What Makes a Paragraph Citeable

Most advice about writing for AI systems traces back to one paper. Almost none of it quotes the paper correctly.

I went back to the primary source last month while building the editorial standard Trovance runs on, expecting to confirm the received wisdom and move on. Instead I found that the single most-cited number in the entire field describes an experimental condition that nobody mentions when they quote it. That discovery changed how we structure content, and it is the reason this piece exists.

What makes a paragraph citeable?

A paragraph is citeable when it contains a verifiable, self-contained claim that a model can lift without carrying surrounding context along with it. The strongest empirical signals are the presence of statistics, direct quotations from credible sources, and explicit citations. Fluent, plainly-written prose helps independently of content. Keyword density does not help at all.

That is the short version, and it is well supported. The longer version is more useful, because three of the four things you have probably been told about citeable paragraphs are either misattributed, overstated, or invented.

What did the founding GEO study actually test?

The paper that started the field is GEO: Generative Engine Optimization, presented at ACM SIGKDD in Barcelona in August 2024. It is worth being precise about the authorship, because the field is not: this is routinely called "the Princeton study," but lead author Pranjal Aggarwal was at IIT Delhi, with co-first-author Vishvak Murahari, Karthik Narasimhan, and Ameet Deshpande at Princeton and two independent researchers. A collaboration, not a Princeton paper. In an article about sourcing discipline, that distinction is not pedantry. It is the whole point.

The researchers built GEO-bench, a benchmark of 10,000 queries drawn from nine datasets across 25 domains, split 8,000 / 1,000 / 1,000 for training, validation, and test. They simulated a two-stage generative engine: Google returned the top five sources for a query, then GPT-3.5-turbo synthesized a cited answer from them. Five responses were sampled at temperature 0.7 to reduce statistical noise, across five random seeds.

They then applied nine content modifications to one randomly-chosen source per query and measured what happened to that source's visibility in the generated answer. The primary metric, Position-Adjusted Word Count, weights how much of the answer traces to your source and where in the response it appears.

Two things about that setup deserve your attention before any of the findings do. The engine was GPT-3.5-turbo, a model that is now several generations old. And only the top five sources entered the context window, because of cost. Both facts constrain how far the results travel.

Which changes moved the needle, and by how much?

Against an unoptimized baseline of 19.3, three modifications clearly outperformed: Quotation Addition reached 27.2, Statistics Addition reached 25.2, and Cite Sources reached 24.6. The authors report these as 30–40% relative improvement on Position-Adjusted Word Count and 15–30% on their subjective-impression metric. The best-performing method improved on baseline by 41%.

The full picture is more interesting than the headline:

Three findings here get far less attention than they deserve.

Keyword stuffing made things worse. At 17.7 against a 19.3 baseline, the most durable habit in twenty-five years of SEO practice actively suppressed visibility. The authors note it performed 10% below baseline in their Perplexity.ai replication as well.

Fluency optimization scored 24.7 without adding a single new fact. This method only improved the readability of existing text. It landed within a rounding error of Cite Sources, which required adding real citations. Clear writing is not the soft, unmeasurable part of this work. It performed nearly as well as the hard part.

Sounding authoritative did almost nothing. The Authoritative modification, which rewrote content in a more persuasive and confident register, produced no significant improvement. The authors read this as evidence that generative engines are already fairly resistant to tonal manipulation. If your content strategy rests on sounding like an expert rather than citing like one, this is your result.

The combination data points the same direction. Pairing Fluency Optimization with Statistics Addition produced the best result of any pair tested at 35.8%, and Cite Sources averaged 31.4% across combinations despite being the weakest of the top three in isolation.

Why does the "+115% for page five" number mislead nearly everyone who quotes it?

Because it comes from a separate experiment in which every source was optimized simultaneously, and in that same experiment the top-ranked page lost 30.3% of its visibility.

This is the finding I referred to at the start. Section 5.2 of the paper describes a deliberately different condition. The authors wanted to know what happens when GEO becomes universal, so they optimized all source contents at once rather than one at a time. Under those conditions:

Read the first row again. When everyone optimizes, the incumbent loses ground and the fifth-place page gains it. The +115.1% is not a return you earn by adding citations to your page. It is a description of how a fixed pool of visibility gets redistributed once the whole field is doing the same thing.

The paper frames this optimistically, as evidence that generative engines could democratize a space where backlinks and domain authority have long favored incumbents. That reading is fair. But the industry has taken a finding about redistribution under saturation and marketed it as individual uplift, which is a different claim with a different expected value. If you are the fifth-ranked page today, the honest version is: this helps you most in a world where your competitors are doing it too, and it helps you least in a world where only you are.

