Ahrefs analyzed 174,048 pages behind 560,346 Google AI Overviews in December 2025 and found almost no relationship between word count and getting cited. The correlation was 0.04, statistically close to zero. Yet the search marketing world keeps repeating the same advice: write 2,000, 3,000, sometimes 10,000 words if an AI engine should quote you. That habit survives because it borrows from classic SEO, where longer pages once correlated loosely with topical depth and rankings. Generative engines don't rank whole pages. They synthesize an answer from fragments, and where a fragment sits inside that answer matters more than how long the page around it runs.
What Ahrefs found when it checked 174,000 citations
Across the Ahrefs study of 174,048 cited pages, the average length was 1,282 words, and length barely moved with citation likelihood: the Spearman correlation between word count and appearing in an AI Overview was 0.04. More than half of citations, 53.4%, went to pages under 1,000 words. Only 16% came from anything over 2,000. Position inside the overview told the same story. Sources ranked first averaged 1,270 words; sources ranked fourth through tenth averaged 1,690, a gap too small to plan a content strategy around.
The variation that did show up came from content type, not length. Listing pages ran a median of 315 words, core product or service pages 317, full articles 1,166, and audio transcripts 1,226. Most cited content sits well under the 2,000-word mark that "comprehensive guide" advice usually targets, a habit that made more sense back when classic SEO rewarded depth signals a crawler could count. It lines up with what VizibleAI has written about answer-first content: a citable paragraph is one that answers a specific question completely, not one padded to look thorough.
Position-Adjusted Word Count: the metric researchers built instead
A response's position-adjusted word count weighs each cited source by how early it appears in the generated answer, not by the total length of the page it links back to.
That's the working definition behind Position-Adjusted Word Count, introduced in the original Generative Engine Optimization paper published by researchers at Princeton and IIT Delhi, Pranjal Aggarwal, Vishvak Murahari, and colleagues, in their GEO study. The authors built GEO-bench, a benchmark of 10,000 real user queries across 25 domains split into 8,000 training, 1,000 validation, and 1,000 test queries, specifically because a normal search-ranking metric breaks once a single response blends and cites several sources at once. The metric discounts a source exponentially the later it appears in the synthesized answer, which is why PAWC, not raw word count, is what the paper's optimization experiments actually measured.
The techniques that moved the needle, and the one that didn't
Testing nine content-optimization techniques against GEO-bench, the researchers found three delivered the largest PAWC gains: quotation addition, statistics addition, and citing sources, each producing roughly 30 to 40% higher position-adjusted visibility. Fluency optimization, tightening grammar and readability without changing the substance, delivered a smaller 15 to 30% lift. Keyword stuffing, the one technique borrowed directly from old SEO habits, produced minimal or negative improvement.
Those top three techniques share a trait: each adds new, checkable information rather than restating what's already on the page. A quote is new. A cited statistic is new. A named source is new. Length was never the variable doing the work; information density was, and that's close to what VizibleAI found from the retrieval side in its piece on how AI engines chunk content: engines break pages into passages before they look at total length at all, so a page's real unit of competition is the passage, not the article.
Where the research gets pushed back on
The GEO paper isn't without critics. SEO analyst Tylor Hermanson's review of the study points out that the three best-performing techniques all involved adding new content to a page, while most of the weaker techniques only edited existing text, which makes it hard to separate "this specific technique works" from "adding any new, unique information works." He also flags that the researchers allowed fabricated statistics and quotes in their test prompts, an approach that likely inflates results in a way real content can't replicate, since a genuine quote or stat usually already exists somewhere else a model has already seen. Both points are fair, and they argue for treating the 30 to 40% figures as directional rather than a guarantee that will hold on any given site.
What separates cited content in 2026, beyond word count
A separate analysis from AI-visibility tracker OtterlyAI, covering more than a million citations across ChatGPT, Perplexity, and Google AI Overviews in January and February 2026, points at the same conclusion from a different angle. Editorial, analysis-style content dominates citations on every platform the report covers, ahead of brand-published material. Pages using schema markup and clear content chunking earned three to five times more citations than pages without either. The report's authors are blunt about the ordering: a well-written page an AI crawler can't parse or reach doesn't help, whatever its length.
That's consistent with a pattern VizibleAI has covered on the content-strategy side in its piece on content cannibalization and AI citations: adding more pages, or more words to an existing page, doesn't substitute for making the content that already exists easier to extract and verify.
What this means for your content calendar
None of this makes length irrelevant. Some topics genuinely need more explanation, and cutting a page down to hit an arbitrary short-form target is its own kind of padding. The finding is narrower than that: length is not the lever to pull. If a page isn't getting cited, adding a named statistic with its source linked, a real quote, or a specific, checkable claim will move a metric like PAWC more reliably than doubling the word count would.
A GEO audit that scores a page against those factors, rather than against a word-count target, is a better use of an editorial calendar's time than another ultimate-guide rewrite.
Frequently Asked Questions
Does longer content rank better in AI Overviews?
No. Ahrefs' December 2025 analysis of 174,048 cited pages found a Spearman correlation of just 0.04 between word count and citation, which is effectively no relationship. More than half of citations, 53.4%, went to pages under 1,000 words, and only 16% came from pages over 2,000 words.
What is Position-Adjusted Word Count?
Position-Adjusted Word Count (PAWC) is a metric from the original Princeton and IIT Delhi GEO paper. It weighs a cited source by how early it appears inside an AI-generated answer, using an exponential decay function, rather than by the raw length of the page the source links back to.
What content changes actually improve AI citation rates?
Testing against the GEO-bench benchmark, researchers found quotation addition, statistics addition, and citing sources each produced roughly 30 to 40% higher position-adjusted visibility, the largest gains of any technique tested. Fluency optimization added a smaller 15 to 30% lift, and keyword stuffing produced minimal or negative results.
Is there an ideal word count for content that gets cited by AI?
Not a fixed one. Ahrefs found average cited content sits around 1,282 words, but medians vary heavily by content type: about 315 words for listing and product pages versus over 1,100 words for full articles. Match length to what the topic requires, not a target number.
How reliable are the 30-40% visibility-lift numbers from the GEO paper?
Treat them as directional rather than exact. SEO analyst Tylor Hermanson's critique notes the top-performing techniques all added new information to a page, and the study's test prompts allowed fabricated statistics and quotes, both of which likely inflate the measured gains compared with real content.
What should content teams optimize for instead of word count?
Extractability: whether a page contains specific, checkable, quotable information positioned where a generative engine's retrieval step can find and lift it into an answer, rather than optimizing for length or keyword coverage the way classic SEO once rewarded.