# Word Count Is a Weak Predictor of AI Citations

> Ahrefs found near-zero correlation between length and AI Overview citations. Where other studies differ, and which page changes have evidence behind them.

URL: https://llmagnet.com/blog/word-count-ai-citations  
Author: Ben Cohen, Founder & CEO at LLMagnet  
Published: 2026-10-28  
Updated: 2026-11-10  
Category: Research

Word count is a weak predictor of AI citations: Ahrefs reported a near-zero correlation (0.04 Spearman) for pages cited in AI Overviews, and 53.4% of the 174,048 pages it studied had under 1,000 words. So don’t pad pages to hit a length target. Other studies find a small edge for longer pages, and the better-supported changes are where you put the answer, tables, and first-hand detail.

## Contents

1. What the length data shows
2. Length depends on the page type
3. Completeness isn’t length
4. Put the answer early
5. Tables and headings
6. Facts, sources and quotations
7. First-hand detail and freshness
8. A worked example: cutting a guide
9. Common mistakes
10. A checklist for this week
11. FAQ

## What the length data shows

[Ahrefs’ December 2025 study](https://ahrefs.com/blog/short-vs-long-content-in-ai-overviews) looked at pages cited in 560,346 AI Overviews, and extracted usable text from 174,048 of them. This is vendor data.

![Share of pages cited in AI Overviews by length: under 350 words 16.6 percent, 350 to 1,000 words 36.8 percent, 1,000 to 2,000 words 30.6 percent, over 2,000 words 16.0 percent.](https://llmagnet.com/images/blog/charts/wordcount-length-share.png)

*Source: Ahrefs, December 2025. Vendor data.*

Ahrefs reports a Spearman correlation of 0.04 between word count and being cited, and the same 0.04 between word count and citation position. Both are close to zero. Because every page in the set was cited, the position figure is the one the dataset can directly support (our reading). The longest cited blog post was about 3,500 words, so the data says little about very long guides.

Cited pages averaged 1,282 words, against 1,188 for ranking content in Ahrefs’ figures. One caution on the sample: text could be extracted for about 10% of the 1,677,876 cited URLs, and Ahrefs doesn’t discuss how that filter shaped the result. Don’t read the 53.4% under 1,000 words as “short wins” either. Ahrefs itself notes the chicken-and-egg problem: pages may be cited at a given length because that’s what exists.

Two other studies complicate this. [Machine Relations Research](https://machinerelations.ai/research/do-ai-engines-cite-longer-pages-2026) paired 706 cited pages with other pages on the same sites, across five engines. Cited pages had a median 1,560 words against 968 (a paired median difference of 226 words), though they warn that those pages may simply rank better. [TJ Digital](https://tjrobertson.com/ai-citation-study/), an agency, found longer pages did “a little better” in AI Overviews and AI Mode, and no difference in ChatGPT.

In Machine Relations’ data, length was no minimum: 24.5% of cited pages ran under 800 words, and 51.8% ran 1,500 or more. Among the 520 pairs where both pages had at least 300 words, the gap narrowed to a median 1,812 against 1,462.

| Study | Compares | Length finding | Why it differs |
| --- | --- | --- | --- |
| Ahrefs | Cited pages only, AI Overviews | 0.04 correlation | No uncited pages to compare |
| Machine Relations | Cited pages against a random page on the same site, five engines | Cited pages longer | Comparison pages can be thin product pages or archives |
| TJ Digital | Cited pages against Google top-10 pages AI passed over | A little better in AI Overviews and AI Mode, no difference in ChatGPT | Comparison pages already rank |

The studies aren’t really in conflict. Ahrefs looked only at pages that were cited, so it can show that short pages get cited, but not whether a longer version would do better. The fair conclusion is that length has a small effect at most, and that it shouldn’t be the goal.

## Length depends on the page type

Ahrefs’ medians show why one target can’t work. Only articles and audio have medians above 1,000 words.

