What is a CRO agent?
A CRO agent is software that runs the conversion optimization loop by itself: it learns a website, watches visitor behaviour, identifies where conversions are being lost, decides what response fits the moment, deploys it, measures the incremental effect, and feeds the result back into its next decision. A CRO platform gives a person the means to do those things; an agent does them.
Last updated
What is a CRO agent?
A CRO agent is software that performs conversion optimization rather than assisting with it. The test is simple: if a human still has to decide what to change and when, it is a tool; if the software decides and is held to account for the outcome, it is an agent.
The word agent is doing real work here, and it is worth separating from the marketing use of it. An agent has a goal it is accountable for, some latitude in how it pursues it, and a feedback signal that tells it whether it is succeeding. Software that produces suggestions for a person to approve has none of those things — the goal, the latitude and the accountability all still belong to the person.
In conversion optimization the goal is unusually well defined, which is what makes the category tractable. There is a conversion, it either happened or it did not, and the agent's contribution to it can be measured. Very few domains give an autonomous system a scoreboard that honest.
What does a CRO agent do?
Seven things, in a loop: understand the website, understand the visitor, identify the opportunity, decide the intervention, deploy it, measure the incremental impact, and learn from the result.
Understand the website
What each page is for, what it offers, what objections it raises, where the conversion happens, and what the brand sounds like. An agent without a model of the site can only apply generic tactics, which is how it ends up showing a discount banner on a support page.
Understand visitor behaviour
Not events in a table — a sequence with a shape. Someone who read the pricing page twice, opened the FAQ and returned to the plan selector is telling a story, and the story is what an intervention has to respond to.
Identify the conversion opportunity
Most sessions contain none: the visitor is browsing, or has already decided, or was never going to convert. The agent's job at this step is mostly to rule sessions out, and an agent that finds an opportunity in every session has not understood the step.
Decide what intervention is appropriate
Which of a bounded set of formats, saying what, at which moment. Doing nothing is one of the options and should usually win. The wrong intervention at the right moment is worse than nothing, because it also spends the visitor's patience.
Deploy the intervention
Rendered into the live page, in the site's own styling, without a release cycle and without anyone approving it individually. This is where an agent separates from a recommendation engine — the decision reaches the visitor it was made for, while it is still their session.
Measure the incremental impact
Against a control group that never sees an intervention, held back permanently. Without it the agent grades its own homework: it chooses which visitors to help, so comparing helped with unhelped measures its selection, not its effect.
Learn from the result
Which moment, which format, which message, on which page, actually converted — accumulated as evidence that changes future decisions. An agent that does not close this loop is a rules engine someone else has to keep tuning.
How is a CRO agent different from a CRO platform?
A CRO platform is an instrument: it gives a team the means to run experiments they design. A CRO agent is an operator: it designs, runs and evaluates them itself. The platform's output is capability; the agent's output is conversions.
Neither is better in the abstract, and a team with a dedicated CRO function and a testing backlog gets real value from a platform — the control is the point. The agent is for the far more common situation: the backlog exists, everyone agrees it matters, and no one has run a test since spring.
| Step | CRO platform | CRO agent |
|---|---|---|
| What it gives you | The means to run tests | The outcome of running them |
| Who finds the opportunity | Your team | The software |
| Who writes the variation | Your team | The software |
| Who decides who sees it | A rule set up in advance | A decision per visitor |
| What happens when nobody has time | Nothing runs | It keeps running |
| Where the expertise lives | In the people using it | In the system, accumulating |
| What you are buying | Capability | Work |
What makes a CRO agent autonomous — and what should never be?
Autonomy means the loop runs without a person in it. It does not mean the agent decides what counts as success, what it is allowed to claim, or where its limits are — those stay with the site owner, and an agent that assumes them is not more autonomous, just less accountable.
Autonomous: the loop
Observing, predicting, deciding, deploying, measuring, learning. Requiring approval per visitor would remove the only advantage the approach has.
Not autonomous: the definition of success
Whether a form submission is a conversion is a business question. A newsletter unsubscribe and a checkout are both form submissions; an agent that decides for itself will eventually count the wrong one and report a lift that is really a leak.
