What is LLM Optimization?

A short, usable definition

LLM Optimization covers the methods that maximise three outcomes: the correct understanding of a site by a large language model, the citation of that site in responses generated by AI engines, and the faithful reuse of its content.

The term is not yet standardised. You will encounter it as GEO (Generative Engine Optimization), AEO (Answer Engine Optimization) or LLM SEO. These labels largely overlap. The terminological differences reflect communities, not substantive disagreements about what the work consists of.

The engines involved in 2026

According to the Goodie Wave 2 study (published 21 May 2026, 41 B2B brands, March-April 2026 data, 25.77 billion SimilarWeb visits), the distribution of AI referrals across engines is:

• ChatGPT Search: 62.6% of AI referrals (down from 89.1% in May-August 2025)
• Claude: 18.5% (up from 1.4% in the same period)
• Gemini: 10.6%
• Perplexity: 7.3%
• Microsoft Copilot: approximately 4%

Google AI Overviews is absent from AI referral tracking because its traffic appears under organic search in most analytics tools. In terms of raw impression count for informational queries, it is the highest-volume surface.

ChatGPT's dominance is real but declining rapidly: its AI referral share dropped from 89% to 63% in eight months. Claude's jump from 1.4% to 18.5% is the most significant shift. Optimising for a single engine in 2026 would be a coverage error.

How AI engines actually select sources

Most AI answer engines that incorporate real-time web content operate on a RAG (Retrieval Augmented Generation) pipeline with four stages:

1. Crawl and indexing: the engine's bot crawls web pages. ChatGPT Search uses OAI-SearchBot, Perplexity uses PerplexityBot, Google AI Overviews uses Googlebot and Google-Extended, Claude uses ClaudeBot. If a bot is blocked in your robots.txt, the page is not eligible for citation.
2. Chunking and vectorisation: extracted text is split into passages (typically 200 to 500 tokens), converted to numerical vectors and stored in a vector database.
3. Retrieval: when a query arrives, the engine retrieves the passages with the highest semantic similarity to the query.
4. Generation: the retrieved passages are passed as context to the language model, which synthesises a response, sometimes with citations, sometimes without.

The critical implication: it is the passage, not the page, that gets selected. A 2,000-word article may be split into 8 to 12 chunks. Only the chunk most relevant to a given query will be retrieved. A paragraph that starts with "as we saw above" or that depends on context from a previous section is unusable by the retrieval system and will not be cited.

For conversational models without real-time search (ChatGPT without search enabled, Claude base), the source selection mechanism is different: the model cites from its training corpus. The exact criteria are opaque, but domain authority, source prevalence in the training data and content quality all appear to matter.

Which surfaces, which queries

LLM Optimization does not apply uniformly across all query types. The impact is strongest on informational, comparative and definitional queries.

Transactional and navigational queries are less affected: someone searching to buy a product or navigate to a brand will still click through. The discipline is most critical for organisations producing informational content: publishers, B2B SaaS companies, professional services, educational platforms.

GEO, AEO, LLM SEO: the terminology untangled

For a full comparative breakdown including measurement metrics, see SEO vs GEO vs AEO. For the academic foundation behind GEO specifically, see GEO: Generative Engine Optimization.

The main levers

The reference study (Aggarwal et al., Princeton, KDD 2024) tested nine editorial interventions across 10,000 queries. The most impactful on citation visibility (Position-Adjusted Word Count metric):

• Adding quotations and expert attribution: +41%
• Adding dated, sourced statistics: +31%
• Citing sources explicitly: +28%
• Fluency and readability improvement: +17%
• Keyword stuffing: -8 to -10% (the only negative intervention tested)

These findings translate into six operational levers:

1. Passage self-containment: each paragraph must be understandable without surrounding context. For full detail, see Content structure for AI engines.
2. Entities and disambiguation: name things precisely, repeat your entity with co-occurring domain markers, provide schema.org structured data (Organization, Article, FAQPage).
3. Topical authority: cover a subject in depth, build a coherent internal link structure, accumulate relevant external links over time.
4. Citations and sourcing: cite studies with authors, dates and sources; use specific, dated figures rather than vague claims.
5. Technical operations: crawl, rendering, AI bot allowance in robots.txt, schema.org JSON-LD, llms.txt. See Technical optimisation for AI visibility.
6. Freshness and maintenance: explicitly date pages, update them when facts change, favour depth over volume.

These levers are expanded into a step-by-step implementation on the method page and organised into a four-dimension framework on the LOOP page.

