The Maze: AI discovery is at least two channels. In an NP Digital test of 1,000 matched prompts, only 2.6% of unique source domains appeared in both Google AI Overviews and ChatGPT. ChatGPT alone accounted for 60.3% of the source pool; Google alone accounted for 37.1%. The result turns “rank in AI” from a single objective into a portfolio problem.
The headline is the missing middle. The three shares form a complete source pool: 60.3% ChatGPT-only, 37.1% Google-only and 2.6% shared. This does not mean ChatGPT answered 60.3% of queries better, or that Google sent 37.1% of the traffic. It says the two systems drew from almost separate sets of domains in this test. A brand can be visible in one answer environment and absent from the other—even when the user asks the same question.
More cited domains is not automatically a win. NP Digital notes that ChatGPT tended to cite more domains, which helps explain its larger exclusive share. But breadth is not quality, prominence or commercial value. A domain cited once at the bottom of an answer is not equivalent to a source that repeatedly shapes recommendations. Teams need to track citation presence, position, context and downstream visits separately. Otherwise, a growing citation count becomes the AI-era version of celebrating impressions.
Independent research points in the same direction, with different numbers. Parse ran 18,206 identical buyer-intent prompts through ChatGPT Search and Google AI Mode and found 6.2% source overlap. Only six domains appeared in both platforms' top-25 source lists. Yet Parse found Google AI Mode cited 14.0 sources per answer versus 3.7 for ChatGPT Search—the opposite directional pattern from NP Digital's high-level note. That is not a contradiction to smooth away. It is a warning that product surface, prompt mix, model version and counting rules materially change the result. “AI citations” is not one stable market statistic.
Google's own sourcing already reaches beyond classic page-one winners. A 2026 preprint studying 7,583 Google AI Overviews collected 61,212 reference URLs across 7,479 hostnames. Roughly 30% of cited domains did not appear on the first page of the matching standard results. That weakens the comforting assumption that strong conventional rankings will automatically carry into synthesized answers. Search authority still matters, but retrieval, passage clarity, original evidence and query fit can reshuffle the shortlist.
Commerce teams need two operating loops. Start with an engine-by-engine prompt set built around category discovery, product comparison, problem solving and brand choice. Record which domains are cited, what claims they support and whether the brand appears directly, through retailers, through reviews or through third-party research. Then create assets that earn references: original datasets, transparent comparisons, stable product facts, expert explanations and pages that answer one decision cleanly. The goal is not to manufacture citations. It is to become useful evidence in each system's separate source market.
Why it matters: Google and ChatGPT may look like adjacent boxes in the same discovery funnel, but their evidence layers can be radically different. A single “AI visibility” score will hide the gap. Brands should measure each surface separately, test the content formats each one rewards and connect citations to qualified visits and sales. The 2.6% overlap is not a universal law. It is a sharp operating signal: visibility in one answer engine does not insure visibility in the other.
