Questions about this study

It means the two vendors showed up named together in the same AI-generated answer to a query. This study tracked such pairs across two different query framings, one carrying commercial intent (like a buying or comparison question) and one that did not, to see whether the pairing was a fixed association or something that only appeared under certain conditions. A pair that shows up together no matter how the question is framed is treated as a more durable co-mention than one that only appears when the question sounds like a purchase decision.

Mostly, yes, but unevenly. Across the full sample of 1,390 pair-by-group observations, 61.9% of pairs held together whether the query was commercial or not, while 38.1% appeared together only under commercial framing. That looks like a solid majority holding shape, but the headline number hides a big split depending on how tightly the underlying group of vendors was bound to begin with, which the segment breakdown addresses directly.

These are segments describing how strongly a set of vendors clusters together overall, independent of intent, before you even ask whether that clustering holds across framings. Loosely grouped pairs come from vendor sets with weak, diffuse association. Tightly grouped pairs come from sets with strong, consistent association. It matters because the intent-stability finding is almost entirely explained by this variable: loosely grouped pairs held across both intents only 50.8% of the time, essentially a coin flip, while tightly grouped pairs held 97.7% of the time, nearly categorical. The overall 61.9% figure is a blend of these two very different realities.

That's a fair concern and worth stating plainly. The tightly grouped segment has 88 units, versus 898 for the loosely grouped segment, so it's noticeably smaller. The 97.7% figure is real within that sample, but with a smaller denominator each individual pair carries more weight in the percentage, and the finding would benefit from replication on a larger tightly grouped sample before treating it as a fixed law rather than a strong pattern.

That's the right skepticism to bring, and the study can only partly answer it. Three mitigations were applied to the dataset before analysis specifically to reduce known sources of measurement bias, so the segment gradient is not simply the product of one uncorrected artifact. But mitigation is not the same as proof of a causal mechanism. An honest alternative explanation is that tightly grouped vendor sets are tightly grouped precisely because they compete on the same well-defined dimension regardless of intent, meaning tightness and stability could both be downstream of a third factor, like category maturity, rather than tightness causing stability. Distinguishing these would require tracking the same pairs as their group tightness changes over time, which this snapshot design cannot do.

It varies quite a lot. Looking at pairs that held across both intents, the share doing so ranged from 0% to 100% across the 28 groups counted, with a median of 78.9%. So some individual groups have almost none of their pairs surviving an intent change while others have essentially all of them. The segment-level averages are useful for describing the overall pattern, but any single group can sit far from that average, so a group-specific check is worth doing before acting on the segment number alone.

If a vendor's co-mentions with a competitor only appear under commercial-intent phrasing, that pairing is more likely a shallow, transactional association than a durable competitive set, and treating it as a stable rival relationship for positioning or messaging purposes would be a mistake. The practical move is to check where a given pairing sits: if it comes from a tightly grouped set, at 97.7% stability, it's likely a real, fixed category association worth planning around. If it comes from a loosely grouped set, at closer to 50.8% stability, treat any single co-mention as noisy and don't overreact to it appearing in one commercial-intent query.

The study covers 1,390 rows with none excluded, organized into 43 groups of vendors averaging 32.33 units each, for a total of 1,390 pair-by-group observations across 2 classes and 3 tightness segments. No single group dominated the sample, the largest one accounting for only 8.3% of all units, so the overall findings aren't being driven by one oversized vendor cluster.

The measurement window ran from 2026-05-01 to 2026-09-15. That's roughly four and a half months, long enough to catch some drift in how AI systems answer queries but not long enough to rule out a seasonal effect tied to that specific stretch of the year. If buying intent itself shifts seasonally, for instance more commercial-intent queries near a budget cycle, that could interact with the clustering pattern in ways this single window can't separate out. Repeating the measurement in a different season would be the direct way to check.