Ideas & Explain
Two features help you decide what to test and understand what a change does, both part of AvsB Copilot and both available right in the editor.
Ideas and Explain use the same AI layer as the copilot, so they only work when AvsB Copilot is enabled for the project (Project settings → Configuration → AvsB Copilot). While the copilot is off, both features are unavailable and point you to the setting: nothing about your page is sent until it's on and you ask.
The Ideas tab
Not sure what to test? Open the Ideas tab in the copilot panel and it suggests three to seven test ideas for the page you're on. Each idea comes with:
- A title: a short name for the test.
- The metric it targets: one of your experiment's own metrics, named exactly as you set it up, so you know what the idea is trying to move. (If the experiment has no metrics configured yet, the idea still comes, just without a metric named.)
- A hypothesis: what you expect to happen and why, in one line.
- Why it should move that metric: the mechanism, the concrete reason this change could shift the metric it targets, not hand-waving.
- A guardrail to watch: a metric the change could hurt, if there is one (a faster checkout that might dent average order value, say), flagged so a gain in one place isn't a quiet loss in another.
- An effort estimate: low, medium, or high, so you can weigh quick wins against bigger bets.
The ideas read the page's structure and its design tokens (and, on the browser extension, a screenshot) and your experiment's configured metrics, so each suggestion stays specific to what's actually on screen and tied to something you're actually measuring, rather than generic advice.
The copilot ties each idea to the metrics you actually configured, but it never makes up statistics. You won't see a predicted lift, a required sample size, or a "this will win by X%": those numbers would be fiction. What an idea is likely to do is a judgement; what it did comes only from your results.
Each idea has a Use this idea button: one click hands the idea's prompt to the copilot and starts a run, so you go from "what should I try" to a staged set of changes without retyping anything. From there it behaves like any other copilot response: review the changes, tweak them, and apply what you want. If the idea builds a new section in a named spot, the copilot confirms where it goes first, just like a typed Build request.
Ideas are prompts for your judgement, not a ranked plan. Pick the one that fits your goal and audience: the effort estimate is there to help you sequence, not to tell you which will win. Only your experiment's results can do that.
Explain any change
Every row in the changes list has an ⓘ button. Click it and the copilot explains, in plain English:
- What that change does: a readable description of the edit.
- Its possible impact: what it might mean for the people who see the variation.
It works for both changes you made by hand and ones the copilot produced, including richer copilot changes like a whole styled section, which it describes in plain English rather than raw markup. It's a quick way to understand a change you inherited from a teammate, or to sanity-check one before you save. Explanations are remembered per change while the editor is open, so re-opening the same one is instant and doesn't spend your daily budget twice.
"Possible impact" is the model's plain-English read of what a change could affect: it is not a forecast of your experiment's outcome. Trust your results for that; use Explain to understand what a change is.
Related
- AvsB Copilot: the copilot these features build on: prompt-to-changes, whole-variant proposals, and the daily budget.
- Getting started: your first edit, end to end.