LLM SEO: what changes from classic SEO, where citations come from, and what does not work

What is LLM SEO, and is it different from SEO?

LLM SEO is classic SEO plus three extra jobs: being retrievable for queries you never see, being worth quoting once retrieved, and being mentioned on the sites the models already trust. The foundations do not change. Google says there are "no additional requirements" to appear in AI Overviews or AI Mode, and no special schema to add^[5]. If a page cannot be crawled, indexed and shown with a snippet, nothing below will rescue it.

What changes is the distance between a ranking and a citation. Ahrefs ran 15,000 long-tail queries through Google, Bing and four AI assistants in mid-2025 and found that on average only 11.9% of assistant citations also ranked in Google's top 10 for the same prompt. For ChatGPT's in-text citations the figure was 8.0%^[2]. Our piece on how GEO differs from SEO covers the strategic split; this one is the working checklist. People also search for this as LLM optimization, SEO for LLMs or ChatGPT SEO. They describe the same work.

What actually changes compared with classic SEO?

The query, the unit of competition and the measurement all change. The clearest evidence so far comes from a preprint published on 11 October 2026 by researchers at the National University of Singapore and Xi'an Jiaotong University, who ran 80 questions 20 times each through two production ChatGPT models^[1]. Three findings matter for practitioners:

• The model rewrites your customer's question. An exact match to the user's query appeared in 5.5% of GPT-5.2 Instant runs and in none of the GPT-5.3 Instant or GPT-5.5 Instant runs^[1]. You are ranking for the model's sub-queries, not the prompt.
• Google rankings are a weak proxy. Google's organic results alone contained only 9.4% and 6.4% of the exact URLs ChatGPT cited, depending on the model^[1].
• The candidate pool is unstable. 40.8% and 44.5% of the URLs ChatGPT retrieved as candidates appeared in only one of 20 runs^[1]. A single spot-check tells you almost nothing.

The table below is how we summarise the shift, with an honest label on each row.

Where do LLMs find the sources they cite?

From a live search index for anything current, and the sources they prefer are often not yours. That is the citation surface: every page a model might retrieve when your category comes up, including forums, encyclopaedias, review sites and the press.

Ahrefs' Brand Radar data for ChatGPT in the US, September 2026, puts reddit.com first with 16.8% of citations among the top cited domains, followed by en.wikipedia.org at 7.0%, consumerreports.org at 3.7% and forbes.com at 3.1%^[3]. Read those shares carefully: they are percentages of the citations going to the top sources, not of all ChatGPT citations^[3].

The practical consequence is that LLM SEO has an off-site half. If the threads, list articles and comparison pages a model retrieves for your category never mention you, a perfect page on your own domain still loses. We map that surface before writing anything: run the category prompts, log every cited URL, and sort them into ones you can influence (your site, partner pages, honest community participation) and ones you cannot.

Read more: what about ChatGPT's training data?

OpenAI separates its crawlers. GPTBot collects data to improve its models, while OAI-SearchBot is used to surface websites in ChatGPT's search features, and sites that opt out of OAI-SearchBot "will not be shown in ChatGPT search answers", though they can still appear as navigational links^[6]. Blocking GPTBot is a training decision; blocking OAI-SearchBot is a visibility decision. A firewall or bot-protection rule can block a crawler that robots.txt allows, so check your server logs too.

What content structure gets extracted and cited?

Pages that answer one specific question per section, in the first sentence, with a number and a named source, are the easiest for a model to lift. The strongest controlled evidence is still the GEO paper by Aggarwal and colleagues, accepted at KDD 2024. On a 10,000-query benchmark, adding quotations raised a source's position-adjusted word count in the answer from a baseline of 19.5 to 27.8, and adding statistics raised it to 25.9^[4]. The authors report gains of up to 40% in visibility overall^[4].