I have watched teams build quarterly content plans on the misread version of that number. The plan is not wrong so much as it is priced incorrectly.

Does front-loading actually matter?

Yes, and the reason is mechanical rather than editorial. Language models do not read their context window evenly. Liu et al., published in Transactions of the ACL in 2024, documented a U-shaped performance curve: models handle information best at the beginning or the end of the input context and measurably worse in the middle.

The sharpest number in that paper is the one worth internalizing. When the relevant passage sat in the middle of the context, GPT-3.5-turbo's accuracy on multi-document question answering fell below its closed-book score of 56.1%. The model performed worse when handed the document containing the answer than when handed nothing and asked to rely on memory. Being retrieved is not the same as being used.

This holds across models, across context lengths, and even for models explicitly built for long contexts. It is a property of how attention distributes, not a quirk of one system. Follow-up work has replicated the primacy and recency bias consistently.

Note what this does and does not license. It supports putting your most extractable claims early. It does not support the specific percentages that circulate about what share of citations come from the first third of a page. Those figures come from vendor blog analyses of proprietary datasets, not from published research, and I would not build a standard on them.

Where does the 40-to-60-word rule come from?

Practice, not research. No peer-reviewed source establishes 40–60 words as an optimal length for extractable answers, and the GEO paper does not test passage length at all.

I want to be direct about this because I have repeated the number myself, and it is worth knowing what kind of claim it is. The convention is a reasonable synthesis of three real constraints: an answer needs enough context to stand alone, short enough to sit inside a synthesized response without truncation, and self-contained enough to survive being lifted from its surroundings. Forty to sixty words happens to satisfy all three most of the time. That makes it a useful working rule and a poor thing to cite as a finding.

The underlying test is better than the number anyway. Cover everything above and below a paragraph. Does it still make complete sense, and does it still answer a question someone would actually ask? If yes, it is extractable at whatever length it happens to be. If it opens with "this means that" or "as we saw above," it is not, and no word count will fix it.

Do these tactics still work now that everyone knows them?

Less than the 2024 results suggest, and the best evidence for that comes from the most rigorous attempt to replicate them. C-SEO Bench, presented at the NeurIPS 2025 Datasets and Benchmarks track by Puerto, Gubri, Green, Oh, and Yun, tested conversational-SEO methods across two tasks and six domains, and formalized something the original work did not: varying rates of adoption among competing actors.

Their finding is blunt. Most current C-SEO methods are largely ineffective, and frequently have a negative impact on document ranking, the opposite of what practitioners expect. Traditional retrieval-side improvements, the unglamorous work of being the source that actually best answers the query, had a significantly larger effect on citation ranking than any content-modification trick.

And the adoption curve behaves exactly as you would fear. As more actors adopt these methods, marginal gains shrink. The authors describe the dynamic as congested and zero-sum, which matches how the original paper's rank-1 penalty behaved under universal optimization. Two independent research groups, working from different benchmarks, arrived at the same structural conclusion: this is a tactic with decaying returns, not a lever with fixed output.

That does not mean stop writing citeable paragraphs. It means stop treating structure as the strategy. Structure is the floor. It makes good evidence extractable. It cannot manufacture evidence that is not there, and once every competitor has cleared the same floor, the only thing left to differentiate on is whether your page actually contains a better answer.

What should you actually change on the page?

Six changes, ordered by the strength of evidence behind them rather than by how satisfying they are to implement.

1. Add real statistics with named primary sources. Statistics Addition was among the top three methods in the original study, and the effect held in the Perplexity replication at +37% on subjective impression. A number without an attributable source is not a statistic; it is a claim.

2. Quote credible sources directly. Quotation Addition was the single strongest individual method at 27.2 against a 19.3 baseline, and the strongest in the real-engine replication at +22%.

3. Cite your sources explicitly. Weakest of the top three alone, but the best team player, averaging 31.4% in combination with other methods.

4. Make the prose clearer. Fluency Optimization reached 24.7 with no new information added. Editing is a visibility lever.

5. Put your most extractable claim in the opening third. Justified by the U-shaped attention curve, not by any specific circulating percentage.

6. Delete the keyword density work. It scored below baseline in both the primary experiment and the Perplexity replication.

What is conspicuously absent from that list: adopting an authoritative tone, adding technical vocabulary, and adding unusual words. All three were tested. None produced meaningful gains.

How do you tell whether it worked?

Not by a single score, and not by whether the page went live. The honest measurement question is whether the market's explanation of your company changed, and that requires preserving what the answer looked like before you shipped anything.

