Does Where a Company Gets Ranked Depend on Which AI Engine You Ask?

This study tracked two separate things about how AI assistants handle buying questions: whether a company gets named at all, and if it does, where in the answer it lands. The two turn out to tell very different stories, and conflating them is an easy way to misread this kind of research.

The setup: three engines, ChatGPT, Gemini, and Claude, each on a single tested model version, answered 8,188 distinct prompts about a cohort of 58,803 companies between September 4 and September 25, 2026, producing 536,301 usable answers in total. Every time a company was named in an answer that listed multiple companies, the study recorded its position in that list, first, second, third, and so on. Across 197,893 such ranked mentions, the average position overall was 4.94.

Broken out by engine, the averages are close: 5.08 on ChatGPT, 4.97 on Claude, and 4.72 on Gemini. Gemini's average sits a bit earlier in the list than ChatGPT's, with Claude in between, but the spread across all three is well under one full position. Compare this to the finding on which companies get named at all, where 88.1% of named pairs were named by only one engine. Position, once a company clears the bar of being named, behaves nothing like that. It is remarkably consistent.

This distinction matters because it tells you where the real variance in AI search outcomes lives. It is not primarily in how an engine orders a list of companies it has already decided to mention. It is in the prior step, deciding which companies belong on the list in the first place. Once you're in, the position you get doesn't depend heavily on which engine is doing the listing. Getting in is the hard, inconsistent part. Where you land once you're in is comparatively predictable.

Why might this be true mechanically? One plausible explanation is that all three engines, despite drawing on different retrieval and ranking processes that produce very different name-selection outcomes, are converging on similar heuristics for ordering a final list once a candidate set has been chosen, perhaps something close to perceived prominence or relevance strength, which doesn't vary as sharply across model architectures as the underlying retrieval and recall step does. Another explanation, harder to rule out with this data alone, is that position compresses toward the middle of typical list lengths almost mechanically: if most answers list a handful of companies, the average position across many mentions will cluster near the middle of that range regardless of engine, simply because there isn't much room for it to spread out. Distinguishing a genuine ranking-logic convergence from a ceiling-and-floor artifact of short list lengths would require looking at the distribution of list lengths by engine, which this study's summary statistics don't break out.

For a company evaluating its own AI search performance, the practical implication is to treat these as two separate metrics with two separate fixes. If a company is absent from an engine's answers entirely, the lever to pull is about earning retrieval and recall in the first place, areas like consistent, well-structured public information, strong category association, and the kinds of signals the engine's underlying model drew on during training. If a company is appearing but landing late in the list, that's a narrower problem, closer to competitive differentiation within an already-cleared bar, and the near-identical averages across engines (5.08, 4.97, 4.72) suggest that fix should transfer across engines rather than needing to be solved three separate times.