Do AI Engines Cite the Same Sources? Not Even Close

A team that tracks its brand on ChatGPT and assumes the result generalizes is making a measurable bet, and the measurements keep coming back against it. A 2026 academic audit accepted at EMNLP's Web Agents and Answer Engines workshop ran the same 500 questions through Google and Baidu in two languages and found the engines' answers were nearly interchangeable — median semantic similarity up to 0.81 — while the domains they cited barely intersected at all: a Jaccard overlap of 0.032 between Baidu and Google answering identical Chinese queries, and 0.034 between Google's own English and Chinese overviews. The engines agree on what to say. They disagree, almost completely, on whom to credit for it.

Key facts

Key takeaways

The EMNLP workshop audit — by researchers including Politecnico di Milano's Francesco Pierri — is the first to run a controlled cross-platform, cross-language comparison of answer-engine sourcing. The design is simple: 500 queries sampled from MS MARCO, translated into Simplified Chinese, submitted to Google and Baidu in both languages between 20 June and 10 July 2026, for 1,998 observations across four platform-language settings.

The first finding is that the surfaces do not even fire equally. Google produced an AI overview for about 97% of queries in both languages; Baidu answered 82% of Chinese queries and just 4% of English ones. Whether an AI answer exists for your buyers' question at all is already platform- and language-specific.

Where overviews did appear, the citation pools barely touched:

Two engines answering the same 500 questions in the same language shared about 3% of their cited hosts. The same engine answering in two languages shared about 3% too — which matches the direction of CiteLens's earlier Turkish–English comparison (22% shared sources) while landing even lower on a stricter metric and a broader query set.

Concentration diverged as sharply as membership. Baidu's citations were dominated by its own properties — baike.baidu.com, baijiahao.baidu.com, zhidao.baidu.com — with a Gini coefficient of 0.895 and 27.1% of all visible exposure going to a single host. Google spread the same questions across roughly 2,700–2,800 unique hosts with a top host share under 7%. One engine runs a closed loop over its own ecosystem; the other samples widely from the open web. A brand can be structurally excluded from one pool and comfortably present in the other while doing nothing differently.

And here is the part that should reframe how you read every engine comparison: the answers converged anyway. Median answer-text similarity across matched queries ran from 0.701 (Google English vs Baidu Chinese) to 0.813 (Google Chinese vs Baidu Chinese). The paper's conclusion is that answer similarity and source exposure are distinct dimensions of AI search — engines can tell users nearly the same thing while crediting completely different sources for it. If your concern is whether your brand is the one being credited, measuring the answer text on one engine tells you close to nothing about the citation set on another.

If cross-engine divergence were just Google-versus-Baidu ecosystem politics, you could dismiss it. It is not. The SIGIR 2026 benchmark — 11,500 queries against Google's traditional results, AI Overviews, and Gemini, with the dataset published — measured source overlap between Google's own two generative surfaces at a Jaccard of 0.11. AI Overviews and Gemini agree with each other less than either agrees with the ranked list (0.18 and 0.16), despite sharing an index, a parent model family, and an owner. We covered that study's rank-overlap findings in our analysis of rankings and AI citations; the cross-surface number is the one that matters here.

Off Google, the gaps are the same shape. Measured against a common reference — Google's top 10 for the same query — Ahrefs' four-assistant study found the assistants landing at very different distances from it: 8% of ChatGPT's cited links ranked there against 28.6% of Perplexity's, a 3.5x spread that cannot happen if the assistants draw from one shared pool. Profound's ~680-million-citation corpus shows the pools directly: each engine has its own source diet. Wikipedia is 7.8% of ChatGPT's citations and 0.6% of AI Overviews' and Perplexity's. Reddit is 6.6% of Perplexity's and 1.8% of ChatGPT's. These are not noisy estimates of one underlying citation distribution. They are different distributions.

Source membership is only the first axis of disagreement. The Semrush and Kevin Indig ghost-citations study found ChatGPT and Gemini failing brands in mirror image: ChatGPT cited a domain in 87% of its brand appearances but wrote the brand's name into the answer only 20.7% of the time, while Gemini named brands in 83.7% of appearances and cited in 21.4%. Muck Rack's corpus adds citation frequency itself: citations appeared in 96% of ChatGPT responses, 82% of Gemini's, and 55% of Claude's — and Claude, when it does cite, lists roughly 13 sources against ChatGPT's 5, diluting any one citation's prominence.

A September 2026 preprint built on one monitoring vendor's data (Tannenbaum, arXiv:2609.23162) is consistent from yet another angle: across 34,960 prompt-engine observations, the same retrieval condition produced different brand-mention rates per engine — with the brand's own domain cited in the retrieval path, GPT mentioned the target brand 49.0% of the time and Gemini 58.4%. Same brand, same evidence in front of the model, ten-point behavioral gap.

Our own tracking shows the divergence day to day. In Elmo's 42-day citation study, Google AI Mode cited about 24 distinct domains per day for a tracked prompt set while ChatGPT cited about 17, and a single broad prompt drew on 445 distinct domains over six weeks on Google's surface. Engines also move on their own clocks: within one 13-week window, Semrush watched Reddit's presence in ChatGPT responses fall from roughly 60% to about 10% while other engines' mixes held — a single-engine model update that would have silently dragged any blended score with it.

The academic audits count domains, not brands. A Jaccard of 0.032 over host domains does not directly say that brand mentions disagree at the same rate. Answers converge semantically, so it is possible engines agree on which brands to name more than on which pages to cite — the ghost-citations data suggests naming and citing are partly independent behaviors. Nobody has published a brand-level cross-engine mention correlation. Until someone does, the safe reading is that citation presence — the thing your pages can earn — is engine-specific, and mention agreement is unmeasured.

MS MARCO queries are not buyer prompts. The EMNLP audit's 500 queries are general informational questions, not the comparison and recommendation prompts that decide shortlists. Evaluation-stage prompts pull from a different, more vendor-heavy pool — Ten Speed's data puts product pages at 24% of citations on those prompts — and cross-engine overlap on that slice has not been measured. It could be higher; the category's own pages are a smaller pool to draw from.

Both academic studies are snapshots. The Baidu–Google audit is three weeks in mid-2026 collected from Milan, which the authors flag may not reproduce mainland-China Baidu; the SIGIR collection is two days in December 2025, before the model change Ahrefs credits with shifting AI Overview sourcing. Overlap numbers move when engines do. The direction — low everywhere, on every pair, in every study — is what has held.

The practical question is scope: which engines does your tracking need to cover, and can any of them proxy for the rest? The data supports one rule with three parts.

The cost objection is real — every added engine multiplies prompt runs — but it prices the wrong risk. The expensive failure is not tracking four engines; it is spending two quarters optimizing against ChatGPT data while Gemini, drawing on a citation pool your dashboard has never seen, quietly recommends your competitor to the other half of your buyers.

Elmo is an open-source, self-hosted AI visibility platform that runs your prompt sets across ChatGPT, Claude, Gemini, Perplexity, and Google's AI surfaces, storing mentions, citations, and competitors per engine on every run. Because results carry the engine on every row and the data is yours to query, per-engine baselines and cross-engine comparisons are a GROUP BY, not a feature tier.

For the fundamentals, start with AI citations and AI share of voice, then how to track your brand in AI search. For why two tools disagree even on one engine, see why AI visibility scores differ; for how sources churn within an engine, citation volatility; for the cite-without-naming problem, ghost citations. For the vocabulary, see the AI search glossary.