Two caveats. The test engine was the authors' own setup built on gpt-3.5-turbo, not today's ChatGPT, and the work predates the query rewriting described above. Treat it as directional. With that in mind, this is the structure we use:

1. Question-shaped headings. Phrase each h2 as a question someone would ask, because the model searches subtopics, and a heading that matches one is easier to retrieve. That is our inference, not a tested result.
2. Answer first. The first sentence under each heading should stand alone as a quotable answer.
3. Numbers with a year and a source. Statistics and citations scored well in the GEO tests^[4], and an undated number is useless to a model answering a current question.
4. Tables for comparisons. Comparisons are a common follow-up, and a table puts the answer in one place.
5. Cover the fan-out. Google says AI Overviews and AI Mode may issue "multiple related searches across subtopics"^[5]. Answer the next three questions a buyer would ask on the same page, or on pages linked from it.

The October 2026 preprint adds a modest signal: cited candidates tended to have longer retrieval snippets, and higher retrieval position was the clearest citation signal among page attributes. Recency was weak, with confidence intervals that included no effect^[1]. Updating a date on an unchanged page is not a strategy.

What does not work for LLM SEO?

Keyword stuffing, special AI files and one-off checks have no evidence behind them, and some evidence against. In the GEO tests, keyword stuffing lowered position-adjusted word count to 17.8 against the 19.5 baseline, and the authors write that such methods "offer little to no improvement"^[4].

• llms.txt and other AI-only files, for Google. Google states you don't need to create "AI text files, or markup" to appear in its AI features^[5]. We have not found a public study showing an llms.txt file increases citations in any engine as of October 2026. It is cheap, so add one if you like, but do not report it as a win.
• Special schema for AI. Google says there is no special structured data required^[5]. Normal schema still helps ordinary search features.
• Writing for the exact prompt. If the model never searches the user's wording^[1], a page built around one long prompt string targets a query that is not run.
• Judging success from one chat. With 40% or more of candidate URLs appearing in only one of 20 runs^[1], a single screenshot is noise. Our guide to measuring AI visibility explains sampling.
• Blocking AI search crawlers by default. A blanket bot block removes you from ChatGPT search answers^[6]. Decide crawler by crawler.

How should you start LLM SEO this quarter?

Fix access, map the citation surface, rewrite your highest-value pages for extraction, then measure with repeated runs. In order:

1. Audit access. Check robots.txt and firewall rules for OAI-SearchBot and the other AI search crawlers, and confirm nothing important carries nosnippet or noindex^[5].
2. Write 30 to 50 category prompts a buyer would ask, and run each several times across the engines that matter to you.
3. Log every cited URL and sort the list into your pages, competitor pages and third-party pages.
4. Rewrite five pages using the structure above, and track them against five you leave alone. That is not a controlled trial, but it is better than a before and after with no comparison.
5. Work the third-party half: earn honest mentions in the threads, reviews and lists the models already cite.

If you would rather see the starting position first, our free AI visibility audit shows where you are cited today and which sources you are missing from.

Frequently asked

Sources

1. Jiang, Zhang, Xing, Chang and Xiao (National University of Singapore and Xi'an Jiaotong University), "Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations", arXiv preprint, 11 October 2026. 80 queries, 20 runs each, 1,600 runs per model for GPT-5.3 Instant and GPT-5.5 Instant. Not peer reviewed.
2. Ahrefs, "Only 12% of AI Cited URLs Rank in Google's Top 10 for the Original Prompt", Louise Linehan with Xibeijia Guan, 11 August 2025. 15,000 long-tail queries across Google, Bing, ChatGPT, Gemini, Copilot and Perplexity.
3. Ahrefs, "The 50 Most-Cited Websites in ChatGPT", Si Quan Ong, updated 2 September 2026. Brand Radar data, broad US queries, September 2026. Shares are of citations to the top sources only.
4. Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, "GEO: Generative Engine Optimization", KDD 2024, arXiv v3 28 June 2024. GEO-bench, 10,000 queries; results from the 1,000-query test split on a gpt-3.5-turbo engine.
5. Google Search Central, "AI features and your website", last updated 10 December 2025, checked October 2026.
6. OpenAI, crawler documentation (OAI-SearchBot, GPTBot, ChatGPT-User), OpenAI developer documentation, checked October 2026.