TL;DR: Almost every claim that GEO raises AI visibility by 40% traces back to one paper: Aggarwal et al., GEO: Generative Engine Optimization, presented at KDD 2024 by researchers from Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi. It found that adding citations, quotations and statistics to content raised its visibility in generated answers by up to around 40% on its own benchmark. The finding is real, the number is a ceiling rather than an average, and the test conditions date from 2023 and 2024.
Key Takeaways
It is a real peer-reviewed study. Published at ACM SIGKDD 2024, with a public benchmark of 10,000 queries across nine domains. That is a stronger evidence base than anything else in this field.
Three interventions did the work. Citing sources, adding quotations from credible sources, and adding specific statistics, each outperforming the unmodified baseline consistently.
40% is the upper end, not the expected result. The paper's own phrasing is up to 40%, measured within its benchmark, its engines and its time period.
Keyword stuffing performed badly. The study tested legacy tactics alongside the new ones, which is part of why it is worth reading.
The engines have changed since. The benchmark predates most of what people now use. Treat the direction as sound and the magnitudes as historical.
Secondary reporting of the detail is inconsistent. Different sites quote different per-method figures. Read Table 1 of the paper rather than anyone's summary, including this one.
Anchor targets assume your CMS generates slug IDs from heading text. Confirm on staging and adjust the hrefs if your platform slugifies differently.
What the study is
Worth establishing precisely, because a great deal of marketing content cites this paper without naming it.
Title | GEO: Generative Engine Optimization |
Authors | Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, Ameet Deshpande |
Institutions | IIT Delhi, Princeton University, Georgia Tech, Allen Institute for AI |
Venue | KDD '24, the 30th ACM SIGKDD Conference, Barcelona, August 2024 |
Preprint | arXiv:2311.09735, first posted November 2023 |
Benchmark | GEO-bench, 10,000 queries across nine datasets |
It is peer-reviewed and published at a major conference. That matters, because almost everything else written about generative engine optimisation is practitioner observation, including most of what this site publishes.
Why it exists
The paper's framing is worth noting. The authors observed that generative engines are useful to users and valuable to the engines, and leave a third group with almost no control: the website and content creators whose material is being synthesised.
So the research question was not how to game anything. It was whether content creators have any levers at all in a black-box system, and which ones measurably move the result.
It also coined the term
Generative engine optimisation as a named discipline comes from this paper. When you see GEO defined anywhere, the definition traces back here, usually through several intermediaries.
Why you should read it yourself
It is publicly available on arXiv and it is readable without a technical background. Table 1 contains the numbers that everyone quotes, and reading them in context takes about twenty minutes.
Given how much advice rests on this single paper, that is twenty minutes well spent for anyone making decisions on the strength of it.
What it tested
The design is simple and that is its strength. Take a page, modify it in one specific way, and see whether generated answers cite it more.
The method
Start with a query and a set of source websites. Produce a generated answer from them. Then modify one source using a given technique, regenerate, and measure whether that source's presence in the answer changed.
Repeat across thousands of queries and nine domains, so the result is not an artefact of one topic.
The interventions
Nine content modifications, spanning both new ideas and legacy SEO tactics:
Method | What it changes |
|---|---|
Cite Sources | Adds citations from credible sources |
Quotation Addition | Adds quotations from relevant credible sources |
Statistics Addition | Replaces vague claims with specific figures |
Fluency Optimization | Improves readability and flow |
Easy to Understand | Simplifies the language |
Authoritative | Makes the tone more confident and authoritative |
Technical Terms | Adds domain terminology |
Unique Words | Increases vocabulary variety |
Keyword Stuffing | The legacy tactic, included as a comparison |
Including keyword stuffing was a good decision, because it tests the assumption that old tactics transfer.
How visibility was measured
Two metrics, and the distinction matters when reading the results.
Position-Adjusted Word Count weights how much of the answer came from your source and where it appeared, so material quoted early counts for more than a passing reference at the end.
Subjective Impression uses a model to judge how prominently and favourably the source featured, which captures something the word count cannot.
Most headline figures circulating refer to the first metric. They are not interchangeable, and a gain on one is not automatically a gain on the other.
What the design does not capture
It measures presence within a generated answer. It does not measure traffic, conversions, or whether a human read the citation and clicked.
