Generative engine optimization (GEO) is the practice of writing and structuring content so that AI answer engines such as ChatGPT, Perplexity, and Google's AI Overviews retrieve it, synthesize it, and cite it inside a generated answer. The discipline rests on one controlled experiment: researchers at Princeton and collaborating institutions tested nine specific content edits across a 10,000-query benchmark and reported visibility gains of up to 40% (Aggarwal et al., KDD 2024). This piece ranks all nine by the lift that study measured, states the ranking criteria first, and names the one tactic that scored below the unoptimized baseline.
Most explainers on this query stop at the definition. We think the definition is the least useful part, because a marketing team cannot execute a definition. What they can execute is a ranked list of edits with known effects, which is what the research actually produced.
What GEO optimizes for
Traditional search optimization competes for a position in a list of links. A generative engine does something different: it retrieves a handful of candidate passages, writes a synthesized answer, and attaches citations to the sources it leaned on. The unit of success moves from rank to inclusion. Your page either supplies a sentence the model is willing to repeat and attribute, or it does not appear at all.
That reframing is the whole discipline. Aggarwal and colleagues formalized it in 2024 under the term "generative engines" and built GEO-bench, a benchmark of 10,000 queries spanning multiple domains and source types, to test whether website content could be deliberately modified to win inclusion (Princeton University). The headline result, up to 40% higher visibility, is a maximum under favorable conditions rather than an average, and low-ranked sources benefited disproportionately (Blck Alpaca).
The commercial case for caring is no longer speculative. Similarweb estimated that AI platforms drove more than 1.13 billion referral visits to the top 1,000 websites in June 2025, up 357% year over year, against 191 billion referrals from Google Search in the same month (TechCrunch). Small channel, steep curve.
How we ranked the nine tactics
Our read, stated before the list so you can argue with it:
- Measured effect in the original experiment. Position-Adjusted Word Count and Subjective Impression are the two metrics the paper reports. We weight them equally.
- Consistency across domains. A tactic that wins on technical queries and loses on consumer queries ranks below one that holds up everywhere.
- Independence from the writer's luck. Adding a real statistic is a repeatable editorial instruction. "Sound more authoritative" is not.
- Cost to implement inside a normal weekly publishing cycle.
One honest caveat before the ranking. The paper reports that its three strongest methods produced roughly 30-40% relative improvement on Position-Adjusted Word Count over the unoptimized baseline (Aggarwal et al., 2024), but secondary write-ups of the same paper circulate different per-tactic percentages, and they disagree with each other. Where a single-tactic number is widely quoted but not consistent across sources, we give the band and the source rather than a false decimal.
1. Cite credible sources
The strongest and most consistent of the three top methods. Adding citations from reliable sources to your own page raises the probability that a generative engine includes and attributes your passage, and the effect compounds when paired with other edits (Aggarwal et al., 2024).
The mechanism is plain once you accept what the model is optimizing for. A synthesized answer carries reputational risk for the engine, so passages that already carry verifiable provenance are cheaper to repeat. Treat every substantial claim as something that needs an attributed source (Elementera).
2. Add quotations from credible sources
Second of the top three. Direct quotations attributed to a real, identifiable source sit in the same 30-40% band on Position-Adjusted Word Count as the other two leaders (Aggarwal et al., 2024).
A quotation is a pre-packaged extractable unit. It has a speaker, a boundary, and an implicit warranty. The practical rule: quote only what you can trace to a URL where those exact words appear, because an invented quote is the fastest way to make a page unusable to both readers and engines.
3. Add relevant statistics
Third leader, and the easiest to systematize. Replacing qualitative statements with specific numbers gave consistent gains in the experiment. "Many companies struggle with this" carries almost no retrieval value; a named study with a precise figure and a year does (Elementera).
4. Fluency optimization
The surprise in the data. Improving the readability and flow of existing text, adding no new information at all, produced a significant visibility boost in the study (Aggarwal et al., 2024). Clear prose is simply easier for a language model to parse, segment, and attribute correctly, while dense or tangled writing works against the source even when the underlying facts are strong.
It also happens to be the cheapest edit on this list. You already own the text.
5. Easy-to-understand phrasing
Closely related to fluency and separately tested. Simplification performed well, with a noticeable skew: consumer and general-interest topics gained more from plain phrasing and quotations, while technical topics gained more from statistics and citations (Princeton University).
6. Authoritative tone
Authoritative tone is one of the nine strategies the study tested, and it sits outside the three methods the paper singles out for strong, consistent improvement (Aggarwal et al., 2024). Our read, labelled as a read: register is a weak proxy for what engines reward, which is traceability. Treat it as a finishing pass over an already sourced draft.
7. Technical terms
The paper's own conclusion is that the efficacy of these strategies varies across domains, which is why it argues for domain-specific optimization instead of one universal recipe (Princeton University). Our read on technical vocabulary: it earns its place on specialist queries where the term is what the reader came looking for, and adds little on general ones.
8. Unique words
Vocabulary variation is among the nine tested strategies, and the paper's strong results cluster elsewhere (Aggarwal et al., 2024). We rank it eighth as a tiebreaker and would not spend a weekly editorial slot on it.
