Generative Engine Optimization (GEO): how AI chooses the sources it cites
Generative engine optimization, or GEO, refers to the work of optimizing to appear in the answers generated by AI engines: ChatGPT, Perplexity, Gemini, Google’s AI Overview and AI Mode. Classic SEO aims to rank a page in a list of links. GEO aims for a source to be read, retained and cited in a synthesized answer. This article explains how these engines select their sources in 2026, and what you can do to be one of them.
What GEO changes compared to SEO
A classic engine returns a list of links and lets the user choose. An AI engine sorts for them: it reads several sources, synthesizes an answer from them, and cites only a handful. Visibility then depends on being selected as a source in the synthesis. A link’s rank isn’t enough: a page ranked well in Google can be absent from AI answers, and a page that is modest in SEO terms can be cited there.
How an AI selects its sources
We don’t know the exact weightings, and they vary from one engine to another. The factors that keep coming up in 2026 are fairly stable:
- The underlying index. Most AI engines rely on an existing index: ChatGPT and Perplexity go largely through Bing, AI Overview and AI Mode through Google. Being properly indexed remains a prerequisite.
- Semantic relevance. The engine looks for the passage that answers the intent, rather than the page that contains the right keyword. Content that tackles the question head-on is more easily extracted.
- Structure. Explicit headings, direct answers, lists, tables, short self-contained paragraphs. A model more easily extracts a passage that stands on its own.
- Factual data and entities. Dated, sourced figures, proper nouns, clear definitions. Engines favor what is verifiable and corroborated elsewhere.
- Freshness. On fast-moving topics, the date matters.
- Author authority. An identified author entity, with real expertise and external signals, weighs on the trust given to the source.
- Structured data. Schema (Article, FAQ, Person) helps make sense of the nature of the content.
What we actually do
Optimizing for GEO comes down to producing content an AI can read, understand and cite effortlessly:
- Answer the precise question at the start of a section, then expand.
- Give dated, sourced facts rather than general claims.
- Structure with clear headings, lists and tables.
- Name the entities and add the matching structured data.
- Take care of the author entity: who writes, their expertise, their external links.
- Check indexing on Google and Bing.
- Measure which queries cite the site in AI answers, and which ignore it when it should be there.
Common mistakes
- Keyword stuffing, which degrades readability for the model instead of helping it.
- Vague content, with no figure or source: nothing to extract, nothing to corroborate.
- Lack of structure: a long block of text splits badly into a citable passage.
- A neglected author: an anonymous source inspires less trust than an identified expertise.
- Content published then abandoned: on moving topics, a dated page drops out of answers.
Where GEO stands in 2026
The share of traffic coming from AI is rising fast: on the US retail side, Adobe measures triple-digit year-on-year growth, and agents are starting to finalize purchases through protocols like UCP (see my analysis of the UCP protocol). The field is still young: engines change their answers often, measurement is imperfect, and no one masters the weightings. This immaturity opens a window, because clean, structured, sourced content takes its place there more easily than a keyword saturated in classic SEO.
An often-forgotten prerequisite: a technically clean site. An AI, like a crawler, reads a site riddled with technical errors, redirects and missing tags poorly. That’s what SEO Cartograph is there to spot.