because most assistants retrieve pages through a search index, but the factors that predict a citation are weighted differently, with brand recognition and third-party mentions ranking well above backlinks.

Is LLM SEO the same as GEO or AEO?

Generative engine optimization, answer engine optimization, AI search optimization and LLM SEO all describe the same goal. The labels differ by vendor more than by method.

How does ChatGPT decide which sources to cite?

ChatGPT runs one or more web searches through its web.run tool, using queries it formulates partly from what it already knows about the topic and the brands in it. Results come from third-party search APIs, selected pages are fetched by the ChatGPT-User agent, and the model attaches citations to the claims it draws from those pages. Sources it recognizes from training are favored when it builds its queries.

Do I need to rank on page one of Google to be cited in AI Overviews?

Not any more. As of January 2026, only 38 percent of AI Overview citations came from pages ranking in the top 10 for the triggering query. The rest ranked for related sub-queries generated by Google’s query fan-out, many of them long-tail phrases with very low search volume.

Do backlinks help with AI citations?

Only indirectly, through the organic rankings that feed retrieval. In Ahrefs’ 75,000-brand correlation study, raw backlink counts showed minimal correlation with AI visibility, while unlinked brand mentions, YouTube mentions and branded anchor text showed the strongest.

Does llms.txt help?

There is no evidence that it does. Google has said it does not use the file, and crawl-log studies covering hundreds of thousands of sites show AI crawlers almost never request it and no correlation between having one and being cited.

How fresh does content need to be?

AI assistants cite content that is on average about 26 percent fresher than Google’s organic top 10, with cited pages averaging roughly three years since publication. Updating existing pages within a one-to-three-year window is enough; daily publishing is not required.

How do I measure LLM SEO?

Track a fixed set of prompts per platform and record, for each, whether your brand is mentioned and whether a page is cited. Keep mentions and citations as separate metrics. Tools such as Ahrefs Brand Radar, Semrush’s AI toolkit and dedicated AI visibility trackers automate the collection.

The Bottom Line

Language models cite what their retrieval layer hands them, and their retrieval layer hands them pages that rank for the specific sub-questions inside a query, load cleanly, state answers in liftable blocks, and belong to entities the model already recognizes. That last condition does most of the work. If you want one sentence to plan a year of LLM SEO around, it is this: be named, consistently and often, in the places the models already read, and make sure that when they search for the questions behind your category, a page of yours is there with the answer in the first paragraph.