That is appropriate for the research question and worth remembering when the figure gets quoted as a business outcome.
What it found
The paper's own summary: including citations, quotations from relevant sources, and statistics can significantly boost source visibility, with an increase of over 40% across various queries.
Three interventions, one direction, consistent across domains.
The three that worked
Method | What it means in practice |
|---|---|
Cite Sources | Attribute your claims to credible external sources, visibly |
Quotation Addition | Include direct quotations from credible sources rather than paraphrasing everything |
Statistics Addition | Replace vague qualitative claims with specific figures |
All three share a property: they make the content more verifiable. A passage containing an attributed figure is easier for a system to treat as a factual claim worth reproducing than a sentence asserting that something is widely regarded as effective.
That mechanism is more durable than the specific percentages, and it is the part worth carrying forward.
Combinations helped
The researchers tested pairs of the strongest methods and found that combining them outperformed the best single method, with fluency optimisation paired with statistics addition noted as a strong combination.
So these are not alternatives to choose between. Doing two is better than doing one.
Lower-ranked content benefited most
One of the more interesting findings, and the one least often quoted: the gains were largest for sources that were not already dominant. Secondary reporting puts the improvement for lower-ranked content considerably higher than the headline figure.
If that holds, it is encouraging for challengers, since it suggests these techniques help most exactly where you need help. Verify the specific figures in the paper before repeating them.
A caution on the detailed numbers
Secondary sources report the per-method figures inconsistently. Different sites attribute 41%, 31%, 28% and 115% to different methods and metrics, and they do not agree with each other.
That inconsistency is a good reason to read Table 1 directly rather than any summary. The headline direction is solid; the decimal places circulating in marketing content are not reliable.
What the 40% figure means
The paper says up to 40%. Almost every secondary citation drops those two words, and they carry most of the meaning.
What the number is
It is | It is not |
|---|---|
A ceiling observed in the best conditions | An average you should expect |
A relative improvement over an unmodified baseline | An absolute visibility figure |
Measured on GEO-bench, with its queries and domains | A result across the open web |
Measured against engines as they behaved in 2023 and 2024 | A statement about engines today |
A gain in presence within generated answers | A gain in traffic, leads or revenue |
Why the distinction matters commercially
If you promise a client a 40% visibility increase on the strength of this paper, you have promised a benchmark ceiling from a specific experimental setup as a business outcome. That is the sort of claim that is checkable and that damages credibility when checked.
The defensible version is available and still persuasive: peer-reviewed research found that adding citations, quotations and statistics measurably increased how often content was cited in generated answers, by up to around 40% in the conditions tested.
That sentence is accurate, it is citable, and nobody can take it apart.
The baseline question
Relative improvements depend entirely on where you start. A move from a low baseline looks dramatic in percentage terms and may be small in absolute presence.
The paper reports absolute impression metrics alongside the relative improvements, which is why reading the table matters more than reading the abstract.
What survives the caveats
Quite a lot, and it is the mechanism rather than the magnitude.
Content that attributes claims to credible sources, contains specific figures and quotes real authorities is more likely to be used in a generated answer than content that asserts things vaguely. That finding is well supported, intuitively sensible given how retrieval works, and consistent with what practitioners have observed since.
Use that. Leave the decimal places in the paper.
What did not work
The negative results are as useful as the positive ones, and they get quoted far less because they do not sell anything.
Keyword stuffing
The legacy tactic was tested alongside the new ones and performed poorly. That is the single most useful negative result in the paper.
It matters because the instinct when a new channel appears is to apply the old playbook harder. This is direct evidence that keyword density, the defining tactic of early SEO, does not transfer to generative retrieval and can work against you.
The weaker interventions
Several methods produced smaller or inconsistent gains:
Method | Why it underperformed, plausibly |
|---|---|
Technical Terms | Adding jargon narrows rather than clarifies |
Unique Words | Vocabulary variety for its own sake adds no verifiable content |
Authoritative tone | Sounding confident is not the same as being verifiable |
That last one is worth dwelling on. Writing in a more authoritative voice did not deliver what adding an actual citation did. Confidence is a style; attribution is a property of the content.
A great deal of marketing copy is written in exactly the authoritative register that this result suggests does not help.
