Generative Engine Optimization: Being Selected as a Source
Updated 12 October 2026
Generative Engine Optimization (GEO) improves the conditions for AI-powered search and answer systems to find a company and interpret it accurately as a relevant source. We assess crawl access, page structure, content and source references. We then develop clear answers to real user questions and connect them to verifiable facts. We measure regularly whether this content appears in AI answers; a specific mention or position cannot be guaranteed.
The difference from SEO
GEO is not the same as SEO. SEO aims at placement in result lists, GEO at being selected as a source within a generated answer. Both share the same technical foundation: what isn’t crawled and indexed can neither rank nor be cited. SEO therefore remains a precondition, but it doesn’t answer why an AI system draws on a particular source.
This selection is fed from two directions. What is anchored in model training can barely be influenced in the short term. What is retrieved live can be influenced — this is where it’s decided whether a source ends up in the answer. GEO therefore starts with retrieval first: crawlable content, unambiguous statements, verifiable facts.
The selection logic differs by system. Perplexity discloses its source selection, Google relies heavily on classic structure and crawling signals, ChatGPT and Claude work with their own crawlers and their own preferences. That’s why work and measurement happen per system, not averaged across all systems.
Content that gets picked up
What gets picked up is what is self-contained and understandable on its own. AI systems process text in chunks of roughly 100 to 500 tokens — an answer block of 50 to 300 words under a clear heading matches this unit and is cited in full most often. The answer comes in the first sentence, context and evidence follow.
Four formats are drawn on disproportionately often. We align content with the prompt set so it hits the questions that are actually being asked.
Definitions
for terminology questions.
Studies & data
as verifiable facts.
How-to content
as actionable guidance.
FAQs
with real questions and concise answers.
This also includes terminology: the same terms for the same things, across the website, press, and third-party sources. Inconsistent naming dilutes the entity; consistency strengthens the associations that language models learn.
Technical prerequisites
What can’t be read can’t be cited.
Access. AI crawlers are their own bots — GPTBot and OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended. Each can be allowed or blocked via robots.txt. Whether to allow them has no standard answer: blocking protects content but also reduces the likelihood of being cited. Anyone who wants visibility in AI answers opens at least the search and retrieval bots.
Delivery. Many AI crawlers don’t execute JavaScript. Content that only appears in the browser or sits behind a click or scroll doesn’t exist for them. Server-side rendered HTML is the precondition.
Markup. Structured data per schema.org, delivered as JSON-LD, makes relationships between entities explicit. An llms.txt adds a curated overview on top — it isn’t mandatory, and the standard is not yet widely established. The effort is low, the risk is zero, but it doesn’t replace clean structure.
How impact is measured
Impact is measured against the prompt set, not against rankings. Each question is regularly put to the relevant AI systems and the answer is evaluated: does the company appear, is it described correctly, is it recommended, is it linked, is it cited. Because generative answers are not deterministic, the work is done through repetition and time series. The reference point is your own baseline and the gap to competitors within the same prompt set.
Which measures actually influence selection as a source is the subject of ongoing research. We stay in direct professional exchange with the research that studies this field empirically, and we regularly align our assumptions with the state of that work. What cannot be substantiated we treat as a hypothesis, not a rule.
Further reading:
