Markets · Feature
ChatGPT checks your reviews and your prices before it recommends you
We asked ChatGPT to pick a cold email agency eight times and saved every search it ran. It started from names it already knew, then looked each one up on Clutch, G2 and Trustpilot.
- 1. Names from memory Belkins, SalesBread, Cleverly, ColdIQ, written into the first search
- 2. Checks Clutch, G2 and Trustpilot ratings and pricing pages, over 2 to 4 rounds
- 3. Answer Winner picked on review counts and ratings
The names come first
On 31 August we put buyer questions about cold email agencies to ChatGPT, eight runs in total, from a logged-in Plus account. For each run we saved the raw response stream, which includes the searches the model sends before it writes anything.
Most first searches already had agency names in them. One read best cold email agency B2B outbound 2026 Belkins SalesBread Cleverly ColdIQ reviews. The model had not read any search result at that point, so the names came from its training data. In the one run where we told it not to search, it listed Belkins, SalesBread, Growth Rhino, Leadium, Martal Group and Cleverly. Growth Rhino did not come up in any of the seven runs that searched.
Then it checks them
The searches after that were checks. Over two to four rounds the model looked up each name on Clutch, G2 and Trustpilot and read the agencies' pricing pages. Belkins, rated 4.9 from 233 Clutch reviews, won every run that asked for the best agency. Cleverly was also named from memory every time, and every time it was demoted or advised against. It had 4.3 on Clutch and 4.0 on G2.
Who was left out
We expected a strong review profile to carry an agency into the answer. SalesAR had 134 Clutch reviews averaging 4.9, came up in one run and was left out. In that run the model took all its candidates from best-of lists, and SalesAR was on too few of them. Most of those lists were published by agencies ranking their own market.
| Agency | Clutch | Outcome |
|---|---|---|
| Belkins | 4.9 from 233 reviews | Won every "best" run |
| Cleverly | 4.3 (G2: 4.0) | Demoted or advised against in every run |
| SalesAR | 4.9 from 134 reviews | Came up in one run, left out |
Prices and tracking
Two smaller findings apply to any company that sells to businesses. The model's search strings included literal prices such as $500 and "starting at $999", so a price written as plain text on a public page can match a search directly. And every link it cited carried ?utm_source=chatgpt.com, so these visits show up by name in analytics.
Limits
Eight runs on one question is a small sample. The answers also moved with location: our account's Berlin timezone sent three runs off to look up German cold-email law.
Method
8 runs on 31 August 2026 · ChatGPT Plus, logged in, thinking model · Berlin timezone · raw response streams saved