![Median words of pages cited in AI Overviews: audio transcripts 1,226, articles 1,166, documents 676, category pages 534, interactive tools 507, video 407, user-generated content 387, core pages 317, product listings 315.](https://llmagnet.com/images/blog/charts/wordcount-page-types.png)

*Source: Ahrefs, December 2025. Vendor data.*

For ecommerce, a 315-word listing is normal, so spend the effort on specifications, price and availability rather than extra paragraphs. For B2B, core pages are short too, while guides and research carry the longer explanations. Ahrefs also found that over 95% of short-content citations land in the top three positions, so a short page isn’t confined to the bottom of the list.

## Completeness isn’t length

The more useful question is whether the page answers the question. In TJ Digital’s blind ratings by AI raters, half of passed-over pages never answered the question in the rated portion, against 23% of cited pages. Cited pages scored 2.3 for completeness on a five-point scale, against 1.7. A long page can fail this test, and a short one can pass it.

An [October 2026 preprint](https://arxiv.org/abs/2610.15033) by Mahe Chen and Nick Koudas points the same way. It covers 158,099 answers on ChatGPT, Gemini and Perplexity. On ChatGPT, page properties explained 2.8% of the variation between pages on one site, while which pages its search retrieved explained 32%. Every page ChatGPT cited had been retrieved by its own search. Content helped mainly by matching those searches in meaning. It isn’t peer reviewed, but it suggests length matters less than covering the narrow questions behind a prompt.

## Put the answer early

[Kevin Indig’s analysis of ChatGPT citations](https://www.searchenginejournal.com/the-science-of-how-ai-pays-attention/567597/) matched 18,012 citations to their place on the page. 44.2% came from the first 30% of the text, 31.1% from the middle and 24.7% from the last third. Within paragraphs the pattern is less tidy: 53% of citations came from mid-paragraph, 24.5% from the first sentence and 22.5% from the last.

![Share of 18,012 ChatGPT citations by where the cited text sits on the page: first 30 percent of the text 44.2 percent, middle 31.1 percent, final third 24.7 percent.](https://llmagnet.com/images/blog/charts/wordcount-citation-position.png)

*Source: Kevin Indig, Search Engine Journal, February 2026.*

This covers ChatGPT only, and shows where citations occur, not why. It does support the old rule of leading with the answer. Give each section a clear first paragraph, and don’t force every key fact into a first sentence.

TJ Digital’s blind ratings support this: a direct answer in the opening appeared on 20% of cited pages against 5% of passed-over pages (23% against 14% in an earlier round, among pages pulled up by the same answer). Indig’s method has a limit worth knowing. It matches sentences by similarity to the answer, which shows they resemble it, not that the model used them.

## Tables and headings

[TJ Digital’s October 2026 study](https://tjrobertson.com/ai-citation-study/) compared 21,874 cited pages with 4,507 Google top-10 pages that AI answers passed over, adjusting for Domain Rating and page type. Pages with a table had 69% higher odds of being cited, and 88% higher in ChatGPT. Lists had 22% higher odds and FAQ sections 26%.

On headings, pages whose headings matched the narrower searches the AI generated were cited more, and pages with more H2 sections did a little better. These are associations from agency data. Use a table where your content is already comparative, such as plans, specs or options, and write headings as the specific question a buyer would ask. In Indig’s data, cited text contained question marks 18% of the time against 8.9% for skipped text, and 78.4% of those citations came from headings.

## Facts, sources and quotations

The strongest controlled evidence comes from the [Princeton GEO study](https://arxiv.org/abs/2311.09735) (Aggarwal et al., KDD 2024), which tested nine methods on the 1,000-query test split of a 10,000-query benchmark. Against baseline, on Position-Adjusted Word Count, Quotation Addition gained about 41%, Statistics Addition about 31%, and Cite Sources about 28%. Keyword Stuffing lost about 8%. These percentages are calculated from the paper’s Table 1, and they are tests on a benchmark, not proof for live engines.

