Ask an AI model whether a company is any good and watch where the answer comes from. It rarely quotes the company's own website. It quotes strangers: a comparison thread, a long reply from someone who switched tools last year, a review that complains about one thing and praises three others.
We wanted to know how consistent that pattern is, so we ran a simple test. Five major models. Forty questions of the kind buyers actually ask: "what's the best tool for X", "is Y worth it for a team our size", "what do people say about Z". For every answer, we catalogued every source the model cited or clearly drew on.
Reddit was a top-three source for all five models. Not for every question, but often enough that ignoring it means ignoring a large part of how your category gets described. This post is about what the cited threads had in common, and what the uncited ones did wrong.
Why a model reaches for a thread at all
A model answering "which one should I use?" has a problem. It needs a judgement, and the most detailed information available about any company is written by that company. That information is, by definition, not neutral.
So models do what a careful person would do. They discount what you say about yourself and give more weight to what other people say about you, especially when several of those people agree. A Reddit thread is close to the ideal source for that: many independent voices, arguing in public, with replies that challenge weak claims.
Your homepage says you're the best. So does every competitor's. To a model, that cancels out.
What the cited threads had in common
Across the forty questions, the threads that kept getting cited shared a handful of traits. None of them are tricks. They're the marks of a genuinely useful conversation.
They answered a question someone actually asked
The cited threads almost always started with a real person asking a real question: "We're about to choose between these three, what are we missing?" They didn't start with a product announcement or a list of features. The question framed the thread as a search for an answer, which is exactly what the model is doing too.
They contained disagreement
Threads where everyone agreed were cited less than threads with a bit of argument in them. A reply saying "we tried it and the onboarding was rough" next to one saying "the support made up for it" reads as a real assessment. Models seem to treat visible disagreement as a sign that the conclusion was tested.
They were specific
The replies that got quoted described something concrete: a migration that took an afternoon, a limit someone hit at a certain size, a support answer that arrived on a Sunday. General praise ("great product, highly recommend") carried far less weight than one specific detail.
They named alternatives
Cited threads compared. They put two or three options side by side and said what each was good and bad at. That's the format of the answer the model is about to give, so these threads are easy for it to draw from.
They came from accounts with history
The replies that carried weight came from accounts that had been around, posting about other things, for a long time. A brand-new account praising a product is easy to discount, for people and models alike.
They had been around long enough to rank
Many of the cited threads were months or years old and still ranked well in search. A good thread doesn't decay. It keeps getting read, and keeps getting re-quoted every time a model's sources are refreshed.
What the uncited threads did wrong
We also looked at threads that mentioned the companies in question but never made it into an answer. The reasons were consistent.
- They were promotional. Posts from a brand account announcing a launch, or replies that read like a sales pitch, were almost never cited. The communities often downvoted or removed them anyway.
- They had no engagement. A single post with no replies gives a model nothing to weigh. One voice is a claim, not evidence.
- They were removed or locked. Threads that broke a subreddit's rules disappeared, and with them any chance of being cited.
- They were vague. "Anyone used X?" followed by "yeah it's fine" doesn't help anyone choose, so it doesn't help a model either.
- They were obviously coordinated. A cluster of new accounts saying the same thing in the same week stands out. Communities notice, and the threads rarely survive.
Why your homepage loses this contest
None of this means your website is useless. Models do read it, and it's often where they get basic facts like what you do and where you're based. But on the question that matters most, "should I choose them?", your own site is the least persuasive source available, because it's the only one with an obvious stake in the answer.
Three strangers agreeing in public will beat your best landing page every time. That's not a flaw in the models. It's the same instinct that makes people check reviews before they buy.
What to do with this
If your category gets discussed on Reddit, and most do, the useful questions are these:
- Which threads already shape the answer about you? Ask the models, then read the sources they cite. Some of those threads will mention you; many won't.
- Where do your buyers actually decide? Usually a handful of specific communities, not Reddit as a whole.
- Who on your side could take part honestly? Someone who knows the product, posting under their own name and disclosing who they work for, answering questions whether or not your product is the answer.
- Can you leave criticism standing? The threads that get cited include the bad along with the good. Trying to remove criticism tends to backfire, and an all-positive thread reads as bought.
What this doesn't mean
It doesn't mean you should flood Reddit with mentions. That's the fastest way to get banned, and the threads it produces are exactly the ones models ignore. It doesn't mean a single viral thread fixes everything either. The pattern we saw rewards steady, useful presence over months.
And it doesn't mean Reddit is the only source that matters. Reference pages and reviews play their own roles, especially for hard facts and recent verdicts. But if you want to know why a model recommends your competitor and not you, start by reading the threads it quotes. The answer is usually right there.