AI Engine Source Preference Correlation: Do ChatGPT, Perplexity, and Gemini Agree on Which Sources to Cite?

AI answer engines agree on which sources to cite between 16% and 59% of the time, depending on the engine pair. This means a source that earns consistent citations from one engine has, at best, a coin-flip chance of being cited by another — and at worst, roughly a one-in-six chance. The practical consequence: a single "AI optimization" strategy aimed at one engine will miss most of the citation surface across the rest.

This analysis examines cross-engine source preference correlation using data from BrightEdge's citation analysis, Conductor's 7-month multi-engine study, a 412-query overlap study, Aether AI's six-engine comparison, academic research on latent LLM source preferences, and Machine Relations Index (MRI) citation data across 7,184 domains and six engines.

How Much Do AI Engines Actually Agree on Source Selection? #

The headline finding across multiple 2026 datasets: AI engines disagree on sources far more than they agree.

BrightEdge's analysis of hundreds of millions of citations found pairwise source overlap ranging from 16% to 59% across major engines. The highest agreement — 59% — occurs between Google AI Mode and Google AI Overviews, which share retrieval infrastructure. The lowest overlap — as low as 16% — appears between engines with fundamentally different retrieval architectures, such as ChatGPT and Gemini.

Conductor's 7-month study across seven engines and 1,056 data points confirmed that each engine maintains what the researchers called a distinct "editorial identity" — a consistent source preference pattern that holds across intent categories and time periods.

A separate study of 412 client-facing queries across four engines found that all four engines cited the same source only 12% of the time. Engine-unique citations — sources cited by only one engine — accounted for 41% of all citations. In a six-week longitudinal follow-up, all-engine overlap fluctuated between 9–16% week to week.

Aether AI's analysis of six engines found an average cross-engine citation overlap of just 23%, meaning roughly 77% of sources cited by one engine do not appear in another's response for the same query.

ZipTie's cross-platform analysis measured the fragmentation at the domain level: only 11% of domains are cited by both ChatGPT and Perplexity for the same query, 71% of all cited sources appear on just one AI platform, and only 7% of sources achieve citation across all major engines. Even within Google, AI Overviews and AI Mode cite the same URLs only 13.7% of the time.

The correlation is even weaker at the URL level than at the domain level. At the brand level, agreement rises to 36–55%, according to BrightEdge — still below majority consensus for most pairs.

Cross-Engine Source Preference Patterns #

Each engine has measurable source-type biases that persist across queries and time:

Aether AI's analysis adds that Perplexity skews 62% of citations toward .edu and .org domains, while Gemini cites non-top-10 organic results approximately 34% of the time compared to 12% for Google AI Overviews — confirming that even Google's own products apply different ranking logic to citation selection.

The divergence is not random. Research on 12 LLMs across six providers found that latent source preferences are predictable, context-sensitive, can outweigh content quality, and persist even when explicitly prompted to avoid them. A study of 13 open-weight LLMs confirmed that models prefer institutionally-corroborated information (government and news sources) over social media and personal sources — but this preference can be reversed by simple repetition of lower-credibility content, suggesting source selection is more fragile than it appears. A separate audit using the ChoiceEval framework across Gemini, GPT, and DeepSeek found that U.S.-developed models show systematic favoritism toward American entities, while DeepSeek shows more balanced but still detectable geographic preferences — evidence that training data geography shapes source selection at the model level.

Why Google's Own Engines Disagree With Each Other #

One of the most counterintuitive findings: Google operates three AI answer products — AI Overviews, AI Mode, and Gemini — and they cite different sources for the same queries.

For Purchase intent, AI Overviews cites YouTube content while AI Mode routes to Google Shopping properties. For Education intent, AI Mode uniquely surfaces LinkedIn content, which neither AI Overviews nor Gemini selects. Gemini anchors to YouTube across nearly all intent categories, while AI Overviews distributes citations more evenly across Reddit, YouTube, and Quora.

The source overlap between Gemini and Google AI Mode is only 27% — lower than some cross-company engine pairs.

This intra-company divergence confirms that source preference is driven by retrieval architecture and training data, not by corporate policy or content quality standards applied uniformly across products.

MRI Evidence: Per-Engine Citation Distribution for Elite Sources #

Machine Relations Index data across 7,184 tracked domains shows how dramatically per-engine citation counts diverge even for sources with Elite-tier consensus scores.

Source: Machine Relations Index, 30-day citation window, MRI Score v1.1 six-engine methodology.

Several patterns emerge from MRI data:

• Google AI Mode consistently dominates citation volume for market database sources, producing 40–47% of total citations for Crunchbase and G2.
• ChatGPT contributes the fewest citations — 1.7% to 5.1% of totals — despite being the most widely used consumer AI tool.
• Claude shows disproportionate citation of analyst sources (64 citations for Crunchbase vs. 22 for Deloitte), suggesting a preference for structured company data over consulting analysis.
• Google AI Overviews cites sparingly across all three sources (2–4% of totals), consistent with its more constrained citation format.