Cite Sources also shows why results depend on your position: it gained 115.1% for pages at rank 5 and lost 30.3% for pages at rank 1.

In Indig’s ChatGPT data, heavily cited text had 20.6% proper nouns, against 5% to 8% in normal English, and definitive language (verbs such as is, are, refers to) in 36.2% of cited intros against 20.2% of skipped ones. Name the product, the person, the date and the source, and say plainly what something is.

## First-hand detail and freshness

In the same [TJ Digital data](https://tjrobertson.com/ai-citation-study/), first-hand experience went with 45% higher odds of being cited, and being published or updated within the past 12 months with 88% higher odds. The author says these are associations.

Real updates are the point here: add the new number, drop what’s out of date, and change the date only when the content changed. Replace claims such as “many teams find” with what you saw: how many customers, over what period, with what result.

## A worked example: cutting a guide

The company and figures below are made up for illustration. Fennick has a 2,600-word guide to choosing a payroll provider.

| Before | After |
| --- | --- |
| The answer arrives in paragraph nine, after a history of payroll | The first two sentences say how to choose |
| Headings such as “Key considerations” | Headings such as “What does multi-country payroll cost?” |
| Pricing described in prose | One table of plans, countries and prices |
| Says nothing about contractors, a common buyer question | A short section on contractors |
| 2,600 words | About 1,500 words |

The page lost about 40% of its words and answered more questions. Whether it earns more citations is something to measure, so compare it with untouched pages after 60 days.

## Common mistakes

- **Setting a minimum word count.** It rewards filler, and no study here supports a threshold.
- **Cutting every long page.** Ahrefs’ data says little about blog posts above about 3,500 words, and a technical reference may need its length.
- **Bolting on an FAQ to add bulk.** TJ Digital’s 26% higher odds for FAQ sections is an association, so add one only for questions readers ask.

## A checklist for this week

1. Pick your five most important pages. Remove any section that exists only to add length.
2. Rewrite each opening so the first two sentences answer the page’s main question.
3. Turn one comparison in each page into a real HTML table.
4. Rewrite vague headings as the specific questions buyers ask.
5. Give every statistic a source, a date and a link, and add one quotation from a named expert you actually spoke to.
6. Add one first-hand detail per page that a competitor couldn’t copy.
7. Check citations for those pages after 60 days, alongside five untouched pages.

None of this requires more words. LLMagnet measures AI visibility across seven engines and its agent prepares fixes for approval. Its [free audit](https://ai-visibility.llmagnet.com) is one way to get a baseline before you start.

## FAQ

### Does longer content get cited more by AI?

Not by much. Ahrefs reported a near-zero correlation (0.04) between word count and being cited in AI Overviews. Two smaller studies found a modest edge for longer pages, which may reflect ranking.

### How long should a page be for AI citations?

There’s no target length. In Ahrefs’ data, 53.4% of cited pages had under 1,000 words and 16% had over 2,000. Write as much as the question needs.

### Where on the page do ChatGPT citations come from?

In Kevin Indig’s analysis of 18,012 ChatGPT citations, 44.2% came from the first 30% of the text, 31.1% from the middle and 24.7% from the final third.

### Do tables help AI citations?

They’re associated with it. TJ Digital’s study of cited and passed-over pages found 69% higher odds for pages with a table, or 88% in ChatGPT, after adjusting for Domain Rating and page type.

### Do statistics and sources help?

In the Princeton GEO benchmark, Statistics Addition gained about 31% and Cite Sources about 28% on Position-Adjusted Word Count. These are benchmark tests, not proof for live engines.

### Is there a proven fact-density ratio?

We found none, so we don’t recommend targeting a ratio.

### Should a product page be as long as a blog post?

No. In Ahrefs’ data the median cited product listing had 315 words and the median cited article had 1,166. Match length to the page type and the question.