GEO vs SEO: What actually changes in your content

GEO is SEO with one extra job: getting cited inside AI assistant responses, not only ranked on a Google results page. It is not a new discipline. The original GEO paper (Aggarwal et al., 2023) tested nine content interventions against generative engines. Four moved the citation needle reliably, three gave moderate gains, and two flatlined or actively hurt visibility, including the classic SEO move of keyword stuffing.

Most online comparisons treat GEO as a brand-new craft requiring a brand-new workflow. The empirical evidence says the opposite: most of what works for GEO is what a good editor would do anyway.

What GEO actually is

GEO stands for Generative Engine Optimization, a term introduced by Aggarwal et al. in a 2023 arXiv paper that was accepted to KDD 2024. The paper formalises the question that came up the moment ChatGPT, Perplexity, Claude, and Gemini started answering queries with citations: how do you raise the probability that your page is one of the cited sources?

The paper measures two visibility outcomes. Position-Adjusted Word Count (PAWC) is roughly 'how prominently your content appears in the answer, weighted by where it appears.' Subjective Impression (SI) is a model-rated assessment of how much the answer relies on your source. Both are imperfect proxies, but together they give a measurable target.

The headline result: the paper's composite GEO approach boosted visibility by up to 40 percent across the test set. Per-intervention results from the paper's Table 1 are more useful than the composite, because they show which specific moves matter and which do not. The numbers below come from that table.

The four strongest interventions

These four moves produced the largest gains on both metrics in the paper's evaluation.

1. Quotation Addition: +42.6% PAWC, +23.3% Subjective Impression

Adding short, properly attributed quotations from authoritative sources had the largest effect of any intervention. Quotations from credible publications, papers, or official documentation are treated by generative engines as high-signal evidence.

Practical version: when you can paraphrase a source or quote them directly, quote them. Pick the sentence that says the thing most precisely. Attribute inline with the source name and date.

2. Statistics Addition: +32.8% PAWC, +16.6% Subjective Impression

Replacing qualitative language with quantitative claims tied to specific data sources came second. The effect is strongest in factual queries and comparison content.

In practice, every claim that can take a number should take a number, and every number should name its source. Replace 'a lot of users complain' with the actual percentage from the actual survey, named. Generic 'studies show' phrasing is the opposite of what works here.

3. Fluency Optimization: +28.7% PAWC, +9.3% Subjective Impression

Improving the readability and linguistic quality of the page came third. The takeaway is not 'use simpler words.' It is that messy, hard-to-parse prose is less likely to be selected by the model than clean prose that says the same thing. The Subjective Impression gain is much smaller than PAWC, suggesting fluency helps the model find and include your content more than it changes how the model judges your authority.

Working version: shorter paragraphs, active voice, sentences that resolve on a clear claim rather than trailing into hedging. Most editors already know how to do this; GEO is the empirical confirmation that it pays off.

4. Cite Sources: +27.7% PAWC, +10.9% Subjective Impression

Adding hyperlinked citations to credible sources came fourth, narrowly behind Fluency. The model treats outbound links to credible publications as a quality signal for the citing page.

What this looks like: when you make a non-trivial claim, link to where you got it. The anchor text should describe the destination, not say 'click here.' If you cannot find a credible source for a claim, downgrade the claim, do not link to a weaker source.

The moderate-gain interventions

These three gave smaller but still meaningful gains, with effects that vary by content area.

Technical Terms: +18.5% PAWC, +8.3% Subjective Impression

Adding domain-specific terminology gave a larger boost than the paper's narrative framing implied. It works because generative engines index by entity; precise technical vocabulary helps the model match the page to the query. The lift is real but smaller than the top four, and the catch is that forcing technical terms into a general-audience piece probably reverses the gain by making the prose harder to parse.

Easy-to-Understand: +13.8% PAWC, +4.7% Subjective Impression

Simplifying language while keeping accuracy improved PAWC but barely moved Subjective Impression. The model rewards clarity for content selection but does not particularly judge simpler content as more reliable. Effect is intertwined with Fluency, and the paper found it hard to fully separate the two.

