Why self-service fails: what Gartner found about self-service
Gartner surveyed 5,728 customers and found that only 14% of customer service issues were fully resolved in self-service. The most common reason it failed was that customers couldn't find content relevant to their issue. This page reads that survey closely: what it measured, what it left out, and where help-center screenshots and AI assistants come in.
The survey
Gartner published the result on 19 August 2024, in a press release titled "Gartner Survey Finds Only 14% of Customer Service Issues Are Fully Resolved in Self-Service". The survey of 5,728 customers ran in December 2023. It is one of the few self-service statistics with a published sample size and date. That is why people quote it so often.
Gartner is an analyst firm. Its full report, "Accelerate Self-Service Success by Diagnosing Customer Needs", is for its clients. We read only the public release. It doesn't say which countries or industries the customers came from, how they were recruited, or how the questions were worded. Treat its numbers as one survey's headline results, not as a benchmark for your own help center.
We checked every number below against the release on 15 November 2026.
What it found
| What Gartner reported | Number |
|---|---|
| Customers who use self-service at some point in their service journey | 73% |
| Customer service and support issues fully resolved in self-service | 14% |
| Issues customers described as "very simple" that fully resolved in self-service | 36% |
| Customers who started in self-service and said the company didn't understand what they were trying to do | 45% |
| Cases where customers couldn't find content relevant to their issue, the most common reason self-service failed | 43% |
Read the units with care, because they differ. The 14% and 36% are shares of issues. The 73% and 45% are shares of customers. The 43% is "of cases", and the release doesn't say whether that means all self-service attempts or only the failed ones. Either way, it is the top reason Gartner gives.
"Fully resolved" is a high bar. An issue that a help article half solved before the customer phoned anyway counts as a failure. So the self-service resolution rate of 14% doesn't mean 86% of help articles are useless. It means most issues don't finish in self-service.
Simple issues fail too
The 36% figure is the one we find hardest to explain away. These are issues the customers themselves called very simple, and most of them still didn't finish in self-service. That makes "the problem was too hard for an article" a weak excuse. Something between the customer and the answer went wrong.
Gartner's own reading is about matching. Eric Keller, a senior director in its customer service practice, says customers find self-service journeys too rigid for their issues, and that companies need to work out what the customer is trying to do and match them with the solution that fits. The 45% who said the company didn't understand them point the same way.
What Gartner did not measure
The release is short, and several things people attach to it aren't in it.
- Screenshots. The release doesn't mention screenshots, images or video. Gartner did not study them.
- Whether content was out of date. "Couldn't find content relevant to their issue" covers several failures: the article didn't exist, search didn't show it, or the customer found it and decided it didn't apply. The release doesn't separate them.
- Software products. The survey covers customers in general. The release doesn't say how many were using a software product's help center or knowledge base.
- Behavior. These are answers customers gave, not logs of what they clicked.
Nothing in the release shows that outdated screenshots cause self-service to fail. If you see that claim credited to Gartner, it is wrong.
Where screenshots come in
What follows is our reasoning, not Gartner's finding. Help-center and knowledge base articles for software usually carry screenshots, and readers compare the picture with the screen in front of them. If the button in the picture has moved or gone, the reader has a reason to decide the article is about another version and doesn't apply to them. That is one plausible path into Gartner's 43%. Nobody has measured how much of it.
There is separate evidence that the mismatch confuses people. A 2026 study of VS Code's documentation found 210 issue reports about outdated screenshots, some from programmers with more than ten years' experience. In 24 of the 44 fixes the authors sampled, the text changed along with the picture. A stale screenshot is often a sign that the steps around it are stale too. For the rest of the evidence, see what the research says about product screenshots.
GitHub's 2017 Open Source Survey adds a wider, softer number. It sampled about 5,500 open-source users at random from GitHub, plus about 500 from elsewhere, and 93% of respondents had seen incomplete or outdated documentation. It measures what people noticed, and it puts incomplete and outdated in one answer. Its readers are open-source users, not help-center customers. It says outdated docs are common. It doesn't say they are the reason anyone gave up.
The same content now answers AI questions
One of Gartner's recommendations is to make a single "digital concierge, such as a GenAI chatbot", the most prominent entry point to the customer journey. That chatbot answers from the help center. If an article is missing or wrong, the chatbot inherits the gap. And AI agents now read help centers directly, so the same pages serve readers who never see your site.
GitBook, which hosts documentation, published numbers on this in October 2026. This is vendor data from its own product. Across questions asked of documentation AI assistants on GitBook in September 2026, one in three couldn't be fully answered from the docs. Around 24% could be answered only in part, and just under 10% not at all. Among sites getting at least 1,000 assistant questions a month, 83% had at least one page flagged as possibly incorrect or outdated.
GitBook gives its own caveats, and they are fair ones. A flag is something to look into, not a confirmed error. Busier sites get more flags because more people use them. The sites are GitBook's customers, not help centers in general. The post ends by introducing GitBook's quality monitoring features.
Screenshots matter here in a roundabout way. An assistant can't hold a screenshot up against the live product. It reads the steps and the alt text. When a screenshot has gone stale, the VS Code study suggests the steps next to it often have too. A person might notice the picture doesn't match and stop. An assistant repeats the steps.
What to do about it
Gartner's last recommendation is to assess self-service content all the time, and to let customers and reps flag content that doesn't work. Flags help, but they wait for someone to notice and bother. For screenshots there is a check that doesn't wait. Compare them with the product every time the product changes.
- Take screenshots from the build. If your end-to-end tests capture the screens your help center shows, each build makes new ones from the current app. Screenshot automation shows the test.
- Embed a URL that follows the build. Screenery publishes each screenshot at an @latest URL that a push on the default branch moves, so an article shows the newest capture without anyone uploading it. See How @latest moves.
- Re-read the steps when a picture changes. On a pull request, Screenery comments with the screenshots that changed. That list doubles as the list of articles whose text may need a look. See Pull-request comments.
The longer version is in how to keep help-center screenshots current. For what stale articles cost a support team, see the cost of stale docs.
Sources
We opened every page below on 15 November 2026 and checked each number on this page against it.
- Gartner, "Gartner Survey Finds Only 14% of Customer Service Issues Are Fully Resolved in Self-Service", 2024
- GitBook, "By the numbers: Where documentation quality breaks down in 2026", 2026 (vendor data)
- GitHub, Open Source Survey, 2017
The VS Code figures come from Tang, Yan, Liu, Meng and Zhong, "Detecting Outdated Screenshot from GUI Document", ACM TOSEM, 2026. Our write-up of that study links the paper.