By the numbers: Where documentation quality breaks down in 2026

Teams can track documentation traffic, page updates, and popular searches without knowing whether their docs help someone finish a task. Missing information rarely announces itself, and a page can pass review while still leaving readers stuck.

We looked at our latest GitBook product data from September 2026, State of Docs 2026 research, and independent studies to understand where documentation quality breaks down. Together, they offer a picture of the problems teams face and the signals that help uncover them.

The GitBook findings reflect the sites and reader activity in our dataset. They give us a starting point for measuring quality, with an important limitation: how much we discover depends partly on how much people use the docs.

1. Keeping docs current is still the biggest operational challenge

In State of Docs 2026, 30% of respondents named keeping docs in sync with the product as their biggest workflow challenge. Among engineers, that rose to 43%.

Automation can reduce that maintenance burden. Systems that detect product or code changes and update the relevant documentation help teams keep pace with releases, especially when engineers ship faster than writers can review every change.

2. One in three questions exposes a gap in the documentation

Across questions submitted to documentation AI assistants in GitBook in September 2026, one in three could not be fully answered from the documentation. Around 24% could only be answered partially, while just under 10% could not be answered from the docs at all, based on our own internal product data.

These results point to gaps in what the documentation covers. A team may have written a page for every feature and still lack the information readers need to connect those features, troubleshoot a problem, or complete a particular task.

Those gaps can be difficult for authors to spot because they already understand the product. Reader questions provide a way to test that understanding against what the documentation explains. When the same unanswered topic keeps appearing, teams have a concrete reason to expand or clarify their content.

Feature coverage remains useful, but completeness is also determined by the questions people bring to the docs.

3. Broken paths are normal, and many go unnoticed

Among GitBook sites with broken links, the average site had five dead URLs in September 2026. Across the broader dataset, that worked out to roughly 50 dead inbound URLs per 1,000 published pages.

Around half had an obvious existing page that could serve as a redirect target. The other half had nowhere useful to send the visitor, suggesting that fixing the route alone wouldn’t resolve the problem.

Independent research shows how widespread broken links can become. Pew Research Center’s 2024 study found that 23% of news webpages contained at least one broken link, while 54% of Wikipedia pages contained a dead reference link.

An older study analyzing 11.9 million external links on Stack Overflow found that viewers flagged broken links in comments on only 1.67% of posts containing them.

Teams that rely on readers to report these failures can miss many of them. Monitoring the URLs people actually request helps reveal dead ends that an author reviewing the current site might never encounter.

4. More usage reveals more quality problems

Among GitBook documentation sites receiving at least 1,000 Assistant questions per month, 83% had at least one page flagged as possibly incorrect or outdated in September 2026. That fell to 51% for sites receiving 100–999 questions, with lower rates on less active sites.

Content-gap detection followed a similar pattern: sites with more reader activity surfaced more missing topics per 1,000 published pages.

While you could that to mean popular docs are worse, we can’t draw that conclusion because higher usage offers more opportunities to detect problems. A flag also identifies something to investigate, rather than confirming an error. The pattern suggests that reader activity can expose weaknesses that otherwise remain hidden.

That complicates comparisons between documentation sites. A low issue count may reflect good content, limited reader activity, or fewer opportunities to observe a failure. Teams need to consider how much usage sits behind the number before treating it as a measure of quality.

5. AI is becoming another quality test

More people now use agents at work. In the 2026 State of Docs survey, 35% of docs teams said readers find their docs through AI-powered search, 18% through coding assistants, and 16% through MCP servers. With content gaps being amplified by AI tools, that’s a bigger concern than ever.

And those numbers align with increased usage across the industry: comparing its 2026 pulse survey with its 2025 Developer Survey, Stack Overflow reported that workplace AI agent use rose from 31% to 59%, while daily use increased from 14% to 37%.

AI systems also retrieve information directly from websites. In Fastly’s May 2026 network data, 51% of AI requests required origin access, compared with fewer than 9% of human requests. That measures server access across its network, rather than documentation quality, but it shows how frequently AI systems reach beyond cached content.

Research on API use offers a more direct test. In a study presented at ICSE 2025, providing documentation increased GPT-4o’s valid calls to low-frequency APIs from 39% to 48%.

For documentation teams, this adds another reason to investigate missing or misleading content: it can affect both the person reading a page and the agent using that page to answer a question or perform a task.

Quality needs to be measured from both sides

Reviews, release checks, and working links remain useful measures of documentation quality. And reader activity adds evidence those checks can’t provide: unanswered questions, suspected errors, and requests for content that has disappeared.

For the team here at GitBook, these findings support measuring documentation quality over time. Teams need ways to observe what happens after publication and use those signals to guide maintenance.

A practical starting point is to review recurring unanswered questions, investigate flagged pages, and monitor dead URLs alongside the existing editorial process. That gives teams a clearer view of where readers struggle — and specific work they can do to help them.

It’s findings like these that led us to build GitBook’s new quality monitoring layer. Content gaps and Broken links bring reader signals into ongoing checks, while Analytics and Style guide help teams understand usage and keep updates consistent. GitBook Agent can then draft changes for review, giving teams a practical way to act on the problems this research reveals.

→ Introducing GitBook’s quality monitoring layer for reliable docs

→ How to audit documentation quality: a practical framework and checklist

→ How to keep technical documentation accurate as code and products change