CRO agent vs CRO platform
A CRO platform is an instrument: it gives a team the means to run experiments they design, and its value is the capability it adds. A CRO agent is an operator: it finds the opportunity, chooses the response, deploys it and measures the effect itself. The simplest test is what happens when nobody logs in for a month.
Last updated
What is the difference between a CRO agent and a CRO platform?
What you are buying. A platform sells capability — you supply the hypotheses, the variations and the attention. An agent sells work — it supplies those itself and is accountable for whether conversions rose.
The last row is the one to hold on to. An agent operates within the page it is given: it can answer an objection, point someone at the right place, reduce a hesitation. It cannot decide the offer is wrong, the price is too high, or the flow has a step too many. A platform, in the hands of a team, can be used to investigate exactly those things.
| Step | CRO platform | CRO agent |
|---|---|---|
| What you buy | Capability | Work |
| Who finds the opportunity | Your team | The software |
| Who writes the variation | Your team | The software |
| Who decides who sees it | A rule configured in advance | A decision per visitor |
| Unit of change | An experiment | A session |
| Output when unattended | Nothing | The same as when attended |
| Where expertise accumulates | In the people using it | In the system |
| What it cannot do | Run without you | Change what the page fundamentally offers |
How can you tell which one a product actually is?
Ask what happens with no human input for a month. A platform produces nothing — correctly, that is what it is for. An agent produces the same output it produced in week one, plus whatever it learned since.
Does it decide, or suggest?
A recommendation waiting for approval leaves the decision, the latitude and the accountability with the person. That is a platform with a generator attached, whatever it is called.
Does it deploy, or hand you a variation?
If a change still needs someone to publish it, the cycle time is a human's, and the per-visitor granularity is lost — the decision arrives after the session that prompted it has ended.
Does it decide per visitor, or per segment?
Segment rules configured in advance are a platform feature and a useful one. An agent's unit is the session, which is why it can treat two people on the same page differently.
Can it say what it caused?
An agent that acts continuously needs a control group that runs continuously. If the only evidence is before-and-after, or treated-versus-untreated with no random assignment, it cannot separate its effect from its own selection — and that question sorts real agents from ambitious ones just as sharply as the first three sort agents from platforms.
When is a CRO platform the right choice?
When you have people to run it and questions worth testing. A team with a dedicated CRO function and a backlog of real hypotheses gets more from a platform than from an agent, because control is precisely what they need.
Structural questions belong here: pricing, positioning, information architecture, whether to remove a checkout step. They are expensive, occasional, contested, and each deserves an experiment designed around it — with a number at the end that survives an argument.
Regulated surfaces belong here too. Where every variant must be reviewed before a visitor sees it, autonomy is the wrong property to want, and a platform's approval step is the feature rather than the friction.
When is a CRO agent the right choice?
When the honest answer to 'when did you last run a test?' is months ago. An agent is for the far more common case where the backlog exists, everyone agrees it matters, and no one has the hours.
It also fits sites below the traffic floor a testing programme needs. If a variation would take three months to reach significance, the programme was never going to work, and per-visitor decisions do not have that threshold — though the aggregate lift number still needs volume like any measurement.
And it fits the case a population-level answer cannot handle: two visitors on the same page who need opposite things. A platform must choose one answer for both; an agent never treated them as one group.
Can you run both?
Yes — they operate at different altitudes and interfere less than expected. Tests decide what the page fundamentally is; an agent works within whatever the page currently is. Keep the two random assignments independent and both measurements stay clean.
The practical rule is that an agent's holdout must not be correlated with any test's variant split. Assigned at random and independently, they are uncorrelated by construction, and each measures its own effect. The mistake is letting one system's assignment depend on the other's — at which point neither number means what it says.
Cromanion is an agent, by that test
It decides rather than suggests, deploys into the live session, decides per visitor rather than per segment, and measures itself against a permanent 10% holdout — the fourth question, which is the one that separates a real agent from an ambitious one.
There is no dashboard to visit for it to work. A crawl builds the site model, the agent watches sessions in Learn mode, and once the owner switches it to Act it decides and renders on its own — at most one intervention per page, from a bounded catalog, in the site's own brand. Nothing waits for approval, because a decision that waits arrives after the visitor has gone.
What is deliberately not autonomous: what counts as a conversion, which facts may be stated, and what must never be said. Those are the owner's, and they are enforced in code rather than requested in a prompt — the difference shows up exactly when it is inconvenient, which is the only time a constraint matters.
And the accountability is not optional. Ten percent of sessions receive 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 the owner already runs.
Common questions
Is a CRO agent just a personalization engine?
No. A personalization engine applies rules based on who a visitor is — segment, source, history — and is usually judged on relevance. An agent decides from what the visitor is doing now and is judged on whether conversions rose against a control group.
Do I lose control with an agent?
You lose per-decision control and keep boundary control: what counts as a conversion, what may be said, which surfaces are enabled, and whether it acts at all. Per-decision approval is the thing being traded away, because a decision that waits for it misses the session it was for.
Can a CRO platform add an agent later?
Several are moving that way, and the interesting question is not whether the feature exists but whether the loop closes: does it decide, deploy and measure without a person, and can it separate its effect from its own selection.
What happens if the agent is wrong about a visitor?
Usually nothing visible — the default decision is to do nothing, so being unsure resolves as silence rather than as a wrong intervention. When it does act and the intervention does not help, the holdout comparison shows it in aggregate.
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.