How Cromanion works

The six stages

Site Brain, Learn, Predict, Act, Measure, Learn again. The first runs once and is refreshed; the middle four run on every session; the last runs nightly.

How does Cromanion learn a website?

It crawls the site and profiles it page by page — role, offers, objections, calls to action — detects the likely conversion goals for the owner to confirm, and extracts the brand so anything it renders later belongs to the page it appears in.

This stage is why the agent can be specific rather than generic. Without a model of the site, every intervention is a template: the same nudge on a pricing page, a support article and a checkout, which is how automated tools end up offering a discount to someone trying to cancel.

Goal detection is a proposal, never a decision. The crawl finds the forms and flows that look like conversions and presents them for confirmation, because what counts as a conversion is a business definition and getting it wrong corrupts everything downstream — a newsletter unsubscribe form and a checkout are both form submissions, and only one of them is a success. A form the owner never confirmed is surfaced as a candidate, not counted.

Brand extraction is the part visitors notice by not noticing it. Palette, typography and voice come from the site itself, so an intervention reads as part of the page rather than as something bolted onto it.

What happens in Learn mode?

The agent does everything except act. It compiles each session into a narrative, predicts, records what happened, and shows the owner all of it — while showing visitors nothing.

Learn mode exists because trust runs in one direction. An agent that starts intervening on day one is asking the owner to believe it understands their site before it has seen a single visitor on it. Learn mode inverts that: the owner watches it read sessions, checks whether its reading matches what they know about their own customers, and switches it to Act when they are satisfied.

It does not start from nothing, either. Cross-site priors mean the agent already knows the general shape of intent before the first visitor lands — warm-started, but not yet acting.

Nothing about this stage is a demo. The predictions are scored against what visitors actually did next, so by the time the owner flips the switch there is a real record of how often the agent was right.

How does it read a live session — and what does the model actually see?

Raw events are compiled into a readable narrative in plain code, at no inference cost, and only that narrative is sent for a decision. The model never receives the event stream.

This is the design decision the rest of the system rests on. Compiling events into a story is ordinary code — free, fast, and running at the edge — while inference is the expensive part. Sending a compressed narrative instead of an event stream is what makes it affordable to think about every session rather than a sample.

It is also what makes the privacy claim structural rather than a policy. The narrative describes behaviour, not content: nothing a visitor typed is stored or transmitted, there are no cookies, and the visitor's IP is stripped at the edge before anything is written down.

What comes back is a fixed shape, not free text: intent, funnel stage, frustration, the predicted next move and how confident the agent is, a conversion probability, and an action. `none` is a valid action and the default one.

When does it decide to act — and what can it show?

Only when the moment justifies it, at most once per page, from a bounded catalog of formats rendered in the site's own brand. The most common decision, by a wide margin, is to do nothing.

What an intervention may say is constrained more tightly than how it may appear. Every fact shown to a visitor must trace to one of exactly two sources: the site's own pages, read into a per-page claim ledger, or something the owner stated themselves in the console. A number, policy or guarantee that neither source ever produced is dropped rather than estimated, however plausible it looks. And an instruction never to say something outranks everything, including the site's own pages.

Both rules are enforced in code, not requested in a prompt. The difference matters exactly when it is inconvenient — which is the only time a constraint is worth anything.

How is the impact measured?

Ten percent of sessions receive inference and memory but never an intervention, permanently and on every plan. The difference in conversion rate between that group and the rest is the incremental lift.

The holdout is not a trial setting or a first-month verification step. It runs forever, because the alternative is an agent whose reported results are a description of which visitors it chose to help. It selects the sessions it acts on; comparing acted-on sessions with the rest measures that selection at least as much as it measures the help.

Every decision is recorded with its assignment — control or exposed — so the comparison is available at the level of the individual decision rather than only in aggregate.

The numbers land where the owner already looks. Four events are pushed into the analytics they already run: every intervention rendered, every one a visitor engaged with, every conversion one helped along, and one event per decision carrying the control or exposed assignment. The proof lives in their tools, not only in ours.

Predictions are scored too, and only on decisive moves — converting, engaging a form, clicking a call to action, or leaving — resolved against what the visitor actually did next. Scoring an agent on its confident guesses about non-events would flatter it.

How does it get better over time?

Three layers of memory. A nightly job distils observed rules with real statistics, finished sessions are remembered with their outcomes for reasoning by analogy, and patterns that hold across enough sessions harden into rules the agent reaches for first.

The nightly playbook is site-specific and evidence-based: which moment, which format, which message actually converted here, with the numbers behind each. It is rewritten every night from the day's sessions, so a rule that stops working stops being reached for.

Episodic memory does something different. Every finished session is stored with what happened at the end of it, and a live session recalls the most similar past ones as worked examples — reasoning from real cases on this site rather than from a general prior.

The compounding is the point. A pattern that proves itself across enough sessions becomes a high-confidence rule, which makes the agent's calls more decisive and more consistent the longer it runs. An agent that is no better in month six than in week one is not learning; it is just running.