LLM SEO: How AI Search Engines Choose Their Sources

LLM SEO is the practice of making your content easy for AI search tools built on large language models (LLMs) to find, choose, use and cite. These tools follow a similar sequence: they turn a question into several searches, retrieve candidate pages, select a few, read them and write an answer with citations. They also draw on what the model already knows, including which brands it has seen often. Research so far shows that the retrieval step decides most of what gets cited, that AI tools lean toward independent third-party sources, and that a well-known brand name can change the answer even when the facts are the same. For an industrial company, LLM SEO means having pages that match the questions buyers ask, that AI crawlers can reach, and that other credible sites back up.

Two places an AI answer comes from

An LLM can answer from two kinds of knowledge. The first is what it learned during training, stored in the model's parameters. The second is information it retrieves at the moment of the question. Lewis and colleagues (2020) formalized the second approach as retrieval-augmented generation (RAG): a model that pulls passages from an external index and uses them to write its answer. They found that RAG models produced more specific and more factual language than models answering from memory alone.

AI search tools such as ChatGPT search, Perplexity, Gemini and Google's AI Overviews combine both. That gives you two jobs. Your pages need to be retrieved when the tool searches, and your company needs to be known well enough that the model mentions it at all.

How AI search engines choose their sources, step by step

Step 1: The question becomes several searches

Google says its AI Overviews and AI Mode may use "query fan-out," issuing multiple related searches across subtopics and data sources to build a response (Google Search Central, 2025). ChatGPT does something similar. In a study of 158,099 AI answers with 468,037 citations, Chen and Koudas (2026, preprint) found that ChatGPT rewrites the question into its own sub-queries for each answer, and that those sub-queries rarely matched what Google or Bing would return.

For you, this means a buyer's single question can trigger searches on materials, standards, failure modes and suppliers. Pages that answer each of those narrower questions have more chances to be found.

Step 2: The engine retrieves candidate pages

Retrieval is where most of the decision happens. In the same study, which pages ChatGPT's search retrieved explained about 32% of the variation in which pages were cited, against 2.8% for page properties such as FAQ blocks or page speed. Every page ChatGPT cited was one its own search had retrieved (Chen & Koudas, 2026, preprint).

Retrieval depends on access. OpenAI says sites that block its search crawler, OAI-SearchBot, will not be shown in ChatGPT search answers (OpenAI, n.d.). For Google, a page must be indexed and eligible to show a snippet to appear in AI features (Google Search Central, 2025), and websites that blocked Google's AI crawler were significantly less likely to be retrieved by AI Overviews (Grossman et al., 2026).

Step 3: The engine selects a few sources

From the candidates, the engine picks a handful to rely on. Several patterns show up in the research:

• Selection is not the same as ranking. Nearly 30% of the domains cited in Google's AI Overviews did not appear anywhere on the first page of normal results, and the cited domains were, on average, more credible than the results shown beside them (Xu, Iqbal & Montgomery, 2026).
• Different engines pick different sources. Google Search, AI Overviews and Gemini retrieved very different sources for the same queries, with an average overlap under 0.2 on a 0 to 1 scale. The generative engines were also more likely to retrieve Google-owned content (Grossman et al., 2026).
• Third-party sources are favoured. AI search showed a strong bias toward earned media, meaning independent, authoritative sources, over brand-owned and social content (Chen et al., 2025, preprint).
• Content that reads like the model's own writing may be favoured. One study found that Google's generative search tended to cite content that was more predictable for the underlying LLM and more similar in meaning to the other sources selected (Ma et al., 2025, preprint).

The idea that some sources carry more authority than others is old. Kleinberg (1999) showed that the link structure of the web could identify "authoritative" pages on a topic, through the hub pages that point to them. AI search seems to apply a similar test with a wider lens: who else, on credible sites, says you are relevant.

Step 4: The model reads and uses the content

Being selected does not guarantee being used. Language models handle long inputs unevenly. Liu and colleagues (2024) found that models performed best when the relevant information sat at the beginning or end of the input, and noticeably worse when it was buried in the middle. Answers that sit at the top of a page, and summaries that close a section, are easier for a model to pick up.

What the page contains also matters. Across ChatGPT, Google and Perplexity, the cited pages that shaped answers most tended to be longer, more structured, closely matched to the question and rich in definitions, numbers, comparisons and procedural steps (Zhang, He & Yao, 2026, preprint).

Step 5: The answer is written and cited, not always accurately

AI answers can misrepresent their sources. An audit of four generative search engines found that only 51.5% of generated sentences were fully supported by citations, and only 74.5% of citations supported the sentence they were attached to (Liu, Zhang & Liang, 2023). A 2026 study of Google's AI Overviews found 11.0% of claims unsupported by the cited pages, most often because something important was left out (Xu et al., 2026). Check what these tools say about your company and your products, and fix the source pages behind any errors.

Brand memory: the part you can't optimize on the page

Retrieval explains much of what gets cited, but the model's own knowledge shapes it too. Chen and Koudas (2026, preprint) found that ChatGPT wrote brands it already knew into its searches, preferred websites it knew and sometimes named brands with no supporting page at all. In controlled experiments, brand names alone reordered its rankings of products with identical facts. The pattern held in a market of small B2B vendors.

That is brand equity at work in a new channel: the difference a name makes to how the system responds. Industrial buyers already lean on strong brands and peer opinions when they research online (Krijestorac, Garg & Konana, 2021), and AI tools appear to do the same. Consistent mentions in trade media, association sites, standards work and customer stories build the kind of memory that pages alone can't.

