Questions about this study

It found that the size of a market category has almost no effect on how many distinct businesses an AI engine names when answering questions about it. Across 193 groups and 752,124 rows, 91% of answer sets named six or more distinct businesses, and this held nearly steady whether the category was in the smallest size quarter (93.2% naming six or more) or the largest (90%). If category size drove naming breadth, you would expect a much wider gap between small and large categories than the study observed.

Categories were ranked by row count, the volume of underlying data behind each one, and split into four equal-sized segments (quarters) from smallest to largest. The smallest quarter held 88,655 rows and the largest held 261,745 rows. This is a relative ranking within the sample, not an absolute size threshold, so 'small' here means small relative to the other categories studied, not small in any external market sense.

Each answer set from an AI engine was read and the distinct, individually named businesses in it were counted. If that count was six or more, the answer set was classed in the 'names 6 or more' group; five or fewer put it in the other group. This is a simple threshold split, not a measure of quality or accuracy. It does not, on its own, tell you whether the named businesses were correct, relevant, or the ones a category actually leads in, only how many distinct names showed up.

It is a chosen threshold, and yes, a different cutoff would shift the exact percentages. But the underlying finding does not depend on picking six specifically: the same near-flat pattern across category sizes shows up in the segment-level breakdown (93.2% to 90%), which is a continuous comparison, not just a single cutoff. If you moved the line to five or seven, you would expect the segment gaps to still be small, though we have not re-run the analysis at other thresholds ourselves.

That is worth checking, and the data says no. The largest single group accounted for only 1.5% of all 752,124 rows, so no single oversized category could be pulling the averages. The result reflects a pattern spread across 193 groups rather than one dominant case. That said, aggregate averages can still hide variation among the many smaller groups, which is why the group-level range matters separately.

More than the segment averages suggest, and this is a real limit on the headline claim. Across 188 individual groups, the share naming six or more ranged from 65.3% to 98.3%, a wide spread. So while the four size segments average out to nearly the same rate, individual categories can differ a great deal from each other. Segment-level flatness and group-level variation are both true at once, and only the first is about category size specifically.

One plausible mechanism is that the AI engine defaults to a roughly fixed answer length regardless of topic, drawing on however many names it has some confidence in, rather than scaling its list to match how large or crowded the actual market is. Another possibility is a genuine ceiling on how many businesses are reliably recallable for any topic, small or large. The study cannot distinguish between these two explanations. Doing so would require a separate test that varies the requested list length or prompt phrasing and checks whether the naming count moves.

0 rows were excluded, meaning essentially the full 752,124-row dataset was used. 3 corrections were applied before analysis, addressing issues like data errors or classification inconsistencies caught during review. The study does not itemize what those three corrections were, so a reader relying on this should treat the dataset as very close to complete but not describe it as entirely untouched.

If you operate in a small or niche category, do not assume you get a pass from name-crowding just because your market is small. The data shows small categories name about as many distinct competitors (93.2%) as large ones (90%), so the practical task, being one of the names an engine recalls, is roughly as competitive in a small niche as in a big market. The group-level range (65.3% to 98.3%) matters more for your specific category than the overall average does, so check where your own category falls rather than assuming it matches the segment figure.

The measurement window was 2026-04-27 to 2026-08-16. AI engines change their underlying models and behavior over time, so a naming pattern measured in this window is not guaranteed to hold indefinitely. The study describes what was true in this window across 193 groups, not a permanent property of these engines. Anyone wanting to know if it still holds would need to re-run the same segment comparison on more recent data.

2 refers to the two answer-naming classes used throughout the analysis: 'names 6 or more' and 'names 5 or fewer'. Every one of the 752,124 answer sets in the study was sorted into exactly one of these two classes. It is not a count of business categories studied, that figure is 193, but a count of the outcome buckets used to measure naming breadth.

No, and this is an important boundary of the method. The study counts how many different businesses appeared in an answer, not whether those businesses are accurate, still operating, or genuinely relevant to the category asked about. A high count of distinct names tells you the engine produced a long list, not that the list is a good one. Assessing correctness would require a separate check against verified business data for each category.