How much does the 'recommended together' vendor cluster change between informational and commercial-intent prompts within the same category?

Why this question matters

When someone asks an AI system a buying question, the answer often names more than one vendor in the same breath. "For project management, tools like Asana, Monday, and ClickUp..." That co-mention is a form of association the vendor did not write and cannot directly edit. It is produced by the model, shaped by whatever the model has absorbed about which companies belong together in a category.

The question this study asks is simple to state and hard to answer without data: does that association hold up when the buyer's framing changes? A person researching a category in general ("best tools for X") is in a different stage of the funnel than a person signaling they are ready to buy ("pricing for X", "X vs Y"). If the same pairs of vendors get named together regardless of framing, that pairing reflects something durable, probably a real market structure the model has learned. If the pairing only shows up under one framing and dissolves under the other, it is more fragile, and possibly an artifact of what kind of content the model was trained to retrieve for that particular query type.

This matters to a few distinct audiences. A vendor doing competitive positioning wants to know whether being named alongside a competitor is a stable fact about the category or a quirk of how they phrased their test query. A category analyst wants to know whether AI-driven co-mention data is even a reliable signal of market grouping, or whether it is too framing-dependent to trust. And anyone building a monitoring pipeline on top of these AI answers needs to know how much weight to put on any single snapshot.

The finding in this study is not "co-mention is stable" or "co-mention is unstable." It is that stability depends on a property of the group itself, how tightly its members already cluster, measured independently of intent. That is a more useful and more falsifiable claim than a single aggregate number, and it is the claim the rest of this page works through.

How the measurement works

The unit of analysis is a pair-by-group observation: one pair of vendors, checked within one group, for whether it clustered together under two different intent framings. Across the whole study there were 1,390 such units, built from 1,390 source rows with 0 excluded, spanning 43 groups of vendors. Each group averaged 32.33 units, and the largest single group accounted for 8.3% of all units, so no one group's behavior can be driving the aggregate numbers by sheer weight.

Each pair-by-group unit was run against two query framings, one aimed at general commercial intent and one broader, and the AI's answer was checked for whether the two vendors appeared together. That produced a binary class per unit, one of only 2 possible outcomes:

A pair that clusters under neither framing is not part of this dataset by construction, because the unit only exists where a cluster was observed at least once, on the commercial framing. That is worth stating plainly: this is not a study of whether vendors get named together at all, it is a study of whether an observed pairing survives a change in framing. It cannot speak to pairs that never cluster in the first place.

Groups themselves were not arbitrary; they represent 3 pre-existing segments of vendor sets, ranging from loosely grouped to tightly grouped, based on how bound the group's members already were to each other independent of this intent test. That segment label is the variable that turns out to matter most, and it is examined in detail two sections down. Three mitigations were applied during construction of the dataset (3 in total) to reduce known sources of noise, though the fact table does not itemize what they were beyond the count.

The headline result, and its limits

Across all 1,390 pair-by-group units, 61.9% clustered together under both intent framings, and 38.1% clustered only under the commercial framing. Read on its own, that looks like a moderate but not overwhelming tilt toward stability, roughly three in five pairings survive a change in framing, two in five do not.

That headline number is real, but it is an average over a population that, as the next section shows, is not homogeneous. Treating 61.9% as "the" answer to "do AI co-mentions survive intent changes" would flatten a much sharper underlying pattern, and would give a false sense that stability is a roughly coin-flip-adjacent property spread evenly across all vendor groupings.

It is also worth being explicit about what "clusters on both intents" does and does not mean. It means the pair was named together in the AI's answer under both framings tested. It does not mean the pair was named together with the same intensity, in the same order, with the same surrounding language, or for the same reason. Two vendors could satisfy this class by both appearing in a long list under one framing and a short, prominent pairing under the other; the measure is binary co-occurrence, not co-occurrence quality. A study designed to measure prominence or ranking, rather than presence or absence, would need a different unit of analysis and is not what this one does.

Finally, the two framings tested here were commercial intent and a second, broader intent; the study does not claim to have sampled every possible way a buyer might phrase a query, only these two. A pairing that is stable across these two framings might still dissolve under a third framing not tested. The finding is about the two framings measured, not about framing-invariance in general.

The shape of the result: it is not evenly spread

The aggregate split of 61.9% versus 38.1% breaks apart cleanly once you cut by segment, which groups vendor sets by how tightly bound they already are.

Among loosely grouped vendor sets, the outcome is close to a coin flip: 50.8% held across both intents against 49.2% that did not. Among firmly grouped sets, the balance shifts substantially toward stability, 78.7%. And among tightly grouped sets, stability is close to universal: 97.7% held across both intents, with only 2.3% failing to.