An LLM SEO checklist for industrial websites

• Access: confirm Googlebot and OAI-SearchBot can crawl your important pages, and that firewall or CDN rules don't block them by accident.
• Text, not files: publish specifications, application notes and comparisons as HTML text. Keep PDFs as downloads, not the only copy.
• One question per page: match the page title and headings to the way buyers phrase the question.
• Answer first: state the answer in the opening paragraph, then the detail and the evidence.
• Extractable evidence: definitions, numbers with their source, comparison lists and numbered steps.
• Consistent names: the same company and product names everywhere they appear.
• Third-party proof: case studies, references and mentions on credible industry sites. See our guide to case studies and customer references.
• Monitoring: a fixed set of buyer questions, run monthly in the main AI tools, with the citations recorded.

For the wider strategy, including measurement and planning, see our guide to generative engine optimization for B2B companies. For the search foundations underneath it, see SEO for manufacturers.

Frequently asked questions

What is LLM SEO?

LLM SEO is optimizing your website and wider online presence so that AI tools built on large language models find, use and cite your content when they answer questions. It overlaps almost entirely with generative engine optimization (GEO) and answer engine optimization (AEO).

How does ChatGPT decide which websites to cite?

When ChatGPT searches the web, it rewrites the question into its own searches, retrieves pages and cites from what it retrieved. Research on 158,099 answers found that which pages its search retrieved explained far more of the citations than page features did, and that ChatGPT also favoured brands and websites it already knew (Chen & Koudas, 2026, preprint).

Do AI search engines just use the top Google results?

No. Nearly 30% of domains cited in Google's own AI Overviews were not on the first page of results (Xu et al., 2026), and different AI tools retrieve very different sources for the same query (Grossman et al., 2026).

Can AI search get facts about my company wrong?

Yes. Audits find a meaningful share of AI-generated statements are not supported by the sources cited (Liu, Zhang & Liang, 2023; Xu et al., 2026). Check regularly and correct the pages the errors come from.

Does page speed or schema markup decide AI citations?

Not on their own. In the largest study so far, FAQ blocks, structured data, summary boxes and page speed helped only when comparing different websites, not pages within one site, and page properties overall explained little compared with retrieval (Chen & Koudas, 2026, preprint).

The bottom line

Ask ChatGPT, Gemini and Perplexity the five questions your buyers ask most, note who gets cited, and build the pages and third-party mentions that would put your company in those answers. AXXEN sets up this kind of AI visibility tracking and content program through our AI & Growth Automation service.

About the author

Mo Farrokhi, PhD, is the founder of AXXEN in Calgary, Alberta. His path runs from electrical engineering through an MBA and a PhD in marketing to industry roles, including VP Marketing & Product at a TSX-listed clean-tech company. AXXEN builds marketing systems for technology, manufacturing and industrial companies that have outgrown ad-hoc marketing.

Sources

• Chen, M.-H., & Koudas, N. (2026). What drives AI citations? Content, retrieval and model knowledge in generative engine optimization (GEO) (arXiv:2610.15033) [Preprint]. https://doi.org/10.48550/arXiv.2610.15033
• Chen, M.-H., Wang, X.-X., Chen, K.-W., & Koudas, N. (2025). Generative engine optimization: How to dominate AI search (arXiv:2509.08919) [Preprint]. https://doi.org/10.48550/arXiv.2509.08919
• Google Search Central. (2025). AI features and your website (last updated December 10, 2025). developers.google.com
• Grossman, R., Liu, S., Chen, M. K., Smith, M., Borcea, C., & Chen, Y. (2026). How Generative AI Disrupts Search: An Empirical Study of Google Search, Gemini, and AI Overviews. Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval, 448–459. https://doi.org/10.1145/3805712.3809667
• Kleinberg, J. M. (1999). Authoritative sources in a hyperlinked environment. Journal of the ACM, 46(5), 604–632. https://doi.org/10.1145/324133.324140
• Krijestorac, H., Garg, R., & Konana, P. (2021). Decisions Under the Illusion of Objectivity: Digital Embeddedness and B2B Purchasing. Production and Operations Management, 30(7), 2232–2251. https://doi.org/10.1111/poms.13363
• Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33. https://doi.org/10.48550/arXiv.2005.11401
• Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024). Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12, 157–173. https://doi.org/10.1162/tacl_a_00638
• Liu, N., Zhang, T., & Liang, P. (2023). Evaluating Verifiability in Generative Search Engines. Findings of the Association for Computational Linguistics: EMNLP 2023, 7001–7025. https://doi.org/10.18653/v1/2023.findings-emnlp.467
• Ma, L., Qin, J., Xu, X.-C., & Tan, Y. (2025). When content is Goliath and algorithm is David: The style and semantic effects of generative search engine (arXiv:2509.14436) [Preprint]. https://doi.org/10.48550/arXiv.2509.14436
• OpenAI. (n.d.). Overview of OpenAI crawlers. Retrieved November 2, 2026, from developers.openai.com
• Xu, H., Iqbal, U., & Montgomery, J. M. (2026). Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact. Proceedings of the 2026 ACM Internet Measurement Conference, 529–546. https://doi.org/10.1145/3777912.3839818
• Zhang, K., He, X., & Yao, J. (2026). From citation selection to citation absorption: A measurement framework for generative engine optimization across AI search platforms (arXiv:2604.25707) [Preprint]. https://doi.org/10.48550/arXiv.2604.25707