# GrowthBook

> GrowthBook is an open source feature flagging and experimentation platform that runs A/B tests against data already sitting in your own warehouse, computing results with a Bayesian or frequentist engine that supports CUPED variance reduction and sequential testing, and showing you the SQL behind every number; it is free for three seats on the cloud, $40 per seat per month on Pro, and free to self-host with no user limit.

- Category: Product Analytics & Session Replay (https://saastracker.org/categories/product-analytics)
- Website: https://www.growthbook.io
- Starting price: $0 (Starter, 3 seats), then $40 per seat per month (Pro)
- Free plan: Starter is free for up to 3 users and 1 project, with unlimited feature flags, unlimited experiments, and unlimited traffic. No credit card required.
- Free trial: No fixed trial; the free Starter tier and the open source build serve as the evaluation path
- Founded: 2020, HQ: Anaheim, California, United States, Ownership: Venture-backed, Y Combinator W22
- Profile last reviewed: 2026-08-22
- Canonical profile: https://saastracker.org/products/growthbook

## Overview

GrowthBook exists because the two common ways to run experiments are both unsatisfying. Buying a closed experimentation platform means shipping your event data to a vendor and trusting a number you cannot audit. Building it yourself means writing the statistics, and most teams get that wrong in ways that quietly invalidate a year of decisions. GrowthBook's answer is to leave the data where it is, in Snowflake, BigQuery, Redshift, Databricks, ClickHouse, Postgres, or a dozen other sources, generate the SQL itself, and show you that SQL alongside the result.

That architecture has a consequence worth stating plainly: only aggregated statistics leave your infrastructure, and GrowthBook collects no data about your users. For a small company with a European customer base, this quietly removes most of the data processing agreement conversation that a hosted analytics vendor forces. It also means GrowthBook is not a source of truth about your product. It has no event pipeline of its own unless you opt into the Managed Warehouse, so a team with no existing analytics data has to solve that problem first.

The statistics are the differentiator and are genuinely better than the category norm. The default engine is Bayesian, reporting the probability that a variation beats control rather than a p-value, with configurable priors and violin plots of the full distribution. A frequentist engine is available with two-sample t-tests, sequential testing to handle the peeking problem, and multiple testing corrections on dimensional breakdowns. CUPED variance reduction works with both engines. Six automatic data quality checks run on every experiment, including sample ratio mismatch detection, multiple exposure alerts, guardrail metric monitoring, and suspicious uplift detection. Most flag vendors in this category ship a naive t-test or nothing at all.

Commercially it is a Y Combinator W22 company founded in 2020 by Graham McNicoll and Jeremy Dorn, both previously at Education.com, with roughly 42 staff and about $23.1M raised including a $22.6M Series A in June 2025. Pricing is $0 for three seats on cloud, $40 per seat per month on Pro up to 50 users, and free forever if you self-host the open source build with unlimited users. The seat meter is the thing to watch: an eight-person team on Pro pays $320 a month, which is more than most flag-only vendors charge for the whole company.

## How it works

1. You connect a data source. GrowthBook supports BigQuery, Snowflake, Redshift, Databricks, ClickHouse, Athena, Postgres, MySQL and MariaDB, SQL Server, Vertica, PrestoDB and Trino, and Mixpanel. If you do not have a warehouse, the Managed Warehouse option gives you one-click setup with built-in event tracking and auto-created metrics, billed at $0.03 per thousand events.

2. You define metrics as SQL or through the metric builder: conversion metrics, revenue metrics, count metrics, ratio metrics, with denominators, conversion windows, and capping rules. This is the step that determines whether your results mean anything, and it is where the hour of care should go.

3. You install an SDK and create a feature flag or an experiment. GrowthBook SDKs evaluate everything locally with no network request at decision time, pulling a payload of rules and computing the assignment in process, which means flag checks add no latency and no outbound call per user.

4. Users get bucketed by a hashed identifier, deterministically, so the same person keeps the same variation. Sticky bucketing keeps the assignment stable even if the targeting rules change underneath them, which is the fix for the classic problem of users flipping between variants mid-experiment.

5. When you look at results, GrowthBook writes the SQL, runs it against your warehouse, and returns aggregated statistics only. You can read the query, export the whole analysis as a Jupyter notebook, and check the statistics engine source on GitHub. Data quality checks flag sample ratio mismatch and suspicious uplifts before you act on a number that is wrong for boring reasons.

