After correcting for Anthropic's duplicate rows, how many distinct businesses does Claude actually name per buyer question, and does it differ from OpenAI or Gemini

Why this question matters, and to whom

When someone asks an AI engine to recommend a plumber, a project management tool, or a mattress brand, the answer can name one business or a dozen. That difference matters commercially. A business that appears in a three-name answer is one of a handful of options a user will actually read. A business that appears in a fifteen-name list is competing for a glance. For a company like BusySeed, deciding where to invest in visibility work, the practical question is not "does AI mention us" but "how crowded is the room we're mentioned in."

There is a widespread impression, repeated often enough in industry commentary to have hardened into received wisdom, that Claude tends to produce longer, more exhaustive lists than ChatGPT, and that ChatGPT is comparatively terse. If that impression is wrong, it changes how a business should read its own visibility reports. Being named by Claude would mean something different, competitively, than the conventional wisdom assumes, and a business optimizing its content or listings toward the wrong engine's tendencies could be solving the wrong problem.

This study set out to settle the question with a direct count rather than an impression, across 926,792 individual answers spanning 193 distinct query groups over the window 2026-04-27 to 2026-09-18. Every answer was placed into exactly one of 4 classes based on how many distinct businesses it named: none, a handful (one to five), a fuller list (six to nine), or an extensive one (ten or more). No rows were dropped from the analysis (0 excluded), so the result describes the full set of answers collected, not a filtered subset chosen after the fact.

The finding matters to three groups differently. For businesses deciding where to invest in AI-visibility efforts, it changes the relative value of being named by each engine. For anyone building tools that monitor or report on AI mentions, it is a caution against summarizing engine behavior with a single adjective like "generous" or "sparse" without checking the actual distribution. And for anyone reading a claim like "Claude names more businesses than ChatGPT" in a report or a marketing deck, it is a demonstration of how such a claim can be produced by counting wrong, and how counting right reverses it.

How the measurement works: what counts as one unit

The unit of analysis here is a single answer: one response from one engine to one query, evaluated once. Across the study, 926,792 such units were collected, distributed across 193 query groups, with an average of 4,802.03 units per group. A query group is a cluster of related prompts, for instance different phrasings of a request for a recommendation in the same category, run across the 3 engines being compared (ChatGPT, Claude, and Gemini) so that each group produces a roughly comparable set of answers per engine.

For each unit, the analysis counted the number of distinct, real business names that appeared in the answer text, then sorted that count into one of 4 classes:

This classing choice, rather than reporting a raw average count of names per answer, is deliberate. An average can be dragged around by a small number of extreme outliers, for instance a handful of answers that name thirty or forty businesses in a dump-style list. Classing puts every answer into a bucket regardless of exactly how extreme its tail is, and lets you see where the bulk of the mass actually sits. It also matches how a human reader experiences the answer: nobody reading a list of eleven names is thinking "eleven," they're thinking "a long list," and the class boundaries are set to track that qualitative jump.

A unit was counted as naming a business only when the answer used something identifiable as a proper business name, not a generic category description like "a local plumber" or "several options in your area." Answers that described options without naming any specific business landed in names none. This distinction is where a large share of the disagreement between casual impressions and careful counts tends to originate: an answer that feels information-dense because it describes many attributes can still name zero actual businesses, and an answer that feels short can still pack ten names into two sentences.

The headline result, and what it does and does not mean

Split by engine, the class distributions look like this:

ChatGPT puts 49.7% of its answers in the names 10 or more class, versus 43.4% for Claude. That is a meaningful gap in the extensive-list direction, and it runs opposite to the conventional wisdom this study set out to test. Claude, meanwhile, puts more of its answers in the names 1 to 5 class (5.8%) than ChatGPT does (1.6%), and slightly more in names none as well (0.9% versus 0.4%).

What this means: on this set of queries, over this window, ChatGPT answers were more likely to be exhaustive lists and Claude answers were more likely to be short, curated ones. That is the opposite of "Claude names more businesses."

