The Assistant

The Assistant is a chat helper that lives on every page of your dashboard. Instead of hunting through menus, you just ask: "is this result a real win?", "what should I test next on this page?", or "create a metric for checkout completion", and it either answers or does the work with you.

Open it from the chat button in the bottom-right corner of any page, or press Cmd+Shift+K (Ctrl+Shift+K on Windows and Linux). It opens as a small panel you can resize, and you can pop it out to full screen when you want more room.

1. Click this button, on any page, to open the Assistant.

1. The header names the page the Assistant is reading, here the results page.
2. Click a starter question, or type your own.
3. Prepare buttons draft the lifecycle action; you still click the real button to do it.

What it can do

• Answer questions about your experiments and results in plain English: how a test is doing, what a result means, whether it's trustworthy. It can list your experiments for itself, so "how are my running experiments doing?" is a question it can actually take on.
• Find things the way you name them. Ask about "experiment 300035", "the Hero test" or "metric 41" and it finds the right one by the number the dashboard shows, by its name, or by its id. If a name fits more than one thing, it lists them with their numbers and asks which you mean, instead of guessing.
• Explain results honestly (see below): it reads your actual numbers and puts them in words, and it never invents a figure.
• Set things up with you: create an experiment, add variations, pick a metric and an audience, or create a metric, audience, or segment. When it's going to change something, it shows you exactly what it will do first and waits for your OK.
• Suggest what to test next: grounded in the experiment or results you're looking at, not generic advice.

What it can read, and what it can change

Launching, pausing, stopping and archiving are on neither side: it prepares them and you press the button yourself.

Some examples you can type as they are:

• "Which metrics does this project have, and which ones are in use?"
• "Explain the results of experiment 300035 in plain English."
• "What would break if I deleted the metric Signup completed?"
• "Make metric 41 the main metric of the Hero test."

The Assistant sticks to what your workspace can actually do. If a feature is switched off for your organization, or is not on your plan, or has been turned off for your account, the Assistant refuses that request in the same words the rest of the app uses, instead of quietly doing it anyway.

Whatever you ask, it either helps or tells you plainly why it can't: it never fails silently or leaves you at a dead end. See When it says it can't help for what the refusals mean.

Nothing changes without your OK

The Assistant is safe to chat with freely. Asking questions changes nothing. When you ask it to create or change something (an experiment, a metric, an audience), it shows you a clear preview of the change. Then it waits for you to approve it. You can approve, reject, or adjust before anything is saved.

And for the biggest moments (launching, pausing, stopping, or declaring a winner), the Assistant will prepare the action and hand it to you, but it never does it on its own. You always make that call and click the button yourself.

Those two kinds of change are never mixed. If a single request would create a metric and stop a test, you get two separate cards. The harmless one keeps its ordinary Apply button. The one that stops something gets a card of its own, marked as serious, with a button that says what it actually does (Stop experiment, Archive experiment) and a line explaining what stops happening and what that means for the test's data. One click can only ever agree to one of them.

It tells you what it changed

Every time the Assistant creates or changes something, it writes it out: a plain sentence naming what it made, the real settings it used, and a link straight to the thing. The reply appears under the approval card the moment you click Apply, so you never have to open the list of steps to find out what happened.

Created the audience "Mobile US and Canada Visitors".

• Targeting: Device is mobile AND Location is US, CA

[See the audience]

The settings are the actual values, never a count like "3 conditions". That is on purpose. A count looks correct no matter what is underneath it, and reading the real settings back is what lets you catch a mistake in the second it is made rather than a fortnight later in the results.

The page behind the panel keeps up as well. Create an audience while you are looking at the Audiences page and it appears there on its own, with no reload.

It removes nothing unless you ask

The Assistant will not remove, clear, archive or stop anything on its own initiative. Not as tidying up, not as a side effect of something else you asked for, and not because it decided something looked unused. Ask it to "clean up this project" and it will tell you what it thinks should go and then wait to be asked. That is a deliberate choice, not a missing feature.

Because of it, there is no Undo button, and that is deliberate too: undoing a creation means deleting it, and deleting is exactly what the Assistant does not do behind you. The link in the confirmation above is the way back. It takes you to the thing, where you can look at what you are about to reverse before you reverse it.

When you do ask, it checks what would break first

Removing something a live test depends on does not produce an error. The test carries on and quietly stops measuring or targeting what it was set up for. You find out when the numbers stop making sense.

So before it proposes any removal, replacement or stop, the Assistant works out what depends on the thing and tells you, naming any test that is running right now, and then waits for you to confirm. The same goes for swaps: changing a running test's main measure resets what it is measuring, and it says so before you approve.

The only thing it can remove itself is an audience that nothing uses. It has no way to delete a metric, a segment or an experiment, and this is where that stops being a dead end. Ask it to delete a metric and instead of "I can't do that" you get the useful answer: which experiments use it, which of those are live, what would stop being measured, and where in the dashboard to do it yourself. You can name the metric by its name or by the number on the Metrics page.

