What a CRO Agent is, and what it isn't

"AI-powered" has been stuck on so many CRO tools that it no longer means much. Usually it means a language model writing headline variations, which is a copywriting feature under another name.

A CRO Agent claims something different. It's software that does the analytical work a conversion specialist does (watching the data, spotting patterns, forming hypotheses, setting up the test) and stops at the point where someone needs to use judgement about your business.

Strip the job of a CRO professional down to its structure and you get a loop: observe, detect, propose, build, target, run, measure, learn, and then round again knowing more than last time.

Most of those steps are mechanical. Observing traffic patterns means reading data consistently. Detecting divergence means comparing segments against a baseline. Building a variant is writing selectors and conditions, and targeting is turning a description of a segment into audience rules. Running and measuring are engine work.

The steps that aren't mechanical are the ones that touch your business. Is a pattern the agent found real, or a side effect of how you buy media? Does the proposed change fit your brand? Should the result change your strategy? An agent that handles the first group of steps and hands the second group back to you is useful. One that pretends it can do the second group is dangerous.

This is the shape of what the agent brings you:

Detected. 71% of traffic to this product page is mobile. New visitors arriving from Meta have an add-to-cart rate 23% below the page average.

Hypothesis. The primary call to action sits below the first viewport on mobile; visitors from paid social arrive with lower intent and leave before reaching it.

Proposed change. Persistent add-to-cart on mobile.

Audience. Spain · mobile · new visitors · source = Meta.

Traffic split. 50/50.

Primary metric. Purchase. Secondary. Add-to-cart. Guardrail. Revenue per session.

Estimated duration. ~12 days.

[ Review blueprint ] [ Launch experiment ]

Each of those fields is a decision someone would otherwise have made by hand, and you can edit every one of them before anything runs.

A company with a CRO team doesn't need the analysis automated so much as sped up. A company without one needs the analysis to happen at all.

The gap is specific. Small brands rarely lack ideas about their website. What they lack is the time to go through traffic data systematically, the statistics to size an experiment, and the engineering capacity to build the variant, and those three gaps line up with the mechanical steps of the loop. So the point isn't to be a cheaper Optimizely; it's to put the work of a CRO team within reach of a company that will never hire one.

It doesn't assign visitors to variants. A deterministic engine does that, and the model has no say in who sees what.

It doesn't decide significance either. The statistics are fixed, testable code with a declared sample size, and nobody can talk an agent into treating a result as ready.

It doesn't publish without approval. Every proposal waits for a person to approve it. That's the design, not a temporary limitation, and the reasoning is in our piece on governance.

And it doesn't know your business. It can see that paid traffic converts worse, but it can't know that you deliberately buy cheap top-of-funnel traffic and expect exactly that number.

An agent that proposes experiments is only as good as the data underneath it. On a site with 200 sessions a day and no custom events there's little divergence to find, and most of what looks like a pattern is noise.

The first useful thing an agent can tell a small brand is therefore often unwelcome: there isn't enough signal here yet; instrument this event and come back. If a tool always has a confident proposal ready whatever the data looks like, it's generating text, and you should treat its proposals that way.