GEO — Research

What twelve studies actually found about GEO

GEO isn't SEO. Here's what the research says.

By Mark Reeves · May 2026

A year ago, Generative Engine Optimization was a concept paper. Today it's a KDD conference topic and the reason your traffic analytics stopped making sense.

Most GEO advice you'll read was written before the measurement existed. Here's what the actual research says — twelve empirical studies, most published in the last six months.

First, the numbers that explain why this matters

Google AI Overviews went from appearing in 6.49% of US searches to 50-60% in twelve months. That's not a trend — that's a structural change in how information moves.

Sites that earn AI citations don't just get a mention. They earn 35% more clicks and five times higher conversion rates than traditional top-ten results that don't appear in AI-generated answers. The Ahrefs two-wave study (July 2025 and March 2026) found something more specific: the overlap between AI Overview citations and organic top-ten results dropped from 76% to 38% in eight months. Being ranked is no longer the same as being cited. The signals that drive one don't reliably drive the other.

This is why a new discipline exists. SEO optimizes for one system. GEO optimizes for a different one.

The distinction that changes how you think about it

The most useful framework comes from a peer-reviewed study published in April 2026 (Zhang Kai, He Xinyue, Yao Jingang — 602 controlled prompts, 21,143 valid citations across ChatGPT, Google AI Overviews, and Perplexity).

The paper formalizes something practitioners had noticed but couldn't measure: getting cited and actually shaping the answer are two different things.

They call these citation selection and citation absorption.

Citation selection is whether your page appears in the retrieval pool at all — whether the platform even considers you as a source. Domain authority, source type, and language all affect selection. Official sources, news sites, and vertical publications account for 79-87% of all citations across the three platforms. If your site isn't in that category, selection is the problem to solve first.

Citation absorption is how much your page actually contributes to the language and structure of the generated answer. A page can be selected as a reference while another page supplies every fact, sentence structure, and conclusion in the response.

Most GEO content optimizes for selection. Absorption is where the actual value lives — and it responds to different signals.

What the research says about absorption

High-absorption pages share specific characteristics. They're longer. They're more modular in structure. They're semantically aligned with the query. Most importantly, they contain specific evidence genres: definitions, numerical facts, direct comparisons, and procedural steps.

These are the content types generative engines extract and use, not just reference.

A useful frame: your page needs to function as an evidence container. A page that defines terms precisely, cites specific numbers, compares options directly, and explains steps clearly is more absorbable than a page that covers the same topic in flowing, narrative prose.

Counter-intuitive finding from the same study: Q&A formatting alone is weak. Pages formatted as question-and-answer scored lower on absorption than non-Q&A pages. The format changes the surface structure but doesn't necessarily add the definitions, evidence, or procedural depth that drives absorption. Packaging content as Q&A signals relevance without supplying the substance the model needs.

The three platforms behave very differently

Perplexity cites an average of 16.35 sources per query. Google AI Overviews cites 12.06. ChatGPT cites 6.88.

More citations doesn't mean deeper absorption. The Zhang Kai study found ChatGPT shows higher average absorption influence per cited page. Perplexity casts wide. ChatGPT goes deep on fewer sources.

For complex, multi-constraint queries, the gap becomes extreme: ChatGPT averages 3.4 citations, Google 12.6, Perplexity 17.7. When a question has multiple conditions, ChatGPT compresses into a small number of deeply absorbed sources. Perplexity decomposes into broad retrieval.

A GEO strategy that treats all three platforms as one audience will underperform on all three.

The finding that most contradicts conventional SEO thinking

The Digital Bloom synthesized 680 million citation events and found brand search volume is the strongest predictor of AI citation — a 0.334 correlation coefficient, stronger than backlink count, domain rating, or content length.

The implication: awareness-building drives AI citations more reliably than traditional link-building. If people search for you by name, AI systems learn to cite you. Building a recognizable brand — through consistency, presence, and direct audience relationship — is not just a marketing goal. It's a GEO strategy.

The Chen, Wang, and Koudas study (September 2025, across ChatGPT, Perplexity, and Gemini) adds to this: AI search systematically favors earned media over brand-owned content. Guest posts, podcast appearances, and third-party mentions outperform your own pages as citation sources. Your site benefits when others talk about you, not just when you publish about yourself.

