Example · Case study
Cartpilot converted well for humans, but AI agents could not compare variants, service levels, and bundle rules. We rebuilt the product evidence, clarified checkout, and left a repeatable agent-readiness test.
Cartpilot supplies industrial and workshop customers with components and consumables — 12,400 SKUs, most with two to seven variants. Human conversion was steady at 4.7%.
In early 2026, AI agents started crawling the catalog: GPT Operator and Perplexity Comet sessions showed up in the logs. They converted at 0.4%. The agents found Cartpilot and bought from competitors.
We reconstructed 47 agent sessions from the last 60 days. Three breaks explained most of the failures:
Marketing named a 'Pro' version, but nothing told the agent which SKU that was or how it differed on voltage, fit, or capacity. Agents picked on shipping speed — often the wrong variant.
Service level and bundle constraints only resolved at the cart. Headless agents bailed after 30 seconds of inactivity, before they ever saw the options.
What combined with what lived in spec sheets the crawler couldn't cite. Agents built invalid bundles and read it as a fault of the shop.
"We didn't have a conversion problem. We had an agent-trust problem. In 30 days it was gone."
Four workstreams in parallel over 30 days. Each with a clear owner and a pass criterion before going live.
We replayed the real buying task — 'find a 110V express-shippable drill compatible with the existing battery system' — and marked every point where agents stalled. The result was a 12-point heatmap per top category, weighted by revenue risk.
Tactic
GPT-4 Action Replay + internal browser telemetry. Output: a hotspot list, not a 60-page audit.
For 11 recurring variant axes we built distinguishing-fact templates and rolled them across 6,214 PDPs as JSON-LD, so agents parse the facts directly instead of scraping prose.
Tactic
Schema.org ProductGroup with variesBy + per-variant Product nodes. 6,214 pages migrated in 8 days, validated against Google Rich Results and a headless agent crawl.
{ "@context": "https://schema.org", "@type": "ProductGroup", "name": "Cordless Impact Drill", // the axes an agent must disambiguate "variesBy": ["voltage", "batterySystem"], "hasVariant": [{ "@type": "Product", "sku": "DRL-110-PRO", "name": "Impact Drill 110V Pro", "additionalProperty": [ { "@type": "PropertyValue", "name": "voltage", "value": "110V" }, { "@type": "PropertyValue", "name": "batterySystem", "value": "M-Core" } ] }]}Service level and bundle constraints moved up to the product page as semantic options agents could evaluate before committing — so the decision happened where the agent still had context.
Tactic
Service level as a labelled <select>, defaults mapped per category, readable in headless browsing without JS.
The 47 agent tasks became a CI check that runs before every release. Below 90% task completion, the deploy blocks — so a weak new product is caught before launch, not after.
Tactic
Playwright + Anthropic Claude API as agent simulator, integrated via GitHub Actions.
Three headline numbers — and the main one breaks down.
−38% agent drop-offs
The 38% broke down into four levers we measured separately:
The durable output wasn't the redesign — it was the test. Cartpilot now runs the 47 agent tasks before every release, so product-readiness is something the team checks itself, the same way it checks that a page renders. Maintenance: about two hours a month.
Example · Case study
Example · Case study
Want the same on your shop?
Check my shop