Running an A/B test costs nothing at the low end. Three of the tools a marketing team is likely to meet give experimentation away on a plan the vendor publishes as ongoing, and none of them is a trial:
- GrowthBook’s free Starter plan lists “Unlimited feature flags”, “Unlimited experiments” and “Unlimited traffic” for up to three users.
- Statsig’s free Developer plan includes A/B tests and multivariate experiments with 2 million events a month, and its plan comparison puts no-code experiments with a visual web editor on that same free plan.
- PostHog charges nothing for the first million feature flag requests a month.
What decides whether you can test is traffic, and traffic has not gotten cheaper. On a page converting at 2%, detecting a 20% improvement takes roughly 42,200 visits split across the two versions. If the page you have in mind gets 4,000 visits a month, there is no tool at any price that will hand you an answer this quarter. That is the whole decision, and it happens before anyone looks at a pricing page.
How much traffic does an A/B test need?
Enough that chance stops being the likeliest explanation for the difference you are looking at. On a page converting at 2%, detecting a 20% relative improvement takes about 21,100 visitors per version, so roughly 42,200 on that page in total. A page already converting at 10% needs about 7,700. Chasing a 10% improvement instead of a 20% one roughly quadruples the requirement.
Every figure below assumes you want to detect a 20 percent relative improvement, at 95 percent confidence and 80 percent power, unless the row says otherwise.
| Conversion rate today | Improvement you want to detect | Visitors per version | Total on the page |
|---|---|---|---|
| 1% | 20% (1.0% to 1.2%) | 42,700 | 85,400 |
| 2% | 20% (2.0% to 2.4%) | 21,100 | 42,200 |
| 2% | 10% (2.0% to 2.2%) | 80,700 | 161,400 |
| 3% | 20% (3.0% to 3.6%) | 13,900 | 27,800 |
| 5% | 20% (5.0% to 6.0%) | 8,200 | 16,300 |
| 10% | 20% (10% to 12%) | 3,800 | 7,700 |
| 10% | 10% (10% to 11%) | 14,700 | 29,500 |
The figures come from the standard two-proportion sample size formula. Put the same conversion rate, lift, confidence and power into any calculator built on that formula and you get these numbers back.
Two things fall out of that table and both of them get missed. The traffic has to be on the page under test, so a site doing 200,000 visits a month across 400 pages does not clear the gate for the pricing page. And you want the answer inside about four weeks, because past that the season changes, the campaign mix changes, and the team stops caring about the question. Take the total column as the monthly visits that one page needs.
Going from a 2% page to a 5% page cuts the requirement by more than half, which is why the highest-converting page you own is usually the only one you can test on.
What to do when the page does not have the traffic
What can you test on a page under 50,000 visits a month?
One thing at a time, and only a big change. Under 50,000 visits a month you are at the 2% row in that table or below it, so a single test consumes the month and only a lift of 20 percent or more will register. A coupon or a price frame can move a conversion rate that far. A headline or a button colour usually cannot, so ship those on judgment and keep the month for the big change.
Most pages do not clear the gate, and that is fine. Our own rule is deliberately more conservative than the arithmetic above: a page without meaningful traffic, which we put at less than 100,000 visits a month, just does not make sense to test. The table gives one test on one page its minimum. The 100,000 is the line below which we tell a team to spend the hours elsewhere.
Below the line, testing is the slowest way to learn something, so spend the hours here instead, in rough order of value:
- Get the number before the argument. Open analytics, filter to the single page, read last month’s visits and its conversion rate. Two minutes, and it settles whether the rest of this is even a question.
- Watch ten session recordings of that page. Below the traffic gate this is the highest-value hour available to you, and it produces the kind of finding a test cannot (“nobody scrolls past the second block”).
- Read the last month of support tickets and sales objections that mention the page. That is qualitative evidence you already own and have not read.
- Ship the change and read the trend for four weeks. Note the date, name the change in your analytics annotations, and resist opening the graph daily.
- Spend your first real test on whichever page already carries the most traffic, even when a quieter page is the one bothering you.
When someone asks us to set up A/B testing, the first thing we say is to settle the experiment criteria and the question they want answered, and to start small rather than spending thousands on enterprise tooling. Write down the question and what you would change if the answer went either way. Teams that skip this step run tests, get results, and do nothing with them, which is a more expensive failure than not testing.