What structural optimization actually looks like

The GEO-SFE paper (Yu, Yang, Ding, Sato, March 2026) tested a macro/meso/micro structure framework across six AI search engines and found a 17.3% citation rate improvement. The hierarchy: clear document structure at the page level, well-segmented section organization at the content block level, and precise, extractable claims at the sentence level.

The GEO-16 framework (Kumar and Palkhouski, September 2025) tested sixteen optimization dimensions and found three with the strongest citation outcomes: metadata freshness, semantic HTML structure, and structured data implementation. These aren't new SEO tactics — they're SEO fundamentals applied specifically to the way AI systems parse and extract page content.

The AutoGEO system (Wu, Zhong, Kim, Xiong, October 2025) learned engine preferences empirically by testing content variants and measuring citation outcomes. It achieved a 36% improvement in GEO metrics. The practical takeaway: AI search engines have preferences that can be learned, and they differ by platform.

What the research doesn't yet tell you

Honest GEO advice includes this.

The Zhang Kai paper is explicit: the study documents observed correlations, not causal mechanisms. A page being long, modular, and full of definitions correlates with high absorption. Whether adding those features to your page will cause higher absorption on a live platform hasn't been established through controlled experiments. The research is strong enough to act on directionally. It isn't strong enough to treat any single tactic as a guaranteed lever.

The Venkit et al. study (ACM FAccT 2025) found hallucination and citation inaccuracy are pervasive across major AI search platforms. Sources get cited for claims they don't make. Pages get attributed with conclusions they don't support. The citation link visible to a user is not always the actual source of the generated text. GEO is optimizing for a system that isn't fully transparent about how it uses what it reads.

None of this makes GEO not worth doing. It makes the measurement obligation clear: track what changes in your actual citation rate, not just what changes in your content.

The practical starting point

Based on the research, in priority order:

Build the brand signal first. If no one searches for your name, the downstream optimization doesn't work. Consistency, presence, and third-party mentions matter more than any single content tactic.

Make your evidence extractable. Add definitions, specific numbers, direct comparisons, and step-by-step procedures to your existing content. These are the evidence genres that drive absorption across platforms.

Distinguish by platform. ChatGPT rewards depth in fewer sources. Perplexity rewards breadth. Google AI Overviews rewards authority signals. A unified strategy works better than three separate ones, but the emphasis differs.

Measure separately from SEO. Your organic rankings will not tell you whether you're being cited. Run manual checks across ChatGPT, Perplexity, and Google AI Overviews for the queries that matter to your business. Track the baseline, then track changes.

Don't wait for the strategy to be settled. The AI search landscape is moving fast — AI Overview coverage went from single digits to majority share in one year. First-mover positioning in specific categories closes as those categories get populated.

Sources: Zhang Kai, He Xinyue, Yao Jingang (arXiv:2604.25707v2, April 2026) · Aggarwal et al. GEO-bench (KDD 2024) · Chen/Wang/Koudas et al. (arXiv:2509.08919) · Yu/Yang/Ding/Sato GEO-SFE (arXiv:2603.29979) · The Digital Bloom 680M citation study · Ahrefs two-wave AI Overview study (2025–2026) · Kumar/Palkhouski GEO-16 (arXiv:2509.10762) · AutoGEO / Wu et al. (arXiv:2510.11438) · Venkit et al. ACM FAccT 2025 · Kai-Cheng Yang (arXiv:2507.05301) · Kirsten et al. (arXiv:2510.11560) · Zhang/Ye/Peng et al. (arXiv:2512.09483)

Want the full system? GEO for Business Owners covers the complete GEO playbook: AI brand audit, entity building, question-shaped content, schema, and citation tracking — written for business owners, not developers. For marketers, GEO for Marketers (Beginner) covers 15 workflows for making brands more visible in AI search. Or start free: get AI for the Curious — a plain-English tour of every major AI tool, including where GEO fits in.

Mark Reeves writes practical guides on using AI tools in real work. AI Field Guide covers Claude, ChatGPT, Perplexity, Cursor, and GEO — for business owners, solopreneurs, marketers, authors, and clinicians. Browse the full catalog.