AI Search

A word-count benchmark is not a writing target

Explore 20 published page-word-count percentiles and learn how to use a benchmark when reviewing coverage, evidence and AI visibility without setting an unsupported length quota.

Two paper documents of different sizes stand beside a magnifier and an evidence folder

Key takeaways

  • HTTP Archive’s July 2025 figures describe rendered page word counts across separate device and page groups. [1]
  • A median or upper percentile describes the sampled pages; it does not establish a writing target or an AI-citation threshold. [1]
  • Google’s people-first content guidance explicitly says it has no preferred word count. [2]
  • Before expanding a page, identify the reader question it leaves unanswered and the evidence needed to answer it.

Use a word-count benchmark to investigate a page, then choose its length from the reader’s needs and the available evidence. A broad web median cannot tell you how many words your next guide requires. It can describe the pages in a sample, but turning that observation into an editorial quota needs evidence the benchmark does not provide.

HTTP Archive’s 2025 Web Almanac offers a useful example. Its SEO chapter reports rendered visible word-count percentiles from the July 2025 crawl, separately for desktop and mobile home and secondary pages. The chapter was published on January 15, 2026 and updated on June 9, 2026. The observations below remain a historical snapshot of those sampled pages. [1]

What the published word counts actually show

The interactive chart retains all 20 selected values: five percentiles for each of four page groups. P50 is the median; P10, P25, P75 and P90 mark other positions in the distribution. These are approximate quantiles, so they should not be read as exact cutoffs that divide every group into precise shares.

Rendered visible page words by percentile

HTTP Archive July 2025 crawl. D = desktop; M = mobile; home = homepage; inner = sampled secondary page. P50 is the median, not an average.

1,313984.75656.5328.250P10P25P50P75P90Percentile
  • D home
  • M home
  • D inner
  • M inner
Full data table
Rendered visible page words by percentile · words
PercentileD homeM homeD innerM inner
P1075697664
P25196177161142
P50413378339323
P75760703668675
P901,2951,1911,2321,313

Approximate quantiles from the published results linked at figures 5.44 and 5.45. Values describe separate sampled page groups. They are not writing targets, article-body counts or AI-citation thresholds. Source: HTTP Archive, 2025 Web Almanac: SEO. © Web Almanac, Apache 2.0. Ansvra normalized the selected data and recreated this chart. [1]

The complete table and chart download use the same published values. D means desktop and M means mobile. Home means homepage; inner means the query’s sampled secondary-page group. Secondary pages are not necessarily articles. The measure is rendered visible words on the page, rather than an isolated count of the article body. Source: HTTP Archive, 2025 Web Almanac: SEO, figures 5.44 and 5.45. [1]

The four medians range from 323 words for mobile secondary pages to 413 for desktop homepages. That range is a comparison between sampled groups. It does not identify a suitable range for a product guide, policy page or support answer, and it is not evidence that a page outside it is too long or too short.

Even the desktop/mobile comparison needs context. At P90, the secondary-page values are 1,232 words for desktop and 1,313 for mobile. These are separate group quantiles. They do not show that an individual mobile page gained 81 words compared with its desktop version. The published summaries do not pair pages for that calculation. [1]

Spread is useful context, not an acceptable writing band

Ansvra calculated two simple spans from the published values. P75 minus P25 describes the distance between the approximate upper and lower quartiles. P90 minus P10 shows a wider distance across the distribution. The table includes every group used in these calculations; all values are in words.

GroupP75 − P25P90 − P10
Desktop home5641,220
Mobile home5261,122
Desktop secondary5071,156
Mobile secondary5331,249

For example, the desktop-home interquartile span is 760 − 196 = 564 words. The mobile-secondary P90–P10 span is 1,313 − 64 = 1,249 words. These are differences between quantiles, not observed changes on the same pages. The breadth of the distributions is a reason to ask what each page does before copying a central value into a brief.

A percentile does not rate content quality. Nor can these five positions produce an accurate average for the entire distribution. Averaging the group medians would create another number without showing which page purpose it should guide. Keep the groups and the original values visible when discussing the benchmark.

What the sample can and cannot represent

The linked published results provide these qualifying query-group counts. They offer important context about the scale of the selected groups, but they are not a count of articles or a list of pages known to receive AI citations.

GroupQualifying pages
Desktop home11,012,179
Mobile home13,902,361
Desktop secondary10,176,398
Mobile secondary12,346,476

The original query counts distinct qualifying pages after filtering its custom-metric records. It does not separately publish how many pages have a non-missing rendered word-count value. The counts above therefore should not be relabeled as exact word-count denominators. That more specific count remains unavailable. Each group count repeats across the five percentile rows in the source worksheet; adding those repetitions would multiply the sample incorrectly. [1]

HTTP Archive’s methodology describes a sampling frame drawn from Chrome UX Report origins, separate desktop and mobile samples, and home and secondary pages. Its lab loads use an empty cache in a logged-out state. That differs from observing everything a returning or signed-in visitor may see. The selected results provide no country, language, industry or article-format breakdown.

Most importantly for an AI visibility team, this selected dataset contains no ranking, citation, engagement or conversion outcome. It cannot demonstrate that one word count performs better than another. It also cannot show a trend from one crawl alone. Those limits do not make the benchmark useless; they define the question it answers: how much rendered visible text appeared at selected positions in these sampled distributions.

