Do ChatGPT, Perplexity and Gemini cite the same sources?

It is tempting to imagine one "AI citation" leaderboard that ChatGPT, Perplexity, Gemini and Google's AI Overviews all draw from. The published data says the opposite. Across 161,286 prompts, any two engines shared only about 17 percent of the sources they cited for the same question, and just 3.8 percent of sources were cited by all four. Being named by one engine tells you almost nothing about the others. Here is what the third-party studies actually show, and why every number in this space needs a date stamp.

The short answer: mostly no

Do the major AI engines cite the same sources for the same question? On the strongest available evidence, the answer is mostly no, and by a wide margin. The clearest data point comes from Writesonic's 2026 study, "Do AI Engines Cite the Same Sources?", which measured citation overlap across 161,286 prompts using Jaccard similarity, a standard set-overlap metric that divides the sources two engines share by the total distinct sources they cite between them. Naming the metric matters: it is what makes the finding reproducible rather than a vibe.

The headline numbers are stark. For the same prompt, any two engines shared only about 17 percent of the sources they cited. Just 3.8 percent of sources were cited by all four engines in the study. And roughly 72 to 73 percent of cited domains appeared on exactly one engine and no others. In other words, the typical source that helps you on one engine does nothing for you on the rest. There is no single leaderboard. There are four largely separate ones.

Even the closest pair barely agrees

It is worth sitting with how low these overlaps are, because they cut against the intuition that these systems are all crawling the same web and should converge. In the Writesonic data, the closest pair was Perplexity and Google's AI Overviews, with a Jaccard similarity of just 0.237. The farthest apart were ChatGPT and Gemini, at 0.119. So even the two engines that agree most still disagree on roughly three-quarters of their sources, and the two that agree least share barely more than one source in ten.

That spread is the practical point for anyone trying to be cited. A strategy that wins on ChatGPT is not a strategy that wins on Gemini. The pair-by-pair numbers tell you that the gap is not a rounding error you can ignore; it is the dominant feature of the landscape. If you only measure your visibility on one engine, you are flying blind on the other three.

"There is no single AI citation leaderboard. Any two engines share only about a sixth of their sources, and just 3.8% of sources are cited by all four."

Want to know which engines actually cite you, and which ignore you? Get a free visibility audit and see exactly how ChatGPT, Gemini and Perplexity describe you today.

Each engine has a source personality

Low overlap is not random noise. The reason the engines diverge is that each one has a distinct taste in sources, and those tastes are visible in the data. The largest window into this comes from Profound, whose analysis of roughly 680 million citations between August 2024 and June 2025 mapped where each platform leans.

The patterns are recognisable once you see them. ChatGPT skews toward Wikipedia, which made up about 7.8 percent of its citations, and toward established media outlets: it behaves like a system that prefers encyclopaedic and editorial sources. Perplexity skews hard toward Reddit, which accounted for about 6.6 percent of its citations and, more strikingly, roughly 46.7 percent of its top-ten cited domains, giving it a distinctly community-and-forum flavour. Google's AI Overviews came out as the most balanced of the set, with Reddit at about 2.2 percent and YouTube at about 1.9 percent, spreading its citations more evenly than the others. Same web, three different appetites.

These personalities are why the overlap is so low. An engine that reaches for Wikipedia and an engine that reaches for Reddit will name different pages for the same question, not because one is right and one is wrong, but because they weigh source types differently. If you want to be cited, you have to understand which appetite you are feeding.

The domains that dominate everywhere

Underneath the per-engine differences, there is a smaller set of platforms that show up at or near the top almost everywhere: Reddit, Wikipedia, YouTube and LinkedIn. Semrush, analysing 230,000 prompts and over 100 million citations collected between July and October 2025, put this cluster of user-generated and reference platforms at the front. Peec AI, examining around 30 million sources, reported the same domains leading, with Reddit, YouTube and LinkedIn among the most-cited, in coverage via Search Engine Land in March 2026.

One honest caveat on Peec AI: it published a ranking of the most-cited domains, not clean per-domain percentages, so treat its finding as an ordering rather than precise shares. The safe, well-supported takeaway across both studies is qualitative and directional: a handful of large community and reference platforms punch far above their weight in AI answers, which is why so much AEO advice points toward earning a credible presence on them. But the ordering shifts by engine and, as the next section shows, by month.

The volatility that undermines every snapshot

Here is the finding that should change how you read every other number in this article, including ours. Citation patterns are not stable. They move, sometimes violently, over a few weeks. Semrush documented a striking example: between August and September 2025, ChatGPT's Reddit citations collapsed from around 60 percent of responses to about 10 percent, and its Wikipedia citations fell from around 55 percent to under 20 percent, in the space of a single month.

Sit with the size of that. A source that appeared in the majority of ChatGPT's answers in August was a minor player by September. Any confident, undated claim that "ChatGPT loves Reddit" or "Wikipedia dominates ChatGPT" was true and then false within weeks. This is why every load-bearing number in this space has to carry a collection window. A citation snapshot without a date is not a fact about how the engines behave; it is a fact about how they behaved once, briefly, and it may already be wrong.

The practical consequence is that citation visibility is not a project you finish. It is a signal you monitor. The engines re-weight their sources on their own schedule, without announcement, and a strategy pinned to last quarter's pattern can quietly stop working. Treat any study, this one included, as a dated reading rather than a permanent law.