The mechanism to run is unglamorous: record the buyer question, capture the raw answers and citation patterns across engines and repeated runs, identify what evidence was actually missing, publish the asset that should exist, then rerun the same question and record what changed and what did not. Repeated runs matter more than most teams assume, because the variance between two runs of an identical prompt is often larger than the change you are trying to detect. The original researchers sampled five responses across five seeds for exactly this reason. A single before-and-after screenshot is not evidence.

This is the discipline Trovance is built around, and the reason we treat visibility as a distribution rather than a number. A mention is not a citation. A citation is not a recommendation. A recommendation is not a shortlist. Collapsing those into one score is how teams end up confident about a result they have not actually observed.

How does Trovance make your content citeable?

By treating citeability as an evidence problem rather than a formatting problem. Trovance observes how AI systems currently explain your company, identifies which claims lack the proof a model would need to repeat them, and turns those gaps into assets built the way the research says extraction works.

Everything in this article is the floor Trovance builds against. Drafts stay tied to an approved claim and a named source, because Statistics Addition and Cite Sources only produced their results when the numbers were real and attributable. Answer blocks are written to survive removal from their surroundings. The strongest claim lands in the opening third, where the attention curve is on your side. None of that is the hard part, though, and the product does not pretend otherwise.

The hard part is knowing what should become publicly true and provable in the first place. A visibility gap is not a content brief. Being absent from an answer might mean your proof is thin, or your positioning is unclear, or a competitor's third-party coverage is stronger, or the prompt describes a buyer you should not be chasing. Trovance diagnoses which one it is before recommending work, because publishing a beautifully structured page against the wrong diagnosis is how teams generate volume and call it progress.

Then the part worth being blunt about: Trovance does not guarantee citations, and I would not trust a tool that claimed to. What it does is preserve the original answer and its citation pattern before you publish, so that when you rerun the same question you can tell whether the market's explanation of your company actually changed. Given how much run-to-run variance sits in these systems, that before-state is the only thing that separates a measured result from a flattering screenshot.

Humans keep the consequential calls throughout. Whether a claim is true, whether the proof is sufficient, what can be public, and what deserves to ship are decisions the system supports rather than makes.

If you want to see the starting point for your own company, run your first market question in Trovance and look at the raw answer before you change anything on the page.

FAQs

What makes a paragraph citeable by an AI system?

A citeable paragraph makes a verifiable, self-contained claim that survives removal from its surrounding context. The strongest empirical drivers are direct quotations from credible sources, statistics with named attribution, and explicit citations. Clear, fluent prose contributes independently. Keyword density measurably reduces visibility.

Did the Princeton study prove GEO increases visibility by 40%?

It found up to 40% improvement, and calling it the Princeton study is imprecise. The paper is a collaboration whose lead author was at IIT Delhi. The 40% figure is a maximum, not an average; the three strongest methods produced 30–40% relative improvement on one metric and 15–30% on another.

Does adding statistics really improve AI citation rates?

In controlled testing, yes. Statistics Addition scored 25.2 against an unoptimized baseline of 19.3 in the GEO-bench experiments, and improved subjective impression by 37% in a replication on Perplexity.ai. The effect requires genuine attributable figures, not numbers presented without a verifiable source.

Should content be structured in 40-to-60-word answer blocks?

It is a useful working convention with no research establishing that specific range. The underlying requirement is that a passage stand alone without surrounding context. Test by covering the paragraphs above and below and asking whether the remaining text still answers a real question completely.

Why did keyword stuffing perform worse than no optimization?

Generative engines do not rely on keyword matching the way ranking algorithms did. In GEO-bench testing, keyword stuffing scored 17.7 against a 19.3 baseline, and performed roughly 10% below baseline in the Perplexity replication. The technique adds noise without adding retrievable substance.

Does research support putting key information early on a page?

Yes, mechanically. Liu et al. documented a U-shaped attention curve in which models use information at the beginning and end of context far better than the middle. Mid-context placement drove accuracy below the closed-book baseline of 56.1%, meaning the model performed worse with the document than without it.

Do GEO tactics still work as adoption increases?

Their returns compress. C-SEO Bench, presented at NeurIPS 2025, found most conversational-SEO methods largely ineffective and often harmful to ranking, with marginal gains shrinking as more actors adopt them. Improving genuine retrieval relevance outperformed content-modification tactics across all six tested domains.

Trovance shows how AI systems explain, compare, and recommend your company, identifies the evidence keeping you off the right shortlists, and turns those gaps into proof-backed content your team can publish. See how AI sees your brand.