The pattern across winners and losers
The methods that worked added something checkable. The methods that did not worked on surface presentation.
Worked | Did not |
|---|---|
A specific figure | A confident adjective |
A named source | A technical term |
A real quotation | A wider vocabulary |
Clear readable prose | Repeated keywords |
That division is the most portable finding in the paper, because it explains why rather than just what, and it is likely to hold as engines change.
What this means for your existing content
If your pages are written in confident marketing language with few specific numbers and no external attribution, this research suggests they are formatted close to the way that performed worst.
That is an uncomfortable finding for most B2B websites, and it is also the most actionable thing here.
The limitations that matter
The authors are clear about the study's scope. Marketing content quoting it generally is not.
Six limitations worth knowing
1. It measures presence in an answer, not business outcomes. Appearing more prominently in a generated response is a reasonable proxy for commercial value and it is not the same thing. No traffic, conversion or revenue was measured.
2. The benchmark is not the open web. GEO-bench is 10,000 queries across nine domains, constructed for the study. Your category may behave differently, and the only way to know is to test it.
3. The engines are from 2023 and 2024. The paper used generative engines as they existed then. Everything about this field has moved since, including which systems people use and how they retrieve.
4. One source was modified at a time. In reality every competitor is optimising simultaneously. An intervention that works when you alone apply it may deliver less when the whole category applies it.
5. Subjective Impression uses a model as judge. A reasonable method and a known limitation, since the evaluator shares failure modes with the systems being evaluated.
6. It has not been widely replicated. As far as I can tell there is no substantial independent replication on current engines. For a finding this heavily cited, that is a real gap.
What the limitations do not undermine
The direction. Three interventions that make content more verifiable outperformed a baseline consistently, across nine domains, on two different metrics, while a legacy tactic underperformed.
That pattern is robust to most of the caveats above. What the caveats undermine is the precision of the numbers and their transferability to your situation.
How to hold it
Treat the paper as strong evidence for a mechanism and weak evidence for a magnitude.
The mechanism tells you what to do: attribute claims, include real figures, quote real sources, write clearly, do not stuff keywords. The magnitude tells you nothing reliable about what will happen on your site, which you can only learn by measuring your own.
That is a less exciting conclusion than a 40% promise and it is the one the evidence actually supports.
What has changed since 2024
The paper's experiments predate most of the current landscape. Some findings should travel well and some are more exposed.
What has changed
Change | Effect on the findings |
|---|---|
New model generations across every provider | The systems tested are not the systems in use |
AI Overviews and AI Mode at Google scale | A surface that barely featured in the original work |
Retrieval pipelines rebuilt | How sources are selected and presented has moved |
Citation formats changed | Which affects how presence is measured at all |
The whole category now optimising | The intervention is no longer unusual |
That last row deserves attention. In 2023 adding citations to a page made it unusual. If every page in your category now does it, the differential advantage shrinks even if the mechanism still works.
That is a familiar dynamic from search generally: a technique's effectiveness decays as adoption rises, without the underlying mechanism becoming wrong.
What should still hold
Finding | Why it should travel |
|---|---|
Verifiable content is preferred over assertive content | Follows from how retrieval and grounding work, not from any model's quirks |
Specific figures beat vague claims | Same reasoning |
Clear prose beats dense jargon | Consistent with how passages are selected |
Keyword stuffing does not help | Nothing since suggests this reversed |
What is most exposed
The magnitudes. A 40% ceiling measured on 2023 and 2024 engines with pre-optimisation baselines is the part least likely to reproduce today, and it is the part most quoted.
The replication gap
The most useful thing anyone in this industry could do is replicate this study on current engines and publish it.
It would be genuinely citable, it would settle which findings survived, and it is within reach of any company already running prompt measurement at scale. That nobody appears to have done it is the most striking gap in this field, and it is an open invitation.
How to use a 2024 paper in 2026
Cite it for the mechanism and the direction, with its date stated. Do not present its magnitudes as current. And where you need a current number, measure your own site and report that instead, which is more persuasive anyway because it is about the reader's situation rather than a benchmark.
What to do with this
The findings convert into editorial practice directly, which is unusual for academic work and is why this paper has travelled so far.
