Playbook · GEO

The GEO playbook: what gets you cited

Generative engine optimization isn't guesswork anymore. A peer-reviewed benchmark of 10,000 queries measured exactly which edits make an AI more likely to cite you — and the winners are concrete, repeatable, and mostly about evidence.

13 min read8 chartsUpdated June 2026Mentionova Research

Visibility lift by GEO tactic% gain in AI-answer visibility

Source: Aggarwal et al., "GEO: Generative Engine Optimization," Princeton / Georgia Tech / IIT Delhi (KDD 2024), measured on GEO-bench.

For a year, "how do I show up in AI answers?" had only folk-wisdom answers. Then a group of researchers from Princeton, Georgia Tech and IIT Delhi did the obvious-in-hindsight thing: they built a benchmark of 10,000 real queries across nine domains, systematically edited the source content, and measured what moved an engine to cite it. They called the discipline Generative Engine Optimization — GEO.

The headline result: the right edits lifted a page's visibility in AI answers by up to 40%. Not by chasing keywords — by making the content more citable.

Baseline vs GEO-optimizedindexed visibility in AI answers

The best combinations of tactics moved an unremarkable page near the top of what engines surface.

The moves that worked

The benchmark tested editorial changes one at a time. Three stood out, and they rhyme: they all make a claim more verifiable.

Add expert quotations — the single biggest lever, around +41% visibility. A named human saying it reads as authority.
Add statistics — +32%. Specific numbers get lifted into answers far more than vague claims.
Cite your sources — +30%. Pages that themselves cite authorities are treated as more trustworthy.

The mirror image is just as useful: classic SEO keyword stuffing did nothing for AI visibility — and in several runs it hurt. The engines aren't counting keywords; they're judging credibility.

Write like a source, not like a landing page. Quotes, numbers, and citations are the currency of being cited.

Position is leverage

Where a fact sits on the page matters almost as much as whether it's there. In analyses of cited pages, 44% of all AI citations came from the first third of the text. The model rewards getting to the point. Bury your best, checkable claim under 800 words of preamble and it may never be read.

Citations by page lengthrelative to the shortest pages (×)

Comprehensive pages (20k+ characters) earned ~4.3× the citations of thin ones. Position matters too: ~44% of citations come from the opening third of a page — front-load the evidence.

4.3×More citations for depth. Pages over ~20,000 characters were cited about 4.3 times as often as short ones — depth and coverage read as authority to the models.

Structure the machine can read

AI engines parse structure, not vibes. Among cited pages, 68.7% followed a clean H1 → H2 → H3 hierarchy and 61% used structured data markup. Headings, lists, tables, and schema turn prose into something a model can lift a precise answer from. A beautiful wall of text is, to a crawler, an unstructured blob.

Structural traits of cited pagesshare exhibiting each trait

Hierarchy and schema aren't nice-to-haves; they're how the answer gets extracted.

The playbook

Translate the research into a checklist you can run on any page you want cited:

Lead with the answer. Put your most quotable, checkable claim in the first third.
Back every claim with a number. Replace adjectives with statistics and dates.
Quote a named expert. One real attributed sentence beats a paragraph of marketing.
Cite your sources — and link them. Borrow authority by referencing it.
Use real structure. H1→H2→H3, short paragraphs, lists, comparison tables, FAQ schema.
Go deep on one thing. Comprehensive beats thin; cover the topic, not the keyword.
Make it crawlable. No content trapped behind JS, logins, or PDFs.
Earn third-party mentions. The page is half the battle; the consensus around it is the other half.

GEO is just good faith at scale: say something true, prove it, structure it, and make it easy to find.

Don't game it — measure it

One caution the research makes clear: the engines are actively tuned against manipulation, and the source list churns without warning. A tactic that wins this month can flatten the next. That's the case for treating GEO as a measured loop, not a one-time checklist — publish, watch what each engine does with it, and adjust.

That loop is the whole product. Mentionova runs your category's questions through all six engines on a schedule, shows which of these moves is actually moving your citations, and flags the day the answer changes. Start with the macro picture in The State of AI Search, or see your own answer below.

Sources