The Princeton KDD Lab published the first peer-reviewed study on generative engine optimisation in 2024. Since then, a body of controlled research has accumulated across nine independent studies. Some tactics are confirmed. Many are not. This is the evidence base for what actually works, scored by source quality.
What generative engine optimisation means
Generative engine optimisation (GEO) is the practice of improving content so that AI systems — ChatGPT, Perplexity, Google AI Overviews, and Gemini — select it as a cited source in their generated responses. The term was coined and defined in a peer-reviewed study from Princeton's Knowledge Discovery and Data Mining lab, published in 2024, which tested nine optimisation strategies across 9,679 queries from a 10,000-query dataset.
GEO is related to AEO (Answer Engine Optimisation), which predates the generative AI wave and covers optimising for featured snippets, voice search, and structured answer formats. GEO is the AI-specific evolution: optimising for retrieval-augmented generation (RAG) systems that synthesise answers from multiple sources and attribute citations.
What the Princeton GEO study confirmed
The Princeton study tested nine strategies: adding statistics, adding citations, adding quotations, adding fluency improvements, adding authoritative tone, adding keyword stuffing, adding easy-to-understand language, adding technical terms, and a combined approach.
The result: citing relevant statistics increased visibility by 40% on average across three AI systems. Adding quotations from experts increased visibility by 20%. All strategies improved citation rates except keyword stuffing, which showed neutral to slightly negative outcomes.
The study confirmed GEO as a valid optimisation practice and established that AI systems respond to content quality signals in measurable ways. It also confirmed that keyword-first thinking — the foundation of traditional SEO — does not transfer to generative AI.
The five most evidence-backed GEO tactics
1. Author credentials (Tier 1 by measured impact)
Presence AI ran a 90-day controlled study tracking 1,200 pages and 3,600 queries across ChatGPT, Perplexity, Google AI Overviews, and Gemini. The finding: pages with expert authors and documented credentials achieved a 72% AI citation rate. Pages with no author attribution achieved 25%. The 2.4x difference makes author attribution the single highest-impact actionable GEO tactic in the database.
Implementation: every published page needs a visible author byline, a credentials statement (specific expertise, not vague), a link to an author page with documented background, and — for Google AI Overviews specifically — Person schema in the article markup.
2. Branded web mentions (Tier 1 by predictive correlation)
Ahrefs analysed 75,000 brands and found that branded web mentions — references to a brand by name across third-party publications, with or without a hyperlink — have a Spearman correlation of 0.664 with AI Overview citation rates. Backlinks correlated at 0.218. That is a 3x difference in predictive strength between the dominant Google signal and the dominant GEO signal.
Implementation: digital PR, expert interviews, being cited as a source in industry publications, brand mentions in aggregator and round-up articles. The goal is text mentions across diverse sources, not link acquisition.
3. Content structure: headings and extractability
AirOps analysis of 815,484 AI-retrieved pages found that 7 to 20 subheadings with 500 to 2,000 words produced the highest consistent per-query citation rate in ChatGPT. Over-structured content (too many headings) and under-structured content (prose without internal navigation) both underperformed.
Microsoft's Bing documentation corroborates this: "strong descriptive headings are signals that help AI know where a complete idea starts and ends." Each H2 section should be independently extractable — it should make sense as a standalone answer to the sub-question it addresses.
4. Specific, sourced claims with statistics
The Princeton study found that adding statistics to content increased AI visibility by 40%. Growth Memo analysis of 21,482 ChatGPT citations found DATE and NUMBER are the two strongest positive entity signals in a page's first 1,000 characters. Specific, verifiable claims create extraction points that AI systems can cite confidently.
The implementation is explicit: cite the source of every statistic, include the study methodology and sample size where known, name the specific percentage or count rather than approximating. Content that reads like it could be cited in a research paper — specific, attributed, falsifiable — performs better than content that reads like marketing copy.
5. Content freshness
Perplexity weights content from the last 30 days most heavily. ChatGPT with browsing enabled shows recency bias: Amsive analysis found 50% of AI-cited content is under 13 weeks old. Content updated with new data or revised findings outperforms equivalent content on the same topic with an older date.
GEO tactics with weak evidence
Three tactics are widely recommended for GEO but are poorly supported by the research:
| Tactic | Recommendation prevalence | Evidence quality | Verdict |
|---|---|---|---|
| FAQ schema markup | Very high | AccuraCast: 1.8% of cited pages | Not a meaningful AI citation driver |
| Meta description | High | Writesonic: 0/6 crawler readability | Not read by AI crawlers |
| Open Graph tags | High | Writesonic: 0/6 crawler readability | Not read by AI crawlers |
| JSON-LD structured data | Very high | Writesonic: 0/6 crawler readability | Indirect benefit via Google only |
| Word count maximisation | Moderate | Mixed (rate vs volume trade-off) | Context-dependent |
Source: AccuraCast (9,000 citation sources), Writesonic (62 elements, 6 crawlers), AirOps (n=815,484)
How GEO differs by AI platform
GEO is not uniform across AI systems. Platform architecture determines which signals reach the model.
Google AI Overviews builds on Google Search infrastructure. Traditional SEO signals — backlinks, domain authority, structured data — transfer more strongly here than to any other AI platform. Person schema and Article schema are relevant for Google AI Overviews in a way they are not for ChatGPT or Perplexity.
ChatGPT (GPT-4o with browsing) shows the weakest correlation with backlinks and the strongest preference for branded web mentions and author credentials. It also requires explicit GPTBot permission in robots.txt.
Perplexity weights recency most aggressively of the four major platforms. It also produces the highest source overlap with Google (15.2%) of the ChatGPT/Claude/Gemini group, suggesting Google ranking transfers to Perplexity better than to other platforms.
Gemini uses Google infrastructure and shows citation patterns closer to Google AI Overviews than to ChatGPT or Perplexity. Standard SEO and GEO signals both apply.
How to measure GEO performance
Google Search Console does not capture AI citation data from ChatGPT, Perplexity, or Gemini standalone. Measuring GEO performance requires platform-specific monitoring.
The available approaches: manual query testing (ask target questions directly in each AI platform and record citation sources), AI answer monitoring platforms (track which sources are cited in AI responses to a monitored keyword set), and share-of-voice tracking adapted for AI search. As of mid-2026, no single tool covers all four major platforms with equivalent depth.
The minimum viable GEO measurement: monthly manual audits of your top 20 target queries across ChatGPT, Perplexity, and Google AI Overviews, recording which pages are cited and which competitors appear. This gives a trend line without requiring specialised tooling.
The bottom line
Generative engine optimisation is a distinct discipline from SEO. The source pools barely overlap. The signals differ. The measurement approach is different. The tactics that drive Google ranking — backlinks, keyword density, FAQ schema — are weak or irrelevant for AI citation rates.
The tactics that drive GEO performance — author attribution, branded web mentions, heading structure, specific sourced claims — are well-evidenced across multiple independent studies and should be treated as the foundational layer of any content strategy aimed at AI search visibility.