The three changes, concretely
Attribute your claims. Every factual assertion that could be sourced, should be, with a named source and a link. Not a vague gesture at research showing something; the specific study, organisation or dataset.
Replace vague claims with figures. Hunt for sentences containing significantly, many, most, leading, or rapidly, and either find the number or delete the claim.
Instead of | Write |
|---|---|
Reduces latency significantly | Reduced median latency from 340ms to 95ms in our own benchmark, March 2026 |
Many companies are adopting this | A 2026 survey of 500 engineering leaders found 34% had adopted it |
Our fastest plan | Processes up to 10,000 records a minute |
Quote real sources. A direct quotation from a named authority, used properly and attributed, rather than paraphrasing everything into your own voice.
A practical audit
Take your ten most commercially important pages and count, per page:
Count | Target |
|---|---|
Specific numbers with units and dates | Several per page, not zero |
External citations to named sources | At least a few, with links |
Direct quotations from named people or organisations | At least one where it fits |
Vague intensifiers (significantly, many, leading) | Drive toward zero |
Repeated keyword phrases | Not a target. Stop counting these |
Most B2B pages score badly on the first three and well on the last. That is the gap this research points at.
What not to do with it
Do not manufacture citations to hit a count. Fabricated or irrelevant attribution is worse than none, because readers check, and a page citing a source that does not support the claim damages exactly the credibility you were trying to build.
Also do not strip your own voice out. The findings favour verifiability, not blandness. Opinions and judgement are part of what makes a page worth citing, provided the facts around them are attributable.
The honest expectation
Doing this will not deliver 40% more citations. It will make your content more verifiable, which is what the research found engines prefer, and it will make it better for human readers too.
That second benefit is certain, which is a reasonable basis for the work on its own.
How it squares with the rest of the advice
Most GEO guidance in circulation is practitioner judgement. It is useful to see which of the common claims this research supports, which it contradicts, and which it simply does not address.
Supported
Common advice | Status |
|---|---|
Cite your sources and link them | Directly supported, one of the three winning methods |
Use specific numbers rather than vague claims | Directly supported |
Quote credible sources | Directly supported |
Write clearly rather than densely | Supported, fluency optimisation performed well |
Do not stuff keywords | Supported by a negative result |
Contradicted or unsupported
Common advice | Status |
|---|---|
Write in a confident, authoritative voice to build trust | The authoritative-tone intervention did not deliver what citations did |
Add technical terminology to signal expertise | Underperformed |
Vary your vocabulary widely | Underperformed |
Not addressed at all
This is the longest list, and worth knowing before anyone cites the paper in support of these.
Common advice | Status |
|---|---|
Publish AI.json or llms.txt files | Not tested. Separately, Google's documentation states no such files are needed |
Optimise your position in vector space | Not what the study did, and not something external parties can do |
Build backlinks for AI authority | Not tested here |
Publish at high volume | Not tested, and in tension with how retrieval selects passages |
Schema markup improves citation rates | Not tested in this paper |
Third-party mentions and corroboration matter | Not tested here, though widely observed |
That schema row is worth noting, since schema is near-universal in GEO advice including on this site. It is well-founded on other grounds, around entity clarity and eligibility, and it is not something this paper establishes.
The useful discipline
When you read a GEO claim, ask what it rests on. Much of it rests on this one paper, some rests on observation, and some rests on nothing but repetition.
The advice in the first two categories is worth following. The third is worth checking before you spend a quarter on it.
Including the advice here
The same test applies to everything else published on this site. Where guidance comes from practitioner observation rather than evidence, it is labelled that way, because the alternative is adding to the repetition.
Six ways this gets miscited
1. Dropping the words up to. The paper says visibility can be boosted by up to 40%. Quoted as a flat 40%, it becomes an expected result rather than an observed ceiling, which is the single most common error.
2. Presenting it as a traffic or revenue figure. The study measured presence within generated answers. It did not measure clicks, sessions, leads or sales, and converting it into a business promise is not something the data supports.
3. Citing it as current. It is a 2024 paper with experiments from 2023 and 2024. Quoting it without the date implies it describes today's engines, which it does not.
4. Attributing precise per-method numbers from secondary sources. Different sites report 41%, 31%, 28% and 115% against different methods and metrics, and they disagree. If you need a specific figure, take it from Table 1 yourself.
