What LLMs Actually Cite: Six Studies, Six Answers

Six large citation studies have been published in the past two years. They disagree with each other so sharply that one ranks Reddit the single most-cited domain while another puts it fourth, and a third finds YouTube ahead of it.

That disagreement is the most useful thing to understand about this topic, because it tells you why the confident percentage tables you'll find elsewhere are unreliable — and what to do instead.

At a Glance

What the studies found

Note that several of these were run by vendors selling AI visibility tools, including two that compete with us. Read the numbers, weigh the interest.

Why they disagree

Different denominators. This is the big one. "Share of overall citation volume" and "share within the top ten cited sources" describe the same underlying data and produce wildly different figures. Reddit appears as roughly 1.8% of all ChatGPT citations under the first measure and roughly 11.3% under the second. Neither is wrong. They answer different questions.

Whenever you see a citation percentage, the first thing to establish is which denominator it uses. Most articles quoting these numbers don't say, and often don't know.

Genuine volatility. Semrush's tracking found ChatGPT citing Reddit in close to 60% of prompt responses in early August 2025, collapsing to around 10% by mid-September. Wikipedia fell from roughly 55% to under 20% over the same period. Those shifts weren't systemic — Wikipedia's share on AI Mode and Perplexity held steady — which means a change in one platform's behaviour rather than a change in the web.

A figure from six months ago may describe a world that no longer exists.

Different engines, different behaviour. ChatGPT leans encyclopedic and editorial. Perplexity leans toward Reddit, LinkedIn and G2 for B2B queries. Google's AI products skew toward Google-owned properties and platforms like YouTube. A single ranking across all engines averages away the thing you need to know.

What is consistent

Strip out the disagreement and a stable picture remains.

Reddit, Wikipedia, YouTube and LinkedIn appear near the top of every study. The order changes, the percentages change, the presence doesn't.

Citation is long-tailed. Evertune's finding that no single domain reliably exceeds 5% of total citations matters more than any ranking. Unlike classic search, where the top ten results capture most clicks, AI citations spread across thousands of domains. There is no winner-take-all position to occupy.

User-generated content punches above its weight. Community discussion and video transcripts appear consistently across studies, which is a genuine departure from how link-based authority worked.

The cited source is usually not the recommended brand. When a system recommends a tool, the citation typically credits a third party — a forum thread, an editorial roundup, a comparison page. The brand gets the recommendation, someone else gets the link.

That last point is the one with strategic consequences, and we cover it in more depth in G2 and Capterra: do review sites still feed AI answers?

The three layers

A useful way to organise the work, given no single number is reliable.

Reference layer. Wikipedia and comparable structured references. Slow, hard to influence directly, disproportionately weighted. Build coverage elsewhere first, since notability rules require independent sourcing, and never edit your own entry.

Community layer. Reddit, industry forums, Slack and Discord communities. Highest variance and highest apparent weight. Participation over months, by people with genuine account history, without dropping links.

Editorial layer. Trade press, analyst coverage, industry publications. Slowest to build, most durable, and the layer most amenable to conventional PR.

Your own site sits underneath all three and gets cited least often. That isn't a reason to neglect it, but it is a reason to stop treating owned content as the whole strategy.

Who owns what

The reason this work stalls is that it crosses functions and nobody owns the crossing.

Community is the one that gets assigned to content marketing and then doesn't happen, because it isn't content production. It needs someone whose job includes being present in places where your buyers talk.

A realistic first 90 days

Weeks 1 to 2. Run twenty buyer prompts across ChatGPT, Gemini and Perplexity. Record which sources get cited in your category specifically. This is the only number that matters, because published averages won't reflect your niche.

Weeks 3 to 6. Fix the owned layer. Answer-first structure, consistent product facts, schema. Cheapest work, fastest to complete, covered in answer engine optimization.

Weeks 5 to 12. Start community participation and editorial outreach. Both are slow, both compound, neither shows results inside a quarter.

Week 12. Re-run the same twenty prompts and compare. Same prompts, same engines, or the comparison means nothing.

Where Strivelabs fits

Everything above depends on knowing what's actually cited in your category, which published studies can't tell you and manual checking rarely gets done consistently.

Strivelabs samples ChatGPT, Gemini, Claude, Perplexity and Google AI Overviews against prompt sets you define, three samples per prompt, recording which sources appear. That gives you your own citation picture rather than an industry average.

It won't participate in Reddit for you or place editorial coverage. Those stay human jobs.

Frequently Asked Questions

Which source do LLMs cite most?

The studies disagree. Reddit ranks first in some analyses and fourth in others, with YouTube and Wikipedia also taking the top spot depending on methodology and date. What's consistent is that those four appear near the top of every study.

---

Why do the percentages vary so much between articles?

Two reasons. Different denominators — share of all citations versus share within the top ten sources produces very different numbers from identical data. And genuine volatility, with one study recording Reddit's ChatGPT citation rate falling from around 60% to 10% in six weeks.

---

Is Wikipedia worth pursuing?

If you qualify. Notability requires independent coverage, which means building editorial presence first. Never edit your own entry, and treat it as a long-term outcome rather than a task.

---

Does Reddit really matter that much?

It appears prominently in every study, though the size of the effect is disputed. Participation over months with real accounts is the only approach that works. Link-dropping is detected and removed.

---

Should I focus on my own site or third-party sources?

Both, in that order. Owned content is cheapest to fix and it's the foundation. But the cited source is usually a third party, so a strategy that stops at your own domain will underperform.

---

How do I find out what's cited in my category?

Run twenty buyer prompts across the major engines and record the sources. Published studies give you averages across all queries; your category may behave nothing like the average.

---

How often should I re-check?

Quarterly. The volatility in the data means anything less frequent leaves you working from a picture that's already changed.