What A/B testing tools cost in 2026
Here is the range, in the currency each vendor prints. Every figure in this table was read on the vendor’s own pricing or documentation pages on 9 October 2026, linked from each vendor’s name.
| Tool | Published entry price | What that covers | Where it runs out |
|---|---|---|---|
| GrowthBook | Free (Starter) | Up to 3 users, 1 project, unlimited experiments, unlimited traffic. Open source and self-hostable. | The AI Visual Editor and the power calculator sit on Pro at $40 per seat per month |
| Statsig | Free (Developer) | 2M events a month, A/B tests, multivariate experiments, no-code experiments in early access, 50,000 session replays | Pro is $150 a month with 5M events included, then $0.05 per 1,000 events |
| PostHog | Free tier, then usage | First 1 million feature flag requests a month free, experiments billed with flags | $0.000100 per request in the 1M to 2M band, dropping to $0.000045 in the 2M to 10M band |
| LaunchDarkly | Free (Developer) | 100K experimentation monthly active users at no additional charge, unlimited seats | Only 1K client-side monthly active users free. Foundation is $8.33 per 1,000 of them per month, billed yearly |
| Convert Experiences | $399 a month (Growth) | 100,000 tested users a month, $299 a month billed annually | Pro is $599 a month, or $420 billed annually. Enterprise is “Price on Request” and annual only |
| Kameleoon | From $495 a month (PBX Starter) | Up to 10 experiments and 50,000 tested visitors a month, after a 30-day free trial capped at 3 experiments | Enterprise is “Custom”. Personalization and mobile app testing are add-ons with no published price |
| Amplitude | Free plan, 2M events a month | “2M Events/month forever” on Free, and Plus “Starts at $0” | Feature Experiment is “Available for Growth and Enterprise”, both custom-priced |
| Wingify (VWO and AB Tasty) | Not published | Growth, Pro and Enterprise tiers per product line | No figures anywhere on the pricing page, “Schedule a Demo”. Priced on “your Monthly Active Users (MAUs), depending on the product” |
| Optimizely | Not published | Agentic Experimentation and Feature Management | “Every Optimizely plan is individually packaged” |
Which A/B testing tools do not charge per seat?
Most of them, and only one on this list uses seats as its main meter. GrowthBook charges $40 per seat per month on Pro, with up to three users free on Starter. Everything else bills a volume: events, flag requests, or monthly tested users. If a growing marketing team is what you are pricing for, that difference matters more than any feature list.
| What the vendor meters | Tools | The number that sets your bill |
|---|---|---|
| Seats | GrowthBook (Pro) | People with a login, at $40 each per month |
| Events | Statsig, Amplitude | Events a month, 2M free on both. Amplitude puts experimentation on Growth and Enterprise only |
| Feature flag requests | PostHog | Requests to evaluate a flag, first 1M a month free |
| Client-side monthly active users | LaunchDarkly | Unique users your front end identifies, 1K free, then $10 per 1,000 month to month or $8.33 billed yearly |
| Monthly tested users or visitors | Convert, Kameleoon | People who entered an experiment, 100,000 and 50,000 on the entry plans |
LaunchDarkly is the most explicit about it: its free Developer plan lists unlimited seats, and the Foundation plan carries no platform or seat fees. You pay for the users your front end identifies instead.
The free plans are real
The four tools at the top of that table are not stripped demos. GrowthBook’s free plan prints unlimited traffic. Statsig’s free plan names A/B tests and multivariate experiments outright, and its plan comparison marks no-code experiments as available there. LaunchDarkly’s free plan carries 100,000 experimentation users a month at no additional charge. If your team already pays for one of these for feature flags or product analytics, your experimentation budget for this year is already spent.
The thing a marketing team should check before anything else is the no-code editor, the part that lets you change a headline without filing a ticket. The vendors handle it four different ways:
- Statsig lists no-code experiments with a visual web editor on all three tiers, the free Developer plan included, so a marketing team that fits inside 2 million events a month gets the editor for nothing. Three conditions its docs name: the editing happens in a Chrome extension, the current version is in early access (“Sidecar v3 is in Early Access.”), and “You also need to be a project admin or have permissions to access console API keys to use all Sidecar functionality.”
- GrowthBook puts its AI Visual Editor and the power calculator on Pro, at $40 per seat per month. For a three-person marketing team that is $120 a month.
- PostHog has no-code web experiments through its toolbar, and its docs are specific about where they work. Small edits to text or basic formatting are “easy to manage and less likely to break anything”, while single-page applications are marked as a poor fit: “Frameworks that frequently re-render or replace DOM elements may overwrite any no-code changes you apply.”
- LaunchDarkly is built around feature flags, so the split lives in code. The 1,000 client-side monthly active users on the free plan is the limit that will actually bite on a public website, and at 100,000 the published Foundation rate of $8.33 per thousand works out to about $833 a month when you commit to a year, or $1,000 a month at the $10 rate the same page prints for paying month to month.