5. Using it to support things it never tested. It says nothing about AI metadata files, schema, backlinks, publishing volume or vector positioning. Those claims need their own basis.
6. Calling it the Princeton study and leaving it there. Four institutions contributed, the first author is at IIT Delhi, and the paper has a title and six named authors. Citing it properly costs nothing and is itself the practice the paper recommends.
The irony worth naming
A paper whose central finding is that properly attributed claims get cited more is routinely cited without attribution, with its qualifier removed and its numbers garbled.
If you want a demonstration of why verifiable content matters, the citation trail of this paper is it.
How to cite it properly
Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., and Deshpande, A. (2024). GEO: Generative Engine Optimization. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. arXiv:2311.09735.
Name it, date it, link it, and keep the words up to. That is the whole requirement.
FAQs
What is the Princeton GEO study?
Aggarwal et al., GEO: Generative Engine Optimization, presented at ACM SIGKDD 2024 by researchers from Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi. It introduced GEO-bench, a benchmark of 10,000 queries, and measured which content changes increased citation in generated answers.
Does GEO really increase visibility by 40%?
The paper reports an increase of up to around 40% from adding citations, quotations and statistics, measured on its own benchmark against 2023 and 2024 engines. That is a ceiling in tested conditions, not an average anyone should expect to reproduce.
What three techniques worked best?
Citing sources, adding quotations from credible sources, and adding specific statistics. All three make content more verifiable, which appears to be the underlying mechanism, and combining methods outperformed using any one alone.
Does keyword stuffing help in AI search?
No. The study tested it alongside the newer techniques and it performed poorly. That is one of the more useful findings, because the instinct when a new channel appears is to apply old tactics harder.
Is the study still relevant in 2026?
The mechanism is, the magnitudes are questionable. The experiments used engines as they existed in 2023 and 2024, and everything has changed since, including which systems people use. Cite it for direction with its date, not for current numbers.
Has anyone replicated it?
Not substantially on current engines, as far as I can establish. For a finding this widely cited that is a real gap, and a replication would be the most useful contribution anyone in this industry could make.
Does it prove schema markup helps AI visibility?
No. Schema was not among the interventions tested. There are good independent reasons to implement structured data, around entity clarity and feature eligibility, and this paper is not evidence for them.
Where can I read it?
Freely available on arXiv as 2311.09735, and in the ACM Digital Library as part of the KDD '24 proceedings. Table 1 holds the figures everyone quotes, and the paper is readable without a technical background.
Using the research honestly
Read the paper. It is free on arXiv, readable in twenty minutes, and it underpins most of what you will be told about generative engine optimisation, including by people who have not read it.
Then apply the three findings that earned their place: attribute your claims to named sources, replace vague language with specific figures, and quote real authorities. Stop stuffing keywords, which the same study found does not help.
The honest framing for your own content
Cite it with its date. A 2024 paper, experiments from 2023 and 2024, engines that have since changed.
Keep the words up to. Forty percent is a ceiling in tested conditions, not a forecast.
Do not stretch it. It says nothing about schema, backlinks, metadata files or publishing volume.
Measure your own site. A number from your own measurement beats a benchmark ceiling, because it is about the reader's situation.
If producing this kind of content is the constraint
Writing pages dense with attributed figures and real quotations is slower than writing marketing copy, which is part of why so few companies do it. If the constraint is production, Blazly SEO covers research, planning, drafting and direct publishing from about $39 a month.
The verification step stays yours. A generated draft can contain a plausible citation to a source that does not say what the sentence claims, and on a page built around being verifiable that failure is the one that matters most.
Summary
Almost every claim that generative engine optimisation lifts visibility by 40% traces to one peer-reviewed paper: Aggarwal et al., KDD 2024, from Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi. Across 10,000 queries and nine domains it found that adding citations, quotations from credible sources and specific statistics increased how often content was used in generated answers, by up to around 40%, while keyword stuffing performed poorly. The mechanism is that verifiable content is preferred over merely assertive content, and that should outlast any particular model generation. The magnitude is a ceiling from 2023 and 2024 conditions, not a forecast for your site. Cite it properly, with its date and its qualifier, and measure your own results for anything you intend to promise.