Certification
Three levels of certification in experimentation: an exam on its own, or the exam plus real work on your own website. Scenario questions, a public syllabus, and not everyone passes.
Associate, Specialist and Expert
One bar for everyone
Pertento certification certifies the experimenter, not the tool. Associate shows what you know; Specialist and Expert show what you can do, with Pertento as the lab you prove it in. Agency CRO people, in-house marketers, analysts and existing customers are all held to the same standard.
Judgement over recall
Every question is a scenario. Nothing asks what a tab does or where a setting lives.
Tool-neutral at its core
The foundations section would still be worth passing if you never used Pertento again.
Real work on a real site
Specialist and Expert are earned on your own website, with its own visitors, not on a sandbox.
Not everyone passes
Each section is passed on its own, retakes wait 14 days, and the credential expires after two years.
Open about what is tested
The syllabus and the capstone rubric are on this page, in full.
01
All three take the same two timed papers. Specialist adds a set of setup tasks in your own account, and Expert adds a real test on your own site. Passed papers count toward the level above, so you can start with Associate and move up without sitting them again.
Associate
Proves you know how a sound experiment is designed and read, and how Pertento runs one. The two timed papers on their own, with no website and no account needed.
Suits anyone starting out, and anyone without a site to test on. Your passed papers count toward Specialist for two years.
Specialist
Proves you can design, set up and read an experiment correctly. Two timed papers, graded automatically, and setup tasks a grader checks in your own account.
Suits in-house marketers and Pertento customers.
Expert
Adds a real test on your own site, planned before launch and graded by a person on how you ran it, not on whether it won.
Suits agency CRO people and analysts. Anyone may attempt Expert directly.
02
Strong product knowledge cannot make up for weak statistics, and strong statistics cannot make up for not knowing the product.
| Section | What it tests | Weight | Grading | Associate | Specialist | Expert |
|---|---|---|---|---|---|---|
| A. Foundations | Experimentation judgement, in any tool | About 40% | Automatic | Included | Included | Harder set |
| B. Pertento functions | Choosing and using the right function | About 30% | Automatic | Included | Included | Included |
| C. Setup tasks | Building correctly in your own account | About 30% | A grader, in your account | Not included | Included | Included |
| D. Capstone | A real test: locked plan, run, decision memo | Pass or fail | A person, against the rubric | Not included | Not included | Included |
03
For Specialist and Expert, Pertento sets up a free account for each candidate. It is not a trial, and it stays active for 60 days. Existing customers take the certification on their current account. Associate needs none of this: you sign up and sit the papers.
Your own, or a client's. Agency candidates confirm they have the site owner's permission.
The first week's baseline from the snippet decides whether the site qualifies. Self-reported numbers are not accepted.
You invite Pertento's grader user into your account, which holds data about your site's visitors. That is how a grader checks your Section C setup.
04
This is the Specialist and Expert window; Associate has no account to run. Specialist candidates sit their sections any time within it. The Expert test is planned to end by day 42, which leaves room for a test of two to four weeks and for grading while the account is still active.
Pertento sets up an empty account. You install the snippet on your own site.
The first week's traffic, measured by the snippet, decides whether the site qualifies and whether the Expert capstone unlocks.
The capstone plan is written in the account before the test launches.
Room for a test of two to four weeks, run to its own stopping rule.
The decision memo is submitted, and grading happens while the account is still active.
The free account closes. You can upgrade to a paid account to keep using Pertento.
05
Everything the exam can ask about. Results screens in the trap cases carry numbers generated for each candidate, so an answer cannot be passed along.
Whether you can design, run and read an experiment well in any tool.
| Topic | What you must be able to do |
|---|---|
| Hypotheses and prioritisation | Write a falsifiable hypothesis tied to evidence, and rank ideas by expected value rather than opinion. |
| Metrics | Choose a primary goal, guardrails and secondaries, and spot a proxy metric that will mislead. |
| Power and sample size | Calculate sample size and MDE, and judge whether a test is feasible on the traffic available. |
| Stopping rules | Explain why peeking inflates false positives, and when sequential testing makes stopping early safe. |
| Data quality | Detect and diagnose sample ratio mismatch, broken tracking and bot traffic. |
| Reading results | Tell "not significant" from "no effect", read confidence intervals, and know when to ship a flat test. |
| Frequentist and Bayesian | Interpret p-values, intervals, probability to be best and expected loss without mixing them up. |
| Multiple comparisons | Recognise when many variants or goals need correction, and what the correction costs in power. |
| Segments and time | Spot Simpson's paradox, novelty and primacy effects, and day-of-week cycles. |
| Running a programme | Handle concurrent tests and interaction, document learnings, and decide what to test next. |
Which function fits a situation and why, never where it sits in the menu. Questions mix scenarios with finding the problem in illustrations of the console, and the question set is reviewed at every major release.