What is the cheapest plan for unlimited experiments?
GrowthBook’s Starter plan, which prints “Unlimited experiments” and “Unlimited traffic” at no cost for up to three users in one project. Statsig’s free Developer plan is the other real answer: unlimited flag and config checks and no cap on experiment count, metered instead at 2 million events a month.
The catch is in the unit each one leaves uncapped. GrowthBook’s limits are the three seats and the single project, and the visual editor sits behind Pro at $40 a seat. Statsig’s limit is the event volume, so an unlimited number of experiments is still bounded by how much traffic they collectively see. For a nonprofit or any team on a fixed budget, both get you to a first real test without a purchase order, and the gate is still the traffic on the page.
How much does enterprise A/B testing cost?
Nobody publishes a figure. What is knowable is the rate per thousand, and which unit it is charged on. Convert and Kameleoon meter monthly tested users or visitors, which is the same number you just read off the traffic table. LaunchDarkly meters the client-side monthly active users your front end identifies, which is a larger number than the one in the table. Here is what the published plans charge per thousand:
- Convert’s Growth plan at $399 for 100,000 tested users is about $4 per 1,000 a month, or $2.99 at the annual rate.
- Convert’s Pro plan at $599 for the same allowance is about $6 per 1,000.
- Kameleoon’s PBX Starter at $495 for 50,000 tested visitors is about $9.90 per 1,000.
- LaunchDarkly’s Foundation plan is $10 per 1,000 client-side monthly active users month to month, or $8.33 billed yearly, which lands in the same band on a different meter.
So the published market sits at roughly $3 to $10 per 1,000 users a month. The suites that publish no figure do not all meter the same thing, which matters before you compare a quote against those four lines.
| Suite | What it publishes instead of a price | What it meters |
|---|---|---|
| Optimizely | “Every Optimizely plan is individually packaged” | Not stated anywhere on the page |
| Wingify (VWO and AB Tasty) | “custom pricing tailored to your needs, based on factors like traffic volume, selected products & plans, implementation scope etc.” | “your Monthly Active Users (MAUs), depending on the product”, and monthly tracked users for its behaviour analytics |
| Amplitude Feature Experiment | “Available for Growth and Enterprise”, both “Custom” | Events |
Wingify is the row to look at twice, because it is two of the three names that used to compete for this shortlist. VWO and AB Tasty announced in January 2026 that they were joining forces, “subject to customary closing conditions”. The merger closed at some point during that year, and in September 2026 they launched as Wingify, “the brand that brings VWO and AB Tasty together as one company”.
Both former pricing pages redirect there now, so a quote that used to come from two desks comes from one.
The January announcement said “All existing client relationships, contracts, pricing, service levels, and support remain unchanged”, which was written before the deal closed and before the two brands launched as one platform. A new quote is a different matter, and it is being written by a company with one fewer competitor on the shortlist.
We went looking for a defensible figure on what the enterprise tier costs in practice and could not get one. The third-party benchmark sites disagree with each other by a factor of five on Optimizely alone, so there is nothing here we would put in front of you as a price. These are real products with capabilities the free tiers do not have, including audience targeting, server-side experimentation across a whole stack, and a support relationship with a person attached. If you are running a dozen concurrent tests across five markets, that is what you are buying.
Walk into the call with your own tested-user count for the pages you actually want to test, ask which unit they meter, and ask for the rate per thousand on that unit. Those two answers are what make a quote comparable to anything.
Is an enterprise experimentation tool worth it at $20m a year?
The decision line is the published cap, not the feature list. Kameleoon’s entry plan covers 50,000 tested visitors a month and Convert’s covers 100,000. A shop doing around $20m a year will clear both on the pages worth testing, and the moment you clear them you are in a quoted conversation whether or not you wanted enterprise features.
So run the arithmetic in that order. Add up the monthly visits on the three or four pages you would actually test, since those visitors are the tested users you will be billed for. If the total tested users sit under 100,000 a month, the self-serve plans cover you and the published rate per thousand is your ceiling. If it sits above, you are buying volume first and features second, and the questions that decide the quote are how many tested users you need and what the rate per thousand is at that volume. Audience targeting, server-side experimentation and a named support contact are worth money. The cap is what gets you into the conversation in the first place.
How do you run an A/B test on a statically generated site?
