What you can ask the assistant

What Claude or ChatGPT can do with Varly, with example prompts, and how launching a test from the chat works.

In this article

The assistant uses your own Varly login: the same clients, sites and role as you. Everything it creates shows up in Varly for your team.

It always asks which client

Varly never picks a client for the assistant and never guesses from part of a name. With more than one client, the assistant asks which one you mean; answer with the exact client name. If the client has several sites, it asks for the site too (by name or domain). With one client and one site, it uses that one.

Every answer starts with the client and site it's about, and one request never mixes two clients.

Ask about your research

  • "Which pages on our shop need attention?" Pages that are slow, frustrating or have JavaScript errors, with Web Vitals.
  • "Give me an overview of the last 30 days." Traffic, devices, top pages, ideas and experiments.
  • "What do visitors click on the home page on mobile? Any rage clicks?" Most clicked elements, rage and dead clicks, scroll depth, speed and errors for one page.
  • "Which clicks on the product page go together with orders?" Click impact per element. This shows a connection, not a cause, so the assistant suggests a test instead of drawing conclusions.
  • "Show me recordings of mobile visitors with rage clicks on the product pages, and tell me what went wrong in one of them." It reads a recording step by step, with a link to watch it.
  • "What did people answer in the exit survey?" Answers per question, NPS, averages and the latest open answers.

Narrow questions by device, browser, operating system, traffic source, new or returning visitors, country, or whether visitors placed an order.

Work on ideas

  • "Which test ideas do we have, and which ones have enough traffic to test?" Ideas ranked by Varly's score.
  • "Write a test idea about showing delivery times on the product page." The assistant gives impact and ease from 1 to 5, and Varly computes the score.
  • "Move the free-shipping idea to the shortlist." Ideas can move to new, shortlist, done or rejected.
  • "Run a research round for this site." Varly's rules always run. If your organization has an AI key, the AI step runs too, which uses credits on that key.

Create surveys and tests

  • "Create an exit survey for the product pages." The assistant creates a draft from a template: why visitors leave, cart objections, product-page clarity, NPS, how easy it was, or how people heard about you. A person reviews it and turns it on in Varly.
  • "Build an A/B test from the delivery-times idea." Varly builds the variant, checks it on desktop, tablet and mobile and makes a preview link, in a few minutes. Ask for up to 3 variants, or a multivariate test with up to 9 combinations. Nothing goes live. This needs an AI key.
  • "I want to change some things on the homepage myself." The assistant creates a test and gives you a link to the visual editor. Select what should change on your live page, then say "Build it" and Varly builds the new version from your selections.

Launch a test from the chat

Launching always takes two steps, and the second one needs your yes.

  1. Preparation. Ask "Launch the delivery-times test." The assistant shows you:
    • the preview link on your real site (only you see the variant and nothing is measured), and screenshots on desktop, tablet and mobile,
    • what the variant changes and any QA warnings,
    • the test plan: weekly visitors, conversion rate, the smallest effect the test can detect and how many weeks that takes. It then asks: dry run (visitors are split but the page doesn't change) or live? Later you can switch a dry run to live in the app with Go live: show the new version to visitors; the count then starts again.
  2. Your yes. Check the preview on your phone and computer. The assistant only launches after you say yes to this test, this client and this mode.

If the test is rebuilt or stopped in between, the assistant must prepare it again. The app's rules apply: a test runs for at least 2 full weeks, the assistant must tell you when your traffic is too low for a reliable result, and you need the Editor or Owner role. See Launch a test.

To stop a test, ask "Pause the delivery-times test" (you can launch it again later) or "Complete it" (it ends for good). Every visitor then sees the original again.

What you see in Varly

  • Ideas made in the chat show the assistant as creator, for example Claude (chat).
  • A test launched from the chat shows who launched it and from which assistant, for example Launched from Claude by and your name.

Common problems

The assistant keeps asking which client. Use the exact client name as it appears in Varly. A partial name or nickname doesn't count, so the assistant asks again.

It says the connection is read-only or may not launch tests. Disconnect it under Account → Connected apps and connect again.

It says nothing can be launched with this plan. The page has too little traffic to detect that effect within 8 weeks. Ask for a bigger smallest effect, or pick a busier page.

Was this article helpful?

Still stuck? Email support@varly.ai and we'll help you out.

Related articles

Updated 5 November 2026