Skip to content
Release log

What shipped, and when.

Every release lands here, newest first. Some entries carry a date, because one day is genuinely the day it shipped. Others carry a month, because the work arrived across that month and picking a single day would be made up.

The copilot tidies up after itself, never after you

  • ChangedAsk the copilot to remove something and it now removes only what it built itself. Anything a person made is hidden instead, and hiding is reversible.
  • FixedA follow-up request now edits the section you meant instead of quietly adding a second copy of it below.
  • FixedWhen you ask for one more change to something the copilot built, its styling is added to rather than replaced with a thinner version.
  • ChangedThe summary of what changed is written in plain English, so you can read a change list without knowing how the page is put together.

A public status page, and billing you run yourself

  • AddedA public status page, checked from outside the platform, so you can see for yourself whether the service is healthy instead of asking us.
  • AddedSubscribe, change your plan, or cancel it from inside the product, and download your own invoices. Only Enterprise needs a conversation.
  • AddedIf you go past your traffic allowance, capacity is added automatically in blocks, and you set the monthly ceiling on what that can cost.
  • ChangedCustom roles, built from a fixed list of permissions, became part of the Enterprise package.

Results that tell you how much to trust them

  • AddedEvery result now carries a trust grade from A to F. It scores how the test was run, not how much you wanted it to work, and nobody can override it.
  • AddedWhen a result crosses the line, you also see the honest, corrected version of the lift, because a number shown to you because it won is biased upward.
  • Added9 one-click analytics integrations: Google Analytics 4, Google Tag Manager, Mixpanel, Segment, Amplitude, Heap, FullStory, Contentsquare, and Adobe Analytics. Events are forwarded from the page into the tool already running there, with no keys to paste in, and this is free on every plan.
  • AddedServer-side delivery on paid plans: forward validated events to a signed webhook, or to the Google Analytics, Mixpanel, and Amplitude server APIs, with a delivery log per destination.
  • AddedFor money metrics with enough traffic, the results page shows what shipping a secretly worse variation would cost you, in currency, next to the upside of holding.
  • AddedGuardrail metrics get a real pass or fail test with three honest outcomes: Safe, Breached, or Inconclusive.
  • AddedWatching many metrics at once no longer manufactures winners. Secondary metrics are corrected as their own family, and your primary metric is corrected across variations.
  • AddedUnder the engine built for daily checking, a correction that would not hold up is automatically upgraded to a stricter one, and the page says so rather than doing it quietly.
  • ChangedThe check for a broken traffic split is now measured from the last time you changed the split, rather than across the whole life of the experiment, so an old change stops raising a false alarm.
  • AddedAlongside the lift, you see the odds that the true effect actually points the other way.

Describe the change. Review what it built.

  • AddedDescribe the change you want in plain English and the copilot edits the real page, producing the same reviewable list of changes a person editing by hand would produce.
  • AddedAny custom behaviour the copilot writes is checked at save time for calls out to anywhere other than your own site, including addresses assembled while the page runs. If it finds one, saving stops until you confirm.
  • AddedA second check catches code that would rewrite parts of the page the copilot did not create, and asks for the same explicit confirmation.
  • AddedWhat it proposes is drawn from your own page: its colours, its type, its spacing. Values close to yours are offered as a one-tap snap, and genuinely new ones are left exactly as they are.
  • AddedUndo works on a copilot change exactly the way it works on one you made by hand.

Write a metric once, use it everywhere

  • ChangedDefine a metric once and reuse it, either inside one project or across your whole organisation, so two teams stop measuring the same thing two slightly different ways.
  • RemovedThe copilot no longer asks you to pick a mode before you type. Every request is routed for you.

Built for sites that sell things, and an API for everything else

  • AddedA purchase metric, so revenue is measured without anyone hand-instrumenting a checkout.
  • AddedRevenue read properly: revenue per visitor, total lift, and ranges that account for it, so a small conversion win that doubles basket size still wins.
  • AddedTarget visitors by cart value, by what they have viewed, and by what they have bought before, matched by product or by category.
  • AddedA documented REST API covering experiments, flags, metrics, audiences, segments, members, roles, and webhooks, with reference pages for every route.
  • AddedRepeat a request safely: send the same one twice and it will not run twice. Rate limits are enforced and written down.
  • FixedAn experiment aimed at several addresses now runs on all of them, instead of only the first one in the list.

The packages went public

  • AddedReact, Next.js, Vue, Svelte, Solid, Angular and React Native each get a real installable package, not a snippet of glue code to paste.
  • AddedA framework-free browser client, a server package for Node, and one package covering the major edge runtimes.
  • AddedTest helpers, so you can assert on flag behaviour inside your own test suite instead of hoping.
  • AddedA command line tool that clones an experiment’s code onto your machine, pushes your edits back, pulls the latest down, and spots a conflict before it overwrites anything.
  • AddedThree linting levels for variation code: off, advisory, or strict. Strict blocks saving on a type error.
  • AddedShared type definitions at two levels: one file shared across an experiment, and one shared across a whole project. Types only, so nothing extra is shipped to the browser.
  • AddedPython, Go, Ruby, PHP, Java and .NET clients exist and are available on request. They are not on the public package registries.

