What Is AI CRO? How AI Conversion Optimization Works in 2026

Updated 9 November 2026

What is AI CRO?

AI CRO (AI conversion rate optimization) is the use of machine learning and large language models to find, prioritise and test changes that get more website visitors to take a desired action, such as buying, signing up or booking a call. The AI speeds up parts of the work. People still decide what to change, and experiments still prove what worked.

In practice, "AI CRO" covers four different jobs: reviewing a page for conversion problems, summarising visitor behaviour, drafting test ideas and variations, and shifting traffic towards the version that is winning. Each one uses AI differently and has different limits, so it helps to know which one a tool means before you trust its output.

Key Takeaways

• AI CRO uses AI to speed up conversion work: reviewing pages, summarising behaviour data, drafting test variations and allocating traffic. It does not replace measuring what changed.
• It works in four steps: collect inputs, find patterns, suggest changes, then validate them with data or a test.
• AI page reviews are fast but miss a lot. In Baymard Institute's 2023 test, GPT-4 found 14% of the UX issues human experts found on 12 live pages, and 80% of its suggestions were false positives. Tests reported in 2025 put AI tools at 50% to 75% accuracy.
• AI does not replace A/B testing. Bandit algorithms shift traffic towards a winner, but GrowthBook's documentation recommends standard experiments when you need accurate effects and want to learn about customer behaviour.
• Use AI as one quick evaluator: check each finding on the live page, fix what is clearly broken, and test what is uncertain.

Jump to: how it works · the four kinds of AI CRO tool · what AI can and can't do · can AI replace A/B testing? · what AI CRO tools can't detect · how to use AI CRO well · AI CRO vs AI search.

How does AI CRO work?

AI CRO follows the same loop as ordinary conversion rate optimisation. The difference is that a model does the reading, sorting or allocating at one or more of these four steps.

1. Collect inputs. The tool takes in the page itself (its HTML, a screenshot or the rendered page), behaviour data (clicks, scrolls, session recordings, funnel steps) or past test results. What it can see sets the ceiling on what it can find.
2. Find patterns. A language model compares the page against usability and persuasion conventions, a summarisation model condenses hundreds of sessions into themes, or a statistical model estimates which variation is ahead.
3. Suggest a change. The output is a ranked list of findings, a plain-language summary, draft copy or variations for a test, or a new traffic split.
4. Validate and learn. A person checks each suggestion against the live page and the analytics. Clear problems are fixed directly; uncertain ideas go into an A/B test, and the result becomes the next round's input.

Step 4 is where most of the value is decided. Steps 1 to 3 produce hypotheses quickly. Only step 4 tells you which of them were true.

The four kinds of AI CRO tool

Most products sold as "AI CRO" do one of four jobs. The table names a documented example of each, checked on 9 November 2026. It is not a ranking; our best AI CRO tools page compares products in more depth.

AI page reviews

An AI page review applies a checklist the way a usability expert would, but in seconds. It is a fast version of a heuristic evaluation, the method Nielsen Norman Group describes as evaluators judging a design "against a set of guidelines (called heuristics)". NN/g recommends three to five independent evaluators, because each one "is likely to miss some of the potential usability issues". An AI review is best treated as one more evaluator, not the whole panel. Our AI CRO audit page explains how one works on a single URL.

Behaviour summaries

Microsoft's Copilot overview for Clarity says Session Insights help you understand the user journey "without requiring you to watch time-consuming recordings". It also warns that "Generative AI models can sometimes misinterpret or generate incorrect information", and the Session Insights page notes that insights are not available for sessions under five seconds. Summaries save viewing time; watch the recordings behind any summary you plan to act on.

Test ideas and variations

Optimizely's documentation for Mark, its agent platform, lists AI features for Web Experimentation that get test ideas, summarise experiment results and generate copy suggestions for variations in the visual editor. Its Variation Generator agent does the same for feature flags, turning a plain-language hypothesis into suggested variations. Tools like this shorten the time from hypothesis to a test that is ready to run. They do not tell you which hypothesis is worth testing; that still comes from your analytics and research.

Adaptive traffic allocation

A bandit is an experiment that changes its own traffic split. GrowthBook's documentation explains that Thompson sampling "allocates traffic proportionally to the probability that an arm is best", while still sending some visitors to weaker variations so it keeps exploring. GrowthBook keeps at least 1% of traffic on each variation in case behaviour changes over time.

What AI can and can't do in CRO

AI is good at anything that can be judged from the page or summarised from data you already collect. It is weak at anything that needs knowledge it cannot see: your customers' motives, your business constraints and proof of cause.

Can AI replace A/B testing?

No. AI can generate test ideas and run adaptive tests, but it cannot replace a controlled comparison when you need to know how much a change helped. A suggestion from a model is a hypothesis, however confident it sounds.

Bandits are the closest AI comes to "automatic" testing, and their own documentation sets limits on them. GrowthBook's comparison gives standard experiments the goal of "obtaining accurate effects and learning about customer behavior", and bandits the goal of "reducing cost of experimentation when learning is less important than just shipping the best version". It suggests bandits for five or more variations and standard experiments for two to four, and notes that bandits do not support multiple goal metrics.

The same page gives a practical warning. An urgency message might make people buy sooner without making more of them buy. That variation looks better early, so the bandit sends it more traffic "even though its final conversion rate is no higher". GrowthBook's remedy is to update the traffic split no faster than people take to convert. A standard A/B test avoids the problem by keeping the split fixed and running long enough to cover that conversion window. Our guide to landing page testing covers how to size one.

So the useful split is: let AI produce and pre-screen ideas, use a bandit for short-lived choices such as a seasonal banner, and use standard A/B tests for decisions you want to learn from.

