# Test more. Guess less. > Bixlow A/B tests compare a page's Original with one or more variants that an AI agent builds as injected code, with deterministic in-browser assignment and a Bayesian verdict that only declares a winner when the evidence is there. - Source: https://bixlow.com/en/platform/ab-testing - Language: en - Publisher: Bixlow Turn an idea into a live experiment the same day. The agent builds every variant on your real site, Bixlow splits the traffic and measures, and you decide with a verdict you can trust. - **Test types:** A/B and A/B/n - **Who builds it:** The agent, no developer - **Traffic split:** Deterministic, weighted - **Verdict:** Bayesian, with guardrails ## Why almost nobody tests as much as they would like The idea is the easy part. The hard part is everything between the idea and a result you can defend in front of leadership. ### Every variant needs a developer Moving a block or changing a line of copy becomes a ticket competing with the roadmap. The test ships weeks later, if it ships. ### Data gets mixed Someone changes the split mid-test or a visitor lands in two experiments at once, and the result no longer means anything. ### Winners are called too early Three good days and a screenshot are enough to crown a variant that never repeats its result. ### Win one metric, lose another The variant lifts add-to-cart clicks and nobody notices revenue per session went down. ## From hypothesis to verdict A test in Bixlow is a mode of the workflow: the Original is still your page, and each variant is a draft of its own. 1. **Frame the hypothesis.** Describe the change in your own words or let the agent propose it from your data: problem, change, audience and primary metric. 2. **The agent builds the variants.** Each variant is built as injected code on your real site. You review it in the preview before anything is published. 3. **Split and launch.** Choose each variant's weight and how much of your traffic enters the test. Every visitor always sees the same version. 4. **Read the verdict.** The statistics engine tells you when there is a real winner. Declare it and that variant becomes the page's base version. ## Everything a test needs to be trustworthy The part you never see in a screenshot — assignment, exclusion, restarts — is what makes the result mean something. ### Variants: A/B and A/B/n from the same workflow The Original plus as many variants as you need, each with its own draft and version history. - Each variant is edited separately, by the agent or a person. - Preview any variant before publishing. - Declaring a variant the winner makes it the base version. ### Assignment: Every visitor, always in the same group Assignment happens in the browser with a first-party identifier, so it is stable across visits and never breaks your site's caching. - Configurable weight per variant. - You decide what share of traffic enters the test. - Mutual exclusion groups for tests that overlap. ### Clean measurement: Data that never mixes Change the split or a variant while the test runs and measurement starts again from that moment. - Every conversion is attributed to the variant the visitor saw. - A primary goal plus guardrails that must not get worse. - Automatic sample ratio mismatch (SRM) check. ## The agent builds the test; you decide if it ships It is not an assistant that helps you write CSS. It builds complete variants, sets up measurement and reads the result. ### You decide - Which page and which goal matter now. - Which metric must not get worse. - The split and the traffic that enters the test. - Publishing: nothing reaches your site without your approval. ### The agent does - Proposes the hypothesis and the variant from your data. - Builds the variant in your brand's look. - Sets up goals and guardrails. - Reads the verdict and proposes the next test. ### A test in Bixlow - **Page:** Product page - **Variant B:** Sticky add-to-cart bar on mobile - **Split:** 50 / 50 · 80% of traffic - **Primary metric:** Purchase - **Guardrail:** Revenue per session - **Verdict:** Waiting for minimum sample · day 4 of 7 ## Bixlow Studio: Want a testing program, not a one-off test? Bixlow Studio designs the roadmap, builds and launches tests every week and walks you through them, with a strategist approving every step. - Funnel audit and a prioritized list of hypotheses. - New tests every week on your site. - Results read with a senior strategist. ## FAQ ### Do I need a developer to run an A/B test? No. You install the script once, and from then on the agent builds each variant as code injected into your site. You can review and edit what it builds before publishing. ### How many variants can I test at once? A test can have the Original plus several variants (A/B/n). The more variants, the more traffic you need to reach a verdict, and the statistics engine accounts for that. ### Can a visitor see different variants on each visit? No. Assignment is deterministic and stored in the browser, so each visitor always sees the same variant of the same test. ### What happens if I change the split while a test runs? Measurement restarts from that moment so data from two different setups never mixes. ### Will it slow my site down? The script loads asynchronously and fails silently: if Bixlow can't be reached, your site renders exactly as it would without it. ### Which plan includes A/B tests? A/B tests are part of the Premium plan. Goals and analytics are in every plan, so you can start by measuring and run tests when you need them. ## Your first test, today Install the script, tell the agent what you want to test and review the variant before it goes live.