What this does not mean: it does not mean Claude is more selective in a quality sense, or that its shorter lists are better curated by some editorial judgment. A shorter list could equally reflect a more conservative response style, a narrower training preference for hedging, or a tendency to describe categories rather than enumerate options. This study counted names, not judged reasoning. It also does not mean the gap holds in every query category. That is the subject of the next section, because a pooled percentage can hide a result that is really driven by one or two categories rather than being a general property of the engine.

The shape of the result: concentrated or general

A pooled percentage answers "on average, across everything," but it can be produced two very different ways. It could reflect a mild, consistent tendency present in nearly every query category, or it could reflect a small number of categories with an extreme skew that drags the average, sitting alongside many categories where the engines behave almost identically. These have different implications: the first means the finding is a general property of the engine worth acting on broadly, the second means it is a property of a few specific use cases and should be acted on narrowly.

The class-spread data, measured across 188 query groups, points toward the general-tendency explanation, at least for the two classes that matter most to this finding. For names 10 or more, the share of answers falling in that class ranges from a minimum of 5.7% up to a maximum of 73.6% across groups, with a median of 38.3%. That is a wide range, meaning some query categories push toward exhaustive lists far more than others, but the median sitting well above the minimum suggests the tendency toward long lists is broadly distributed rather than confined to one or two categories.

The names 1 to 5 class shows a narrower and lower spread: a minimum of 1.5%, a maximum of 28.8%, and a median of 6%. Short-list answers are consistently a minority outcome across groups, not a phenomenon concentrated in a few.

The names none class is rare everywhere, from 0.2% to 6.9%, median 2%, meaning answers that name zero businesses are consistently uncommon rather than clustered in a subset of query types where the engines simply refuse to name anyone.

The practical reading: the engine-level gap in list length is not an artifact of one unusual query category dominating the pooled figure. It shows up, to varying degrees, across a broad swath of the 193 groups studied. That said, "varying degrees" is doing real work in that sentence, since a range from 5.7% to 73.6% means the size of the gap a business will actually encounter depends heavily on its own category, not just on which engine it's watching.

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

A second, related concern is different from the spread question above: even if a tendency shows up across many groups, the pooled headline figure could still be dominated by one or two unusually large groups if those groups contribute a disproportionate share of the total units. A finding driven by one oversized query group masquerading as a population-wide result is a common way studies mislead without any single number being wrong.

Here the data works against that concern. The study spread 926,792 units across 193 groups, averaging 4,802.03 units per group, and the single largest group accounted for only 1.5% of all units. No group comes close to being large enough to single-handedly produce the pooled result. The headline percentages in the class table are a genuine aggregate across a wide base of roughly-equal-sized groups, not a figure inflated by one dominant category of query.

This matters for how confidently the finding generalizes. If ninety percent of units had come from a single query group, say a cluster of prompts all about restaurant recommendations, then "ChatGPT names more businesses than Claude" would really mean "ChatGPT names more restaurants than Claude," and a business in a different category would have little basis for assuming the same gap applies to them. With the actual distribution here, where the largest single group is 1.5% of the total, that risk is much smaller. The finding is built from breadth rather than from one loud group drowning out a quiet population.

It is worth being precise about what "breadth" does and doesn't buy you. It rules out the specific failure mode of one group dominating the average. It does not by itself prove every group behaves identically, and the previous section already showed real variation in the names 10 or more share across groups, from 5.7% to 73.6%. What the even group sizes establish is that the pooled figure is an honest average across many contributors, each with modest weight, rather than a number that looks like an average but is actually a proxy for one category's behavior. Combined with the spread evidence, the overall picture is a real, moderate, broadly-if-unevenly distributed tendency, not a statistical illusion created by uneven sampling.

How much to trust the classification

Every measurement study carries a question behind the numbers: how reliably were units actually assigned to their classes? A classification scheme that sounds precise, names none versus names 1 to 5 versus names 6 to 9 versus names 10 or more, still depends on a judgment call for edge cases. Does a business mentioned by name but only in passing, not as a recommendation, count as "named"? Does a franchise mentioned once but implying multiple locations count once or many times? These edge cases are exactly where a classification scheme earns or loses trust.

