Cursor Tab's suggestion acceptance rate fell 10% week over week. How would you investigate?

Introduction

Cursor Tab is the autocomplete that predicts a developer's next edit, often across several lines, and can jump to the next place that needs changing. It is used thousands of times a day by each active developer, so its acceptance rate (suggestions accepted divided by suggestions shown) is one of Cursor's most watched quality signals. In this scenario, acceptance rate fell 10% week over week, for example from 30% to 27%.

Because Tab runs on a fast, specialised model with a tight latency budget, small changes in the model, the client, the context it receives or the users themselves can move this metric. I'll work through it systematically.

Step 1

Clarifying Questions (3 minutes)

• Is this a relative drop of 10% (30% to 27%) or 10 percentage points? I'll assume relative.

• Was there a release in the past week: a new Tab model, a client update, or a change to how suggestions are shown? I'll ask for the full change log for the model, editor and backend.

• Did suggestions shown also change? If suggestions shown rose sharply, the rate can fall even when accepted suggestions are flat.

• Is the drop global, or concentrated in a language, platform, plan or region?