The per-engine distribution for a single source can vary by more than 20:1 (Crunchbase: 94 Google AI Mode citations vs. 4 ChatGPT citations). A source's total citation count masks where those citations actually come from.

Citation Breadth vs. Citation Depth: A Second Axis of Divergence #

Source preference correlation is not the only axis where engines diverge. Research analyzing 602 prompts across ChatGPT, Google AI, and Perplexity found that citation breadth and citation depth are inversely correlated across platforms.

Perplexity and Google cite more sources per response but absorb less content from each. ChatGPT cites fewer sources but achieves stronger "citation absorption" — the degree to which a cited page's language, evidence, and structure actually appear in the generated answer.

This means two sources can have identical citation counts but radically different influence. A source cited by ChatGPT may contribute more to the actual answer than a source cited by Perplexity, even if Perplexity cites it more frequently.

For Machine Relations strategy, this distinction matters because it separates visibility (being listed as a citation) from authority (having your content shape the response).

What Drives Source Preference Convergence #

Despite the low overall correlation, engines do converge on certain source characteristics:

1. Structured, extractable content. Sources with clear definitions, comparison tables, and numbered data points earn citations across engines more consistently than narrative-heavy content. Conductor's analysis found that engines spanning different architectures all favored structured evidence when available. The 412-query overlap study confirmed that sources achieving full four-engine overlap were more likely to answer the prompt directly in the first 150 words and use structured schema markup in visible HTML. ZipTie's content-type analysis measured the effect: original research earns a 38–65% citation rate across engines, compared to 6–15% for standard blog posts and 3–8% for product or marketing pages.

2. Brand authority signal. Ahrefs' study of 75,000 brands found that branded mentions correlate 0.664 with AI visibility across engines, compared to 0.218 for backlinks. The top quartile of brands by mention volume earned 10x more AI citations than the next tier.

3. Commercial domain dominance. Across 680 million citations, commercial (.com) domains account for 80.41% of all AI citations. Nonprofit (.org) sites represent 11.29%. This concentration persists across all measured engines.

4. Decoupling from Google rankings. Only 12% of URLs cited by ChatGPT, Gemini, and Copilot rank in Google's top 10. Perplexity shows the highest overlap with traditional search at roughly 29%. Google AI Overviews has 38% top-10 overlap — the highest, but still a minority.

Machine Relations Implications #

The low cross-engine correlation rate has direct consequences for how companies should approach AI visibility:

Single-engine optimization is a partial strategy. An approach tuned for ChatGPT's encyclopedic preference will underperform on Perplexity's community-content bias and miss Claude's institutional-source preference entirely. The 16–59% overlap range means any single-engine strategy leaves 41–84% of the citation surface unaddressed.

Source architecture matters more than content volume. The convergence factors — structured content, brand authority, commercial domain presence — represent the shared floor across engines. Building these properties earns a baseline across all platforms, even as engine-specific preferences determine marginal citations.

Per-engine citation tracking is necessary. Aggregate citation counts obscure which engines are actually citing a source. MRI data shows that a source with 234 total citations can have a 23:1 ratio between its highest-citing and lowest-citing engine. Without per-engine decomposition, companies cannot diagnose where their citation authority is strong and where it is absent.

The correlation will likely decrease. As engines differentiate their retrieval architectures, fine-tune on different data, and respond to different user bases, source preferences will diverge further. Early Machine Relations measurement that captures per-engine baselines now will be more valuable as the ecosystem fragments.

FAQ #

What is the source preference overlap rate between AI engines? #

Pairwise source overlap ranges from 16% to 59% depending on the engine pair, according to BrightEdge's analysis. The highest overlap (59%) occurs between Google AI Mode and AI Overviews. At the brand level, agreement rises to 36–55%.

Do Google's own AI products agree on which sources to cite? #

No. Google AI Overviews, AI Mode, and Gemini show as low as 27% source overlap with each other. Each product uses different retrieval logic, so they select different sources even for the same query.

Does ranking well in Google Search help with AI engine citations? #

Minimally. Only 12% of URLs cited by ChatGPT, Gemini, and Copilot rank in Google's top 10. ChatGPT cites pages ranking position 21 or lower 90% of the time. Google AI Overviews shows the highest overlap at 38%, but this is still a minority of citations.

What predicts cross-engine citation more than backlinks? #

Branded mentions. Ahrefs found that brand mention volume correlates 0.664 with AI visibility, compared to 0.218 for backlinks. Top-quartile brands by mention volume earned 10x more AI citations than the next tier.