| Area | Functions covered | Example judgement |
|---|---|---|
| Building variants | Visual editor, AI editor, custom CSS and JS, URL redirect tests, server-side tests and the server SDK, feature flags, component library, dynamic content | Which build method fits a pricing change, a new checkout, or a backend algorithm |
| Delivery | Anti-flicker rendering, sticky variant assignment, activation | What a visitor sees when each one is missing |
| Targeting and traffic | URL targeting conditions, device and cookie targeting, weights per variant, automatic rebalancing, concurrent tests | When rebalancing weights mid-test invalidates the result |
| Goals and statistics | Primary and secondary goals, revenue and AOV per variant, frequentist and Bayesian results, expected loss, sequential testing, Holm, Benjamini-Hochberg and Sidak corrections, SRM alarms | Which correction suits five variants on one goal |
| Before launch | Test runs, variant screenshots, power analysis, MDE, run-length plan, scheduling | Reading a run-length plan that says the test cannot finish |
| While running | Health monitor, minimum-runtime guard rails, trust gates, stop on significance, collision and interaction detection, alerts | When stopping on significance is safe |
| Programme | Attention queue, experiment debrief, programme report, hypothesis tracking, PXL scoring, pipelines, Learnings, one-click winner deployment, implementation handoff | Ordering a backlog with PXL and defending the order |
| Research inputs | NPS, CSAT and CES surveys, triggers and frequency capping, surveys inside experiments, Insights funnels and journeys | Turning a funnel drop and survey answers into a hypothesis |
You build things in your own account, and a grader opens it and checks the saved setup against your brief. It grades the setup, never the results, so it does not depend on how much traffic arrives. Every task can be done in draft or QA, and nothing has to reach a visitor. Details are generated for each candidate; a typical set looks like this.
For example two variants on the product page, revenue as the primary goal, add to cart as a secondary, mobile only, 50/50, in QA. Every field is checked.
For example exit intent on the checkout path, shown once per visitor, with two questions.
With the build method matched to what the brief asks for.
Saved on the experiment from the first task, using your site's own baseline.
06
One real test on your own site. A grader marks how you ran it, not whether it won.
Hypothesis, primary goal, guardrails, MDE with its justification, run length and stopping rule, written in the account by day 14. The activity log timestamps it, and a plan edited after launch fails.
For at least two full weekly cycles, following your own stopping rule, ending by day 42.
When the planned run ends: ship, iterate or stop, with your reasons.
If the power analysis shows your site cannot detect a realistic effect in time, saying so and changing the design is the right answer: a bolder change, a busier page, or a goal earlier in the funnel. Running anyway and declaring a winner fails.
| Criterion | Fail | Pass | Strong |
|---|---|---|---|
| Hypothesis | No mechanism or evidence | States a mechanism and the evidence behind it | Evidence from the site's own data, such as Insights or a survey |
| Metrics | Primary goal does not match the hypothesis | Primary goal and at least one guardrail fit | Explains why each secondary is there and how it could mislead |
| Power and duration | No power analysis, or the arithmetic is wrong | MDE and run length follow from the baseline | Redesigns the test when the first design is underpowered |
| Discipline | Plan changed after launch, or stopped early against its own rule | Plan followed | Deviations documented with reasons before they happen |
| Validity checks | SRM, tracking or interactions not checked | Checks done and reported | Checks interpreted, with what would change the decision |
| Decision | A conclusion the data does not support | Decision follows from the result and its uncertainty | Names what the result cannot tell them, and the next test |
The rubric weights the decision over the writing. Your data is unique to your site, so AI can polish the prose but cannot supply the right call. Every submission is read by one grader, borderline ones by a second, and a 15-minute call is offered when a result is borderline.
07
The question formats keep the exam hard. The retake and expiry rules keep the credential worth something.