What is Generative Engine Optimisation?GEO
Generative Engine Optimisation (GEO) is the practice of building brand presence inside the training corpus of large language models so that ChatGPT, Claude, and base Gemini recommend a brand even when no live web search is performed.
GEO is distinct from AEO. AEO targets engines with live retrieval (Perplexity, ChatGPT Search, Google AI Overviews); GEO targets engines responding from training data alone (ChatGPT default mode, Claude, base Gemini). You cannot fake GEO from your own site. You earn it by being mentioned across the surfaces the model was trained on: Wikipedia, Wikidata, Reddit, GitHub, top-tier industry publications, and Common Crawl-indexed content. The compounding loop runs on a 12–24 month cycle aligned with model retraining cadences. Brands starting today are 18 months ahead of those starting next year.
What it includes
- Wikipedia and Wikidata entries (where notability allows)
- Mentions in established industry publications
- Active presence in topical subreddits and forums
- GitHub repositories or contributions for technical brands
- Citations in academic or research material
- High-quality on-site content that gets crawled into Common Crawl
How it works
Establish entity surface
Create or claim Wikipedia, Wikidata, LinkedIn, and Crunchbase entries. These are the primary entity sources LLMs ingest.
Earn third-party citations
PR campaigns, podcast appearances, guest posts, and original research that gets republished by industry sites.
Build category-defining content
Long-form definitional pages on your domain that get crawled, embedded, and surface-citied.
Track training-data inclusion
Test queries in default ChatGPT (no web search) every quarter. Brand mentions there mean you are inside the corpus.
Frequently asked
Can I optimise GEO from my own site only?
No. On-site signals affect AEO (live retrieval) but not GEO directly. GEO requires earning mentions on third-party domains the model was trained on.
How long until GEO compounds?
Model retraining cadences are 12–24 months. Brands published today appear in next-cycle training runs. Compounding is real but slow.
Is GEO measurable?
Indirectly. Run quarterly query sets in default-mode LLMs (no live search). Brand mentions there are the GEO signal.