Not autonomous: the facts
An agent may repeat what the site says, or what its owner told it. Anything else — a stock level, a delivery date, a review, a count of people viewing — is invention, and the fact that it is plausible is what makes it dangerous rather than what makes it acceptable.
Not autonomous: the boundaries
What must never be said, what must never be shown, which surfaces are off. These are policy, they belong to the owner, and enforcing them in code rather than in a prompt is the difference between a constraint and a suggestion.
Cromanion: an autonomous CRO agent
Cromanion is an autonomous CRO agent installed with one script tag. It runs the seven steps as Site Brain, Learn, Act and Measure — and the measuring is permanent, not a trial period, because a permanent 10% holdout is what keeps the other three honest.
The design decision underneath all of it is what reaches the model. Raw visitor events never do: they are compiled into a readable narrative in plain code, at no inference cost, and only that narrative is sent for a decision. It is what makes thinking about every session affordable, and it is also why nothing a visitor typed is stored or transmitted.
And the boundaries above are enforced where they cannot be talked around. A claim traces to the site's own pages or to something the owner stated; an instruction never to say something outranks everything, including the site's own pages. Both are code, not prompt text.
Site Brain
A crawl builds the model of the site: every page's role, its offers, objections and calls to action, the conversion goals for the owner to confirm, and the brand — palette, typography, voice — so anything rendered later belongs to the page it appears in. Steps 1 of the loop, done once and refreshed.
Learn
The agent watches real sessions and shows visitors nothing. It narrates, predicts, and banks what it learns while earning the owner's trust before it spends a visitor's attention. It does not start from zero either: cross-site priors mean it already knows the shape of intent before the first visitor lands. Steps 2 and 3.
Act
Switched on by the owner, never by us. Doing nothing remains the default; at most one intervention per page, drawn from a bounded catalog — inline hint, injected block, contextual card, sticky bar, mobile bottom sheet, desktop exit modal — rendered in the site's own brand. Steps 4 and 5.
Measure
Ten percent of sessions get inference and memory but never an intervention, permanently, on every plan. Every decision is recorded with its control or exposed assignment, and the resulting lift is written into the analytics tools the owner already uses rather than kept in a dashboard of ours. Step 6.
Learn again
A nightly job distils the day's sessions into observed rules with real statistics — which moment, which lever, which page actually converted on this site. Finished sessions are remembered with their outcomes, so a live session can be reasoned about by analogy with the ones most like it. Step 7, closing the loop.
Common questions
Is a CRO agent the same as an AI CRO tool?
Not necessarily. Many tools described as AI CRO add a generative feature to a workflow a person still drives. An agent runs the loop itself and is accountable for the outcome. The distinguishing question is whether anything happens when nobody logs in.
What is a holdout group, and why does a CRO agent need one?
A holdout is a randomly assigned share of visitors who never receive an intervention. It is necessary because the agent chooses whom to help: comparing helped visitors with unhelped ones measures that choice rather than the help. Only a group assigned before the decision isolates the effect.
Does a CRO agent replace a CRO team?
It replaces the part of the work that never gets done — the hundred small decisions per day nobody has time to make. Strategy, positioning, pricing and what the site is fundamentally offering stay firmly human, and an agent works better on a site where those are handled well.
How quickly does a CRO agent start working?
It can watch and learn from the first session. Acting usefully depends on having a site model and enough observed behaviour to tell moments apart; the causal lift number needs enough sessions on both sides of the holdout to be worth reading, which is a matter of traffic rather than of time.
What stops a CRO agent from saying something untrue?
Architecture, if anything does. A system that generates copy freely and is asked in its prompt to be truthful will eventually assert a delivery date or a stock level nobody gave it. The constraint has to be a check the claim passes or fails, not an instruction it is encouraged to follow.
See it on your own site
Paste one tag. Cromanion crawls your site, watches real sessions in Learn mode, and only acts when you switch it on — with a permanent 10% holdout proving what it caused. Free to start, no credit card: a 14-day or 1,000-session live trial, then keep measuring for free.