Documentation outranked every vendor blog in our citation data, and what a B2B company does about it

A B2B company with a content budget has to decide where it goes. Our own citation data points somewhere most content plans do not, and it also contains the detail that complicates the story, so both are below.

What the citation data shows

On 22 October 2026 we scored 635 readings of ChatGPT and Gemini against 318 United States buyer questions and collected 19,793 citation URLs across 3,307 distinct hosts. Half of every citation came from 51 of those hosts. Eight of the twenty five largest hosts are documentation, help centres or developer portals.

• developers.google.com, 1,360 citations.
• support.google.com, 695.
• help.openai.com, 467.
• developers.google.cn, 447, kept as its own row rather than merged into the .com host.
• developers.openai.com, 191.
• help.otterly.ai, 143.
• business.google.com, 119.
• learn.microsoft.com, 112.

Against that, review and comparison directories were thin, with g2.com at 125. The full labelled table is in the citation host study.

The detail that complicates it

Vendor marketing domains are cited, and heavily. In the same run semrush.com took 744, ahrefs.com 354, peec.ai 328, otterly.ai 255, tryprofound.com 227 and hubspot.com 119. So a vendor site is plainly not excluded from the citation ground.

Look at who those vendors are. In the same 635 readings, Semrush was named in 259, Profound in 154, Ahrefs in 145, Otterly in 124 and Peec in 91, out of 1,053 namings across a 24 name roster. The vendor domains being cited are almost exactly the vendors already being named. The naming table is in which brands the assistants put forward.

Our own contribution to this is the control. We added 507 recommendation shaped pages between 20 and 22 October 2026, going from 134 live to 641, and our domain appears in 0 of the 19,793 citation URLs and our name in 0 of the 635 readings. The method is in the 507 page null result. So publishing more marketing pages on a domain the answers do not already name did nothing for us, while marketing pages on domains the answers do name are cited constantly.

What the specification asked for and what is missing

The engine that specified this question asked for a B2B citation rate as a numerator over a denominator, by engine and by buyer intent category, with the test dates and confidence intervals. We hold the citation URLs and the hosts. We do not hold a per engine, per intent citation rate for B2B companies, so that figure is absent here rather than estimated. A host count tells you where citations landed. It does not tell you that publishing in that shape causes them, and no page should pretend otherwise.

What to do with a content budget, in order

1. Make the facts fetchable. Prices, limits, integrations, regions and support terms in the HTML the server sends, not assembled by a browser afterwards. Our own pricing page failed this earlier in October 2026 and was repaired the same day, which is how we know it is worth checking rather than assuming.
2. Write the help centre and the developer docs as publishable pages, with headings that match the questions a buyer types and one checkable fact per section. In this run those are the hosts that held the top of the citation table.
3. Publish one number nobody else holds. Of the 224 page specifications the engines wrote, 212 named a single most citable number the page had to carry and most wrote it out as a fraction with its denominator. The requirement counts, with the counting rule attached, are in what makes an assistant cite a vendor website.
4. Then the blog, and treat it as an experiment with a review date rather than a visibility mechanism. Ours was an experiment and the result was zero.

What a citable documentation page looks like

The hosts at the top of that table have a shape in common and it is not a design system. Each page answers one question, states the fact in the first sentence under a heading a buyer would type, and carries a figure or a limit rather than a description of a benefit.

• One question per page or per section, with the question as the heading. An answer an engine can lift has to sit under a heading that matches the query rather than under a clever title.
• The fact first, the explanation after. 212 of the 224 page specifications the engines wrote named a single most citable number the page had to carry, and most wrote it out as a fraction with its denominator rather than as a claim.
• Numbers that are limits, not adjectives. A prompt allowance, a rate cap, a region list, a retention period. These survive being quoted six months later and a benefit sentence does not.
• A date on the page, and a date on every external fact in it. 152 of the 224 specifications asked for a date or a date range, counted by word boundary.
• No login between the fact and the fetcher. A documentation page behind authentication is a page that cannot be cited, whatever else is true about it.

How to check this on your own category rather than ours

Ask fifteen buyer questions in your category, your company name in none of them, three runs in each engine on one morning, and keep the sources block from every answer. Then count the hosts. Two things fall out immediately: whether documentation leads your category the way it leads ours, and which named competitors own the cited vendor domains. The fifteen question version of this test is set out in how to test whether an assistant recommends your business.

Three runs rather than one, because on the same day we measured a brand that appeared at all appearing in all three runs 53.0 per cent of the time on ChatGPT and 26.1 per cent on Gemini. One pass over fifteen questions produces a host list with a known amount of luck in it, quantified in the run to run measurement.

What this does not cover

Two engines, one day, one category of buyer questions about marketing and AI visibility, and a host count rather than a page type experiment. A B2B category with mature documentation on every vendor site could distribute differently and nothing here says which way. We sell AI visibility measurement, and the figure on this page that costs us most is the zero.

On our own limits: we read five surfaces. Our pricing page renders no SSO, SAML, SOC 2, ISO, DPA or white label, checked by word boundary over 119,203 bytes on 22 October 2026, and a loose search for sso on it matches crossOrigin seven times and subprocessors twice. Fetched by us the same day, tryprofound.com/pricing publishes SSO and SAML with SOC 2 compliance and up to 9 answer engines tracked on Enterprise, with the price given as Custom and no figure in any currency on the page.

Questions about where B2B citations come from

Should we move budget from the blog to the docs

On our data the documentation hosts held the top of the citation table and our own blog experiment returned zero. That is a reason to try it, measured, not a proven causal lift.

Do review directories matter

They were thin in this run. g2.com took 125 of 19,793 citations while documentation hosts took thousands.

Why are competitor blogs cited when ours is not

The cited vendor domains belong to the vendors the answers already name. Semrush appeared in 259 of 635 readings and its domain took 744 citations. We appeared in 0 and took 0.

What is the single cheapest change

Putting your prices and limits in the served HTML. It is a one off engineering fix rather than a content programme, and it is checkable from outside in a minute.

Two companion pages from the same run. The three page briefs we refused because they asserted things we do not have are in three pages the AI engines told us to publish, and why we did not publish them, and the accuracy test nobody in this category has run is in has anybody measured whether AI visibility tools report accurately, and what that test needs.