What’s changed in Probant.
Every change shipped, day by day, from newest to oldest. Fixes are labeled as such.
Version 1, before public launch
From Nov 6, 2026 to Nov 13, 2026: 99 changes shipped while preparing the first version.
4 changes
- Editor“Edit on my site” no longer sends you back to “Plan and launch”: a test that isn’t saved yet is saved as a draft when you open your site, and you finish setting it up afterward without leaving the editor.
- EditorDrag the selected element in the preview to its new position; “Find an element” locates it by the page’s text or by a CSS selector, so you don’t have to look for it.
- EditorA change can apply to certain devices only: a shorter headline on mobile, reviews moved up on tablet, a block hidden on desktop. Use the “Devices” option in the “Element” tab and on each change in the list; the device is the one used for targeting, and elsewhere the change counts neither as not found nor as not displayed.
- EditorEditing on your site: the “Insert” tab offers the same ready-made blocks as the “Element” tab, in the colors of your page’s buttons, and what you typed stays in place if an error appears.
4 changes
- Editor“Element” tab: insert a block next to the selected element, with no code and no assistant (heading, paragraph, button, image, reassurance list, divider, or custom HTML). The button and check marks take the color of your page’s buttons, with the most readable label color.
- Editor“Copy appearance” then “Paste appearance”: one element takes another’s colors, font, padding, border, corner radius, and shadow. Only the settings that change something are applied.
- EditorSelecting the right element: the trail of blocks that contain it appears in the “Element” tab, and the arrow keys move to the parent block, the first child, or the siblings.
- EditorStyle settings grouped (text, background and border, spacing and size), with font (your page’s fonts suggested), case, underline, letter spacing, border color, height, max width, and spacing between elements, also when editing on your site.
16 changes
- IntegrationsFixLeads and Subscriptions pages: the API call examples now point to your app’s address instead of a placeholder address.
- IntegrationsYour tests’ exposures are now also sent to Contentsquare, Piano Analytics, Amplitude, and Mixpanel, as they are to GA4 and Segment, and always after the visitor consents.
- PersonalizationImported audiences: your customer data platform sends an audience’s members (API or Segment Engage), and a test or personalization can target them. No list goes through the script’s public configuration.
- EditorNew “Survey” widget: a closed-ended question or a 0-to-10 rating. The report shows how answers are distributed in each variant, with their ranges.
- ExperimentsPer-device optimization test: on mobile, tablet, and desktop, the shares each follow what works on that device. Results are read device by device, with no declared winner.
- IntegrationsAn optimization test can also be launched through the management API and the MCP server, with the same validations as in the app. Its report can be exported as a CSV and shared by link, period by period.
- ExperimentsZone-based and switchback tests: checkpoints set in the test plan, an automatic stop when a guardrail is crossed, and a place in the library once completed.
- SecurityMCP server: an assistant such as Claude or ChatGPT connects to it through OAuth, with no key to copy. Each connection stays within the permissions of the person who authorized it and can be revoked from Settings, Integrations.
- ReportsSurveys: answers are included in the CSV export, and an answer can become a metric, counted over all exposed visitors and not just those who answered.
- ExperimentsMeasurement survey: the same survey, asked in every branch, original included. The report can now compare the original with the variants.
- ExperimentsOptimization test on an amount: revenue or order value per visitor, extreme values capped, read period by period.
- InstallationPer-device optimization tests served without flicker by the server helper, @probant/edge, feature flags, and the PHP, Python, Ruby, and Go SDKs.
- ExperimentsZone-based and switchback tests: guardrails declared and read through the API and the MCP server. At each checkpoint, Probant tells you when the target effect can no longer be reached.
- PersonalizationImported audiences: full sync in one call, verified Segment signature, MCP tools, and the estimated size of the personalizations that target them.
- SecurityConnected assistants: their connections appear in the workspace export, and an assistant connected with read access can request write access, which you grant on the consent screen.
- InstallationFixServer-side device targeting: an iPad in “desktop site” mode is no longer mistaken for a desktop. When the header isn’t enough to tell, the test is left to the script, which knows the device.
