Coframe is an AI service that writes and builds variants of your web pages (copy, layout code and images), then lets a multi-armed bandit shift traffic toward the winners after you approve them.
Coframe, Evolv AI, Webflow Optimize, Wingify's Wingz (formerly VWO Copilot) and Kameleoon's prompt-based testing can all make variants faster than a person can.
None of them replaces the research that decides what is worth testing, and the big lifts on their sites are vendor case studies, not independent results.
This roundup explains what each tool actually automates, what it costs where prices are public, and when a person with a standard testing tool is the better bet.
We have not used these tools ourselves; everything below comes from vendor sites, docs and the tests in our database.
Disclosure: we may earn a commission if you buy VWO or Convert through our links. We have no deal with Coframe, Evolv, Webflow or Kameleoon.
What "AI A/B testing" means: variant generation, auto-allocation and bandits
"AI A/B testing" covers three different jobs, and most tools only do one or two of them well.
- Variant generation. A language model writes new headlines, CTAs or whole sections, or builds a layout change from a prompt. This replaces the copywriter or developer, not the tester.
- Auto-allocation. Instead of a fixed 50/50 split, the tool sends more traffic to whichever version is converting better as the test runs.
- Bandits and personalization. A multi-armed bandit is the math behind auto-allocation. Contextual bandits go one step further and pick a different winner per segment (device, source, new vs. returning).
The difference matters. A classic A/B test tells you how much better B is than A, with a known error rate (see statistical significance).
A bandit maximizes conversions during the test but gives a fuzzier answer about why something won. You trade learning for earning.
The test types these tools automate best are the ones our database is full of: copy and headline changes.
Dave Diederen changed a product title to an "Achieve [outcome] with [product]" format and got +16.92% revenue per visitor.
Source Patrol added one word, "supplement," to a landing page headline and saw +89.97% purchases.
A model can draft dozens of variants like these in seconds.
Picking the right angle is still the hard part.
@DaveDiederen
Optimizely@DaveDiederen@DaveDiederen
Clean Commit| Tool | What the AI does | How traffic is split | Pricing (as of November 2026) | Best for | Not for |
|---|---|---|---|---|---|
| Coframe | Generates copy, code and images; you approve | Multi-armed bandit | Not published; demo only | High-traffic landing pages, teams without designers | Low traffic, teams that want to own the test design |
| Evolv AI | Recommends and builds ideas, tests many combinations | Bandits plus Bayesian models ("Active Learning") | No public pricing found | Enterprise funnels with lots of traffic | Small stores, simple A/B questions |
| Webflow Optimize | "AI Optimize" picks the best variant per visitor | A/B or AI auto-allocation | From $299/month, priced by page views | Webflow marketing sites, B2B personalization | Shopify stores, teams that need deep stats |
| Wingify (VWO) Wingz | Prompt-to-variation, AI agents for analysis and ideas | A/B, MVT or bandit | Custom quote, free trial | Teams already on VWO or AB Tasty | Buyers who want a list price |
| Kameleoon PBX | Builds experiments from a chat prompt | A/B, MVT; bandits as paid add-ons | Starter $495/month; free trial | Teams that want AI build speed with standard stats | Sites under a few thousand weekly visitors |
Coframe
Coframe sells itself as a "managed AI platform" rather than software you run.
It generates variants across code (UI and layouts), copy (headlines, CTAs, full pages) and on-brand images.
AI content has to be approved before it goes live, and the company says it is SOC 2 Type II certified.
Testing runs on multi-armed bandits, which Coframe explains in its own post on A/B testing vs. MABs: better versions get more traffic, and some visitors still see new or less-tested variants.
You can add your own human-written variants alongside the AI ones.
The homepage claims "10x+" more experiments at "10%" of the cost per experiment, a 410% conversion lift for Replit and +59% for StartEngine.
Those are Coframe's numbers, with no sample sizes or test lengths shown on the page.
Pricing: Coframe does not publish prices. The pricing URL returned a 404 when we checked, and the site points to a demo.
Third-party sites list tiers, but we could not confirm them on Coframe's site, so we leave them out.
Downsides: it is a managed service, so you depend on their team and their judgment about what to test.
Bandits need steady traffic to settle, which rules out most small stores.
Evolv AI
Evolv AI calls itself an "AI-led experience optimization platform" that recommends, builds and optimizes test ideas.
Its technical overview describes an "Active Learning" engine that uses multi-armed bandits for traffic allocation and Bayesian models to judge results, rather than one fixed A/B comparison.
The appeal is testing many element combinations at once and letting the system keep evolving the page.
The cost is interpretability: when dozens of changes are mixed, it is harder to say which one caused the lift, and that is the lesson you would carry to the next page.
Pricing: we found no public pricing. Evolv targets enterprise buyers, so expect a sales process.
Not for: stores that need a clear answer to one question, such as "does a sticky add-to-cart help?" A standard tool answers that more cleanly.
Webflow Optimize (formerly Intellimize) and other ways to test on Webflow
Webflow bought Intellimize, an AI personalization and CRO platform, and folded it into Webflow Optimize.
