446 of 1,260 readings cited one documentation host, and that is a fact about the questions
On 16 October 2026 developers.google.com was a source in 446 of our 1,260 readings of ChatGPT and Gemini, which makes question selection, not page quality, the first thing to fix.
Published by AI Knows Us · Updated 19 October 2026
developers.google.com turned up as a source in 446 of 1,260 readings we took from ChatGPT and Gemini on 16 October 2026, a run of 630 buying questions across three markets with every cited domain written down. Nobody is going to take that position off it. The move available instead is to find out which of your own questions drag a platform's documentation into the answer, then put your writing budget on the ones that do not.
This is a count of citations, meaning a domain that appeared in an answer's source list, deduped to one domain per reading. It is not a count of businesses named in answer text, which is a separate measurement with a separate denominator.
What a single host at 446 actually tells you
Two readings of the same figure are available and only one of them is worth acting on.
The first reading is competitive: Google's developer documentation outranks us on these questions. That reading leads nowhere, because the documentation is the primary source for the thing it documents. When the question is what a crawler directive does, the company that wrote the directive publishes the answer.
The second reading is about what we asked. A host reaches 446 only if a large share of the 1,260 readings put a question that a platform's own documentation already answers in full. The figure is describing the shape of our question bank at least as much as it is describing the engines. That is a diagnosis you can act on this week, because you choose your questions.
The four kinds of question a platform's own documentation cannot answer
Documentation describes a product to the person using it. There are four things it structurally does not do, and all four are the questions a buyer actually asks.
- Price. What a thing costs, in the reader's currency, with what is included at each level. A platform documents its own billing and nothing about what a supplier in your category charges.
- Who to hire. Which supplier fits a named situation. No operator publishes a comparison of the firms that work on its platform.
- Which option for which case. A recommendation that depends on business size, sector, team or constraint. Documentation gives capabilities; it does not choose between them for a reader.
- Anything tied to a place. A supplier in a named city, a market's own rules, a local cost. Documentation is written once for everybody.
Every one of those four is a commercial question, which is the part that matters. Sorted by the kind of page they landed on, the 20,143 citations in that run put 12.3% on official or platform documentation. That share is concentrated in a kind of question nobody was ever going to pay you to answer.
How to sort your own question list in an afternoon
You need your buying questions written down, the answers from two named apps, and a column for the top source. The sorting rule is one question: does the company that publishes the top source also own the subject of the question?
| Question you asked | Top cited source | Does the publisher own the subject | What to do |
|---|---|---|---|
| How does an engine choose which pages to cite | A platform's own documentation | Yes | Answer it for your readers, expect no citation |
| What does this work cost for a firm of my size | A supplier's own page | No | Publish your figures with a date |
| Who does this in my city | A small local site or a directory | No | A page specifically about that place |
| Which of these two tools suits my situation | A vendor's own comparison | No, and it is interested | Publish the comparison with the method shown |
The yes rows are the ones to stop measuring yourself against. The no rows are where the 82.0% sits, and the deciding question there is whether the page is worth writing at all, which is the same judgement described in how to tell a long tail keyword worth a page.
What the 446 does not mean
It is one host in one run on one date. The run read two apps, ChatGPT and Gemini, and our product reads five surfaces, so nothing here is a statement about Perplexity, Claude or Google AI Overviews. There is no per industry figure inside it, because the questions named no industry. And a citation count is a count of what the answers were built from, not a count of which companies the answers put forward to a buyer.
We sell the measurement and the writing that follows it, which is a reason to show you our own numbers rather than only the flattering ones. Out of the 1,260, six readings had a page of ours anywhere in the source list. The Google Search Console side of the same business, two separate domains with their export dates and windows printed, is on our proof page.
Questions about one host appearing in 446 readings
Was developers.google.com the top domain in the whole run?
It was, for this bank of questions read on 16 October 2026. The host sat in 446 of the 1,260 readings, with support.google.com next at 328 and developers.google.cn third at 325.
Does that mean Google decides what gets cited?
It means the questions in this bank often asked about platform mechanics, and the platform is the primary source for those. In the same run the largest slice of citations by far sat on vendor and agency own sites, at 82.0%, and that is where the commercial questions landed.
Should I write a page about how search engines pick sources?
Write it if your readers ask you, and judge it on whether they read it. Measured against a citation, that page is competing with the documentation of the company that built the thing, and the count above is what that competition looks like.
How big does a question bank have to be before this triage is worth doing?
Big enough that every kind of question you sell into has several entries of its own. We ran 102 per market first and 210 per market second, and the wider bank surfaced 4,947 distinct domains against 2,708, so the size of the bank moves the table more than anything the engines did.
Which engines produced this count?
ChatGPT and Gemini, read on 16 October 2026, 210 questions per market in the United States, the United Kingdom and India. Two apps. The full method and the question design for the earlier run of the same programme are published as 597 readings across ChatGPT and Gemini in three markets.
The number to write on your own list
Not 446. The figure worth having is the share of your own questions whose top source is published by the company that owns the subject, because that share is the part of your content budget currently aimed at a wall. Ours came down once we looked at it, and the questions left over were all the ones a buyer types. Counting that split is the first thing we do for a customer, and it costs an afternoon to do yourself.