## Best for

Product and growth teams at small software companies who already have event data in a warehouse and want to run experiments with defensible statistics, plus engineering teams who want feature flags and the measurement of those flags in the same system without shipping user data to a vendor.

## Not the right fit for

- Teams with no analytics data and no warehouse who expect the tool to answer product questions on its own; without a data source GrowthBook has nothing to compute, and the Managed Warehouse is a starting point rather than a full product analytics suite.
- Larger teams shopping purely on price; the $40 per seat Pro meter means a fifteen-person team pays $600 a month, where a flag-only vendor like Flagsmith charges $250 for the whole company.
- Anyone who wants session replay or heatmaps; GrowthBook does not record sessions, so a diagnostic question about why users are confused needs a second tool.
- Marketing teams wanting to test landing pages without engineering; the visual editor is Pro-only and still expects a developer to have installed the SDK, and this is not a substitute for a dedicated CRO platform.
- Companies that need an experimentation platform to also be their event pipeline, identity resolution layer, and dashboard tool; GrowthBook deliberately reads rather than owns your data.

## Features

### Statistics engine

The part that actually separates GrowthBook from a flag service with a percentage split.

- **Bayesian engine by default**: Reports the probability that a variation beats control rather than a p-value, using an uninformative Normal prior with mean 0 and standard deviation 0.3 by default, with proper priors configurable per organisation.
- **Frequentist engine**: Two-sample t-tests on relative percent change, for teams whose stakeholders expect confidence intervals and p-values or whose statistical review process demands them.
- **Sequential testing**: Available on the frequentist engine to solve the peeking problem, so checking results early does not inflate your false positive rate. Very few products at this price ship this.
- **CUPED variance reduction**: Uses pre-experiment data to reduce noise, available on both engines. In practice it shortens the runtime needed to detect a given effect, which matters enormously for a product with modest traffic.
- **Multiple testing corrections**: Applied on dimensional breakdowns in the frequentist engine, which is the guard against slicing until something looks significant.
- **Automatic data quality checks**: Six checks run on every experiment including sample ratio mismatch, multiple exposures, guardrail metric monitoring, and suspicious uplift detection, so broken experiments surface before anyone acts on them.
- **Visible SQL and notebook export**: Every result shows the query that produced it, and the full analysis exports as a Jupyter notebook. The statistics engine source is public on GitHub, which is the strongest auditability claim in this category.
- **Multi-armed bandits**: Pro tier adds bandit experiments that shift traffic toward the winning variation as evidence accumulates, for cases where maximising outcomes matters more than a clean read.

### Feature flags

A complete flag product, not an afterthought bolted onto an analytics tool.

- **Unlimited flags and experiments**: Unlimited on every tier including the free one, with unlimited traffic, so the meter is seats rather than usage.
- **Local SDK evaluation**: SDKs pull a rules payload and evaluate in process with no network request per decision, so flag checks add no latency and a GrowthBook outage does not affect your application path.
- **Targeting rules and saved segments**: Attribute-based targeting with force rules, percentage rollouts, and reusable segments across environments.
- **Sticky bucketing**: Keeps a user in the same variation even when targeting rules change, preventing the mid-experiment reassignment that silently corrupts results.
- **Safe rollouts**: Pro tier monitors guardrail metrics during a rollout and surfaces regressions rather than waiting for a human to notice, which turns a flag into a release safety mechanism.
- **Holdouts**: Keep a slice of users out of a set of launches for a long period to measure the cumulative effect of everything you shipped, an enterprise-tier feature and a rare one.
- **Ramp schedules and approval workflows**: Scheduled traffic increases and change approvals for teams that need production toggles under change control. Enterprise tier.

### Data sources and warehouse integration

Your data stays where it is, which is the architectural decision that defines the product.