This study applied 3 distinct mitigations during construction of the dataset, aimed at reducing exactly this kind of measurement noise before classification, rather than correcting for it afterward with a statistical adjustment. That approach, catching problems at the counting stage, is preferable to a post-hoc calibration figure because it means the reported percentages are closer to being direct counts rather than estimates with a modeled correction layered on top. The tradeoff is that this study does not publish a single calibration statistic, such as an inter-rater agreement rate, that would let a reader independently size the residual uncertainty in the counting itself. That is a real limit worth naming rather than glossing over: a reader should treat the class boundaries, especially the boundary between names 6 to 9 and names 10 or more, since it is the one carrying the headline finding, as reflecting careful construction rather than as independently re-verified.

What would increase confidence further is a published sample of borderline units, for instance answers that named exactly five or exactly six businesses, with the reasoning for why each landed where it did. Absent that, the honest position is that the 0 excluded rows and the 3 mitigations describe a careful process, but the process is not the same thing as an audited error rate. The headline result, that ChatGPT's names 10 or more share of 49.7% clears Claude's 43.4% by a wide enough margin to survive ordinary counting noise around any single boundary, is robust to small classification disagreements. A gap that size would not flip from a modest number of borderline units being reclassified. Smaller, secondary comparisons in this dataset should be read with correspondingly more caution than the headline gap.

What someone acting on this finding should actually do differently

The practical implication depends on why a business or analyst cares about AI-named mentions in the first place.

If you are trying to be named at all: the names none share is small in every engine (0.4% for ChatGPT, 0.9% for Claude, 5.7% for Gemini), so the binary question of "does the engine ever name businesses in this category" is rarely the bottleneck. The more useful question is which class of list you're likely to land in, since that determines how much competition you share the answer with.

If you care about being one of a few names rather than one of many: Claude's answers are relatively more likely to land in the names 1 to 5 class than ChatGPT's (5.8% versus 1.6%). A mention inside a Claude answer is, on these figures, somewhat more likely to be one of a handful than a mention inside a ChatGPT answer. That changes the relative value of a mention by engine, and it argues against treating a Claude mention and a ChatGPT mention as equivalent units of competitive exposure in any tracking dashboard.

If you are building or reading an AI-visibility report: stop describing engines with a single adjective. "Claude is thorough, ChatGPT is terse" is precisely the impression this study set out to check and found reversed on the specific dimension of list length. Report the class distribution, not a mean count of names per answer, since the class table survives the outlier problem that a mean does not.

If you are deciding where to invest content or listing effort: recognize that the gap varies by category. The spread across groups for names 10 or more ranged from 5.7% to 73.6%, so the size of the ChatGPT-versus-Claude gap in your specific category is not guaranteed to match the pooled figure. Before reallocating effort based on this study, check whether your own query category sits near the low or high end of that range, which requires either replicating this method on your own category or requesting a category-level cut of the published dataset.

How to read the published dataset yourself

The dataset behind this study is organized around the same units used throughout: 926,792 answers, 193 query groups, 3 engines, 4 classes, with 0 rows excluded from the analysis. Anyone re-examining the data should start from three checks before trusting any percentage pulled from it.

First, check the denominator behind any cut. The engine-level table has real, comparable denominators, 329,786 for ChatGPT, 298,968 for Claude, 298,038 for Gemini, all large enough that a percentage built on any of them is meaningfully stable. A group-level cut is a different matter: with 193 groups and 4,802.03 units per group on average, and a most-imbalanced group carrying only 1.5% of the total, individual group percentages will have far more sampling noise than the pooled or engine-level figures. Treat any single-group percentage as a rougher estimate than the headline numbers in this piece.

Second, distinguish the class-share table from a raw average. If you see a claim built from an average number of names per answer rather than a class distribution, ask how it was computed. The class boundaries used here, none, one to five, six to nine, ten or more, were chosen to resist outlier distortion, and any downstream reuse of the data that collapses back to a single average count reintroduces the exact distortion this study designed around.

Third, when comparing engines on any specific query category, pull the group-level breakdown for that category specifically rather than assuming the pooled engine gap applies uniformly. The spread data in this study, ranging as widely as 5.7% to 73.6% for the names 10 or more class alone, is the clearest evidence in the whole dataset that category matters, and it is the first thing to check before generalizing the headline finding to a category not directly represented in the 193 groups studied here.