# How a test run works: AI agents on real store tasks.

> How a test run works: a named AI agent tries a real task on your store, with step screenshots, checked outcomes and fixes.

- Canonical URL: https://aiscovery.io/product/agent-readiness-audit
- Markdown version: https://aiscovery.io/product/agent-readiness-audit.md
- Last updated: 2026-10-21

Named AI agents try key tasks on your public pages. Each run keeps a step log, step screenshots and the checks it was held to.

- [Request a demo](https://aiscovery.io/demo)
- [Replay a test run](https://aiscovery.io/product/agent-readiness-audit#replay)

## Illustrative run

> Interactive demo on the HTML page: a run player that steps through one test run with a step log, elapsed time, attempts and layout shift. Its content is written out below.

> Illustrative run on an example store. Step screenshots, not video. Store: Northwind.

Agent: OpenAI GPT-6 Luna. Task: Compare delivery options for the chosen size.

| Step | Elapsed | Page | Step log | Status | Layout shift |
| --- | --- | --- | --- | --- | --- |
| 1 | 0.0 s | / | Opened the home page | passed | 0.00 |
| 2 | 3.1 s | /search?q=trail+shoes | Searched for trail shoes and opened the first result | passed | 0.00 |
| 3 | 7.8 s | /products/ridge-runner | Picked size US 9 | passed | 0.02 |
| 4 | 11.2 s | /products/ridge-runner | Stock loaded late and pushed the delivery options down | warning | 0.18 |
| 5 | 15.6 s | /products/ridge-runner | Read Standard and Express delivery for US 9 | passed | 0.18 |
| 6 | 18.4 s | /products/ridge-runner | Stopped before cart. Assertion checked. | passed | 0.18 |

The delivery control moved 64 px while the agent read it.

Attempt: 2 of 2. Result: Worked, with friction.

## A scoped audit answers operational questions.

The audit starts from a customer goal and ends at a verified state, not a checklist count.

1. Can the agent find the current facts it needs?
2. Can it interpret prices, policies, inventory and controls?
3. Can it complete the allowed action without losing state?
4. Can it recover from validation, expired sessions and unavailable inventory?

## Every conclusion stays connected to the run.

This illustrative chain shows how an observed problem becomes an engineering action.

> Interactive demo on the HTML page: pick a link in the chain to read it.

1. **Scenario.** Compare delivery options on a product page.
2. **Run.** An AI agent opens the product page, picks a variant and reads the delivery options. It stops before cart.
3. **Assertion.** The delivery options and their prices are visible for the chosen variant.
4. **Evidence.** Step log, masked step screenshots and layout-shift measurement.
5. **Finding.** Inventory hydration moves the delivery control during the action window.
6. **Remediation.** Reserve the control height and keep the accessible name stable during hydration.

## Review the session, not only the final label.

Timing, attempts and layout stability explain how reliable a success really was.

> Illustrative run

| Signal | Example value | What it shows |
| --- | --- | --- |
| Step screenshots | 6 | Follow the agent path step by step, with masked screenshots and the checks each step was held to. There is no video. |
| Completion time | 18.4 s | Measure wall-clock duration and time spent waiting for navigation, rendering and recovery. |
| Attempts | 2 | Each test run is one attempt by one agent. Repeated attempts show whether a success holds. |
| Cumulative Layout Shift | 0.18 | Measure layout shifts on each page the agent visits. It is a lab measurement in our browser, not field data. |

## The audit package is built for action.

Results are useful to leadership, engineering, QA, product and security without hiding uncertainty.

1. A plain verdict limited to the tasks we tested
2. Results per task for each agent that ran
3. Critical and High fixes with the step that failed
4. Step screenshots and step logs
5. Remediation backlog with acceptance criteria

## Boundaries are agreed before the first run.

AI agents stop before login, cart, checkout, payment, booking or any form that sends data. Full journeys on an approved staging site are coming later.

- Login
- Cart
- Checkout
- Payment
- Booking
- Forms that send data

## What the audit does not claim

- A single agent result does not represent every agent product.
- A detected technology is not proof that a business task works.
- A passed task is not a guarantee of future model behavior.

## A partner audit starts small and stays reproducible.

1. **Scope the journey.** Choose one to three business-critical tasks, environments and stop gates.
2. **Run and review.** Run each task, check the outcome against the evidence and have Aiscovery review it.
3. **Prioritize and retest.** Ship selected fixes, rerun the same contract and compare the evidence.

## Audit questions

**Q: Can you test authenticated workflows?**

A: Not yet. Tests of logged-in pages are coming later. Today site checks read public pages only.

**Q: Do you record the agent session?**

A: Yes, as evidence: each run keeps a step log and step screenshots with sensitive data hidden. We don't record video.

**Q: How many times is a scenario run?**

A: Each test run is one attempt by one agent. Run a key task again, or with another agent, to see whether a success holds.

## Bring one difficult journey into focus.

Leave three details. We pick the task with you, run it on your store and walk you through the evidence.

[Request a demo](https://aiscovery.io/demo)

Have a real workflow to test now? [Apply as a launch partner](https://aiscovery.io/partners)

---

This is the Markdown version of the page for AI agents. The HTML page at the canonical URL is authoritative. Site index: https://aiscovery.io/llms.txt