Perplexity never touches 89.1% of the websites ChatGPT cites for the exact same question. Run it the other way and ChatGPT misses 90.1% of Perplexity's sources on those same questions. That's the headline from Wellows' analysis of 22.7 million citations across 1.15 million questions, published July 29, 2026 and updated August 6, 2026: the five major AI engines are not reading from the same shortlist. Each one has built its own, almost entirely separate pool of sources, even when answering the identical prompt.
If you've only ever checked your ChatGPT citation status, this data says that number tells you almost nothing about Gemini, Perplexity, Google AI Overviews, or AI Mode.
What did the Wellows study actually measure?
Wellows analyzed 22.7 million citations pulled from 1.15 million questions answered by ChatGPT, Gemini, Perplexity, Google AI Overviews, and Google AI Mode between January and June 2026. The citations trace back to 441,946 distinct websites. Traffic sources were 84% US, with the remainder spread across 27 other markets including Australia and the UK, and every question was in English.
This is a size and scope that's hard to argue with. It isn't five people running fifty test prompts and eyeballing the results. It's a systematic pull of what each engine actually linked to, question by question, across half a year.
Why don't ChatGPT and Perplexity cite the same sources?
They don't cite the same sources because each engine runs its own retrieval pipeline, built on different indexes, different ranking logic, and in some cases different underlying search partnerships, so the shortlist of "plausible sources for this question" barely overlaps before either model even starts writing an answer.
The gap runs almost exactly both ways: ChatGPT cites a source Perplexity skips 89.1% of the time, and Perplexity cites a source ChatGPT skips 90.1% of the time, on the same question. That symmetry matters. It rules out the easy explanation that one engine is simply "worse" at finding good sources. Both are finding plenty of sources. They're just not finding the same ones.
How much citation overlap is there across all five engines?
Across the full five-engine set, on questions where all five returned an answer, 79.6% of cited sources show up on only one engine. Another 14.5% make it onto exactly two engines. Only 0.31% of sources are cited by all five.
| Engines citing the same source | Share of citations |
|---|---|
| Exactly 1 engine | 79.6% |
| Exactly 2 engines | 14.5% |
| 3 engines | ~4.4% |
| 4 engines | ~1.2% |
| All 5 engines | 0.31% |
(Wellows reports the 1, 2, and 5-engine figures directly; the 3- and 4-engine shares are the approximate remainder implied by those numbers, since the reported categories don't sum to 100% on their own.)
Wellows also checked whether this was a matching-methodology artifact, loose URL comparisons inflating the apparent gap. It wasn't. Overlap stays low even under stricter matching: 85.1% of sources fall outside the overlap when excluded at the exact-page level, and 93.2% when matched at the raw-URL level. Tightening the definition doesn't close the gap. If anything it confirms it.
What does the brand-vs-page gap actually mean for GEO strategy?
It means the raw 30.3%-vs-6.8% gap measures how easy each target is to hit, not how strong a signal either one is. Once Wellows ran a permutation test to strip out pool-size effects, the order flipped: exact-page agreement is 42x above what random chance would produce, brand agreement is 17x. Page agreement is the statistically stronger signal. Brand agreement just has a smaller field to converge on, a handful of plausible companies for most categories versus thousands of candidate pages, so two independent retrieval pipelines land on the same brand more often by pool size alone.
That's a real distinction for how you sequence GEO work, not a reason to skip page-level optimization. A single page is competing against every other page an engine's retrieval system could plausibly surface, and the 6.8% figure says that competition rarely resolves the same way twice, which is exactly why it's the harder target and the more meaningful one to win. Brand-level signals (consistent entity data, third-party corroboration across independent sites, a name that shows up in the same breath as a category regardless of which page gets linked) are worth building because they're the lower bar, a reasonable place to start, not because they outperform page-level work once you control for pool size.
In practice that's a mix of GEO optimization work on your highest-value pages, which is where the harder, more durable win sits, plus consistent entity signals everywhere else your brand shows up online.
Citation status on one engine doesn't predict citation status on the others: 79.6% of citations across ChatGPT, Gemini, Perplexity, Google AI Overviews, and Google AI Mode are exclusive to a single engine. If you only monitor ChatGPT, you have no idea what's happening on the other four. Pair multi-engine monitoring with GEO work at both levels: brand-level signals (entity clarity, structured data, third-party mentions) are the easier win because there's a smaller pool of plausible brands to converge on, while page-level optimization is the statistically stronger signal, a 42x-above-chance event versus 17x for brand, once you control for that pool-size difference.
Does this mean single-engine citation tracking is useless?
