Two things are true about this material at the same time, and separating them matters. The teaching is very good. The corporate situation behind it is a mess. On the teaching: most vendor documentation about A/B testing is a description of how to click the buttons, with a paragraph of hand-waving about statistical significance and a link to a p-value calculator.
Statsig's is not that. The statistics section works through the things that actually determine whether your experimentation programme produces knowledge or noise. CUPED, which uses pre-experiment data to reduce variance and can cut required sample sizes substantially, is explained with enough detail that you understand what it assumes rather than just that it exists. Sequential testing is covered properly, including the real reason it matters, which is that engineers peek at results and peeking at a fixed-horizon test inflates false positives in a way that feels like nothing is wrong.
Sample ratio mismatch gets its own treatment, which is right, because SRM is the single most common signal that an experiment is broken and the single most commonly ignored one. Switchback tests and interference between units get discussed, which almost nobody bothers with despite marketplace and social products violating the independence assumption constantly. If you work on experimentation and you have not read a proper treatment of these ideas, this is a free one. The corporate situation is where I would urge caution.
OpenAI acquired Statsig in September 2025 for around 1.1 billion dollars, and founder Vijaye Raji became CTO of Applications there. The acquisition was, fairly transparently, about the people. OpenAI wanted an experimentation and release-control capability in-house as it turned from a research lab into a company shipping consumer products, and it bought a team that had built exactly that. What it did not want, it turned out, was the enterprise software business bolted to the side.
In May 2026 Amplitude took on the Statsig brand, platform and customer base, with the team staying at OpenAI. Read that sequence carefully, because it describes a product now maintained by an organisation that did not build it, with the engineers, product leaders and statisticians who did build it employed somewhere else entirely. That is a specific and unusual risk. It is not the ordinary acquisition risk where a roadmap slows down.
It is the situation where the accumulated judgement about why the system works the way it does has left the building. The statistics documentation in particular reads like it was written by people with strong opinions who had argued these questions internally. Those arguments are not repeatable by the new owner. There is a second-order problem too.
Amplitude already has its own experimentation product, its own academy and its own certification track, all of which this catalogue has reviewed. It now owns two overlapping products and two overlapping bodies of educational material, and nothing published so far explains which one a customer should be learning. If you are choosing what to invest study time in this quarter, that is a live question with no answer. My recommendation splits along those lines.
Treat the statistical writing as what it is, a free and unusually rigorous explanation of experimentation methodology that happens to be hosted by a vendor, and read it for the ideas rather than the product. Those ideas transfer to GrowthBook, to Eppo, to LaunchDarkly, to an in-house setup on top of your warehouse, to anything. Treat the platform-specific training as provisional. The gaps are worth naming as well.
Metric design gets less space than the statistical machinery, and that is backwards for most teams, because the common failure in experimentation is not that the maths was wrong, it is that the team optimised a metric that did not represent anything they cared about and shipped a change that improved clicks and worsened retention. There is also nothing that tells a small team the honest thing, which is that below a certain traffic volume a formal experimentation programme costs more in engineering time and delayed decisions than it returns, and that they should be doing qualitative research instead. No vendor will ever write that, because it argues against buying the product, which is the same structural blind spot that shows up in every vendor academy in this catalogue. Four point one.
Excellent free teaching on a genuinely difficult subject, with a rating held back not by the material but by the unusual fact that the people who wrote it no longer work on it.