Five AI Engines, One Question: Do They Cite the Same Sources?

This is the first post from GetCited Labs. The idea is simple: instead of repeating what everyone says about AI citations, we run controlled experiments on our own measurement engine and publish the data — method, numbers, and the caveats that keep us honest. Every post is one experiment.

So here's the first question. When five different AI engines answer the same question, do they cite the same sources? If they did, you could optimise for one and ride along on the rest. If they don't, you have to earn each engine separately.

What we did

We took five buyer-style questions in one category (someone shopping for an AI-augmented SEO service) and ran each one through five answer engines:

• ChatGPT (web search on)
• Claude (web search on)
• Perplexity (sonar-pro)
• Google Gemini (gemini-2.5-flash, grounded)
• Google AI Overviews

Five prompts, two repetitions each, five engines — 50 answers. For every answer we recorded each cited URL and reduced it to a domain. Then we measured how much the five engines' cited-domain sets overlap.

How much each engine cited (unique domains across the run):

• Google AI Overviews — 88
• Gemini — 73
• Perplexity — 41
• ChatGPT — 39
• Claude — 20

Right away, one thing is clear: the engines differ enormously in how widely they pull. AI Overviews cited four times as many distinct domains as Claude.

Finding 1 — They overlap less than you'd hope, more than "disjoint"

There are two honest ways to measure set overlap, and they tell different halves of the story.

Jaccard (shared ÷ combined) is brutal here: mean 0.08 across all pairs, ranging from 0.01 to 0.19. By that measure the engines look almost disjoint. But Jaccard is unfair when one set is 88 and another is 20 — the big set inflates the denominator and crushes the score.

So we also report the overlap coefficient (shared ÷ the smaller set), which asks the more useful question: of the domains this engine cites, how many does the other engine also cite? By that measure the mean is 0.25, and the picture is very different at the top:

• Claude × Google AI Overviews — 0.50
• Perplexity × Google AI Overviews — 0.51
• Claude × Gemini — 0.50

Half of what Claude cites also shows up in Google AI Overviews. That's not disjoint. AI Overviews behaves like a broad net that catches a big slice of what the others surface.

The takeaway: AI engines are not citing from sealed, separate universes — but they're not interchangeable either. Optimise for one and you'll get partial coverage of some others, and almost none of one in particular. Which brings us to the outlier.

Finding 2 — ChatGPT is the loner

ChatGPT shares almost nothing with anyone. Its overlap coefficient with the other four engines:

• vs Claude — 0.05
• vs Perplexity — 0.03
• vs Gemini — 0.08
• vs Google AI Overviews — 0.08

Where Claude, Perplexity, Gemini and AI Overviews form a loosely connected cluster, ChatGPT sits off on its own, citing domains the others mostly ignore (brandboost.ink, rankboostagency.com, thecmo.com). If your AI-visibility plan implicitly assumes "get cited and you're cited everywhere," ChatGPT is where that assumption breaks first.

Finding 3 — A small "universal" core does exist

Six domains were cited by four of the five engines:

thriveagency.com · eesel.ai · get-ryze.ai · snezzi.com · gomega.ai · respona.com

None were cited by all five (ChatGPT, again, opts out of most). But this is the most useful slice in the whole dataset: a handful of domains carry enough authority to be cited almost everywhere. Earning a mention on one of those compounds across engines — which is exactly the kind of placement that's worth chasing.

What this means if you want to get cited

1. Measure every engine. No single one proxies the rest — and the one most people care about (ChatGPT) is the least like the others. A one-engine check will mislead you.
2. Chase the universal core. The domains cited by 4+ engines are the high-leverage placements. Get referenced there and you show up in multiple answers at once.
3. Treat ChatGPT as a separate campaign. It rewards a different set of sources. Plan for it explicitly.

The honest caveats

We'd rather you trust the next ten posts than oversell this one.

• Small sample. Five prompts, two reps, one category (AI/SEO services), one locale (Australia). Treat the contrasts as the signal, not the exact decimals. Per-prompt Jaccard ranged 0.02–0.08 — the direction is stable, the precise number is not.
• Mixed measurement harness. ChatGPT and Claude were queried through their own APIs; Perplexity, Gemini and AI Overviews through a third-party SERP/LLM API. Part of why Perplexity and AI Overviews overlap most may be a shared retrieval stack, not a shared "mind." (Worth noting the opposite cuts too: ChatGPT and Claude share a harness and are the most disjoint pair — so harness isn't the whole story.)
• A bug we caught in the act. Our first pass reported Gemini citing exactly one domain and we nearly published "Gemini is near-disjoint from everything." It was wrong — Gemini returns citations as Google grounding-redirect links, with the real publisher domain tucked in a different field. Our extractor was reading the wrapper, not the domain. We found it because every finding goes through an adversarial review before it ships; fixing it turned Gemini from an outlier into one of the more connected engines. The corrected numbers are the ones above. We're telling you this because a lab that never shows its corrections isn't a lab.

That's experiment 011. The measurement engine, the five-surface coverage, and the review step that caught our own mistake are the same machinery we point at customer brands. Next experiment: which on-page changes actually move citations — measured before and after, not asserted.