Authoritative: +11.8% PAWC, +15.5% Subjective Impression

Switching to a more confident, definitive voice stands out because it is the only intervention where Subjective Impression gains exceed PAWC gains. Definitive prose changes how the model judges your reliability more than it changes how often the model picks your content. Position-taking matters, but the lift is moderate, not transformational.

Across all three moderate-gain interventions, the move is the same: take a clear position in entity-precise prose, written for the audience the topic implies. The bigger lift still comes from quotes, statistics, and citations; tone and clarity are smaller multipliers on top.

The interventions that flatlined or hurt

Two interventions produced near-baseline or actively negative results.

Unique Words: +6.2% PAWC, +6.2% Subjective Impression

Sprinkling rare vocabulary across the page produced near-baseline gains on both metrics. The model is not impressed by lexical diversity for its own sake. Skip this move entirely.

Keyword Stuffing: -8.7% PAWC, +2.6% Subjective Impression

The classic SEO move of increasing search-keyword density actually decreased Position-Adjusted Word Count by 8.7 percent. Subjective Impression saw a small +2.6 percent bump, but the headline finding is clear: keyword density hurts citation visibility, not helps it. Generative engines do not rank documents the way a keyword-based search index does; they read for meaning and citation-worthiness. Keyword density is not a citation signal, and aggressive stuffing is a citation anti-signal.

In short, do not write GEO content like 2012 SEO content. Density is not depth; in the paper, density was a penalty.

Where GEO and SEO overlap and where they diverge

Most of the empirically supported GEO interventions are also SEO best practice. Citing sources, using statistics, writing clearly, taking positions: these have been in the SEO playbook for over a decade, even if the rationale was different.

The divergences are smaller than the GEO marketing suggests:

• Keyword density actively hurts in GEO. Classical SEO rewards keyword presence and proximity to natural phrasing. GEO penalises high density. A page that says the right thing once with good context outperforms a page that repeats the same thing five times with awkward stuffing.
• Outbound links matter more, not less. Classical SEO worried about 'link juice' leaking through outbound links. GEO rewards citations to credible sources at roughly +27.7 percent PAWC. Outbound links to authoritative work are a feature, not a leak.
• Quote density matters. SEO does not care whether you paraphrase or quote your source. GEO rewards direct quotation with attribution at the largest effect size in the paper (+42.6 percent PAWC). This is the single largest divergence between the two playbooks.

Structural signals carry similar weight across both. Headings, scannable structure, direct-answer paragraphs near the top help in both regimes. The recent rise of llms.txt is a separate access-layer concern, not a content-layer one; see the WebPixie llms.txt adoption post for that side of the picture.

Net effect: GEO does not require you to throw out the SEO workflow. It requires you to dial up two practices (quotation and citation) and dial down one (keyword stuffing, which the paper shows is now actively harmful). The rest is recognisable.

A practical GEO checklist

Distil the empirical findings into something you can hold next to a draft:

1. Quote your sources directly when paraphrasing weakens the claim. Pick the sentence that says it most precisely. Attribute inline with source name and date. This is the single highest-leverage move per the paper.
2. Convert vague claims to specific numbers with sourced provenance. Replace 'most users' with the actual percentage from the actual survey, named.
3. Cite credible sources with descriptive anchor text. If you cannot find a credible source for a claim, downgrade the claim instead of citing a weaker source.
4. Take a position. Hedging across every claim signals to the model that the page is not the best source on the topic.
5. Write entity-precise prose. Use named tools, exact version numbers, official acronyms ('TLS 1.3', not 'modern encryption'). Generative engines index by entity.
6. Keep the prose clean. Short paragraphs, active voice, sentences that resolve. The 'fluency' effect in the paper is partly about not making the model work harder than necessary.
7. Skip the SEO-era keyword stuffing entirely. The paper shows it actively hurts citation visibility, not just fails to help.

The 2023 GEO paper is the empirical baseline, not the final word.

For now, the practical workflow is short. Quote, cite, count, write clearly, take a position. None of that requires a new content team or a new editorial calendar. It requires an editor who treats every paragraph as a possible answer in someone else's AI search.