No, but it means single-engine tracking answers a narrower question than most teams think it does. Knowing your ChatGPT citation rate tells you your ChatGPT citation rate. Given a 79.6% engine-exclusivity rate on citations generally, and an 89-90% one-way gap specifically between ChatGPT and Perplexity, it does not reliably tell you anything about Gemini, Perplexity, or either Google surface. A business that looks well-covered in one engine and invisible in the other four would look "fine" on a ChatGPT-only dashboard. That's the blind spot this study puts a number on, and it's the same blind spot AI SEO services built around single-engine tracking tend to miss.
Why do engines converge more on brands than on pages?
Engines converge more on brands than on pages because the candidate pool is smaller. For most questions, there's a limited set of companies that plausibly belong in the answer at all, maybe five or ten real contenders, while the set of individual pages that could serve as the cited source for any one of those companies is much larger: a homepage, a pricing page, a blog post, a help doc, a third-party review. Two independent retrieval systems are more likely to land on the same company by chance (or by shared signal) than on the exact same page from that company. Wellows frames this as a structural effect of pool size, not evidence that some brands are inherently more citable everywhere.
Is this specific to English-language, US-heavy results?
Mostly, yes, and that's worth flagging rather than glossing over. Wellows' dataset is 84% US traffic, with the rest spread across 27 markets, and every question sampled was in English. The overlap findings hold up well within that scope (22.7 million citations is a large sample), but they shouldn't be assumed to hold identically in non-English markets or regions with different search infrastructure and different engine market share. Treat the specific percentages as US/English-market findings and the general pattern (low cross-engine overlap, higher brand-level than page-level convergence) as the more transferable claim.
FAQ
What is AI citation overlap?
AI citation overlap measures how often two or more AI answer engines (ChatGPT, Gemini, Perplexity, Google AI Overviews, Google AI Mode) cite the same source for the same question. Wellows' July 2026 study found this overlap is low: 79.6% of citations across all five engines are exclusive to just one engine.
Why does ChatGPT cite different sources than Perplexity for the same question?
Each engine runs its own retrieval pipeline with a different index and different ranking logic, so the shortlist of plausible sources rarely matches before either model generates an answer. Wellows found ChatGPT misses 90.1% of Perplexity's sources and Perplexity misses 89.1% of ChatGPT's, on identical questions.
Should I track my citations on more than one AI engine?
Yes. Given that 79.6% of citations are exclusive to a single engine, tracking only ChatGPT (or any one engine) leaves a large blind spot on the other four. A business can look well-cited on one engine and be effectively invisible on the rest, and there's no way to know without checking each engine directly, which is what a multi-engine AI visibility audit is for.
What's the difference between brand-level and page-level GEO?
Page-level GEO optimizes a single URL's structure, schema, and content for citability. Brand-level GEO builds consistent entity signals (structured data, third-party corroboration, category association) so the company itself, not just one page, gets recognized and mentioned across engines. Wellows found engines agree on the same brand 30.3% of the time versus the same exact page only 6.8% of the time in raw terms, but when they ran a permutation test against random chance, the order flipped: page agreement is a 42x event above chance versus 17x for brand. Brand agreement is easier to achieve because there's a much smaller pool of plausible brands than plausible pages, which makes it a reasonable starting point, but page-level agreement is the stronger signal once you control for that.
How is this different from the Google-rank-vs-AI-citation research?
That's a separate comparison. Research on Google rank vs. AI citation overlap measures how often a page ranking well in Google also gets cited by AI engines, a search-engine-to-AI-engine comparison. The Wellows study covered here measures overlap AI-engine-to-AI-engine, whether ChatGPT and Perplexity (for example) cite the same sources as each other. Both point to fragmentation, but they're measuring different gaps.
Is this the same as the "mentions vs. citations" distinction?
No. Why isn't my business cited by AI covers the gap between a business being mentioned by an AI model from training knowledge versus being cited with an actual source link, a distinction about citation mechanics. This Wellows study assumes citations are already happening and measures whether different engines cite the same sources as each other. They're related but distinct problems, and worth reading together if you're building a full picture of your AI visibility.
Does a high citation rate on ChatGPT mean I'm visible everywhere?
No. Given the 89.1%/90.1% one-way gap between ChatGPT and Perplexity specifically, and the 79.6% engine-exclusivity rate overall, a strong ChatGPT citation rate says very little about Gemini, Perplexity, or either Google AI surface. Each needs to be checked and optimized on its own terms, combining brand-level entity work (the easier target, since there's a smaller pool of plausible brands to converge on) with page-level optimization, which Wellows' own permutation test found is the statistically stronger signal of the two.
Sources: Wellows, "AI Citation Overlap Study", published July 29, 2026, updated August 6, 2026.
This post is part of our GEO Optimization guide. Related reading: Multilingual GEO, ChatGPT vs Gemini, Product Pages Beat Blogs for B2B AI.
Abdelmoghit drives the content engine at AY Rank. He researches keywords, plans content clusters, and produces citation-optimized articles that rank in both Google and AI search engines.
Full Bio →
