Generative Engine Optimization (GEO) is the practice of getting your brand cited in the answers AI assistants give to user questions. It's adjacent to SEO but operates on a different ranking signal: not links, but consensus. Here's the playbook for AI startups specifically — the cohort with the highest leverage and the worst hygiene.
What GEO actually optimizes for
Classic SEO optimizes for a click. GEO optimizes for a sentence in an LLM response. The unit isn't a ranking position — it's a citation footnote. You win GEO by being the source the model trusts when summarizing your category.
The four GEO surfaces
1. Your own site — has to be machine-parseable (semantic HTML, JSON-LD, llms.txt). Without this, even high-authority links won't translate to citations. 2. Developer communities — GitHub README files, MCP server registries, Hacker News threads, dev.to posts. LLM training corpora pull heavily from these. 3. Independent directories — There's an AI for That, Futurepedia, AI tool roundups. These show up disproportionately in retrieval-augmented answers. 4. First-party docs — Open-source repos, API references, public changelogs. These are dense, recently-updated, and trusted by name.
The AI-startup advantage
Your customers are already technical. Your product naturally produces docs, SDKs, and CLIs. Every npm package, every GitHub stars page, every MCP server you publish becomes training data and retrieval context. The work isn't unnatural — it's productizing the artifacts you already generate.
The AI-startup mistakes
- A homepage that's pure React shell with no SSR — the bot sees nothing
- A "Product" page hidden behind a waitlist with no text describing what it does
- Pricing locked in a "Contact us" form — no model will ever cite a price it can't read
- A blog with one launch post from 18 months ago — no signal of recency
The 90-day GEO sprint
Week 1-2: ship semantic HTML, JSON-LD, llms.txt, and a sitemap. Audit with our checker at /check.
Week 3-4: publish 3 long-form posts targeting "what is X" queries in your category. Plain text, real examples, opinionated takes.
Week 5-8: open-source one tool — a TypeScript SDK, an MCP server, an eval harness. Document it like a product, not a side project.
Week 9-12: submit to the 5 directories your category is searched on. Pitch one podcast and one dev newsletter.
Then measure. Search your category in ChatGPT, Perplexity, and Claude monthly. The first time your brand appears unbidden, you've won a citation. Compound from there.
What this is not
GEO is not prompt injection, not cloaking, not stuffing llms.txt with marketing copy. Models are getting better at filtering manipulation and worse at tolerating it. The boring answer — structured, factual, well-attributed content — is the only one that compounds.