The only community host in our top ten was a code repository, cited in 243 readings
github.com reached 243 of 1,260 readings while no forum, no question site and no subreddit came near the top, which says something specific about what counts as community content.
Published by AI Knows Us · Updated 19 October 2026
Our top ten source table from 16 October 2026 holds exactly one row classified social or community, and it is not a forum. A code host, github.com, sat in 243 of 1,260 readings of ChatGPT and Gemini. No discussion platform, no question and answer site and no subreddit is anywhere in that top ten. The label says community. The thing that got cited is a place where files, specifications and version histories live.
The useful distinction is between an opinion and an artefact. A thread carries what somebody thinks. A repository carries a thing: a file, a licence, a specification, a commit with a date on it. An answer that needs something to quote reaches more easily for the second.
What a repository has that a thread does not
- A canonical object. The file is the thing being discussed, not a description of it.
- A version and a date. Both are machine readable and both are the fields a question about what is current actually needs.
- A stable address. The same path keeps pointing at the same thing, which a thread that gets edited, locked or deleted does not.
- Attribution written in. Who changed what, and when, is part of the record rather than a claim in prose.
- No requirement to be persuasive. A specification does not have to win an argument to be quotable.
Put those five together and a repository looks less like a forum and more like documentation that happens to be hosted somewhere social. That is the reading our count supports.
What this does not say about community content generally
It does not say forums are worthless, and it does not say posting in them is wasted. It says one thing, narrowly: in a bank of 630 buying questions read on two apps on one date, the community shaped source that cleared 243 readings was a code host, and the rest of that category did not reach the top ten. Community and social sources together took 1.9% of the 20,143 citations recorded in the run.
Three more limits, because the shape of the measurement sets them. The questions were about one category, ours, so a bank about kitchen equipment or conveyancing would surface a different community layer. Two apps were read and the product covers five, so the three we did not read are unmeasured here. And a citation records a domain the answer was assembled from, which is a different event from a business being named in the text of the answer.
The practical version of a community strategy, given this count
| What you were going to do | What the count suggests instead |
|---|---|
| Post answers in threads that mention your category | Publish the artefact the thread keeps asking for: a file, a template, a worked specification, with a version and a date |
| Chase upvotes on a discussion platform | Put the thing somewhere with a stable address and a change history |
| Write an opinion piece about the category | Write the reference the opinion pieces would have to cite |
| Treat community as one channel | Split it: places that hold objects, and places that hold conversation. They behave differently in a source list |
Where this gets concrete is anything a reader can download and run. A repository is also where a model's own files and licence terms actually sit, which is the practical difference between a product and a download, worked through in what is still available and how to build one.
We get paid to put businesses into tables like the one this count comes from, so our own row belongs on the page: of those 1,260 readings, 6 had a page of ours in the source list. The Google Search Console half of our own record, two domains with their windows and the export date printed, is on our proof page.
Four tests for whether your own category has a citable artefact
Most categories have one and most companies have not published it. Run your category through these four and the answer usually arrives in ten minutes.
- Is there a thing people ask each other for? A template, a checklist, a sample contract, a calculation, a mapping table, a list of codes. If the same request comes up in your inbox, that is the artefact.
- Can it carry a version number? Anything that changes when a rule, a price or a standard changes can be versioned, and a version is what makes it quotable about the present.
- Does it stand alone? The test is whether a stranger can use it without reading your marketing. If it needs your sales call to make sense, it is not yet an artefact.
- Does it have one stable address? One page, one path, no redesign that moves it. A thing that keeps moving cannot accumulate anything.
A category where all four answers come back yes has an unwritten reference page in it, and the company that publishes it first is the one the question resolves to afterwards. That is slower work than commenting on a thread and it compounds, which is the opposite trade of the one most content plans make.
Questions about code hosts and community sources in AI answers
How often was github.com cited?
In 243 of 1,260 readings on 16 October 2026, which places it sixth in the top ten behind four Google hostnames and OpenAI's help centre.
Did any forum make the top ten?
Not in this run. The social or community classification appears once in the top ten and it is the code host. The whole category came to 1.9% of citations.
Should I stop posting in community threads?
Keep doing it for the reasons it was always worth doing, which are the people reading it. As a route into a source list on these questions, the count above is what the category produced, and an artefact with a version and a date travels further than a comment.
Why would a code host answer a buying question?
Because several of the questions in this bank turn on specifications and file formats, and the repository holds the specification itself. The closer a question gets to a named file or a named standard, the more that holding matters.
Which apps and which date produced the 243?
ChatGPT and Gemini, 210 questions per market across the United States, the United Kingdom and India, read on 16 October 2026. The earlier run of the same programme, with its question set and full source mix, is 597 readings across ChatGPT and Gemini in three markets.
Publish the object, not the opinion about it
That is the whole of what 243 readings on a code host supports, and it is enough to change a quarter's plan. If your category has a file, a format, a checklist or a calculation that people keep asking for, the version of that thing with a date and a stable address is the asset. The commentary about it is the easier thing to write, which is precisely why it is already everywhere.