Do ChatGPT, Gemini, and Claude agree on which companies to recommend?

If you are trying to figure out whether an AI assistant's recommendation reflects some underlying consensus about who the good vendors are, the first thing to check is whether the assistants agree with each other. This study ran 14,951 distinct buyer-style prompts against 100,966 companies across three engines, ChatGPT, Gemini, and Claude, and counted how often a company named by one engine was also named by the others for the same prompt.

The answer: rarely. Looking at every company-prompt pairing where at least one engine named the company (337,500 such pairings), all three engines agreed on only 2.8% of them. The far more common outcome is a sole mention: one engine names a company and the other two do not. That happened 88.7% of the time. Two engines agreeing, with the third silent, happened in 28,662 pairings, or 8.5% of the total. All three agreeing happened in 9,541 pairings, 2.8%.

Why would this be true? Each engine draws on a different training corpus, a different retrieval or browsing mechanism, and a different ranking logic for deciding which handful of companies to surface out of a much larger field of candidates. Even if all three were drawing on roughly the same underlying set of facts about who exists and what they do, the step where the model picks which three, five, or ten to actually name is a lossy compression of a much longer list. Small differences in how each model weighs recency, brand familiarity, or category fit can easily produce a different short list each time, even without any disagreement about who is objectively good.

There is also a simpler, more mundane explanation worth naming: these are three separate products built by three separate companies, each with its own index of the web, its own training cutoff, and its own system prompt. Disagreement here does not necessarily mean one engine is wrong and the others are right. It means the three products are not drawing from a shared, stable ranking the way a search engine's organic results might be benchmarked against each other. A company absent from Claude's answer is not necessarily being penalized by Claude; it may simply not have been among the names Claude's generation process surfaced for that specific phrasing.

This matters for anyone treating a single AI assistant's answer as a reliable signal of market position. If you ask only ChatGPT and conclude your company is or is not a contender in its category, this study suggests that asking Gemini or Claude the same question would very likely produce a different short list. The practical implication is that visibility in AI search needs to be measured across engines, not sampled from one, in the same way that a company would never gauge its search visibility from a single Google query on a single day.

What would distinguish "these engines fundamentally disagree about quality" from "these engines sample a large qualified field differently each time" is harder to establish from aggregate agreement rates alone. It would require tracking whether the same company reappears across repeated runs of the same prompt on the same engine, to see how much of the variation is engine-to-engine versus run-to-run noise within a single engine. This study's window, August 4 to August 31, 2026, captures one run per prompt per engine, so it measures cross-engine agreement but not within-engine repeatability. A company that wants to know whether its absence from an answer is systematic or incidental would need the latter test too.