Tell the right people, and start on your own schedule

  • AddedRoute alerts into Slack, Microsoft Teams, or Jira, per project, with keyword filtering so the channel does not turn into noise everyone mutes.
  • AddedSchedule an experiment or a rollout to start and stop at a wall-clock time in the time zone you name. Daylight saving is handled, and there is a one-hour catch-up window if something goes wrong at the moment of launch.
  • AddedAPI tokens scoped to your organisation, with per-resource read and write permissions, optional expiry, rotation that keeps the audit trail intact, revocation, and a record of when each one was last used.
  • AddedSegment lift: we test whether a result genuinely differs by segment, instead of leaving you to eyeball a table and talk yourself into it.

Metrics that survive harder questions

  • AddedRatio metrics get proper variance handling, and a relative-lift range switches to a more careful method when the baseline is weakly known. Where it genuinely cannot tell you, it says so rather than printing a misleading range.
  • AddedComposite metrics combine several measures with weights you set, and show which part contributed what.
  • AddedExtreme values are capped before analysis, worked out from non-zero values only, so one enormous order stops swinging the whole result.
  • AddedSplit URL tests, where the two versions are separate pages. Match an address exactly, by path, by substring, by pattern, or by regular expression.
  • AddedOne primary metric, as many additional metrics as you want, and guardrails kept separate from both.

A visual editor you can actually work in

  • Added8 point-and-click change types, with no code at all: text, style, visibility, image, reorder, insert, move, section.
  • AddedAim a change at every width, or at mobile, tablet, desktop, or a custom range you define yourself.
  • AddedA command palette and a keyboard shortcut reference, inside the editor, so you can work without hunting through panels.
  • AddedA page-structure tree for picking exactly the element you mean, rather than the one nearest your cursor.
  • AddedShare any variation with a preview link. Whoever opens it needs no account, the link expires on its own, and you can revoke it at any time.
  • AddedDraft history with restore: every save is a version you can go back to, with the author and the version it was based on.

Three ways to read a result, and a plan you cannot quietly change

  • AddedTwo more ways to read a result joined the Bayesian engine. Bayesian, Frequentist and Sequential are now all available, and the results page tells you which one produced the number in front of you.
  • AddedOne of the three is built for people who check results every day. Looking early does not inflate your false-alarm rate.
  • AddedAnalysis plans are written down in advance and seal the moment traffic starts. After that they are read-only.
  • AddedEvery change to a sealed plan is recorded field by field: what changed, from what to what, by whom, when, and why.
  • AddedThree interlocking guards against stopping a Frequentist test too early: a progress banner, a confirmation that blocks, and a permanent record of the override. You can always overrule it. It is always written down.
  • AddedA sample-size calculator that answers the question on all three engines, so you know what a test will cost you before you run it.
  • AddedA/A test mode, which runs the original against itself, so you can prove your setup is honest before you trust anything it tells you.
  • AddedAutomatic variance reduction that uses the parts of your traffic that never change to filter out noise, and only switches itself on when your data supports it.
  • AddedSet a band inside which a difference is too small to care about, so a technically real but commercially pointless result stops being called a win.

Platform launch

  • AddedA guided experiment builder in five steps: targeting, variations, metrics, analysis, and a review before you launch.
  • AddedWrite your variation in JavaScript or TypeScript, with CSS or SCSS, in a real editor with autocomplete rather than a plain text box.
  • AddedFeature flags holding a true or false value, a piece of text, a number, or structured data, delivered to an audience or wired to an A/B test rule.
  • AddedResults read by the Bayesian engine: the probability a challenger beats the original, a credible interval around the difference, and a check that your traffic split is not broken.
  • AddedAudiences built by combining conditions with AND and OR, nested as deep as the question needs, saved once and reused across experiments and flags.
  • AddedSend exposures and conversions straight into the analytics tools you already run: Google Analytics, Google Tag Manager, Mixpanel, Segment, Adobe Analytics, Amplitude, Heap, FullStory, Contentsquare. Anything not on that list is handled by a custom webhook.
  • AddedOne script tag with no dependencies, which runs nothing and tracks nothing until your consent tool says go.
  • AddedThe same visitor always sees the same variation, decided in the browser with no round trip to a server.
  • AddedAn audit log covering a sign-in, an experiment launch, a permission change and everything in between: who did it, from where, and what it looked like before and after.
  • AddedTwo-factor sign-in using a one-time code sent to your email, with a short expiry, a limited number of attempts, and rate limiting.
  • AddedAs many environments as you need, each with its own key, so production and staging never read the same configuration by accident.

Everything on this page is already running.

Nothing here is a plan. Start free and use it today.

Start freeSee how pricing works