AI agents are shopping your store. Can they check out?

Serge runs a real Claude agent on your live e-commerce storefront and shows where it fails to complete a buying task. You type a prompt ("find a black leather backpack and add it to the cart"), Serge dispatches the test, and the report comes back with screenshots, the exact failure step, and a fix snippet.

A real Claude agent today (Anthropic computer-use) — the same class of software your customers delegate purchases to. Operator and GPT Agent support is on the roadmap as they ship stable browser capability. Each test runs against the actual agent, not a generic headless browser, so the failure modes match what a real customer would hit.

You give Serge a buying task in plain English. Serge dispatches a real AI agent against your live storefront and records every step — the URL the agent navigated to, the elements it tried to click, the accessibility tree it parsed, the moment it gave up. The report streams back live in a minute or two, with screenshots and a paste-ready fix for each blocker.

Playwright tests what your selectors do; QA tests what humans do. Neither runs the actual AI agents your customers use. AI agents read the DOM differently — a hard-coded test script with fixed selectors will pass while a real Claude agent quits at a variant selector that has no accessible name. Serge catches the gap.

No. GEO tools (Athena, Profound, Scrunch) measure whether ChatGPT mentions your brand in its answers. They sit upstream of the customer journey. Serge sits where the customer actually arrives — when an agent lands on your site and tries to find a product and add it to a cart. The two are complementary, not competitive.

Pro starts at CHF 159/month with daily journey tests, 12-month retention, and PAYG add-ons for extra runs. Agency is a contact-us tier for teams running tests across multiple client stores. A free account lets you scan your store and share the results with your team.

A minute or two. The agent runs against your live storefront in real time, streams the steps as it goes, and lands the full report with screenshots, the failure point, and a fix snippet. No staging environment to set up, no scenarios to script — paste a prompt, hit Run.

Product Directors and Front End / Engineering Leads at mid-to-large e-commerce companies. The Product Director feels the pain (silent conversion loss to AI shoppers), the FE Lead owns the DOM where the failures happen. Economic buyers are usually the Head of E-commerce or CMO at the same account.

No. Journey tests run with throwaway test data — synthetic email addresses, test card numbers, no real customer accounts. Serge does not have credentials to your back office, does not see your customer database, and does not store form values from real shoppers.

Test reports and screenshots are stored in EU regions (Neon Postgres + Vercel Blob), retained 12 months on Pro by default. You can delete a test at any time. We do not sell, share, or train models on your data. Full data processing addendum on request.

The free scanner is the entry point. Sign in with Google in one click, paste your domain, get a 0–100 score for how easy it is for an agent to find a product and add it to a cart, plus the top blockers we found. It runs deterministically — no AI calls at scan time, no card required. The full journey-test product runs a real Claude agent against your store and lives behind the paid tier.

The 0–100 score reflects structural conditions that correlate with agent task success — product schema, semantic HTML, accessible variant selectors, bot-protection posture, robots.txt access. A high score means the structural ground is laid; a low score means an agent will struggle before it even gets to the variant selector. Journey tests verify it for real.