Why this matters even less than you think: clicks

There is a sharper reason to care about being the cited source rather than just a ranked link. Pew Research Center, in a study published on 22 July 2025 covering the behaviour of 900 US adults across 68,879 searches, found that when a Google search returned an AI summary, the rate at which users clicked any traditional link dropped to 8 percent, versus 15 percent without a summary. More pointedly, only about 1 percent of users clicked a source cited inside the AI summary itself.

Read alongside the overlap data, this is the whole argument for taking AI citations seriously and measuring them per engine. The click that used to reward a good ranking is drying up as summaries answer the question in place. What is left is being named and described accurately inside the answer, on the specific engine your buyer happens to use. And because the engines share so little, "the specific engine" is not a detail; it is the entire strategy.

We can put a small, honest face on this from our own logs. On 5 August 2026, ChatGPT sent a reader to one of our own pages, our guide to AEO tools, having cited it in an answer. It was a single citation and a single click, from one person, nothing more, and we would be lying if we called it a trend. But it is the exact mechanism this article describes, observed in our own data rather than someone else's study: an engine read a page, decided it was worth quoting, and named the source. One citation is an anecdote, not a number. The point of measuring is to turn anecdotes like that into a dated, per-engine pattern you can actually act on.

A note on where citations sit on the page

One adjacent finding helps explain why structure matters, not just which domain you are on. Analysis by Kevin Indig, reported via Search Engine Land, found that about 44 percent of ChatGPT's citations came from the first third of the source page. The engines are not reading your whole document with equal weight; the top of the page carries disproportionate influence over what gets quoted. That is a reason to state your most important, most quotable facts early, rather than burying them beneath a long preamble. It is engine-specific and should be treated as directional, but it points the same way as everything else: make the citable claim easy to find and easy to lift.

What makes a study like this worth citing

Because this piece is built entirely on other people's data, it is worth being explicit about why we trust some numbers and hedge others. The studies that earn a citation, from us and from the engines, share a short list of properties, and they are the same properties that make any research credible.

• A stated sample size. "161,286 prompts", "680 million citations", "230,000 prompts and over 100 million citations", "68,879 searches". A number you can weigh beats an anecdote you cannot.
• A named, reproducible method. Writesonic naming Jaccard similarity is what lets anyone else run the same test and check the result. A method you can name is a method you can challenge.
• A dated collection window. Profound's August 2024 to June 2025, Semrush's July to October 2025, Pew's July 2025. In a space this volatile, a study without a window is nearly unusable.
• Per-engine breakdowns and honest limits. The useful studies report each engine separately and say what they cannot conclude, rather than blending everything into one flattering average.

Where a source falls short of that bar, we have said so: Peec AI's rankings are treated as an ordering, not percentages, and per-engine "citations per answer" figures that float around online without a traceable primary source are left out entirely. Precision about your own limits is not a weakness in a data piece. It is the thing that makes the rest of it worth believing.

What this means for your visibility

Put the findings together and the instruction for brands is unusually clear. You cannot treat AI citation as one thing to win. You have to treat it as four related but separate problems, monitored over time.

• Measure every engine separately. With any two engines overlapping on only about 17 percent of sources, a strong ChatGPT result tells you little about Gemini, Perplexity or AI Overviews. Test each with the real questions your buyers ask.
• Match the engine's appetite. ChatGPT leans encyclopaedic and editorial, Perplexity leans toward community and forums, AI Overviews spreads more evenly. Where you invest in corroboration should reflect where each engine actually looks.
• Re-measure on a schedule. A month can turn a dominant source into a minor one. Treat visibility as a monitored signal, not a one-off audit, and date-stamp everything you record.
• Be the clearest, most quotable source about your own facts. Across every engine and every window, the sources that get lifted are the ones that state their facts plainly and early, in language an engine can quote without interpreting.

What Stellarcast is measuring next

Everything above comes from published third-party research. We think it is the honest state of the evidence today, and we are not going to dress up someone else's dataset as our own. But the gaps in the public data, especially the lack of a consistent, repeated, per-engine measurement over time, are exactly the ones we are now setting out to fill.

Stellarcast is running an ongoing prompt-panel measurement of AI citations across ChatGPT, Gemini, Perplexity and Google's AI Overviews, and we will publish it. The methodology will follow the same standard we hold others to: a fixed panel of real buyer-style prompts, a stated sample size, per-engine breakdowns, a clearly dated collection window for every snapshot, and repeated re-measurement so the volatility is visible rather than hidden. We have written up exactly how our measurement works in how Stellarcast measures AI visibility, so you can hold our numbers to the same test we apply to everyone else's. We are not going to preview numbers we have not finished collecting, because a made-up figure would defeat the entire point of a piece about credibility. When the data is ready and the method is documented, it will appear here with its window attached. Until then, treat the studies above as the best available reading, and treat every one of their numbers as dated.

The takeaway

ChatGPT, Perplexity and Gemini do not cite the same sources, and it is not close. Any two engines share only about 17 percent of their cited sources, just 3.8 percent of sources reach all four, and roughly 72 to 73 percent appear on a single engine alone. Each has a distinct source personality, a shared cluster of dominant platforms sits underneath, and the whole picture can shift in weeks. The strategy that follows is simple to state and harder to do: measure each engine, match its appetite, date-stamp your numbers, and be the clearest source about your own facts. There is no one leaderboard to win, so win the ones that matter to you, one engine at a time.