How to write a page an AI will quote
The short version
- The one peer-reviewed study of this question found adding quotations, statistics and source citations produced the largest visibility gains — and that keyword stuffing scored below the unmodified baseline.
- It also has real limitations: GPT-3.5-era models, a five-source competitive pool, English only. Treat it as a strong signal, not a rulebook.
- A controlled study of 1,885 pages found adding schema markup produced no meaningful change in AI citations. This is the most confidently over-recommended tactic in the category.
- Where a claim sits on the page matters: 44.2% of ChatGPT citations came from the first third of an article.
Most advice about writing for AI search is someone’s hunch dressed as a checklist. There is, however, one properly peer-reviewed study, and a handful of large controlled industry ones, and they are more interesting than the hunches — partly because of what they rule out.
What the research actually tested
“GEO: Generative Engine Optimization” was published at KDD 2024 by researchers from Princeton University and IIT Delhi, with two independent contributors. They built GEO-BENCH — 10,000 queries drawn from nine sources across 25 domains — and tested nine content modifications to see which ones increased a source’s visibility in a generated answer.1
The headline: “GEO can boost visibility by up to 40% in generative engine responses.”
That number gets quoted constantly and usually without the rest of the sentence. The more useful output is the ranking of what worked.
Three things stand out.
Adding quotations, statistics and source citations were the strongest interventions, producing what the authors describe as “a relative improvement of 30-40% on the Position-Adjusted Word Count metric.”1 All three do the same underlying thing: they make a passage self-contained and verifiable, so it can be lifted out of your page and still stand up.
Keyword stuffing came last, scoring 17.8 against an unmodified baseline of 19.5. The tactic that defined a decade of SEO actively reduced visibility.
And fluency optimisation — simply writing more clearly — placed third. That is a slightly humbling result for an industry that prefers technical levers.
Where this study is weak
Industry analysts have raised fair criticisms: the competitive pool was five sources per query, which likely amplifies relative gains; the optimisations were applied by an LLM rather than by writers; the primary experiments ran on GPT-3.5-turbo and predate current models; and it was English-only. Real-world validation covered a 200-sample subset on Perplexity. It remains the best evidence available, which is not the same as settled.
The tactic that did nothing
If you have read anything about AI search optimisation, you have been told to add structured data. It is the most common recommendation in the category.
Ahrefs tested it properly. They identified 1,885 pages that added JSON-LD schema between August 2025 and March 2026, matched them against 4,000 control pages, and measured citations for 30 days before and after.2
Their summary: “Adding schema produced no major uplift in citations on any platform.”
Two caveats matter. The study looked at pages already receiving 100+ AI Overview citations, so it does not tell you what happens to a page with no visibility at all. And schema has other jobs — rich results, entity disambiguation — that this study did not measure.
But as a lever for AI citations specifically, on pages that already have some, the controlled evidence says it is not one. You will find claims that schema produces multiples more citations; we could not trace those to a study with a disclosed method.
Where on the page a claim sits
Kevin Indig analysed 1.2 million AI answers and 18,012 verified citations to find where in a document ChatGPT tends to pull from.3
Within paragraphs the pattern inverts in a way worth knowing: 53% of citations came from the middle of a paragraph rather than its first or last sentence. Opening and closing sentences tend to be transitions and summaries; the substantive claim usually sits in between.
What this adds up to in practice
- Make each claim survive removal Write sentences that still make sense with no surrounding paragraph. "Retention improved" needs its context. "Weekly check-ins cut first-90-day attrition from 22% to 14% across 340 hires" does not. The second is quotable; the first is not.
- Attach a number and a source to the claims you want owned Statistics addition and citing sources were two of the three strongest interventions in the GEO study. They also happen to be what makes writing trustworthy to humans, which is a rare alignment.
- Put the answer near the top Nearly half of citations came from the first third. If your key claim is in a conclusion, it is in the least-cited part of the page.
- Give each page one job A page covering six topics competes with itself for which passage gets extracted. A page answering one question cleanly gives the model an obvious thing to lift.
- Write it better Fluency optimisation outscored technical terms, authoritative tone and unique-word density. Clear prose is a ranking factor now, which is a strange sentence to type.
- Do not stuff keywords It was the only tactic that scored below the unmodified baseline.
A word about writing for machines
There is a temptation to build pages that exist only to be extracted — thin, keyword-shaped, useless to a person who lands on them.
Set aside whether that is ethical; it is fragile. The measured wins in the research are things that also make content good for readers: quotes, evidence, sources, clarity. The measured loss is the tactic that games the machine at the reader’s expense. That is not a coincidence so much as a design goal on the model builders’ side, and they iterate faster than anyone optimising against them.
Questions we get asked
Should I remove my schema markup, then?
No. It does other jobs — rich results, entity clarity, feeding your own systems — and the study found no meaningful harm. The finding is narrower than it sounds: on pages that already get cited, adding schema did not increase citations. Stop selling it as an AI-visibility lever; keep it for the jobs it does.
Does the 40% figure mean I can expect 40% more visibility?
No. It is an upper bound from a benchmark with a five-source competitive pool on a 2023-era model. Expect the ranking of tactics to transfer better than the magnitudes.
How long should a quotable passage be?
Short enough to lift whole, long enough to carry its own context — usually one or two sentences. The test is not word count. It is whether the sentence still means something with everything around it deleted.
Does this work the same across models?
The GEO study used GPT-3.5 with a Perplexity subset; the citation-position study is ChatGPT; the schema study covered three surfaces and found different signs on each. Assume model-specific variation and measure per model rather than trusting a single playbook everywhere.
The short of it
Write fewer pages, each answering one question, with the answer near the top, stated in sentences that carry their own evidence. Cite your sources. Do not stuff keywords. Treat schema as plumbing rather than as a growth lever.
None of that is a trick, which is rather the point.
Sources
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, "GEO: Generative Engine Optimization", ACM SIGKDD 2024, arXiv 2311.09735. Affiliations on the paper: Princeton University, IIT Delhi, and two independent researchers. GEO-BENCH: 10,000 queries, nine sources, 25 domains; primary experiments on GPT-3.5-turbo with a 200-sample Perplexity validation.
- Ahrefs (Louise Linehan & Xibeijia Guan), "We tracked 1,885 pages adding schema. AI citations barely moved.", 11 May 2026. Matched difference-in-differences, 1,885 treated pages against 4,000 controls.
- Kevin Indig, reported in Search Engine Land, "44% of ChatGPT citations come from the first third of content", 18 February 2026. 1.2 million AI answers, 18,012 verified citations.