How was the vendor-pairing dataset for this study built?

Before trusting a percentage, it helps to know what was counted, what was thrown out, and what adjustments were made along the way. This page describes the construction of the dataset behind the finding that 61.9% of AI-named vendor pairs hold together across a change in buyer intent, versus 38.1% that only appear together under commercial framing.

The study drew on 1,390 raw rows of underlying data. Of those, 0 were excluded, which means the full raw dataset made it into the analysis with nothing dropped for quality or coverage reasons. That is worth noting explicitly, because a study that excludes a meaningful share of its raw rows is implicitly telling you something about data quality in the discarded portion, and a reader can't independently verify that judgment. Here, there is no exclusion judgment to scrutinize.

Those rows were organized into 1,390 units, where a unit is a pair-by-group observation, meaning one specific two-vendor pairing evaluated within one specific competitive group, classed by whether it held together across both intent framings or only under commercial framing. The units are not spread one-to-one with rows because each underlying row can contribute to the classification of a unit alongside other rows covering the same pairing. The 1,390 units span 43 distinct groups of vendors, for an average of 32.33 units per group. Concentration is a common failure mode in studies like this, where one or two unusually large groups can dominate a pooled average and make it look more representative than it is. That does not appear to be the case here: the largest single group accounted for only 8.3% of all units, meaning no group is doing outsized work in the headline numbers.

Each unit was classed into exactly one of 2 classes, the pair clusters on both intents or the pair clusters on commercial intent only. Those two classes are exhaustive and mutually exclusive by construction, which is why the two percentages in the overall split, 61.9% and 38.1%, sum to the full sample. There is no third outcome, such as a pair that clusters only on informational intent, in this dataset. That is a design choice worth flagging: the study set out specifically to test whether commercial framing adds pairings beyond what plain informational framing already shows, so its category structure is built around that specific comparison rather than around every logically possible pattern of intent-sensitivity.

The study also separately grouped units into 3 segments by grouping tightness, loosely grouped, firmly grouped, and tightly grouped, which is the basis for a separate finding on this site showing that the both-intents rate climbs sharply with grouping tightness. And the study notes that 3 mitigations were applied during construction. The fact table available for this study does not specify what those three mitigations were, so this page cannot describe them beyond noting that they exist, and a reader who wants to weigh their effect on the headline numbers would need to consult the study's underlying methodology documentation rather than this summary. What can be said is that after those mitigations, the dataset was used whole, with zero rows excluded, across a base large enough, 1,390 units over 43 groups, that the headline percentages are not resting on a small or lopsided sample.