- **Twelve-plus warehouse connectors**: BigQuery, Snowflake, Redshift, Databricks, ClickHouse, AWS Athena, Postgres, MySQL and MariaDB, SQL Server, Vertica, PrestoDB and Trino, and Mixpanel.
- **Aggregate-only transmission**: Only aggregated statistics leave your infrastructure. GrowthBook does not collect data about your users, which materially simplifies the GDPR conversation compared with hosted analytics platforms.
- **Managed Warehouse**: For teams with no warehouse, a one-click hosted option with built-in event tracking and automatically created metrics, billed at $0.03 per thousand events on top of the plan.
- **SQL metric definitions**: Metrics are defined as queries with conversion windows, denominators, capping, and metric types including binomial, count, duration, and revenue, so business definitions are explicit rather than vendor-decided.
- **Fact tables**: Define a table once and derive many metrics from it, which keeps a growing metric library maintainable instead of turning into hundreds of nearly identical queries.

### SDKs and delivery

How the assignment gets into your application.

- **Broad SDK coverage**: JavaScript and React, Node, Python, PHP, Ruby, Go, Java, C# and .NET, Elixir, plus mobile SDKs for Kotlin and Android, Swift and iOS, React Native, and Flutter.
- **Edge SDKs**: Deployment targets at the CDN edge including Cloudflare Workers, Fastly, and Lambda@Edge, so variant assignment can happen before the page is even rendered and avoid the flicker problem.
- **Visual editor**: Pro tier includes a visual editor for making variant changes to a page without shipping code, plus URL redirect tests. Useful for copy and layout changes, not a replacement for a full CRO platform.
- **CDN payload delivery**: Rules payloads are served through a CDN with published overage rates of $10 per million requests and $1 per gigabyte beyond the plan allowance, which is the only usage-based line item on the cloud plans.
- **REST API and webhooks**: Full API coverage plus webhooks for pushing flag and experiment state into other systems, with an official Terraform-style infrastructure-as-code posture through the API.

### Governance, hosting, and administration

Where the tiers actually diverge.

- **Open source self-hosting**: The full platform runs on your own infrastructure with unlimited users and one project, behind a firewall if you want, with community support. This is the same codebase as the cloud, not a crippled community edition.
- **Advanced permissioning**: Pro adds granular roles and permissions across projects and environments; Enterprise adds advanced access control on top.
- **SSO and SCIM**: Single sign on and SCIM provisioning are Enterprise features on both cloud and self-hosted, alongside exportable audit logs.
- **Custom environments and projects**: Free gives one project, Pro gives three, and Enterprise gives unlimited projects with custom environments, which is the constraint that pushes multi-product companies upward.
- **Power calculator**: Pro includes a sample size and power calculator, so you can decide before launching whether your traffic can detect the effect you care about rather than discovering after four weeks that it never could.
- **Uptime SLA**: A 99.99 percent uptime SLA is an Enterprise commitment; lower tiers have no contractual SLA.

## Use cases

- **Growth team at a Series A SaaS with data in BigQuery**: Experiments are currently run by shipping a variant, waiting, then arguing about a dashboard number that nobody can reproduce. Outcome: GrowthBook connects to the existing BigQuery tables, metrics are defined once as SQL, and every result carries a probability, a data quality check, and a query anyone can read. Arguments become about the decision rather than about the number.
- **Engineering team that wants flags and measurement in one place**: The team has a flag vendor and an analytics vendor, and connecting a rollout to its effect on retention is a manual export and a spreadsheet. Outcome: Flags and experiments live in the same product, safe rollouts watch guardrail metrics automatically, and a launch that hurts a key metric is caught by the tool rather than by a customer complaint.
- **Product team with modest traffic**: Only a few thousand users hit the feature each week, so most experiments never reach significance and the team stops bothering. Outcome: CUPED variance reduction plus the power calculator make the traffic problem explicit and shorten the runtime for detectable effects, and Bayesian reporting gives a usable read on partial evidence instead of a binary verdict that never arrives.
- **European company with a strict data protection posture**: Legal will not approve shipping behavioural event data with user identifiers to another hosted analytics vendor. Outcome: GrowthBook queries the company's own warehouse and receives only aggregated statistics, or the whole platform is self-hosted behind the firewall with unlimited users at no licence cost.

## Pricing

Per-seat subscription on cloud with a free tier, plus a free unlimited-user open source self-hosted edition and quoted Enterprise plans on both. Usage charges apply only to CDN delivery and the optional Managed Warehouse.

