The GEO Paper: How to Boost Visibility in AI Search

Short answer: The Princeton paper GEO: Generative Engine Optimization ran controlled experiments on nine ways to rewrite a webpage and measured which ones got it cited more often by AI answer engines like Perplexity. The winners were not keywords. Adding quotations, statistics, and explicit source citations to your content produced the largest gains — up to a 41% relative lift on the paper's "Position-Adjusted Word Count" metric — while old-school keyword stuffing made visibility worse. This article walks through every method the paper tested, what the numbers actually mean, and how to apply each one to a real page today.

This is a breakdown of the research itself. If you want a plain-language primer first, read what Generative Engine Optimization is or the short GEO introduction. When you are ready to act, the GEO checklist turns the paper's findings into steps, and GEO Checker is a free tool that scores how ready a page is to be cited.

What you will learn

• What a generative engine is and why it broke the old SEO playbook.
• How the paper measures "visibility" when there is no ranked list of links.
• All nine GEO methods, ranked by measured impact — and the exact action to take for each.
• Which combinations work best and how to sequence them on your own pages.
• How GEO and SEO fit together instead of competing.

1. Why generative engines changed the rules

Traditional search returns a list of blue links. Tools like Perplexity, Google AI Overviews, and Bing Chat instead write an answer and cite a few sources inline. The paper groups these systems under the term Generative Engines. A typical one works in three steps:

1. Query rewriting — an LLM breaks your question into smaller, more searchable sub-queries.
2. Document retrieval — a search index returns candidate pages.
3. Answer generation — an LLM synthesizes one answer from several sources and attaches citations.

The catch for creators: a reader can get the full answer without ever clicking through, and you have no dashboard telling you when or how you were cited. The paper's contribution is a way to optimize for that citation layer even though the engine is a black box. You cannot see inside it, so you change the one thing you control — the content — and measure what comes back.

2. How the paper measures visibility

A ranked link has one position. A cited paragraph does not — your sentence might appear early or late in the answer, take up two words or twenty, in any context. So the authors define visibility with three measurable signals:

• Word Count Share — what fraction of the answer's words come from your cited sentences.
• Position-Adjusted Word Count — the same idea, but earlier sentences count for more (readers see them first). This is the paper's headline metric.
• Subjective Impression — an LLM evaluator (G-Eval) scores qualities like relevance, influence, and uniqueness, approximating how a human perceives the citation.

The practical point: citation is not all-or-nothing. Being mentioned earlier, longer, and more prominently is the goal — and that is something you can rewrite toward.

3. The nine GEO methods, ranked by measured impact

The authors tested nine ways to rewrite a page, then measured the change in visibility. Here is the full result, sorted by effectiveness.

The order is the takeaway: the methods that add real evidence win; the method that games keywords loses. The rest of this section is a step-by-step guide to applying each one.

High-impact methods — do these first

1. Add quotations (the single biggest lever, ~41%)

• What it is: quote a credible source directly instead of paraphrasing it.
• Where the paper found it works best: people and society, explanatory, and historical topics.
• How to apply it: find the strongest claim on your page and back it with a short, attributed quote from a named expert, study, or primary document. On Perplexity, the paper measured a +22% lift in Position-Adjusted Word Count from this alone.

2. Add statistics (~31%)

• What it is: swap qualitative descriptions for specific, sourced numbers.
• Where it works best: legal/government, debate, and opinion content.
• How to apply it: every time you write "many," "growing," or "significant," ask whether a real figure exists. The paper's example: rewriting "robots can take over repetitive tasks" to cite a concrete robot-density figure raised that passage's visibility by 65.5%.

3. Optimize fluency (~28%)

• What it is: make the prose clearer, smoother, and easier to read.
• Where it works best: business, science, and health content — and it amplifies every other method.
• How to apply it: cut filler, shorten sentences, and put the answer before the explanation. Fluency on its own helps, and paired with Statistics Addition the paper saw an extra ~5.5% gain on top of either method alone.

4. Cite your sources (~28%, and huge for underdogs)

• What it is: explicitly attribute facts — "according to [source]" — rather than stating them bare.
• Where it works best: expository, factual, and legal/government content.
• How to apply it: add inline attributions and a reference for any non-obvious fact. The effect compounds for lower-ranked pages: in the paper's multi-site test, a page ranking fifth gained 115.1% visibility from citations alone. Example: adding "according to the International Chocolate Consumption Research Group" to a chocolate article lifted its visibility 132.4%.

