Why do tightly grouped vendor pairs hold together almost universally?

The overall finding from this study, that 61.9% of vendor pairs hold together whether the buyer's question is informational or commercial, hides a large amount of variation once you break the sample apart by how tightly each group of vendors was bound to begin with. This study sorted groups of vendors into three segments based on how firmly the group clustered as a set: loosely grouped, firmly grouped, and tightly grouped. That segmentation was done independently of the intent-clustering measurement being reported here, which is what makes it useful as a predictor.

The gradient is stark. Among loosely grouped pairs, 50.8% held together across both intents and 49.2% held together only under commercial framing, drawn from 898 observations. That is close to a coin flip. Among firmly grouped pairs, the both-intents rate rises to 78.7%, out of 404 observations. And among tightly grouped pairs, the both-intents rate reaches 97.7%, out of 88 observations, with only 2 pairs failing to hold across both framings.

That is close to a ceiling. Once a group of vendors is tightly bound, meaning an AI assistant reliably treats them as a set regardless of what else is being asked, the specific two-vendor pairings inside that set are almost unshakeable by a change in query framing. Loosen the grouping and the pairings become roughly as likely to be commercial-only as to be durable.

The mechanism this points to is that grouping tightness is not a separate phenomenon from pairing durability, it is largely the same phenomenon measured at two different resolutions. A tightly grouped set of vendors is, definitionally, a set the model associates strongly and consistently, which is another way of saying the model has a stable internal representation of that category that does not depend much on how the question is asked. A loosely grouped set is one the model does not consistently reconstruct, so which two vendors happen to land next to each other in an answer is more sensitive to incidental factors, including the specific angle of the query. Commercial framing, which tends to pull in comparison and review content, is one such incidental factor.

An honest caveat: the tightly grouped segment is the smallest of the three, with 88 observations against 898 for loosely grouped and 404 for firmly grouped. The near-ceiling rate in that segment is a real pattern, not noise from a handful of pairs, but it does mean fewer distinct groups are contributing to that number, and a single unusually stable group of vendors could pull the rate up. The study did not report how many distinct vendor groups fall into each segment, only the pair-level counts, so this page cannot say how many separate groups the tightly grouped rate rests on.

For a company deciding what to do with this, the actionable distinction is not "are we mentioned with a competitor" but "is our category, as a whole, one the assistant treats as a firm set." If it is, a pairing earned under one kind of query is likely to persist under others, and effort spent shaping that pairing under one framing should carry over. If the category is loosely grouped, a pairing observed today under a commercial query is roughly a coin flip to still be there under an informational one, and a company should check both framings separately rather than assuming one result generalizes.