- **Starter (Cloud)**: $0 per month, up to 3 users. Unlimited feature flags and experiments; Unlimited traffic; 1 project; Full Bayesian and frequentist statistics engines; Managed Warehouse available. The statistics are not crippled on the free tier, which is unusual; the constraint is three seats and one project.
- **Pro (Cloud)**: $40 per seat per month, up to 50 users. 3 projects; Visual editor and URL redirect tests; Multi-armed bandits and safe rollouts; Advanced permissioning and customisable dashboards; Power calculator and premium support. Seat-metered, so cost scales with team size rather than traffic. Model this before assuming it is the cheap option.
- **Enterprise (Cloud)**: Custom quoted. Unlimited seats and custom environments; SSO and SCIM provisioning; Approval workflows, ramp schedules, holdouts; Exportable audit logs and advanced access control; 99.99 percent uptime SLA with dedicated support.
- **Open source (Self-hosted)**: $0 plus your own infrastructure cost. Unlimited users; 1 project; Bring your own warehouse; Default environments only; Community support. The most generous self-hosted offer in this category: unlimited users at no licence cost, provided you can run it.
- **Enterprise (Self-hosted)**: Custom quoted. Everything in open source plus custom environments; SSO and SCIM provisioning; Visual editor, bandits, safe rollouts; Ramp schedules, approval workflows, sticky bucketing, holdouts; Dedicated support and training.

Billing notes:

- The cloud meter is seats, not events and not monthly tracked users, so a high-traffic product with a small team is cheap and a low-traffic product with a big team is not.
- Usage charges are limited and published: CDN requests at $10 per million beyond the included allowance, CDN bandwidth at $1 per gigabyte, and Managed Warehouse events at $0.03 per thousand.
- Self-hosting with unlimited users and no licence fee is a real option, not a trial version, and it is the reason GrowthBook shows up in cost comparisons against LaunchDarkly.
- Project limits, one on Starter and three on Pro, are the quiet constraint for companies running several products from one account.
- SSO, SCIM, audit logs, approval workflows, and holdouts are all Enterprise-only on both cloud and self-hosted, so a compliance-driven buyer will end up quoted rather than self-serve.

Value assessment: Judged on statistics per dollar, GrowthBook is the best value in this category by a wide margin: sequential testing, CUPED, proper Bayesian inference with configurable priors, multiple testing corrections, and six automatic data quality checks are all present on a free tier. Judged on total bill, it depends entirely on headcount. Because the meter is seats rather than events, traffic volume barely matters: a product with 10,000 monthly users and one with 100,000 monthly users both cost the same, which is either a bargain or an irrelevance depending on your shape. A five-person team on Pro pays $200 a month; a twenty-person team pays $800, which is where the free self-hosted build with unlimited users starts to look attractive if you have the operational capacity. Against Statsig, whose free tier is enormous and whose statistics are comparably sophisticated, GrowthBook wins on warehouse-native architecture and auditability and loses on out-of-the-box analytics. Against Flagsmith or Unleash it costs more and does far more.

## Strengths

- The statistics engine is genuinely serious: Bayesian by default with configurable priors, a frequentist option with sequential testing, CUPED variance reduction on both, and multiple testing corrections on dimensional breakdowns.
- Every result shows its SQL and exports to a Jupyter notebook, and the engine source is public, which makes results auditable rather than a vendor assertion.
- Warehouse-native architecture means your event data never leaves your infrastructure and only aggregate statistics reach GrowthBook, which shortens the data protection review considerably.
- Six automatic data quality checks including sample ratio mismatch catch the broken experiments that quietly poison decision-making elsewhere.
- SDKs evaluate locally with no network request per decision, across a broad language set plus edge runtimes on Cloudflare Workers, Fastly, and Lambda@Edge.
- The free self-hosted build supports unlimited users with the same codebase as the cloud, which is a real cost escape and a real exit.
- The free cloud tier does not cripple the statistics, so you can evaluate the thing that actually matters before paying anything.