14 changes
- MetricsSubscriptions: Stripe connects directly, with no intermediate app. Tests are read on net revenue, MRR, and retention, and a test’s cohort keeps being tracked after the decision.
- MetricsLead generation sites: your CRM sends qualified leads, meetings held, and closed deals. A test is read on these delayed conversions, not just on submitted forms.
- AppNew workspace from the menu: one workspace per site, each with its own plan, free to create.
- AppThe menu shows the domain and icon of the site where your script runs.
- InstallationA visitor keeps the same identifier across your site’s whole domain, subdomains included.
- MarginWithout a connected back office, the line items of a purchase sent by the script give a gross margin at catalog cost. The back office’s net margin takes precedence as soon as it exists.
- DataAn objection button to place in your privacy policy: a visitor who objects is no longer measured, and what was linked to them is deleted.
- InstallationA visitor who accepts cookies mid-page keeps the page as they see it: their variant starts on the next page.
- Test planThe test plan is prefilled with the traffic and baseline rate measured by Probant. When launch isn’t possible yet, the reason appears under the button.
- InstallationA single script to install: it loads session recording itself when the workspace turns it on.
- SecurityA variant’s code loads from an address specific to your workspace, so your content security policy only allows your own variants.
- LibraryHistory import from AB Tasty, VWO, Kameleoon, or a CSV file: Probant recalculates each verdict, and an import can be undone.
- SecurityPage changes can’t touch scripts, the base tag, or form destinations: all JavaScript goes through the Code tab and its checks.
- SecurityDeleting your account revokes the invitations you sent that are still pending.
31 changes
- TeamCustom roles, built from named permissions. No one can grant more than their own rights.
- EditorNo-code widgets: banner, countdown, pop-in, and badge, kept in the workspace library for other tests.
- Metrics“Non-inferiority” and “equivalence” goals, retention metrics, internal traffic filtered out on receipt, and dated annotations on the report.
- Editor“Edit on my site”: the visual editor opens on the real page, scripts included, with no browser extension.
- ExperimentsDynamic allocation (bandit): the split follows the results, day after day, without reassigning a returning visitor.
- ReportsOptional Bayesian reading, with a prior drawn from your past tests. The decision is still made by the plan.
- Test planMultivariate test: sections and levels, full factorial design, main effects, and interactions.
- LibraryDecision quality, criterion by criterion, and an agency-branded client report, shared through an expiring link.
- MarginActual margin: order lines, product costs read from Shopify or imported as a CSV, and margin per visitor calculated on the server.
- Feature flagsTyped feature flags, with rules, gradual rollout, kill switch, server helper, and evaluation API.
- EditorCode tab: hand-written variant CSS and JavaScript, reviewed at approval and served from a dedicated domain.
- ReportsPage speed by variant (LCP, INP, and CLS), tracked as a guardrail.
- InstallationOn Safari, the server helper resets the visitor’s cookie so they keep their variant.
- ReportsTraceable decisions: sealed and verifiable plan, displayed false positive risk, test rated testable or not, fading effects flagged, and a certified impact report.
- SessionsClick and scroll maps, variant by variant.
- ExperimentsConcurrent tests on the same zone without splitting traffic: conflicts are blocked and interactions are checked.
- IntegrationsReport CSV export and share link, Slack alerts and signed webhooks, management API, and MCP server.
- TeamTwo-factor authentication, audit log, launch approval by another member, contributor role, and staging key.
- ExperimentsTargeting: audiences, triggers, redirect tests, scheduling, bots excluded, and exposures sent to GA4 or Segment.
- ReportsSeveral metrics per test: one primary, secondaries corrected together, ratios such as average order value, and guardrails that stop the test on their own. Differences adjusted with CUPED, extreme values capped.
- MetricsNo-code goals: a click on an element, a form submission, or a page visit, with the element picked on your site’s real page.
- MetricsEvents your site already sends are suggested as metrics, with the form prefilled.
- PersonalizationMeasured personalization: an experience shown to an audience, with a holdout kept on the original to check that it really helps.
- SessionsSession recordings: the journey of a test’s visitors, with pages, clicks, and errors, and no video or typed input.