Google’s guidance gives a different basis for the decision

Google’s people-first content guidance asks creators to examine usefulness, completeness, original value and whether the content helps the intended audience. In its warning signs for search engine-first content, it explicitly rejects the premise that Google has a preferred word count. The English documentation version checked for this article was updated on October 5, 2026. [2]

That is first-party guidance about Google Search. It is not a study of every AI assistant, and it does not establish that length has no effect on usability or comprehension. A detailed decision may need substantial explanation. A straightforward policy question may need a concise answer. The useful editorial test is whether the material helps the reader complete the task and whether its claims have support.

The archive data and Google’s guidance serve different roles here. One describes page-word-count distributions. The other explains how Google advises creators to evaluate content. Neither supplies a formula that predicts whether an assistant will cite a particular page.

An editor arranges question cards around a page and clips a supporting-evidence card into place
Decide which question needs an answer before deciding to add more words.

Decide what is missing before deciding to add words

Consider two fictional editorial situations. A product-care page explains the cleaning steps but omits a material-specific restriction. A comparison guide names two alternatives but does not explain which fits the reader’s use case. Both may need a revision, while the reason is the missing information, not their distance from a benchmark.

For the care page, the next step is to verify the relevant restriction against reliable product guidance and make it clear where the reader needs it. For the comparison, the team needs supported criteria and an explanation of the trade-offs. Extra background would not automatically solve either problem. These examples illustrate editorial decisions; they contain no measured performance results.

The reverse can also happen. A page may already answer its intended question with adequate evidence. Discovering that its count falls below a broad median gives the team a reason to inspect coverage, but does not itself justify adding another section. The decision should survive a simple challenge: which reader question will the proposed material answer?

This distinction also helps when reviewing an AI-generated draft. A longer version may look more complete while repeating the same point or adding unsupported claims. Review each proposed section for its contribution, then verify material factual statements. Keep an open evidence need visible rather than filling the space with confident wording.

A practical review sequence for your team

The following diagram is Ansvra’s suggested editorial workflow. It starts with the reader’s need, checks the benchmark’s scope, then moves to coverage, claim verification and review. It is a decision aid, rather than a measured causal model. Select a step to inspect its explanation and supporting references.

Use a benchmark to ask a better editorial question

Ansvra’s original suggested review process, not an observed causal model.

  1. Define need

    Identify the specific reader decision. A web-wide word-count quantile does not define the job of an individual page. [1] [2]

    Next: Check benchmark

  2. Check benchmark

    Check observation date, page group, device and metric scope. Rendered page words are not an isolated article-body target. [1]

    Next: Inspect coverage

  3. Inspect coverage

    Review the questions and evidence the page needs. Add detail for a supported information need rather than to reach a percentile. [2]

    Next: Verify claims

  4. Verify claims

    Check material factual claims and unresolved evidence before approving the page. A longer draft is not proof of better support. [2]

    Next: Review outcome

  5. Review outcome

    Assess whether the page answers the need, and inspect later saved-answer evidence separately. This workflow does not guarantee rankings or citations. [1] [2]

HTTP Archive supports the descriptive benchmark; Google supports people-first evaluation. The five-step workflow is Ansvra editorial interpretation, not measured performance guidance. [1] [2]

AI Content Briefs brings the reader’s purpose, essential questions, suggested structure and open information needs into a writing plan. Use that plan to explain what a section must contribute. A word-count estimate can still help a team plan its work, provided it remains an editorial estimate rather than a claim about a platform’s preferred length.

AI Content Optimizer connects the selected question, available research, brief, draft and review for new content or an existing-page improvement. Keep the final decision grounded in the evidence and the reader’s task. A structured draft is work for review, not proof that every claim is correct or that future citations will follow.

When an available saved answer references a page, Citation Source Analysis helps connect that reference with the tracked question. Open the relevant source to investigate its actual information. Its length can be a descriptive detail; its appearance does not reveal a source-selection formula or establish that copying its length would reproduce the result.

Before approving an expansion, record the unanswered question, the evidence that supports the addition and the intended reader benefit. After review, assess whether the revised page now answers that question. Examine later saved-answer evidence separately, with a defined comparison, rather than attributing any visibility change to the number of words added.

Summary

Keep word-count percentiles as context for investigation. Preserve their page groups, measurement scope and July 2025 observation period, then choose coverage from the reader’s question and the evidence available. Add useful, supported information where it is missing, and review subsequent AI visibility separately. The benchmark does not establish a length target or promise citations.

Sources

  1. 2025 Web Almanac: SEOAccessed
  2. Creating helpful, reliable, people-first contentAccessed

Discussion

Join the discussion

Start the conversation

Share a question or tell us what worked for you.

Comments appear after review. How we handle comments

The Ansvra journal

Keep reading

Browse all articles

Measure AI shopping by task

Separate discovery, purchasing and service questions using Salesforce’s historical shopper survey, complete response tables and a practical AI visibility review workflow.

Ansvra

A clearer next step.

MEET YOUR AI GUIDE

A little clarity.
A clearer next move.

Ask about Ansvra, AI visibility, or your next step with content.

Opening chat…