If you searched for Intellimize, this is where it lives now.
Optimize combines three modes: classic A/B testing, rules-based personalization, and AI optimization that "learns what content works best for each visitor" and allocates traffic automatically.
It segments by device, location, traffic source, UTM and behavior.
Enterprise plans add CRM and account-based tools such as Salesforce, HubSpot, 6sense and Demandbase.
Pricing: as of November 2026, the Optimize add-on starts at $299/month, based on page views, with up to 5 concurrent optimizations.
The Enterprise version (1M+ page views a year, unlimited optimizations) is "contact us."
Optimize also works on non-Webflow sites through a JavaScript snippet; that pricing is sales-only.
Other ways to A/B test on Webflow: any snippet-based testing tool can run on a Webflow site, including Convert, VWO and Kameleoon.
That is usually cheaper if you only need straightforward split tests, and it gives you standard reports.
Intellimize's own case studies (via Webflow's announcement) claim a 53% conversion lift for Sumo Logic and 60% more conversions for Snowflake's ABM team. Treat those as vendor examples.
VWO Copilot (now Wingz) and Kameleoon AI
Wingify Wingz (VWO Copilot)
VWO and AB Tasty announced on January 20, 2026 that they were joining forces under Wingify, subject to closing conditions, as a combined company with more than $100M in revenue and over 4,000 customers.
VWO Copilot now redirects to Wingz, the AI layer across the platform.
Wingz turns a prompt, a Figma file or a screenshot into a variation, and its agents read funnels, replays and surveys to propose ranked test ideas.
Wingify says nothing goes live without approval, changes are checked on "synthetic visitors" first, and results run against a permanent holdout.
Autonomy is "a dial," which means a person can stay in charge.
The main difference from Coframe: Wingz sits on a full testing platform with A/B, multivariate and bandit campaigns, so you can choose a classic test when you want a clean answer.
Pricing is a custom proposal with no fixed plans, plus a free trial.
See our VWO review for the testing side.
Before the merger, AB Tasty reported an AI product recommendations rollout at Bricomarché that it credits with +21% homepage conversion.
That was a before-after comparison, not a controlled split, which is typical of AI case studies.
Kameleoon prompt-based experimentation
Kameleoon's Prompt-Based Experimentation (PBX) builds a test from a chat prompt: you describe the change, the audience and the metric, and it creates the variant, targeting and traffic split.
An Ideate feature scans a page and suggests prompts.
Kameleoon frames it as human-led: a person decides what to test, the AI builds it.
Pricing: as of November 2026, Kameleoon's plans page lists a free 30-day trial (up to 3 experiments) and a Starter plan at $495/month for up to 10 experiments and 50,000 tested visitors a month, with PBX included.
Each prompt uses a credit. Contextual and multi-armed bandits are add-ons, and Enterprise is unlimited. More in our Kameleoon review.
What the claims rest on
Every big number on these vendors' sites is a case study the vendor chose to publish.
That does not make them false, but it does mean you are seeing the winners.
Our own archive shows the same bias: of 1,724 results, 1,447 are marked as wins, because people share wins.
Our A/B test win rate research and losing tests pages show what the full picture looks like.
Three things to check before trusting an AI lift claim:
- Was there a control? Bandits and before-after rollouts don't give the same evidence as a fixed split. Optimizely's greeting card example reports +24% paid card sales from a bundle of pricing and add-on changes, so no single change can take the credit.
- How many variants were tried? Generate 40 headlines and one will look great by chance. More variants need more traffic to tell luck from lift.
- Did it hold up? Clean Commit's homepage tailored to new vs. returning visitors showed +241% conversion rate, marked unproven. Its separate test of personalized product pages for returning visitors came in at +1.0%, flat.
Copy changes also lose. Clean Commit's product page redesign with new copy dropped conversion 28.3%, and Dave Diederen found that plain "Add to Cart" beat every creative button label tested.
An AI that writes more variants writes more losers too, so you need a testing setup that can catch them.
When to use an AI testing tool (and when a person wins)
An AI tool makes sense when variant production is your bottleneck, not ideas or traffic:
- You have high, steady traffic on many similar pages (landing pages, long-tail product or location pages).
- You'd rather earn during the test than get a clean read, for example on a short promotion.
- You have no designer or developer to build variants, and you're comfortable approving AI output.
A person with a standard tool such as Convert or VWO is the better choice when:
- Your traffic is low. Bandits and multi-variant generation need volume; see A/B testing with low traffic.
- You need to know why something won, such as testing a pricing or offer change you'll roll out sitewide.
- Your copy is regulated (health, finance) and every claim needs review anyway.
Either way, the inputs still come from people.
The strongest tests in our database started from customer insight.
Oliver Kenyon's test of a pain-relief vs. longevity angle on dog supplement advertorials lifted CTR 28%, and that angle came from knowing the buyer, not from a prompt.
Do customer research first, then let AI speed up the build.
For the wider market, see our A/B testing tools roundup, website personalization tools and AI landing page builders.