- MarginOrders received through the API or Shopify webhooks, matched to visitors, with refunds deducted.
- EditorIn your workspace, the editor starts from your real page: you enter its address, variant B starts empty, and the draft keeps the tested page.
- EditorThe preview shows your page without running its scripts.
- EditorIn your workspace, the assistant doesn’t reply with a demo example: if it’s unavailable, it says so, and manual editing remains available.
- DataYour visitors’ events are kept for 13 months, then deleted automatically every night.
- AppFixThe menu no longer shows the demo store while your workspace is loading.
- AppThe app overview has a new address: /dashboard.
30 changes
- ReportsEach report tab (Verdict, Metrics, Segments, Sessions, Plan) shows its own panel. The selected tab can be shared by link.
- ReportsNew Plan tab: metric, target effect, sample size, and checkpoints, as locked at launch.
- DataComputations run in Europe, where your visitors’ events are stored.
- AppSign-up with email and password. Each account gets its own workspace; the demo stays open to everyone, read-only.
- ExperimentsIn your workspace: drafts saved from the editor, launch, pause, resume, and stop with a recorded decision.
- ReportsA real test’s report is calculated from your events: difference, range, sample ratio mismatch (SRM) check, verdict, and checkpoints. While a variant has fewer than 5 converted visitors, the difference isn’t shown.
- InstallationThe overview tracks your installation: script key and code to copy, events received in the last 24 hours, and first test launched.
- IdeasYour workspace’s ideas: creation, statuses, filters by zone and theme. The expected effect only appears once two tests have been completed and measured.
- MetricsYour workspace’s metrics: events received over 7 days, baseline value, smallest effect detectable in 4 weeks, the probant.track(…) call to place on your site, and CSV export.
- LibraryYour workspace’s library: each stopped test with its decision and measured effect, or the reason it’s missing.
- RolloutsA winning variant rolls out in steps from its test: move up a step, suspend, resume, go to 100%, or revert to the original.
- RolloutsA rollout that crosses a guardrail reverts to the original on its own.
- InstallationThe script and the server helper apply rollouts: a visitor who has seen the new version keeps it at the following steps.
- AppCommand palette (⌘K or Ctrl+K) to open a page, an experiment, or a completed test.
- ReportsDedicated screens for tests without a winner: running, completed with no detectable effect, paused for sample ratio mismatch, stopped by a guardrail.
- ExperimentsThe experiment list can be filtered by status.
- AppOn mobile: a report you can read without horizontal scrolling, a drawer menu, and experiments shown as cards.
- App“Not found” and “error” pages with a Try again button.
- InstallationFixAn invalid selector no longer stops the script from applying the following changes.
- InstallationBrowser script under 7 KB compressed: consent first, no page hiding while it loads, and changes reapplied after every re-render.
- InstallationServer-side assignment helper, for a variant known before rendering and a flicker-free display.
- InstallationPublic API: workspace configuration and event ingestion, each event validated individually.
- EditorVariant editor: changes checked on the preview, mobile, tablet, or desktop view, side-by-side A/B comparison, and assistant.
- EditorHTML and CSS suggested by the assistant are sanitized on the server; each selector is rated stable or fragile.
- LibraryLibrary screen: lessons by theme, timeline by zone, and distribution of completed tests’ effects.
- IdeasIdeas screen: each idea is estimated from past tests (expected effect, time to conclude, gain) and ranked by how worthwhile it is to test.
- MetricsMetrics screen: detectable effect in 2 and 4 weeks for each metric, with and without CUPED.
- Test planPlan and duration screen: duration, visitors, and checkpoints calculated before launch.
- ReportsReport calculated by the statistical engine: winner’s curse correction based on past tests, segments corrected with Holm’s method, and monthly gains rounded to the nearest hundred.
- ReportsStatistical engine with no third-party library: delta-method effects, power and duration, SRM, CUPED, always-valid intervals, O’Brien-Fleming boundaries, Holm, and Benjamini-Hochberg.
Conclusive Or not yet…
Create your workspace, or explore the demo first with data from a fictional store.
- Open demo, no sign-up
- Free to start, no overage charges
- Database and compute in Amsterdam, with Scaleway