The split has to happen somewhere, and on a statically generated site there is no server rendering the page per visitor to do it. You get two places instead: in the browser after the HTML arrives, or at the CDN edge before it ships. That choice decides what your marketing team can change alone and what needs a deploy.
| Where the split happens | How it works | What the visitor sees | What it needs from you |
|---|---|---|---|
| In the browser | A script asks the vendor which version this visitor gets, and changes the page | The original version is briefly visible, which is the flicker every no-code editor works around | The vendor’s snippet, and bootstrapping if you want the flicker gone |
| At the CDN edge | A worker decides the variant and rewrites the HTML on the way out | One version, with nothing to flicker | A developer once, to put the worker in front of the site |
Both vendors publish something for this. PostHog documents bootstrapping: “Bootstrapping Feature Flags makes precomputed flag values available as soon as a client-side PostHog SDK initializes. This prevents flicker and enables startup logic, such as redirects, to use flags before the SDK finishes its first /flags request.” GrowthBook publishes edge apps for Cloudflare Workers, Fastly Compute and Lambda@Edge, and describes the Cloudflare one as a way to “cut flicker before HTML ships”.
The edge is the better answer on a static site, and it is the one that needs a developer once.
For the browser-side route a statically generated page is in a good position, for the same reason PostHog gives server-rendered apps a tick in its own no-code limitations table: “Because the page doesn’t typically re-render on the client, your changes are less likely to be overwritten.”
A statically generated page arrives as finished HTML. Nothing re-renders it out from under the edit. A React or Vue single-page application is the case PostHog marks as a poor fit.
What that leaves a marketing team is a usable division of labour:
| Change | Who can ship it |
|---|---|
| Headline, body copy, image swap, button label on a static page | Marketing, in a visual editor, once the snippet is live |
| A second version of the copy held in the CMS | Marketing, if the content model has somewhere to put it |
| Starting, stopping and reading a test | Marketing, in the vendor’s UI |
| The snippet, the SDK and the bootstrapping | A developer, once |
| An edge worker doing the split before the HTML ships | A developer, once, and it is the fix for flicker |
| Anything that changes layout structure or adds a component | A developer, and a deploy |
The one-time developer job is small and it is the thing that decides whether the rest of the list belongs to marketing. On an Astro site, where edge rendering is where the fresh-page work happens, that worker has somewhere obvious to live.
Why this got cheap
Google Optimize was the free option, and Google shut it and Optimize 360 down on September 30, 2023, pointing users at AB Tasty, Optimizely and VWO. All three of those quote by phone, so the immediate effect was a category that had a free entry point suddenly not having one. Two of the three merged during 2026 and launched as one company, Wingify, in September.
What filled the gap was a different kind of vendor. Feature-flag and product-analytics tools already had the hard parts built:
- the traffic split
- keeping a returning visitor on the version they saw the first time
- the statistics
- the event pipeline
Experimentation was a feature they could add to a plan they were already metering. They price by usage and publish the rates, because self-serve signup is how they sell. That is most of the reason a small site can now run a real test for nothing. The rest of it is that wiring a traffic split into a site stopped being weeks of specialist work, and we made the same argument about a different piece of marketing plumbing in the AI control center marketing teams need.
The one part your CMS has to handle
Your CMS needs somewhere sane to keep the second version of the copy, and any serious headless system does this already. Storyblok and DatoCMS are the two we work in most and both handle it comfortably. If the variant text ends up hard-coded in the front end instead, the test still runs, and your editors just lost the ability to change it. That is the failure mode from the invisible time trap of poor editor experience, showing up in a new place.
What to do this week
- Pull the monthly visits and the conversion rate for the single page you most want to test. Sitewide numbers will flatter you.
- Find your row in the traffic table. If the total is higher than that page gets in a month, stop, and work the list above instead. Come back when the traffic does.
- If you clear the gate, check what you already own. PostHog, Statsig, GrowthBook and LaunchDarkly all include experimentation on plans teams already pay for, and finance has probably already approved one of them.
- If nobody in-house will wire a split for you once, price the visual editor first. Statsig includes one on its free plan inside 2 million events a month, in early access and through a Chrome extension, GrowthBook charges $40 a seat, and a dedicated tool starts at $399.
- Before any sales call, write down your monthly tested users for that page, then ask the vendor which unit they meter and what the rate per thousand is at your volume. Going in without your own number means the quote gets built on someone else’s assumption.
If the wiring is the part that has been sitting undone for two years, that is exactly the work a fractional marketing engineer does. It never reaches the top of a product roadmap, which is how a team ends up paying a licence for a capability it could have had for the price of a few days of work.
Start with the traffic number for one page. It takes ten minutes and it decides everything that comes after it.