ChatGPT search research · Categories

Categories compared

The same counts for every category, side by side, on one ChatGPT surface at a time.

Wave 2026-115 categories100 answers
API Autos →Beauty & fitness →Business →Finance →Home & garden →
How much it searches
Answers that searched the web
100% 20 of 20 100% 20 of 20 100% 20 of 20 100% 20 of 20 100% 20 of 20
Search steps per answer
One step is one search call; it can send several queries.
2.3 45 in all 2.5 49 in all 2.1 42 in all 3.1 61 in all 3.3 66 in all
Queries per answer
8.7 173 in all 9.3 186 in all 8.2 163 in all 11.4 228 in all 13.0 260 in all
First search names a brand
Among answers whose first search is known. The first step as the stream shows it.
85% 17 of 20 65% 13 of 20 90% 18 of 20 85% 17 of 20 80% 16 of 20
Search journeys
Names brands, searches once
The first search already names the brands and there is no second one: it checks its own picks in one go, then writes.
10% 2 of 20 5% 1 of 20 25% 5 of 20 5% 1 of 20 0% 0 of 20
Names brands, then searches again
The first search names brands; after the results come back it searches again.
75% 15 of 20 60% 12 of 20 65% 13 of 20 80% 16 of 20 80% 16 of 20
Starts open, names brands later
The first search names no brand. Brands appear only in a later search, after it has seen results.
0% 0 of 20 15% 3 of 20 0% 0 of 20 15% 3 of 20 20% 4 of 20
Never searches for a brand
No search names a brand: it takes the options from the lists and pages it found.
15% 3 of 20 20% 4 of 20 10% 2 of 20 0% 0 of 20 0% 0 of 20
Query kinds
Mix of all queries
Hover a segment for its share. The legend is under the table.
173 queries 186 queries 163 queries 228 queries 260 queries
Queries in follow-up searches
Share of the queries that came after the first search, once it had seen results.
59% 102 of 173 62% 116 of 186 55% 90 of 163 68% 155 of 228 73% 189 of 260
What the web can move
Answer brands the first search did not name
Brands the web had a chance to put there. Answers whose first search is known.
57% 51 of 89 58% 32 of 55 57% 51 of 89 49% 43 of 87 43% 30 of 70
Answers with an open list search
35% 7 of 20 20% 4 of 20 30% 6 of 20 35% 7 of 20 50% 10 of 20
Cited pages that are "best …" articles
10% 8 of 77 16% 12 of 75 1% 1 of 91 1% 1 of 83 19% 14 of 72
Answer brands that no search named
Answers that searched.
50% 43 of 86 33% 18 of 55 36% 32 of 89 29% 24 of 83 21% 13 of 63
Cited pages from brands' own sites
38% 29 of 77 47% 35 of 75 81% 74 of 91 72% 60 of 83 54% 39 of 72
Searches that added no cited page
No page it cited first came back in that search step.
16% 7 of 45 18% 9 of 49 12% 5 of 42 25% 15 of 61 33% 22 of 66
What reaches the answer
Cited pages per answer
3.9 77 in all 3.8 75 in all 4.6 91 in all 4.2 83 in all 3.6 72 in all
Search results from big platforms
Pages from Reddit, Wikipedia, YouTube, Amazon, Quora, Facebook, Instagram, TikTok, LinkedIn, X, Pinterest, Medium, Yelp, Tripadvisor, among all the pages its searches returned.
24% 249 of 1017 16% 150 of 962 15% 123 of 809 14% 162 of 1167 18% 244 of 1353
Cited pages from big platforms
The same sites among the pages it cited.
0% 0 of 77 0% 0 of 75 2% 2 of 91 0% 0 of 83 0% 0 of 72
Answers with a link in the text
A link ChatGPT wrote into the answer, cited or not.
0% 0 of 20 0% 0 of 20 0% 0 of 20 0% 0 of 20 0% 0 of 20
Linked pages per answer
0.0 0 in all 0.0 0 in all 0.0 0 in all 0.0 0 in all 0.0 0 in all
Linked pages on sites it did not cite
No page of that site is among the answer's citations.
– 0 of 0 – 0 of 0 – 0 of 0 – 0 of 0 – 0 of 0
Brands named per answer
4.5 89 in all 2.8 55 in all 4.5 89 in all 4.4 87 in all 3.5 70 in all
Query kinds: Topic searchOpen list searchList search with its own picksBrand checkFact lookupComparisonReviews and forumsOther: news, official rules, not labeled

Each column is the category page's “At a glance” for this surface; the category's prompt cards add up to it. Every prompt is a commercial search typed word for word as US searchers type it, so the numbers hold for prompts like these, not for ChatGPT in general. One run per prompt and surface, so small differences between categories can be chance. In the API the citations are themselves the links in the text, so its linked pages count only bare URLs outside them.