What can't AI CRO tools detect?

AI CRO tools can't detect what isn't in their input. A page review can't see your analytics, your customers' reasons or the pages behind a login, and a screenshot-based review misses most interaction problems. They also can't tell you whether a fix will move your conversion rate before you measure it.

How accurate are AI page reviews?

The most detailed public test is Baymard Institute's. In Testing ChatGPT-4 for 'UX Audits' (published 18 October 2023), Baymard gave GPT-4 screenshots of 12 ecommerce pages, asked what UX improvements could be made, and compared its 178 suggestions with 257 issues found by six human UX professionals who spent 2 to 10 hours on each page.

Most of the wrong suggestions were harmless but wasted time. About one in eight were "likely harmful", such as suggesting infinite scrolling for a site that uses pagination. GPT-4 also suggested making every page "mobile responsive" despite being given desktop screenshots.

The study is from 2023 and tested one model on screenshots only. Baymard notes that other prompts "gave largely the same responses", and it expected that navigating the live site would lift discovery "from 14% close to 26%".

Newer tests are better but still short of human reviewers. In a 2025 follow-up article (updated 28 May 2026), Baymard reports that two Microsoft UX researchers tested AI tools for heuristic evaluation in March 2025 and found accuracy rates of 50%, 62%, 67% and 75%. The 75% tool only got there by reporting far fewer issues: it missed 13 of the 16 that the human expert found on the same page.

Keep in mind that Baymard sells human UX research and its own AI evaluation tool, UX-Ray, for which it claims a 95% accuracy rate. That is a vendor claim we have not checked. The safe working assumption is that an AI review finds some real issues, misses many, and includes suggestions you should reject.

What AI page reviews can't see

• Your numbers. Without traffic and revenue data, a tool ranks findings by how bad they look, not by how many visitors they cost you.
• Motivation. Why a visitor hesitated is a research question. NN/g's heuristic evaluation guide says the method "cannot replace user research", and the same applies to AI reviews.
• Gated steps. Logged-in areas, carts with items in them and checkout steps that need payment details are usually out of reach.
• Cause. A review can say a call to action is weak. Only a test can say that changing it helps.

Our post on where AI falls short in CRO covers these limits in more detail, with a table of which questions to give to AI, analytics, experiments and people.

How to use AI CRO without being misled

Use AI where it is fast and checkable, and keep people and experiments where judgement and proof are needed.

1. Start with exposure. List the pages that carry the most traffic and revenue before you review anything. An AI finding on a page few people see is low priority, however severe it looks.
2. Run an AI review as your first evaluator. Treat the output as a list of hypotheses, not a verdict.
3. Check every finding on the live page. Reject anything that is already fixed, generic or about a page the tool did not see. Baymard's results suggest that many suggestions will need rejecting.
4. Fix what is clearly broken. Errors, missing calls to action and confusing forms don't need a test.
5. Watch the evidence behind summaries. If an AI summary of recordings points at a problem, watch a few of the sessions it is based on.
6. Test the uncertain ideas. Use an A/B test with a fixed sample size for changes you want to learn from; consider a bandit only for short-lived choices with many variations.

If you want to see how AI handles your own page, can AI improve my landing page? walks through what to expect from a single-page review.

Is AI CRO the same as optimising for AI search?

No. AI CRO uses AI to improve how many visitors convert. Optimising for AI search is about being shown or cited in AI answers such as Google's AI Overviews. The two meet when visitors arrive from those answers.

Google's guidance on AI features and your website suggests tracking conversions and time on site for that traffic, and says that when people click from search results pages with AI Overviews, "these clicks are higher quality (meaning, users are more likely to spend more time on the site)". That is Google's own observation, not an independent measurement, but it is a reason to check how AI-referred visitors convert on your landing pages, not just how many arrive.

Try it on one page

The quickest way to judge AI CRO is to run it on a page you already know well and compare its findings with what you know. Run a Croopt audit on your page, then check each finding against the live page and your analytics before you change anything.

Sources and author

• Baymard Institute, Testing ChatGPT-4 for 'UX Audits' Shows an 80% Error Rate & 14–26% Discoverability Rate (published 18 October 2023), retrieved 9 November 2026. Used for the 2023 GPT-4 figures.
• Baymard Institute, AI Heuristic UX Evaluations with a 95% Accuracy Rate (published 7 October 2025, updated 28 May 2026), retrieved 9 November 2026. Used for the March 2025 Microsoft researchers' results, as reported by Baymard, and Baymard's claim for UX-Ray.
• Nielsen Norman Group, How to Conduct a Heuristic Evaluation (Kate Moran and Kelley Gordon, 25 June 2023), retrieved 9 November 2026. Used for the definition, the three-to-five evaluators advice and the limits of the method.
• GrowthBook, Bandits documentation, retrieved 9 November 2026. Used for Thompson sampling, the 1% minimum, the bandit vs experiment comparison and the urgency example.
• Microsoft Learn, Copilot overview and Session Insights for Microsoft Clarity, retrieved 9 November 2026.
• Optimizely, Variation Generator agent and Optimizely Mark and AI features, retrieved 9 November 2026.
• Google Search Central, AI features and your website (last updated 10 December 2025), retrieved 9 November 2026.

Kieron Bertani is the founder of Croopt, an AI CRO audit tool of the kind described in this article, so we have an interest in the topic; no vendor named here paid to be included. He writes about conversion rate optimisation. This article and its hero illustration were created with AI assistance. The four-step model and the tool categories are editorial judgement based on the sources above, not results from Croopt's own research. For corrections or product questions, contact Croopt support.