Reddit, Wikipedia And The Citation Lottery: What LLMs Actually Cite, And How Fast It Changes

Reddit, Wikipedia And The Citation Lottery: What LLMs Actually Cite, And How Fast It Changes

Most writing about AI citations is vibes. Someone runs fifteen prompts, notices Reddit comes up a lot, and publishes a strategy post recommending you go and post on Reddit. I’ve read a dozen of these. I’ve probably half-believed a few.

So when Semrush published a study with a sample size large enough to actually mean something, I read it properly. The headline finding isn’t the one that got quoted. The finding that matters is buried in the middle, and it should change how you think about this entire discipline.

The study, and why the sample size matters

The methodology: **230,000+ prompts**, generating over **100 million citations**, captured in **weekly snapshots across 13 weeks** from **14 July to 12 October 2025**, covering **ChatGPT Search, Google AI Mode and Perplexity**, analysed as the top 25 cited domains per platform per week.

Two things make this useful where most citation research isn’t.

First, the scale. At 230,000 prompts you’re measuring the system rather than your own prompt-writing habits. Small studies mostly measure the researcher.

Second, and more importantly, the **weekly snapshots over 13 weeks**. Almost all citation research is a snapshot — a single point in time presented as though it describes a stable state. Time series lets you see movement, and the movement turned out to be the story.

The September reshuffle, in one paragraph

Partway through the study window, ChatGPT’s citation profile reorganised dramatically.

**Reddit citations fell from roughly 60% to about 10%.** **Wikipedia fell from roughly 55% to under 20%.** Both mid-September, both fast. The gainers were PRNewswire, Forbes and Medium. Among the losers alongside Reddit and Wikipedia: TechRadar.

Now the crucial control. **The shift was isolated to ChatGPT.** Google AI Mode maintained a balanced mix throughout. Perplexity showed minimal volatility across the entire period.

So this was not an industry movement, not a change in how “AI” values sources, and not something anyone could have optimised for. It was one platform changing its own configuration, and every brand that had built its visibility on being Reddit-adjacent in ChatGPT lost a large fraction of it in a fortnight, having done nothing wrong.

Three platforms, three different internets

The other thing the study makes clear is that “AI citations” is not a single phenomenon. The platforms disagree profoundly about what a good source is.

**ChatGPT**, post-September, leaned toward press-release wires, established publishers and platform-hosted writing — PRNewswire, Forbes, Medium.

**Google AI Mode** ran a far more balanced mix throughout, with **LinkedIn, YouTube, Reddit and Google’s own properties** leading. Notably, **Wikipedia appeared in only about 2% of AI Mode responses**, against a far higher share in ChatGPT. Google — which has spent twenty years building its own knowledge graph — leaning less on Wikipedia is not surprising in hindsight, but it’s a large divergence.

**Perplexity** favoured **Reddit, LinkedIn, NIH, Microsoft and Google**, and did so with remarkable stability. Wikipedia, Microsoft and Forbes drifted modestly upward.

If you are optimising for “AI search” as a single target, you are optimising for an average of three systems that disagree. That average describes none of them.

Why volatility is the actual finding

Here’s the conclusion I’d draw, and it runs against most of the advice currently being sold.

Citation share on any individual platform is a **volatile asset**. It can halve in a fortnight because of a configuration change you’ll never be told about, on a platform you can’t audit, for reasons that have nothing to do with the quality of your content.

Compare that with classical SEO. Rankings move, sometimes brutally — but core updates are announced, they roll out over weeks, they’re analysable after the fact, and there’s a documented set of guidelines to work against. None of that infrastructure exists here. There’s no changelog, no announcement, no guidelines document, no appeal.

Which means the thing to optimise for is not maximum citation share on the platform that’s currently sending you traffic. It’s **resilience** — being credibly present across enough independent surfaces that any single platform’s reconfiguration is a dent rather than a collapse.

That’s a genuinely different strategy. Maximising for one platform means concentrating effort where it currently pays. Resilience means deliberately accepting a lower peak for a higher floor.

What this means for a brand with no Reddit presence

Let’s make it practical, because I don’t want this to be an interesting-but-useless post.

**Don’t chase the current leaderboard.** The single worst response to this data is a six-month campaign to build Reddit presence, because Reddit was the most-cited domain in a study that also showed Reddit’s ChatGPT share falling 50 points in two weeks. Build presence where your customers actually are. If that’s Reddit, great — but for the reason, not the ranking.

**Diversify the surface types, not just the sites.** The three platforms weight different *kinds* of source: community discussion, professional networks, video, press, encyclopaedic reference, institutional. A brand present in one type is fragile. Present in four, it survives any single reweighting.

**Own the substrate everyone reads.** Whatever the citation mix does, structured product data, accurate specifications and clean schema on your own domain are inputs to all of them. This is the unglamorous layer, and it’s the only part that doesn’t get reweighted overnight — which is precisely why I ended up building a tool for it rather than for citation chasing.

**Track platforms separately, always.** If you’re measuring citation share, never report a blended number. A blended figure would have shown a modest decline in September and completely hidden that one platform halved while another didn’t move.

**Don’t over-react to one month.** Corollary of the above. A single platform swinging is now the expected behaviour, not an emergency. Look for a sustained direction across several months before rebuilding a strategy.

Building citation resilience

If I were starting from zero on a brand today, in order:

1. **Get the owned layer right.** Complete structured product data, valid schema, plain-language factual content. Everything downstream reads this, and it’s the only layer you control.
2. **Establish presence on three or four different surface types** — not four sites of the same kind. Community, video, professional, press.
3. **Build a fixed prompt set and track per-platform**, monthly, from now. You cannot create a baseline retroactively, and in a year the only people who can show a trend are those who started.
4. **Assume the leaderboard changes.** Write it into the plan as an expectation rather than a risk.

The uncomfortable truth in this data is that a meaningful part of AI visibility is, right now, a lottery — decided by configuration choices at three companies who don’t publish their reasoning and don’t owe you an explanation.

You can’t win a lottery by playing harder. You can buy enough different tickets that no single draw ruins you, and you can make sure that when a model does look at your product, everything it finds is accurate, complete and specific.

That’s not as satisfying as a growth hack. It’s what the evidence supports.

*Source: Semrush, “The Most-Cited Domains in AI” — 230,000+ prompts, 100M+ citations, weekly snapshots 14 July – 12 October 2025, across ChatGPT Search, Google AI Mode and Perplexity.*