GEO, Audited: One Study Holds Up, the Rest Is Vendor Folklore
Generative Engine Optimization is sold as a new discipline. The single controlled trial — KDD 2024 — found stats, quotes and citations lift AI-answer visibility 30-40%; keyword stuffing did nothing.
TL;DR — Strip the marketing and "GEO" (Generative Engine Optimization) reduces to one replicable result. The lone controlled trial — GEO, KDD 2024 — measured a 30-40% lift in AI-answer visibility from adding statistics, quotations, and citations to reputable sources. Keyword stuffing registered negative. Claims like "Claude prefers long-form" or "+71% from expert quotes" trace to no primary methodology. Net: cite, quote, quantify; discount the rest.
A new acronym arrives every quarter promising to decode the ranking layer. The 2026 entry is GEO — engineering content so ChatGPT, Perplexity, and Google's AI Overviews cite it. A consulting market is already pricing the premise. The relevant question is narrower than the pitch: which levers survive measurement?
The evidence base is a single paper
Most GEO guidance is output-watching dressed as method. One exception clears peer review: "GEO: Generative Engine Optimization" (Aggarwal et al., KDD 2024), out of Princeton, Georgia Tech, and the Allen Institute for AI. The design is concrete — a 10,000-query benchmark, AI answers generated over live search results, nine content edits tested, each scored by how prominently the source surfaced.
The ranking of effects:
| Edit | Measured effect |
|---|---|
| Add quotations from credible sources | ~+41% visibility |
| Add statistics (concrete numbers over vague claims) | ~+34% |
| Cite sources (inline links to reputable references) | ~+30% |
| Improve fluency / readability | ~+30% |
| Keyword stuffing | negative — it hurt |
Three causal levers clear the bar: quotations, statistics, citations. The classic SEO-spam move — keyword density — tested as useless for AI engines and mildly counterproductive. This article runs dense with numbers, quotes, and links by design, not habit.
Correlational, not causal — treat as priors
Past the trial, the better advice is observational. Reasonable bets, not laws:
- Front-load the answer. A Search Engine Land analysis put ~44% of ChatGPT citations in the first third of the content. A TL;DR or a direct opening line plausibly pays.
- Question-shaped headings. Self-contained FAQ-style
H2/H3s correlate with citation frequency — the prose does more work than the FAQ schema, which Google says is optional. - Self-contained sentences. A claim pulled into a retrieval chunk must stand alone; name the subject rather than leaning on "it" from three paragraphs back.
- Recency. Recently updated pages cite better than stale ones.
- Visible authority. Bylines, bios, a named publisher. Google frames quality as E-E-A-T (Experience, Expertise, Authoritativeness, Trust); AI answers lean on the same signals.
What fails verification
Where to stop spending:
- Suspiciously precise multipliers — "+71% from expert quotes," "19+ statistics doubles citations." The direction matches the research; the figures trace to no published method. Marketing.
- Platform personality lore — "Claude prefers long-form, Perplexity rewards freshness." Vendor-blog inference, not controlled study.
- Fabricated-brand demos. The circulated test that invented a brand ("Xarumei") and watched models echo it was criticized for running in an authority vacuum — it mainly showed that absent real authority, answer-shaped detail wins by default.
- "GEO is a new discipline." Google's stance is flat: optimizing for AI features is still SEO. No
llms.txt, no content "chunking," no AI-specific rewrites required.
Operational checklist
The effective playbook is conventional:
- Be specific and useful — original data, firsthand experience, real numbers.
- Cite inline and quote experts — the proven lever, and it improves the writing regardless.
- Structure for skimming — headings, a TL;DR, lists and tables, standalone sentences.
- Establish authority — real author, real publisher, genuine off-site reputation.
- Stay technically clean — crawlable, fast, mostly-static HTML (AI fetchers often see only initial HTML); don't block AI search bots.
The unglamorous reading: GEO is good writing plus real authority, with one research-backed nudge — stack statistics, quotations, and citations. The framework being sold around it is the part with no evidence behind it.
FAQ
Does GEO actually work?
Partly. The controlled 2024 study measured a ~30-40% gain in AI-answer visibility from adding statistics, quotations, and citations. Most other GEO advice is correlational or unverified, and keyword stuffing was shown not to work.
What separates GEO from SEO?
Contested. Google says optimizing for AI features is still SEO. GEO emphasizes being cited by AI answers — via stats, quotes, citations, structure — rather than ranked in a list, but the fundamentals overlap heavily.
Do I need llms.txt or special AI formatting?
No. Google says llms.txt, content chunking, and AI-specific rewrites are unnecessary. They're low-cost extras some tools adopt — see our AI-readable blog guide — but authoritative, well-structured content carries the load.
Sources: GEO paper, KDD 2024 (arXiv), Google AI optimization guide, Search Engine Land, Search Engine Journal.
Image: Muhammad Rafizeldi, CC BY-SA 4.0, via Wikimedia Commons.
← Back to all posts