What LLM optimisation is not

• Not a prompt hack: hiding instructions or injecting text into page content does not work reliably and violates engine terms of service.
• Not a ranking promise: no method guarantees a citation. AI engines are probabilistic and non-deterministic; the same query produces different answers across sessions.
• Not a SEO replacement: it is an extension. Without correct crawling and a baseline domain authority, no page will be cited by any generative engine.
• Not just an llms.txt file: that file is a declared-intent signal that solves access, not authority. See The complete llms.txt guide.
• Not a black-box algorithm to reverse-engineer: the systems are opaque and under continuous development. Observable behaviour is the only reliable basis for practice.

Real limits of the discipline in 2026

A serious account of the subject requires surfacing what remains partially or fully out of your control:

• Citation measurement is a proxy: answers vary by user, session, model version and query phrasing. No available metric provides exhaustive coverage.
• Training corpus opacity: when a page published today becomes eligible for citation in a conversational model training corpus is not published. Estimates range from 3 to 12 months depending on the model and update cycle.
• Retrieval ranking is not documented: the internal ranking mechanisms of RAG engines (Perplexity, ChatGPT Search) are proprietary and unpublished.
• Corpus biases favour established sources: Wikipedia, major media and high-authority publishers have structural advantages that take time to build against.
• Engine plurality is increasing: each engine has its own bot, its own retrieval criteria and its own user base. Optimising for one does not guarantee optimisation for all.

Business stakes

For a brand or publisher, being cited in AI engine responses produces three distinct effects:

1. Asymmetric visibility: present in an answer at the moment a prospect is searching, without paying per click.
2. Proxy authority: being cited by an AI engine acts as a credibility signal to users, particularly in B2B contexts where sourcing matters.
3. Long acquisition tail: a single well-structured, well-sourced page can be cited across many related queries for months.

According to the Seer Interactive study (53 brands, 2.43 billion impressions, January 2025-February 2026), pages cited in a Google AI Overview achieve an average click-through rate of 2.1%, versus 0.9% for pages on the same SERP but not cited in the AIO. The difference is +133%. Being present in the generated answer, not just on the page below it, determines whether organic traffic is captured or lost.

For a step-by-step implementation plan, see How to optimise your site for AI and the 40-point audit checklist.

Frequently asked questions

Is LLM Optimization the same as GEO?

In practice, yes. GEO (Generative Engine Optimization), AEO (Answer Engine Optimization) and LLM Optimization refer to the same set of editorial, technical and semantic practices. The differences are terminology conventions: GEO originates from academic literature (Aggarwal et al., Princeton/IIT Delhi, KDD 2024), AEO predates LLMs (featured snippets, voice assistants), and LLM Optimization is the technical community preferred umbrella. Choosing one is a signal of school, not a substantive difference in the underlying work.

Does LLM Optimization replace classic SEO?

No. LLM Optimization is built on top of SEO, not in its place. A page that Google cannot crawl will not appear in ChatGPT Search or Perplexity either. The disciplines share foundational elements (crawlability, domain authority, content quality) but diverge at second-level optimisations. A site well-optimised for SEO is almost always better positioned for LLM citation. The reverse is not guaranteed.

Which AI engines does LLM Optimization apply to?

In 2026, the main AI answer engines are: ChatGPT Search (62.6% of AI referrals, Goodie Wave 2, May 2026), Claude (18.5%), Gemini (10.6%), Perplexity (7.3%) and Microsoft Copilot (around 4%). Google AI Overviews (deployed broadly in France from June 2025) also intercepts informational queries. Each engine has its own retrieval stack, but the foundational optimisation principles apply across all of them.

How long does it take to appear in AI engine responses?

It varies by engine type. For search-augmented engines (ChatGPT Search, Perplexity) that crawl the live web, a well-structured page can become eligible within days to weeks of publication, provided the AI crawler has indexed it. For conversational models relying on training data (ChatGPT without search, Claude base), the delay can be 3 to 12 months, depending on the next training cycle. Google AI Overviews follows standard Google indexing timelines.

Can any site appear in ChatGPT or Perplexity?

Any publicly accessible page, correctly crawled by AI bots, can in principle be cited. In practice, three factors create a barrier: domain authority (low-authority sites rarely outcompete higher-authority rivals on the same topic), content quality (a short or vague page will not displace a well-sourced, structured one), and passage relevance (the right paragraph must be self-contained enough to answer the exact query). LLM Optimization work shifts those factors in your favour, but does not guarantee a citation.