9. Keyword stuffing, which made things worse
The only tactic in the set that scored below doing nothing. Loading a page with repeated query terms came in under the unmodified baseline on Position-Adjusted Word Count, by roughly 8% in one detailed reading of the results (Elementera) and with the same directional finding reported elsewhere (Blck Alpaca).
Carrying a legacy SEO playbook into an AI answer engine can therefore cost you visibility rather than merely wasting effort.
| Rank | Tactic | Reported effect | Best fit |
|---|---|---|---|
| 1 | Cite credible sources | Top band, 30-40% relative lift on PAWC | All domains |
| 2 | Add quotations | Top band, 30-40% relative lift | Consumer, opinion, news |
| 3 | Add statistics | Top band, 30-40% relative lift | Technical, B2B, research |
| 4 | Fluency optimization | Significant, no new content required | All domains |
| 5 | Easy-to-understand | Significant, skews consumer | Consumer, explainer |
| 6 | Authoritative tone | Outside the paper's top three; our read | Finishing pass |
| 7 | Technical terms | Domain-dependent; our read | Specialist queries |
| 8 | Unique words | Not among reported leaders; our read | Tiebreaker |
| 9 | Keyword stuffing | Below baseline | Nothing |
The researchers also tested pairwise combinations of the top-performing methods, and combinations outperformed single edits (Elementera). Read the table as a stack, not a menu.
Why your SEO authority does not carry over
Ahrefs analyzed 75,000 brands and found that branded web mentions correlate with AI Overview brand visibility at 0.664 on the Spearman scale, while backlink counts correlate at 0.218, with branded anchors (0.527) and branded search volume (0.392) filling out the top three. All three leaders are off-site brand signals (Ahrefs). Ahrefs flags the obvious caveat in its own write-up: these are correlations, and large brands accumulate both mentions and AI visibility for reasons that may have nothing to do with each other.
Take it as a directional finding. The asset that predicts citation is how widely and consistently your brand is discussed in machine-readable text, which is built by publishing and being referenced rather than by acquiring links.
Score each engine on its own line
Platform share is moving fast enough that a blended visibility number hides the story. Similarweb's 2026 landscape data shows ChatGPT's share of generative AI website visits falling from roughly 76% in June 2025 to about 53% by May 2026, with Gemini rising from under 9% to roughly 27-28% over the same period (Similarweb).
A single "AI visibility score" averaged across engines can stay flat while you lose Perplexity and gain Gemini. Run a fixed prompt set weekly and score each engine on its own line. We build our own tracking this way, and it is also the first thing we tell people to check when evaluating a monitoring vendor: see how to vet tools that monitor AI brand mentions.
A three-week start, for a team of one or two
Week 1, roughly three hours. Write 15-25 prompts a buyer would actually type, covering category, comparison, and use-case questions. Run them against ChatGPT, Perplexity, and Gemini. Record which brands get named and which URLs get cited, per engine. That citation list is your real competitive set.
Week 2, roughly four hours. Take your five highest-intent pages and apply tactics 1 through 3 to each: an attributed source for every substantial claim, one traceable quotation where a real one exists, and a specific number with a year replacing each vague qualifier. Then run the fluency pass. Republish.
Week 3 onward, one focused day per week. Publish one or two new pages against the prompts where no page of yours was cited, and re-run the prompt set. Citation movement on retrieval-based engines shows up in weeks, not days, so judge the loop on a monthly trend rather than a single check. Our weekly organic content workflow for small teams breaks the same loop into time-boxed blocks.
The reason we push cadence over one-time optimization is structural. Citation authority accumulates the way domain authority once did, and a page you rewrote in March competes against pages your competitors published last week.
FAQ
What is the difference between GEO and SEO?
SEO competes for a ranked position on a results page. GEO competes for inclusion and attribution inside a synthesized answer. They overlap on crawlability and content quality, and they diverge sharply on tactics: keyword stuffing, a classic SEO lever, scored below the unoptimized baseline in the Princeton experiment (Elementera).
Is GEO the same thing as answer engine optimization (AEO)?
The terms are used interchangeably by most practitioners. GEO is the label from the academic literature (ACM SIGKDD 2024) and tends to be the more precise one, since it names the mechanism, which is generative synthesis with citation.
How long does it take to see results?
For retrieval-based surfaces like ChatGPT Search, Perplexity, and AI Overviews, a republished page can change citation behavior within weeks because the engine fetches live content at answer time. Influencing what a base model knows without retrieval depends on the next training cycle, which is outside any publisher's control. Optimize for the retrieval layer.
Does GEO work for a site with low domain authority?
The study found that lower-ranked sources gained disproportionately from the top tactics (Blck Alpaca), and Ahrefs' correlation table puts backlinks well below brand mentions as a predictor of AI visibility (Ahrefs). A small site with dense, well-sourced pages on a narrow topic is in a better position here than it ever was in classic search.
What should I measure instead of clicks?
Citation rate per engine on a fixed prompt set, brand mention rate within answers, and the share of your cited URLs that you own. Zero-click answers generate no session, so a referral-only dashboard will under-report GEO by design.
Where to start with us
If you want the loop running without building it yourself, Aeolo finds blog topics from your brand URL, tracks how often AI engines name you across a weekly prompt set, and keeps a publishing cadence going so citation authority accumulates instead of resetting. For how we score the tools in this category, including our own, read Best GEO Tools, Scored Across the Citation Chain.