This is the actual finding of the study, and it is a claim about a gradient, not a threshold. Stability rises monotonically with how tightly a group is already bound. That is a mechanistic, not merely descriptive, pattern: it suggests that whatever produces co-mention stability across intent framings is closely related to whatever produces tight grouping in the first place. A plausible mechanism is that tightly grouped vendors are tightly grouped because the model has encountered strong, repeated, categorical association between them across many kinds of source content, of which intent-specific commercial content is only one slice; that same broad exposure would make the association survive a change in query framing. Loosely grouped vendors, by contrast, may be co-mentioned mostly because of shared presence in one type of content, such as commercial comparison pages, which would make their association framing-dependent by construction. This study cannot distinguish these mechanisms from the data alone, but the segment gradient is consistent with it, and a follow-up study varying the content sources feeding each intent framing could test it directly.

Is this a property of the whole population, or of one group?

A gradient across three segments could in principle be driven by a single unusual group sitting in the tightly grouped segment, especially if group sizes are uneven. The fact table lets us check this at the level of individual groups, not just segments.

Across the 28 groups for which the "clusters on both intents" class was measured, the share of pairs in that class ranged from a minimum of 0% to a maximum of 100%, with a median of 78.9%. That is a wide range, some individual groups show no both-intent stability at all, others show complete stability, and the median group sits well above the overall aggregate figure of 61.9%.

The mirror image holds for the commercial-only class: across the same 28 groups, its share ranges from 0% to 100%, median 21.1%.

This spread confirms that the segment-level gradient shown in the previous section is not an artifact of one outlier group dragging a small segment average. It is a genuine population-wide pattern: groups differ substantially from one another in how stable their internal pairings are, and that difference lines up with the loosely-to-tightly grouped ordering rather than appearing as noise scattered randomly across groups. A reader should still note that a median describes the typical group, not every group; a group sitting near the minimum end of the range behaves nothing like the aggregate, and if you care about one specific vendor group, the aggregate percentage is the wrong number to quote. The right move is to look up that group's own figure, or at minimum its segment, before treating the headline percentage as a prediction for it.

What to do differently with this

The practical implication follows directly from the segment gradient, not from the aggregate number.

If you are tracking your own brand's co-mentions. Before treating a single snapshot of "who gets named next to us" as a stable competitive fact, check whether your category behaves like a loosely grouped set or a tightly grouped one. If pairings in your category are unstable across framings, a single query's output is close to a coin flip and should not be reported to stakeholders as a fixed competitive position. It needs to be re-tested across multiple framings, and ideally across multiple points in time, before anyone draws a conclusion from it. If your category behaves like a tightly grouped set, a single well-chosen framing is a much more reliable read, though it is still worth confirming with a second framing given that even the tightest segment here was not at total unanimity (97.7%, not full agreement).

If you are deciding how many query framings to sample. The honest answer, based on this data, is that the number of framings you need depends on how tightly bound the category already is, which is often not known in advance. A defensible default is to always test at least two framings, one general and one commercial, since that is what distinguishes the classes in this study, and to treat disagreement between them as informative rather than as noise to average away.

If you are interpreting someone else's single-framing claim about AI co-mentions. Ask what kind of category it is. A claim like "AI names us next to Competitor X" made from one commercial-intent query is much more likely to be a stable fact if the category is a small, well-established, tightly bound set of vendors, and much more likely to be framing-dependent noise if the category is broad, new, or has many loosely affiliated players.

If you are building monitoring tooling on top of this kind of data. Segment membership, loosely, firmly, or tightly grouped, is worth surfacing as a first-class field alongside any co-mention statistic, because it changes how much confidence a downstream reader should place in any individual pairing observation.

Reading the published dataset yourself

The dataset behind this study is organized around the pair-by-group unit described earlier, so the first thing to orient on is scale: 1,390 units, 43 groups, 1,390 source rows with 0 excluded. If you pull the raw file, confirm those totals reconcile before trusting any downstream cut; a mismatch usually means a filter has been applied silently somewhere upstream.

Three fields matter most for reproducing any number on this page:

1. Class, one of 2 values, either the pair clusters on both intents or the pair clusters on commercial intent only. This is the outcome variable.
2. Segment, one of 3 values, loosely grouped, firmly grouped, or tightly grouped. This is the variable that explains almost all of the interesting variation in the outcome, as shown above.
3. Group ID, one of 43 distinct groups. Use this if you want to look up a specific vendor set rather than a segment-level average; recall that individual groups ranged all the way from 0% to 100% on the both-intents class, so the segment label is a useful prior but not a substitute for checking the specific group.

A reasonable first check on any subset you build is to compare its size against 32.33, the average units per group; a subset built from very few groups will be noisy even if the row count looks large, since 1,390 rows compressed into 43 groups means the effective sample size for a between-group comparison is closer to 43 than to 1,390.

The study records that 3 mitigations were applied during construction to reduce known sources of measurement noise. The fact table does not specify what these were, and that is a real limit on independent verification: a reader who wants to know whether a specific confound, such as list-position effects in the AI's answer or vendor name ambiguity, was addressed cannot confirm it from the published numbers alone and would need to request the methodology notes directly.