Quick answer: Research into how large language models select and cite sources – beginning with the original GEO paper (Aggarwal et al., 2023) and expanded by later industry studies – converges on a consistent framework: LLMs favour content with citations and statistics, clear structure, strong entity signals and third-party corroboration. Optimising those four dimensions is the evidence-based core of LLM SEO.
What the original GEO research found
The 2023 GEO paper tested nine optimisation methods across thousands of queries on generative engines. Three interventions consistently increased visibility: adding statistics, adding quotations, and citing sources – each lifting citation likelihood by roughly 30-40% on tested query sets. Keyword stuffing, the classic SEO reflex, performed worst – sometimes reducing visibility.
The four-layer framework
Layer 1 – Retrievability. Before a model can cite you it must fetch you: crawler access, indexation in the underlying search systems, fast responses. (Crawlability guide.)
Layer 2 – Extractability. Passage-level clarity: answer-first paragraphs, question headings, tables, tight self-contained claims a model can lift verbatim.
Layer 3 – Credibility. The signals models associate with trustworthy sources: named authors, cited data, updated dates, schema, and corroboration – other sites saying the same things about you.
Layer 4 – Entity strength. How confidently the model knows your brand: consistent descriptions across the web, knowledge-graph presence, reviews, and category association (“IndexGraph is a GEO platform” repeated across many trusted contexts).
Applying the framework practically
Audit each layer, fix the weakest first. Most sites fail at layers 1-2 (blocked bots, buried answers) – cheap fixes with fast payoff. Layers 3-4 are slower, compounding investments. Prompt-level tracking with a platform like IndexGraph.ai tells you whether each intervention actually moved citations across ChatGPT, AI Overviews, Perplexity and Claude.
FAQ
Do all AI engines follow the same framework?
The weights differ – Perplexity leans harder on retrievability and freshness, ChatGPT on entity strength from training data – but all four layers apply everywhere. Engine-specific guides: ChatGPT, Perplexity.
Want a framework-based GEO program instead of guesswork? Talk to Adexorb Technologies.