Interface vs API
ChatGPT Plus in the app against the OpenAI API with web search: what the API shows and what people see.
- API runs about 2.1 times as many search steps per answer (1.2 for Plus, 2.6 for API).
- The first search already names a brand in 78 of 100 Plus answers and 81 of 100 API answers.
- Of the brands in the answer, 47% on Plus and 53% on API were not in its first search, so search results had a chance to put them there. The rest were in its first search, named before any results.
- "Best …" articles make up 33% of Plus's citations and 9% of API's; brands' own sites 36% and 60%.
- The API opened a page in 39 of 100 answers; most of what it cites it saw only as a search result.
- For the same prompt they agree on 39% of the brands they name between them.
Every prompt ends by asking ChatGPT to list the searches it ran. The app hides the text of some searches, mostly the second-layer fan-outs it runs while reasoning, so for the app's query numbers this self-report is the default: it is used when it contains every query seen in the stream (Plus 100 of 100 answers), otherwise only what was seen counts. Where the self-report changes a number, the number without it is shown under it. Search steps, page reads, citations and brands always come from the stream.
How much it searches
Counted from search events, which both surfaces show even when the app hides the query text.
Seen: events in the app's stream or calls in the API response.
The first search
Does ChatGPT start from brands it already has in mind, or from an open search? The first search step decides it.
A step can hold several queries. It counts as naming a brand when at least one of them does. The step is taken from the stream on every surface. Round 1 of the app's self-report fills in only a first step whose text was hidden: that round also lists product lookups by name that ran in no visible step, so it would overstate how often the first search names a brand.
In this wave every first search showed its query text. When the app hides it, the answer is counted as unknown here rather than judged by its second search.
From memory to the answer
The first search runs before ChatGPT has seen any results, so the brands it names there come from what the model already knows. How many of them survive into the answer, and how much of the answer they make up?
Brands are matched loosely: “RAV4” in a query matches “Toyota RAV4 Hybrid” in the answer. Answers that did not search, or whose first search hid its text, are left out. On the app, a hidden first step is filled from round 1 of the self-report.
Search journeys
A journey is the searches of one answer in order. Two questions sort them: does the first search already name brands, and does ChatGPT search again once it has seen results?
Most common: names brands, searches once on Plus (57% of answers), names brands, then searches again on API (72%).
A second search is counted from the search steps in the stream, or from a later round of the self-report on the app. Brands are matched loosely, as above.
- 3 queries · List, Facts, Compare
- best websites sell your car online 2026 CarMax Carvana Kelley Blue Book Edmunds private sale
- Carvana CarMax sell my car online offer pickup 2026 official
- Autotrader Cars.com Facebook Marketplace private party sell car listing fees 2026
- 4 queries · Brand check, Facts
- site:carmax.com sell my car online offer valid 7 days appointment
- site:carvana.com sell my car online offer pickup payment
- site:cars.com sell my car private listing fees 2026
- and 1 more
- 4 queries · Brand check, Facts, Reviews
- site:cars.com sell my car private listing free sell to dealer offer
- site:cargurus.com sell my car instant offer private listing fees
- site:facebook.com/help marketplace sell vehicle listing fee
- and 1 more
- 3 queries · Facts
- site:facebook.com/help sell vehicle Facebook Marketplace local pickup listing fees vehicles
- site:bringatrailer.com how it works sellers listing fees 2026
- site:kbb.com instant cash offer sell car participating dealer 7 days
Query kinds
The mix of its queries: each slice is the share of its queries of that kind. Each prompt's answer page shows every query with its kind.
Most common: brand check on Plus, fact lookup on API.
Kinds come from the classifier's query labels. On the app the mix includes the self-report, which also covers the steps whose text the app hid; the table gives the mix without it. Lookups the self-report puts in its first round but the stream shows in no step count under all searches only.
| Kind | Plus | API |
|---|---|---|
| Topic search | 1% 5 1% without the self-report | 0% 1 |
| Open list search | 10% 53 14% without the self-report | 5% 48 |
| List search with its own picks | 12% 63 15% without the self-report | 3% 26 |
| Brand check | 36% 194 11% without the self-report | 28% 281 |
| Fact lookup | 26% 138 34% without the self-report | 40% 403 |
| Comparison | 2% 12 4% without the self-report | 1% 13 |
| Reviews and forums | 3% 18 6% without the self-report | 5% 46 |
| Other: news, official rules, not labeled | 10% 54 16% without the self-report | 19% 192 |
| Kind | Plus | API |
|---|---|---|
| Topic search | 1% 2 1% without the self-report | 0% 1 |
| Open list search | 14% 42 14% without the self-report | 12% 42 |
| List search with its own picks | 15% 47 15% without the self-report | 6% 23 |
| Brand check | 11% 33 11% without the self-report | 15% 52 |
| Fact lookup | 34% 104 34% without the self-report | 38% 136 |
| Comparison | 4% 12 4% without the self-report | 3% 9 |
| Reviews and forums | 6% 17 6% without the self-report | 6% 20 |
| Other: news, official rules, not labeled | 16% 48 16% without the self-report | 21% 75 |
| Kind | Plus | API |
|---|---|---|
| Open list search | 0% 0 – without the self-report | 1% 6 |
| List search with its own picks | 4% 5 – without the self-report | 0% 3 |
| Brand check | 60% 68 – without the self-report | 35% 229 |
| Fact lookup | 29% 33 – without the self-report | 41% 267 |
| Comparison | 0% 0 – without the self-report | 1% 4 |
| Reviews and forums | 1% 1 – without the self-report | 4% 26 |
| Other: news, official rules, not labeled | 5% 6 – without the self-report | 18% 117 |
What it searches for
Share of answers with at least one query of each type.
“Words” rows are read from the text of the query, so anyone can check them. “Classifier” rows come from the query labels. On the app the numbers include the self-report; the number without it follows where it differs.
Pages it reads
Opening a page means reading it beyond the title and snippet in the search results.
The last two rows read the overlap both ways: how much of what it cites it had read, and how much of what it read it cites.
What reaches the answer
What the reader gets: citations, links in the text, the sites behind them and the brands named.
For the same prompt, the two share 26% of their cited sites on average.
What the web can move
How much of the answer is open to what search finds, rather than fixed by what the model already knows. Content cannot change the model's memory, but it can change what these searches return.
The rows below are the ways in, each a share that can be checked on the answer pages. Page types and site types come from the classifier's labels. Cited pages that are among the search results the app shows: Plus 622 of 624.
Every prompt in detail
Each prompt with its searches, citations, links and brands on all three surfaces is on its category page.
Read with care
What these numbers can and cannot say.
- The prompts decide most of what these numbers can say. They are 100 commercial searches such as “best credit cards for travel” or “where to open a roth ira”, typed word for word as US searchers type them, in five categories. There are no comparison, advice or opinion prompts. Read every share as true for prompts like these, not for ChatGPT in general.
- One run per prompt and surface. ChatGPT does not answer the same way twice, so a single prompt can differ by chance; averages over many prompts are what to read.
- The surfaces run different models (Plus: gpt-6-thinking; API: gpt-6-sol, medium effort). A difference between them is the surface and the model together.
- The app hides the query text of some search steps (Plus 19%). Query numbers for the app include ChatGPT's own list of its searches to fill that gap; a list the model wrote can miss or add a search.
- Search results are not compared: the API returns every result it got, the app only what it shows.