Does AI Quote Your Help Center? Two Citation Studies Compared

A help center is the most question-shaped content a company owns. Its pages are written as answers to customer questions, so it should be easy for AI to quote. That assumption is worth checking against the published evidence before anyone spends money acting on it.

I'm Bob Michaels. My work on the web started in 1992, and these days I measure AI visibility for companies: I run locked sets of questions on several AI platforms and record what the answers say and cite. I'm primarily interested in remote work, and I'm open to part-time, full-time, fractional or consulting work. That's paid work, so weigh what follows knowing I have a stake in companies measuring this.

Two studies I could verify measure which pages AI cites, and they disagree about documentation. Below is what each found, what it leaves out, whether anyone has measured an ordinary business help center, and how to get an answer for your own company. Everything here is current as of November 3, 2026.

What did DevTune find about AI citing documentation?

DevTune found that documentation was the second most cited type of page in AI search answers about developer tools, at 17.3% of citation records. Blog posts came first at 34.3%. Product pages took 7.5% and landing pages 2.6%. DevTune, a vendor, published these figures on September 16, 2026 in AI Search Cites Blog Posts, Not Homepages, drawn from its State of AI Search for Dev Tools report.

A citation is a link that an AI answer offers as its source. Every week, DevTune asked the same structured evaluation questions on ChatGPT Search, Perplexity, Gemini Search, Google AI Mode and xAI Search. It then sorted each cited URL, or web address, by content type, such as blog post, documentation or product page. The report covers 43 developer-tool categories, 478 tracked brands and 126,814 citation records, gathered over a 90-day window from April 2 to June 30, 2026.

The overall share hides large swings by category. In API Development & Management, documentation made up 39.6% of citations. In Documentation & Developer Portals, it was 37.5%. DevTune also notes that its blog-post share is only as good as the labels it assigned.

For anyone outside software, the limits matter more than the numbers. The sample covers developer tools only. It does not describe retail or services.

What did Datadab find when it tested B2B buyer questions?

Datadab found that AI cited blog and article pages far more often than documentation: 133 citations pointed to blog or article URLs, against just 5 for documentation. Datadab published the work as original research in May 2026 under the title What AI Actually Cites. B2B means businesses buying from other businesses.

Datadab ran 100 real-world B2B buyer prompts, meaning questions typed into the AI, through OpenAI's search-enabled model (gpt-4o-search-preview) via OpenRouter. It documented every URL the model cited, which came to 328 citations across 239 domains.

Datadab also says 41% of the cited URLs are listicles. That figure, like the rest of the study, comes from a small sample run on one model, so I read it as a snapshot of that model answering those questions.

Why do the two AI citation studies disagree about documentation?

The two studies differ in industry, in platforms, in sample size and in how they classify pages, so their documentation figures describe different things. DevTune tracked developer-tool brands across five AI search products. Datadab sent B2B buyer prompts to a single OpenAI model. A share averaged across five products and a count from one model are answering different questions, even when both use the word documentation.

Classification is the easiest difference to miss. Each study sorted cited pages into types using its own rules, and the share a report gives any page type depends on those rules. DevTune flags this for its own numbers. Datadab's categories deserve the same check, including the one behind its listicle figure.

Before you repeat a citation figure to your team, read how the study picked its questions and how it defined each page type. Then ask whether your customers' questions look anything like the ones the study used.

Has anyone measured how often AI quotes an ordinary business help center?

I did not find a study of how often AI quotes the help center of an ordinary business, such as a bank, a clinic or a retailer. The two studies above come from developer tools and from B2B buyer prompts, so neither one answers the question for a business outside those groups.

What remains is a writing practice. A page written to answer one question directly, with the answer in its first sentence and its facts as text, is easier for any system to quote than a page that buries the answer. When the answer to a refund question sits halfway down a long page, anything quoting that page has to dig it out from among other topics. Facts as text means the fee, the deadline or the steps sit in the words of the page itself, rather than only inside an image or a downloadable file. I go through the PDF version of that problem in my post on fee schedules kept as PDFs. I treat all of this as a reasonable practice, not a measured guarantee.

Say your help center has one billing page that covers refunds, invoices and failed payments together. One edit I'd consider is giving each of those questions its own page, titled in the customer's words, with the answer in the first sentence. Run your locked questions before and after the edit, and compare what AI cites.

How can you test whether AI quotes your own help center?

Test your own pages with your own customers' questions on each AI platform they use, and record whether each answer cites your page. I don't take a citation study as proof for a different company, so this is the instrument I use with clients.

Start with the question list. Write down the questions customers actually ask, in their words, and lock the list so the wording never changes between runs. I explain how I build that list in my post on building the question library first.

Run every question on each platform, and repeat the full set on a fixed schedule. Each result is a plain record: the question, the platform, the date and whether the answer cited your page. When a customer asks how to cancel their plan, that record tells you whether the answer cited your cancellation page. I describe the rest of that instrument in how I measure AI visibility without one magic score.

Record what each answer says as well as what it cites. The wording of the answer is what your customer reads, and you'll need it for the next step. Because the question list never changes, any difference between two runs comes from the answers or from your pages.

Then check each citation, because a cited link can still come with a wrong answer. Open the cited page and compare it with what the answer claims. If the AI states a refund time your page never gives, the citation looks like a win on a dashboard while a customer walks away with the wrong policy.

You can start small. The questions your help center already answers and the AI products your customers already use are enough for a first run, and the list can grow once you've seen the results.

Common questions