Three hostnames, three kinds of writing, 446, 328 and 309 readings each
Developer documentation, product help and a company blog were cited at comparable rates in the same 1,260 readings, and they are not the same kind of page.
Published by AI Knows Us · Updated 19 October 2026
In 1,260 readings of ChatGPT and Gemini taken on 16 October 2026, three hostnames belonging to one company sat near the top of the source table: developers.google.com in 446 readings, support.google.com in 328 and blog.google in 309. One company, three hostnames, and no sense in treating them as a single source. A reference manual, a troubleshooting library and an announcement feed are three jobs for three readers, and the engines reached for all three at rates inside a hundred and forty readings of one another.
That matters to anybody writing a page, because it says the register of the writing is doing work. A publisher who only ever produces one register is competing for one of these three slots and ignoring two.
What each of the three registers actually is
- Reference documentation. Written for somebody implementing a thing, in the present tense, exhaustive, with the parameters and the limits named. It answers what a thing does. It never answers whether you should buy it.
- Product help. Written for somebody stuck, organised by the symptom rather than by the feature, in short task shaped articles. It answers how to do one thing right now.
- The company blog. Written for somebody who wants to know what changed and when, dated by nature, carrying the announcement and the reason. It answers what is new and what it replaces.
A question about a current rule pulls the blog. A question about an error pulls the help centre. A question about what a setting does pulls the reference. Our 16 October bank of 630 questions clearly held plenty of all three kinds, because the lowest of the three hostnames still sat in 309 readings.
The register a business is usually missing
Most companies publish a version of the third one and a thin version of the first. The middle register is the gap, and it is the cheapest of the three to produce because the material already exists in your support inbox.
| Register | Reader's state | What the page has to contain | Where a business usually fails it |
|---|---|---|---|
| Reference | Building or configuring | Every option, every limit, every default, with versions | Marketing adjectives instead of values |
| Help | Stuck on one symptom | One symptom per page, the cause, the fix, the check | No page at all, because it lives in email |
| Announcement | Checking whether something changed | The date, what changed, what it replaced | Undated posts, so the page cannot be used to settle a question about now |
Dating is the recurring failure and it is not a small one. We ran 102 build spec conversations on 15 October 2026, and the tally on that single requirement came back 102 of 102 for a date, a last verified line and a stated refresh frequency. An announcement register with no dates in it cannot do the only job it has.
Which of the three a smaller publisher can actually win
Not the reference register, when the subject belongs to somebody else. If the question is what a platform's own setting does, the platform wrote the manual and the count above is what that looks like.
The help register is different, because it is organised around a symptom rather than around a product, and a symptom usually spans several products. A page about one specific thing going wrong in a real setup is a page no operator writes, because no operator knows your setup. The announcement register is winnable too, in a narrower way: not announcing your own news, but dating somebody else's change and saying what it means for a reader in your market.
Deciding which register a given piece belongs to is a scoping decision before it is a writing decision, and it sits naturally alongside the question of who owns each stage of the work, set out in a workflow standard with six stations and an owner for each.
The edges of a three row reading
Three hostnames, one bank of 630 questions, one date, and two apps: ChatGPT and Gemini. No industry was named in any question, so there is no per industry rate inside it. City figures live in a separate scan of ten named cities and not here. And every count above records a domain that an answer was built from, which is not the same event as a business being named in the answer text.
We sell the measurement that produces counts like these and the writing that follows, so we print our own row too: 6 of those 1,260 readings had a page of ours in the source list. The Search Console side of the same business, with both windows and the export date on the page, is our proof page.
Questions about documentation, help pages and company blogs in AI answers
Which of the three hostnames was cited most?
The developer documentation, in 446 of 1,260 readings on 16 October 2026. Product help followed at 328 and the company blog at 309.
Does a help centre really compete with reference documentation?
They answer different questions, so they rarely compete for the same reading. A symptom question goes one way and a parameter question goes the other, and the two counts above are large at the same time because the bank contained plenty of both.
Should a small business write reference documentation?
Write it for the thing you own. Your own pricing, your own coverage, your own limits and defaults are reference material that nobody else can publish, and they are the facts an answer needs to quote in order to put you forward.
Why does a company blog get cited at all?
Because a dated announcement is the primary record of a change, and questions about what is current need one. The condition is the date. An undated post cannot settle a question about the present.
Which engines produced these three counts?
ChatGPT and Gemini, two apps, read on 16 October 2026 with 210 questions per market in the United States, the United Kingdom and India. The earlier run of the programme, its question design and its own source breakdown are published as 597 readings across ChatGPT and Gemini in three markets.
Sort your own library by register before you write anything new
List what you have published and put each piece in one of the three boxes. Most libraries come out heavily weighted to announcements, lightly to reference and empty on help. The empty box is the one with material already sitting in your support replies, and it is the register organised around the reader's problem rather than around your product, which is why it travels furthest. Getting a business into source tables like the one above is what we are paid for, and this sort is the first thing we ask to see.