## Limitations

- Per-seat pricing at $40 punishes larger teams; a fifteen-person product organisation on Pro pays $600 a month for capabilities that a smaller team gets for $120.
- You need a data source. Without a warehouse or the paid Managed Warehouse, GrowthBook has nothing to compute against, so it is not a first analytics purchase.
- Metric definition is real work in SQL, and the quality of every result depends on getting conversion windows, denominators, and capping right. There is no autocapture shortcut.
- Project limits are tight: one project on Starter and self-hosted open source, three on Pro, unlimited only on Enterprise.
- SSO, SCIM, audit logs, approval workflows, holdouts, and custom environments are Enterprise-only, which means a compliance-driven small buyer leaves the self-serve path.
- No session replay, no heatmaps, no error monitoring, so this sits alongside other tools rather than replacing them.
- Sequential testing is currently a frequentist-engine feature, so teams who prefer the Bayesian default should understand which safeguards apply to which engine before setting a peeking policy.

## Comparisons

- **GrowthBook vs Statsig**: The closest comparison on statistical seriousness. Statsig meters on analytics events with a very large free allowance and brings its own analytics, warehouse-native mode, and a mature experimentation platform. GrowthBook meters on seats, reads your existing warehouse, and shows you the SQL. Pick Statsig if you want the whole platform including analytics and are comfortable with event-based pricing; pick GrowthBook if auditability, warehouse-native architecture, and per-seat predictability matter more.
- **GrowthBook vs Flagsmith**: Flagsmith is a flag service with no statistics engine at all, priced at $40 a month for a whole three-person team. GrowthBook charges $40 per seat but computes experiment results with sequential testing and CUPED. If you want release control, Flagsmith is far cheaper and does that job well. If you want to know whether the release helped, Flagsmith cannot tell you and GrowthBook can.
- **GrowthBook vs Unleash**: Unleash is an enterprise-grade open source flag platform at $75 per seat with governance included and a light experimentation layer. GrowthBook is $40 per seat with a much deeper statistics engine but a thinner governance story until you reach Enterprise. Choose Unleash if flag operations and approvals are the requirement, GrowthBook if measurement is.
- **GrowthBook vs LaunchDarkly**: LaunchDarkly has unlimited seats and meters on monthly active users and service connections, with a large ecosystem and now observability and replay bundled in. GrowthBook is seat-priced, open source, warehouse-native, and considerably more transparent about its statistics. Small teams almost always land cheaper on GrowthBook; large teams with many engineers may find LaunchDarkly's unlimited-seat model cheaper and its enterprise controls more complete.
- **GrowthBook vs Convert Experiences**: Convert Experiences is a marketing-side CRO platform with a visual editor, priced from $299 a month billed annually for 100,000 tested users. GrowthBook is a developer-side experimentation platform priced per seat with a far deeper statistics engine. A marketing team testing landing pages without engineering should use Convert; a product team testing in-app changes with warehouse data should use GrowthBook.
- **GrowthBook vs PostHog**: PostHog gives you analytics, replay, flags, and experiments on one usage-based bill, which is unbeatable for convenience. GrowthBook does not collect your data at all and computes against your warehouse with a more configurable statistics engine. Take PostHog if you want everything in one place and no warehouse; take GrowthBook if you already have a warehouse and want experiment results you can audit line by line.

## Implementation

- Setup time: A day for a first experiment if you already have clean event data in a warehouse. Connect the data source, define two or three metrics, install the SDK, and launch. Without a warehouse, add however long it takes to get events landing somewhere, or use the Managed Warehouse to shortcut it.
- Learning curve: Moderate and front-loaded on metric definition. The flag half is easy; the experimentation half asks you to think about conversion windows, denominators, guardrails, and priors. Teams that skip this get results that look precise and are not. The visible SQL helps, because a data-literate person can check the definition rather than trusting a label.
- Onboarding: Self-serve on Starter and Pro with no credit card required to begin, and self-serve on the open source build. Enterprise adds dedicated support and training. Documentation covers per-warehouse setup, the statistics methodology, and per-language SDK guides.
- Migration: Migrating flags from another vendor is straightforward through the API, and the switching cost is mostly re-expressing targeting rules. The genuinely valuable asset is your metric library, which is defined as SQL in GrowthBook and therefore portable in a way that a closed platform's metric definitions are not. Moving from cloud to self-hosted, or the other way, runs on the same codebase. Because your event data never lived in GrowthBook, leaving costs you the configuration rather than the history.

