We counted the follow-ups.
Twenty everyday requests, each asked twice: as typed, and after one click on Perry. Then we counted the corrections it took until the answer had everything the person wanted. Here is all of it, including where Perry did worse.
What came out
| As typed | With Perry | |
|---|---|---|
| What the person wanted, already in the first answer | 37% | 52% |
| Follow-up messages until everything was there, on average | 3.6 | 2.7 |
| Right on the first answer (of 60) | 3 | 6 |
| Still missing something after 5 follow-ups (of 60) | 11 | 7 |
| Words the person typed, on average (with Perry: including the blanks they filled in) | 32 | 32 |
Our test, run on 22 October 2026: 20 tasks, 3 conversations each way. The chat AI was Google’s gemini-3.5-flash, not ChatGPT or Claude, so results there can differ. These numbers describe this test, not every prompt you will ever write.
The person typed about as much either way. With Perry, those words went into his [blanks] before the first answer instead of into corrections after it.
Task by task
Follow-ups in each of the three conversations; ✗ means something was still missing after 5. Perry needed fewer on 10 tasks, about the same on 7, and more on 3.
| The request | As typed | With Perry | |
|---|---|---|---|
| An email to the landlord about broken heatingAbout the same | 3 · 3 · 2 | 3 · 3 · 3 | About the same |
| A quick dinner ideaFewer with Perry | 4 · 3 · 4 | 4 · 1 · 2 | Fewer with Perry |
| A cover letter for a marketing jobFewer with Perry | 4 · 5 · 5 | 1 · 2 · 1 | Fewer with Perry |
| Explain inflationAbout the same | 2 · 2 · 2 | 2 · 2 · 2 | About the same |
| A workout planAbout the same | 4 · 4 · 4 | 4 · 4 · 5 | About the same |
| An Excel formula for paid invoicesAbout the same | ✗ · ✗ · ✗ | 5 · ✗ · ✗ | About the same |
| Three days in LisbonAbout the same | 1 · 1 · 1 | 1 · 1 · 1 | About the same |
| Summarise these meeting notesMore with Perry | 0 · 0 · 0 | ✗ · ✗ · ✗ | More with Perry |
| Say no to a friend’s weddingFewer with Perry | 2 · 5 · 3 | 0 · 0 · 0 | Fewer with Perry |
| A Python script to merge CSV filesFewer with Perry | 4 · 3 · 4 | 3 · 3 · 3 | Fewer with Perry |
| A LinkedIn post about a certificateFewer with Perry | 5 · ✗ · 5 | 3 · 3 · 3 | Fewer with Perry |
| How to ask for a raiseFewer with Perry | 4 · 4 · ✗ | 2 · ✗ · 4 | Fewer with Perry |
| A bedtime storyMore with Perry | 3 · 3 · 3 | 3 · ✗ · 3 | More with Perry |
| A product description for a candleFewer with Perry | 4 · ✗ · 4 | 1 · 1 · 1 | Fewer with Perry |
| Study for a biology examAbout the same | 3 · 3 · 3 | 3 · 4 · 3 | About the same |
| Complain about lost luggageMore with Perry | 4 · 3 · 3 | 4 · 5 · 4 | More with Perry |
| Gift ideas for DadFewer with Perry | 4 · 4 · 4 | 3 · 3 · 3 | Fewer with Perry |
| A newsletter subject lineFewer with Perry | ✗ · ✗ · ✗ | 0 · 0 · 0 | Fewer with Perry |
| SQL for customers who stopped orderingAbout the same | 1 · 1 · 1 | 1 · 2 · 1 | About the same |
| A toast at a brother’s weddingFewer with Perry | 4 · ✗ · ✗ | 2 · 2 · 2 | Fewer with Perry |
How we tested
Each task is a short, rough request, the way people really type, plus three to five things the person wants but didn’t say: a length, a name, a deadline, a tone. We wrote all twenty and saved them before the first run, so we couldn’t pick tasks that suit Perry afterwards.
Each request went to the chat AI once as typed and once through Perry, exactly as the free extension sends it, with no profile. Where Perry left a [blank], the person’s own details went in. A second AI then read only the answer and the list of wants, without knowing which way it was asked, and ticked off what was there. For the first thing missing, the person sent one follow-up, and so on, up to 5.
What went wrong
Meeting notes: the request had the notes pasted below it, and Perry’s rewrite left them out and asked for them instead. As typed, the AI got it right first time. It was a real bug. We fixed it the same day, and the same request now keeps the notes word for word. The numbers above are from before the fix and stay as they are.
The Excel formula: neither way finished, because our own rule (“explain it in one or two sentences”) was stricter than a real person would be. We left it as it was rather than change the test after seeing the results.
What this doesn’t show
Twenty tasks is a small test. We wrote the tasks and we make Perry. The chat AI, the judge and Perry all run on Google models. And a follow-up here fixes one thing at a time, where real people sometimes fix several at once, which is why the first-answer figure matters most.
All the data
- The method, written before the first run
- The twenty tasks, with what each person wanted
- Our summary
- Every prompt, answer, follow-up and judgement (long)
- The raw data