Cromanion for WordPress

How do you add an AI conversion agent to WordPress?

Install the Cromanion plugin from the WordPress directory, activate it, and paste your key under Settings → Cromanion. The tag then loads on every public page. No theme file is edited, so the install survives theme and core updates.

The manual route still exists — the same tag pasted before </body> in the theme footer — and it is the wrong one on WordPress specifically, because a theme update overwrites it silently. The plugin is a setting, and settings survive.

Does it work with Elementor, Divi or the block editor?

Yes, and the builder is not consulted. The tag is injected on the wp_footer hook, which every WordPress theme is required to call, so whatever rendered the page above it is irrelevant. No block, template, shortcode or builder widget is involved.

This is the practical reason to prefer a plugin over a builder-level script embed on WordPress. A script dropped into a page builder belongs to that builder — it travels with the template, disappears from pages built another way, and has to be re-added when the site is rebuilt on a new theme. A footer hook belongs to WordPress.

Will it slow a WordPress site down?

The tag is under 20 kB gzipped and loads asynchronously, after the page. It renders nothing until there is a reason to, and the analysis that decides that runs at the edge rather than in the browser.

The budget is enforced rather than aspirational: the build fails if the tag exceeds 20 kB gzipped, which is what keeps a size claim on a marketing page from quietly becoming false a year later.

The heavier half of the work — compiling a session's raw events into a readable narrative, deciding whether to act — happens server-side. The visitor's browser carries the collector and the renderer, not the agent.

How does it handle a WooCommerce store?

A completed WooCommerce order is reported as a conversion on the order-received page, with its total, and no goal has to be configured for it. A server-confirmed purchase is certainty, so it overrides whatever the model estimated during the session.

That ordering matters more than it sounds. An add-to-cart is a gesture and can be read as one; a completed order is a fact, and a system that lets a confident model estimate outrank a fact will eventually report a number nobody can reconcile with the store's own totals.

Everything else on a WooCommerce site is a goal like any other: a form, a booking, a signup — proposed from the crawl and counted only once it has been validated, because what counts as a conversion is the merchant's definition rather than the agent's.

How do you know it actually worked?

10% of sessions permanently receive no intervention. Comparing that control group against the rest is what makes the lift causal rather than a correlation, and the comparison is rebuildable inside Google Analytics or Plausible from the events the tag emits.

The holdout is not a trial period that ends once the numbers look good — it runs for as long as the agent does, on every plan. An agent that acts continuously needs a control that runs continuously, or it cannot separate its own effect from the traffic it happened to receive.

Each decision is written into the site's existing analytics with a holdout flag on it, so the exposed-versus-control comparison can be rebuilt natively. The proof is not a screenshot of a vendor's dashboard.