## Platform, API & security

- Platforms: JavaScript and React, Node, Python, PHP, Ruby, Go, Java, C# and .NET, Elixir, Kotlin and Android, Swift and iOS, React Native and Flutter, Cloudflare Workers, Fastly, and Lambda@Edge
- API: Full REST API and webhooks covering features, experiments, metrics, and results, plus notebook export of any analysis and a public statistics engine repository on GitHub.
- Compliance: GDPR posture strengthened by design: only aggregated statistics leave your infrastructure and GrowthBook collects no user data, Self-hosted deployment available behind a firewall for buyers who cannot use a processor at all, Confirm current SOC 2 scope with the vendor during procurement
- Data residency: Because analysis runs against your own warehouse and only aggregates are transmitted, residency is effectively determined by where your warehouse lives rather than by GrowthBook. Self-hosting removes the question entirely. The Managed Warehouse option reverses this, so buyers with residency requirements should stay on their own data source.
- SSO: SSO and SCIM provisioning are Enterprise features on both cloud and self-hosted plans.
- Security notes: SDKs evaluate locally with no per-decision network call, so user attributes used for targeting stay in your application process. Exportable audit logs and advanced access control are Enterprise; Pro provides advanced permissioning across projects and environments.

## Support

- Channels: Community Slack and GitHub issues on free and open source, Premium support on Pro, Dedicated support and training on Enterprise
- Documentation: Detailed documentation at docs.growthbook.io including a statistics section that explains the Bayesian and frequentist engines, sequential testing, CUPED, and the six data quality checks in more depth than most vendors publish.
- Community: Active open source community on GitHub with a public statistics engine repository, plus a Slack community. The willingness to publish the methodology is the community's centre of gravity.

## Company

- Founded: 2020
- Founders: Graham McNicoll, Jeremy Dorn
- Headquarters: Anaheim, California, United States
- Ownership: Venture-backed, Y Combinator W22
- Employees: Approximately 42 (2026)
- Funding: Approximately $23.1M raised across two rounds, including a $22.6M Series A completed in June 2025 and a $500K seed from Y Combinator.

Funding history:

- Seed (2022): $500K. Y Combinator W22 batch.
- Series A (2025): $22.6M. Completed June 2025, bringing total funding to roughly $23.1M.

Timeline:

- 2020: Founded by Graham McNicoll and Jeremy Dorn, who had previously built and scaled the experimentation infrastructure at Education.com and wanted an open source version of it.
- 2022: Goes through Y Combinator's W22 batch and releases the open source platform combining feature flags with warehouse-native experiment analysis.
- 2023: The statistics engine matures with CUPED variance reduction, sequential testing on the frequentist engine, and automatic data quality checks including sample ratio mismatch detection.
- 2024: Adds multi-armed bandits, sticky bucketing, and edge SDK support for Cloudflare Workers, Fastly, and Lambda@Edge, closing the flicker problem for server-rendered and edge-rendered sites.
- 2025: Raises a $22.6M Series A in June and introduces the Managed Warehouse option so teams without existing data infrastructure can adopt the platform.
- 2026: Operating at roughly 42 staff with cloud pricing at $0 for three seats and $40 per seat on Pro, alongside a self-hosted open source edition that remains unlimited on users.

## Integrations

BigQuery, Snowflake, Redshift, Databricks, ClickHouse, AWS Athena, Postgres, MySQL and MariaDB, SQL Server, Vertica, PrestoDB and Trino, Mixpanel, Segment and RudderStack event streams, Slack, GitHub, Webhooks and REST API

## FAQ

### What is GrowthBook?

GrowthBook is an open source feature flagging and experimentation platform. It manages flags and A/B tests in your application through local-evaluation SDKs, then computes experiment results by querying data already in your own warehouse, using a Bayesian or frequentist statistics engine that supports CUPED variance reduction and sequential testing. It is free for three seats on cloud, $40 per seat per month on Pro, and free to self-host with unlimited users.

### How much does GrowthBook cost?

Starter on cloud is free for up to three users and one project, with unlimited flags, experiments, and traffic. Pro is $40 per seat per month for up to 50 users and three projects, adding the visual editor, bandits, safe rollouts, advanced permissioning, and the power calculator. Enterprise is quoted. Self-hosted open source is free with unlimited users and one project. The only usage charges are CDN requests at $10 per million, CDN bandwidth at $1 per gigabyte, and Managed Warehouse events at $0.03 per thousand.