Honest about results

Results are the one place a confident wrong answer really costs you, so the Assistant follows strict rules there:

• It never makes up a number. Every figure it mentions comes straight from your computed results: it will not estimate or invent one.
• Numbers read the way the results page shows them. Rates, lifts and chances come as percentages, like "converts at 9.1%", "+6.3%" or "a 73.9% chance of being better", and a range reads "-1.1% to +13.1%". You never get a raw decimal such as 0.0909 to decode yourself.
• Caveats come first. If a result isn't trustworthy yet (the test is still collecting, the traffic split looks off, or a guardrail slipped), it tells you that before any "winner" talk, never after.
• It won't tell you to ship a loser. It will never call a losing or inconclusive test a win, and it never contradicts the verdict your results already show.
• It knows what's on your screen. A plain "is this good?" refers to exactly the metric, segment, and dates you're currently viewing, not the experiment in the abstract.

One conversation that follows you

• Pick up where you left off. Come back to a project and the Assistant resumes your most recent conversation there. Move to a different project and it switches to that project's conversation: the previous one is saved and waiting.
• The whole conversation comes back, not just your half. Close the panel and reopen it and you get everything: your messages, its replies, the record of anything it created, and the link to each one.
• A question you never answered comes back with its buttons. If it asked you to pick between options and you closed the panel, the choices are still there when you return. You never have to guess what it offered.
• An approval you never answered comes back live. Ask it to create something, close the panel before deciding, and the approve and reject buttons are waiting for you when you open it again. If the conversation list says something is waiting for you, it is reachable.
• It remembers what it did, so it does not do it twice. Ask for an audience and then ask for something completely different, and it does the second thing. It can see the audience it already made, so it does not quietly make another one.
• Start fresh anytime. A New conversation button (and a keyboard shortcut) gives you a clean slate in one click. Your old conversations stay in the history list, where you can search, rename, or delete them.
• Nothing gets lost. If you start a new conversation while something is still waiting for your approval, the Assistant saves it to your history and shows a small "waiting for you" marker on the button so you can come back and finish it.
• A chat that behaves. Hover your last message and click the pencil to edit and resend it; the earlier answer is replaced instead of duplicated below. Every message and code block has a one-click copy button. And while an answer streams, the chat follows along only if you're already at the bottom; scroll up to re-read and it stays put, with a Jump to latest button to bring you back. Your conversation history opens as a slide-over panel over the chat.

Turning it on and off

The Assistant is its own feature, separate from AvsB Copilot in the visual editor, though the two are billed from the same AI actions. It comes on automatically for every project once billing details are on file. There's no per-project switch to find or flip.

AI actions need billing details on file, billed per action with no packs to buy. If the chat button isn't showing up, adding billing details is what turns it on. See Billing for how.

When an answer fails on our side, it is free

Each message you send uses one AI action. There is one exception. Sometimes the Assistant finds the experiment or metric you asked about, and then fails to read it because of a problem on our side. That reply does not use an action, and a line under the answer says so:

This reply did not use one of your AI actions, because the lookup failed on our side, not yours.

The counter in the chat goes back up straight away. A question about something that does not exist, such as an experiment number nobody has used, is a real answer, and it uses an action as normal. So does a reply that went on to help you another way.

Your privacy

The Assistant follows the same privacy promise as the copilot: nothing is sent to an AI provider until you submit a message, prompts and data are routed with zero data retention (not logged, never used to train models), and anything from your account is treated as data for the AI to read, never as instructions it must obey. For the full details of what is and isn't sent, see What leaves your browser on the copilot page.

When it says it can't help

A refusal tells you which refusal it is, and what to do about it. There is no generic "something went wrong" standing in for every different reason.

The most common one by far: your email is not verified yet. If you signed up with Google, your email address is not marked as verified, and AI features stay closed until it is. The Assistant says so:

Please verify your email address before using AI features. Check your inbox for the verification code.

Fixing it takes a minute. Open Account → Profile. Your email address shows a Not verified badge with a Send verification email button next to it. Click it, take the six-digit code out of the email, and enter it on the verification page. Check your spam folder if nothing arrives. The Assistant works normally from the moment your address is verified, and nothing else about your account needs changing.

The other refusals name themselves the same way:

• The feature has been switched off for your organization, or the account is suspended. The message says which. Contact support to have it restored.
• Your organization has no billing details on file. AI actions are billed per action once billing details are on file. Add them under Settings → Billing.
• Your agreement's AI actions for this month are used up. If your usage is covered by an agreement with us, it sets a monthly number of AI actions. Contact support to raise it.
• The daily AI spending limit was reached. It resets at midnight UTC.
• Your sign-in expired. Sign in again.
• Too many requests, or another reply is still finishing. Wait a few seconds.

Try again without retyping

Most refusals clear on their own within seconds: a burst of requests, another reply still finishing, the daily budget. So a failed message now carries a Try again link right next to the explanation. One click resends exactly what you wrote. You never have to type it out a second time.