Medium-impact methods — worth doing, narrower wins

5. Use an authoritative tone (~10%)

• What it is: write with confidence and professional framing instead of hedged, casual phrasing.
• Where it works best: debate, history, and science topics.
• How to apply it: state conclusions plainly and support them; avoid wishy-washy qualifiers. The paper's example raised one passage's visibility 89.1% by reframing it authoritatively and backing it with stats.

6. Make it easy to understand (~14%)

• What it is: simplify wording so a non-expert can follow it.
• Where it works best: almost everything, especially educational content.
• How to apply it: define jargon on first use, prefer common words, and break dense paragraphs into shorter ones.

Lexical methods — small, situational gains

7. Add technical terms (~18% in fit domains)

• What it is: use precise, domain-specific vocabulary where it genuinely belongs.
• Where it works best: technical documentation and academic writing.
• How to apply it: use the correct term of art the first time, not a vague stand-in — but do not force jargon into general-audience pages.

8. Use unique words (~6%)

• What it is: vary your vocabulary with more distinctive wording.
• Where it works best: creative and marketing copy.
• How to apply it: treat this as polish, not strategy. The measured gain is small.

The method that backfires

9. Do not stuff keywords (~ −8%)

• What it is: repeating the query's keywords to raise their density — the classic SEO move.
• What the paper found: on Perplexity it performed about 10% worse than the unedited baseline.
• Why: generative engines reward semantic quality and evidence, not keyword frequency. Stuffing signals low quality. This is the clearest break from traditional SEO in the whole study.

4. Which methods to combine

The methods are not mutually exclusive, and the paper studied how they stack (Section 5.3).

• Best pairing: Fluency Optimization + Statistics Addition — about +5.5% on top of either alone.
• Cite Sources punches above its weight in combination: modest by itself, but it averages a ~31.4% boost when paired with other methods.
• Fluency is a multiplier: it reliably amplifies whatever it is combined with.

A practical sequence for one page: add the missing evidence first (statistics and quotes), attribute every claim (cite sources), then do a fluency pass to make the whole thing read cleanly. Skip keyword stuffing entirely.

5. Match the method to your topic

The paper also breaks down which methods perform best per domain (Section 5.1):

• Authoritative tone — debate, history, science.
• Cite sources — expository, factual, legal/government.
• Quotation addition — people/society, explanatory, history.
• Statistics addition — government/legal, debate, opinion.
• Fluency optimization — business, science, health.

Most methods just reframe information you already have. Quotation Addition, Cite Sources, and Statistics Addition are the exceptions — they may need you to go find a real source or number, which is exactly why they work.

6. Why this helps smaller sites most

When the paper simulated every site adopting GEO at once (Section 5.2), the gains were uneven — and they favored the underdog.

The lowest-ranked page gained the most; the top-ranked page often lost visibility. Traditional search rewards backlinks and domain authority, which entrench incumbents. Generative engines weigh the content of the answer, so a well-evidenced page from a new site can be cited next to an established one. For independent creators, that is the most encouraging finding in the paper.

7. How the researchers tested this

To make the results credible, the authors built GEO-BENCH — 10,000 queries spanning multiple domains, drawn from MS MARCO, Natural Questions, All Souls essay prompts, Perplexity Discover, and GPT-4-synthesized Q&A, and tagged by intent (informational, transactional, navigational). Each query is paired with the body content of the top five Google results, so methods can be tested per vertical and question type rather than on a single example. They then validated the methods on both a self-built engine and the live Perplexity.ai.

8. GEO and SEO are not the same job

GEO does not replace SEO — it optimizes a different layer. SEO gets you retrieved; GEO gets you cited once you are.

Keep doing SEO for retrieval and site health. Layer GEO on top for the high-value pages you most want AI engines to quote.

9. What to do next

1. Pick one high-value page — something you genuinely want AI search to cite.
2. Add the evidence — replace vague claims with statistics and short attributed quotes from named sources.
3. Cite everything — attribute each non-obvious fact inline.
4. Do a fluency pass — lead with the answer, cut filler, shorten sentences.
5. Never stuff keywords — it measurably hurts you here.
6. Check your work — run the page through GEO Checker to see how citation-ready it is, then use the GEO checklist to close the gaps.

The paper's larger message is that the rules of visibility have changed, and they now reward exactly what good writing always should: real evidence, clear sources, and readable prose. The creators who get cited in AI search will be the ones who make their content the most useful thing to quote.

The only statistics in this article come from the cited Princeton GEO paper; we do not use unverified marketing numbers. For the plain-language version, see what GEO is; for more guides, visit the GEO Checker blog.