METHODOLOGY · UPDATED SEPTEMBER 4, 2026

What we test.
What it proves.

Three kinds of evidence answer different questions. We name the method on each report and keep untested steps visible.

1. The free HTTP preflight

We fetch public pages, robots.txt and available product data from our server. We compare a browser-like request with four published user-agent strings: ChatGPT-User, Perplexity-User, Amazonbot and Claude-User. These are request identities, not four consumer agents.

A store URL triggers product discovery; a product URL checks that specific page. The report shows the selected URL and timestamp. Requests have an eight-second timeout, a two-megabyte response cap and a 40-second overall budget. Results may be cached for ten minutes.

We inspect Product / ProductGroup JSON-LD for price, stock, identifiers, shipping, returns and ratings. Missing markup is a data signal. It does not establish that the same information is absent from the visible page, a merchant feed or a platform catalog.

Not tested: JavaScript interactions, real agent network identities, variant selection, cart contents, address entry, payment completion, recommendation ranking or revenue. An empty-cart checkout URL may legitimately return an error. A loaded CAPTCHA script does not establish that an action is blocked. Network errors and timeouts are reported as unassessed.

2. Automated browser observations

A scripted browser can select a product, inspect a cart and look for a payment step. This is a reproducible automation test, not a consumer AI agent. A selector failure can be a harness limitation; we review it before attributing a problem to the store.

Historical browser observations in our research are not presented as proof of real-agent completion or lost sales. Public examples identify their method. Illustrative deliverable formats are labeled separately from measured results.

3. The $500 checkout evidence pilot

We agree on one store, one product with its variant, one US shopping scenario and two available consumer browser agents. Each agent attempts the same scenario twice, producing four recordings. Agents are confirmed before payment; no platform is silently substituted.

Each attempt records the prompt, agent and date, chosen product, variant, quantity, purchase type, cart total when visible, furthest observed step, and any obstruction. We distinguish observed behavior, a likely cause and a confirmed cause.

We stop before submitting payment. Any login, approval or required human takeover is recorded as a handoff, not a merchant defect. We do not bypass security challenges, alter store settings or need admin access for the initial pilot.

The deliverable contains all outcomes, including passes, prioritized developer tickets and acceptance checks. One retest of the same four attempts is included when requested within 14 days of delivery. Implementation and broader testing are quoted separately.

Timing: scope and availability are confirmed before payment; delivery is within two business days after confirmation and payment. If we cannot run the agreed test, we decline it before charging. Passing outcomes do not trigger a refund: the product is documented evidence, not a promised failure or revenue lift.

How to interpret the result

Four attempts are a small diagnostic sample. They do not estimate a population conversion rate. A failure in one dated run does not show that all agents or all shoppers fail. A pass is not a certification. No score here predicts revenue, recommendation placement or agent order volume.

We recommend verifying a finding in the relevant environment before changing access controls, checkout settings or data feeds.

Data handling

The free preflight needs only a public store or product URL. Do not include access tokens, private preview links, passwords or customer details. Results are held in an in-memory cache for up to ten minutes and displayed in your browser; they are not automatically added to a public leaderboard.

Copy and download controls create a report on your device. If you request a pilot through the form, the store, product, email, role, optional notes and preflight summary are sent to the operator’s configured request inbox or storage. They are used to scope and deliver that request, not to subscribe you to marketing.

We log basic funnel events, such as a completed check or pilot request, with campaign labels. The event payload does not include your email, store URL, cookies or a browser fingerprint. The hosting provider may retain standard server request logs. Contact hello@usecreel.com to ask about or delete your submitted request.

Source context

Shopify’s product-page scanner describes its output as informational product-page signals. OpenAI’s product discovery update describes Shopify Catalog as a source of merchant product data. That is why missing page markup alone is not proof that an agent cannot find or buy a product.