### What does GrowthBook cost at 10,000 versus 100,000 monthly users?

Almost exactly the same, because the meter is seats and not users or events. A five-person team pays $200 a month on Pro at either scale, with only CDN request overage and optional Managed Warehouse event charges varying with traffic. This is the opposite shape from Mixpanel or Amplitude, and it makes GrowthBook unusually cheap for high-traffic products and comparatively expensive for a large team on a small product.

### What statistics does GrowthBook actually use?

The default is Bayesian, reporting the probability a variation beats control, using an uninformative Normal prior with mean 0 and standard deviation 0.3 unless you configure proper priors. A frequentist engine computes two-sample t-tests on relative percent change and supports sequential testing to handle peeking, plus multiple testing corrections on dimensional breakdowns. CUPED variance reduction works with both engines. Six automatic checks run per experiment including sample ratio mismatch, multiple exposures, guardrail monitoring, and suspicious uplift detection.

### Does GrowthBook store my user data?

No. The architecture queries your own warehouse and only aggregated statistics are transmitted back, and GrowthBook states it collects no data about your users. That is the reason it is often the easiest experimentation tool to get past a data protection review. The exception is the optional Managed Warehouse, where GrowthBook does host your events, so buyers with residency requirements should stay on their own data source or self-host.

### Which data warehouses does it support?

BigQuery, Snowflake, Redshift, Databricks, ClickHouse, AWS Athena, Postgres, MySQL and MariaDB, SQL Server, Vertica, PrestoDB and Trino, and Mixpanel. If you have none of those, the Managed Warehouse gives you one-click setup with built-in event tracking and auto-created metrics at $0.03 per thousand events.

### Is the self-hosted version crippled?

Less than most. The open source build supports unlimited users, connects to your warehouse, and runs the same statistics engine, with community support and one project. What it does not include is the Pro and Enterprise layer: visual editor, bandits, safe rollouts, custom environments, sticky bucketing, ramp schedules, approval workflows, SSO, and SCIM all require a commercial self-hosted licence. The operational cost is the usual one, since you own the deployment and its upgrades.

### Can GrowthBook replace my product analytics tool?

No, and it does not try. There are no funnels, retention curves, or exploratory dashboards in the Mixpanel or Amplitude sense, and no event pipeline of its own unless you take the Managed Warehouse. GrowthBook answers whether a specific change moved a specific metric. Understanding the product broadly is a job for an analytics tool sitting next to it, or for the warehouse GrowthBook is already reading.

### Does GrowthBook add latency or a network call to my app?

No. SDKs pull a rules payload and evaluate everything locally in process, so there is no network request at decision time and a GrowthBook outage does not affect your application. Edge SDKs for Cloudflare Workers, Fastly, and Lambda@Edge take this further by assigning variants before the page renders, which removes the flicker that plagues client-side testing tools.

### Who owns GrowthBook and how is it funded?

GrowthBook was founded in 2020 by Graham McNicoll and Jeremy Dorn, who previously built experimentation infrastructure at Education.com, and went through Y Combinator's W22 batch. It is venture-backed with roughly $23.1M raised including a $22.6M Series A in June 2025, and employs around 42 people. The open source codebase is the practical hedge against vendor risk, since the self-hosted build runs the same engine.

## Editorial verdict

GrowthBook is the experimentation platform to buy if you care whether the numbers are right. Sequential testing, CUPED, configurable Bayesian priors, multiple testing corrections, and automatic sample ratio mismatch detection are present on a free tier, and every result shows the SQL that produced it. That combination does not exist anywhere else at this price. The architecture is the other half of the argument: your data stays in your warehouse, only aggregates leave, and a self-hosted build with unlimited users is always available as an exit. Buy it if you already have event data in a warehouse and a habit, or an ambition, of running real experiments. Do not buy it as a first analytics tool, because it computes against data it does not collect, and do not assume the seat meter is cheap without multiplying $40 by your actual headcount. The work that determines your return is metric definition, and no vendor can do that part for you.

## SaaSTracker awards

- Innovation (Product Analytics & Session Replay, Summer 2026): "A real statistics engine that queries your own warehouse, so experimentation happens where the data already lives."

---

Source: SaaSTracker (https://saastracker.org), an independent editorial project. This profile is compiled from public information, carries no peer reviews or paid placement, and was last reviewed 2026-08-22. Awards are judged on published criteria: https://saastracker.org/methodology