Free run
NO EMAIL- One scenario × two agents
- Live view with screenshots and the agents’ thoughts
- Stage scores, stuck points and fixes
- A 33-item checklist of what tires the agent, with evidence
- A link you can share
AGENT SIMULATOR · MYSTERY SHOPPER AGENTS · FREE
Send AI shopping agents to your store and see where they get stuck
Two AI agents shop your store live in a real browser: they search, open a product, choose a size or colour, add it to the cart and stop at checkout. You watch every step and get a report on where they got stuck and how to fix it.
- Real browser, your real pages
- Two agents, each driven by a different AI model
- Never orders, never enters personal data
- Free, no email needed
MYSTERY SHOPPER AGENTS
Mystery shoppers for the age of AI agents
Like a mystery shopper, our agents walk into your store the way a customer’s AI assistant would and report what got in their way. Unlike a person, they show you every step, and they stop before paying.
- They enter like a customer: search, product, size or colour, cart, checkout.
- They stop at checkout: no order, no payment, no personal data.
- They report what got in the way, with the step it happened at and a fix.
A REAL RUN, NO SIGN-UP
One recorded run, from requirements to result
A real run of our free simulator on a live fashion store, with its name removed. First what the agent needs to finish the task, then the run step by step, then what the report told the store.
Real run · October 8, 2026 · one agent · store name and product names removed
Task: Find a named polo T-shirt in purple, size M, add it to the cart and go to checkout.
What the agent needs, and what it found
| Requirement | Result | Evidence from the run |
|---|---|---|
| The store opens for the agent’s browser | Met | HTTP 200 on the home page. |
| Site search works | Met | Step 1: typed into the search box. Step 2: a results page. |
| The product page carries price and stock data | Met | Product JSON-LD on every product page opened: 24.90 EUR, in stock. |
| Size and colour can be chosen | Partly | Size M was a labelled button (step 9). The colour swatches had no name, so the agent clicked them by screen position (steps 5 and 17). |
| The agent sees the store your customer sees | Not met | The browser asked for Turkish. The store served its German version with prices in EUR until the agent switched the language itself (step 16). |
| “Add to cart” appears after choosing a size | Not met | Not found in 22 steps: the agent scrolled, reopened the page and changed the size (steps 10-13, 21, 22). |
| Guest checkout | Not reached | The run never got this far. |
The run, step by step
The agent’s own note at each step, translated from Turkish, with product names removed. The badge shows the language of the page the store served.
- 1Let’s type [product name] into the search bar.
- 2The product shows 21 more colour options; opening it to check for purple.
- 3The search results had a lilac (purple) option; going back to pick it.
- 4Opening the product page to see whether purple is among the colours.
- 5Clicking the purple/lilac colour swatch.
- 6Searching for “[product name] purple” to find the purple one.
- 7No results; searching for just the product line’s name.
- 8Clicking the link to the purple polo T-shirt.
- 9On the purple polo page. Choosing size M, which is in stock.
- 10Scrolling down to find the add-to-cart button.
- 11Reopening the product page to look for the button from the top.
- 12The size doesn’t look selected; trying size S so the button appears.
- 13Scrolling down again for the add-to-cart button.
- 14Searching again to check the sports polo model.
- 15Opening the sports polo page to check the other colours.
- 16Switching the site to Turkish to check the add-to-cart button.
- 17Clicking the swatch that may be lilac/purple.
- 18Searching again for the purple colour.
- 19Only one model came up; opening it to check sizes and add to cart.
- 20Choosing size M.
- 21Size M is selected; scrolling down to find the Add to cart button.
- 22There is no “Add to cart” button on the page, or the product isn’t on sale because of a stock or region restriction; it can’t be added to the cart.
✓Find✓Product✓Size✕Cart-Checkout
Stopped on the product page after 22 steps. It never reached the cart.
Before and after
BeforeRECORDED
- German page and prices in EUR, for a browser asking for Turkish
- No add-to-cart button after choosing a size
- Stopped on the product page after 22 steps
After a fixEXPECTED · NOT RECORDED
- The page in the language the browser asks for, or a visible country and language choice that keeps add-to-cart
- Colour swatches with a text name
- Expected path: search → product → size → cart → checkout
The same task wasn’t run again after a fix, so this side is what we would expect, not a measurement. On your own store, run the simulator before and after a fix to see both sides for real.
What the report told the store
Location decides what the agent sees
Serve the language the browser asks for instead of deciding by IP address alone, or show a visible country and language choice that doesn’t hide add-to-cart.
Colour swatches without a name
Give every swatch a text name (visible, or an aria-label) so an agent can pick a colour without guessing a screen position.
Is this a quirk of our test? No. Shopping agents’ browsers also run in data centres, which may be in another country. If a store decides by location alone, an agent can see a version of it your customer never would.
Source: Specoria agent simulator, recorded run · · 1 run, 22 steps
HOW IT LOOKS
Watch an agent shop, step by step
This is what you see live in your report: every page the agent opens, what it was thinking, where it hit friction and where it stopped.
Concept demo: the store, figures and steps are illustrative, not a real customer result.
WHAT THE AGENTS DO, AND WHAT THEY NEVER DO
A customer’s journey, without the customer’s data
The agents behave like a shopper using an AI assistant. The rules below are enforced by our harness on every step, not left to the model.
They shop like a customer
Search box, categories, product pages, size and colour options, the cart: the same interface your customers use.
They stop at checkout
Reaching the address or payment form, or the “log in / continue as guest” screen, counts as success. They never press the order or payment button.
No personal data, no accounts
They never type a name, email, phone, address or card, never log in or create an account, never accept cookie banners and never leave your domain.
Identified honestly
Our browser runs on Cloudflare and is identified as Cloudflare’s signed (verified) bot; your store sees that, not ChatGPT. We don’t disguise it, solve CAPTCHAs or use residential proxies.
A block is a finding
If your bot protection stops the agents, the report says so plainly. Real shopping agents can hit the same wall, so it’s worth knowing.
Our agent, different models
The agents are Specoria’s own agent loop, each driven by a different AI model (for example Gemini). They are not the ChatGPT or Gemini consumer agents themselves.
HOW IT WORKS
From address to report in about five minutes
- 1Pre-check (seconds, no AI)
We check that the store is reachable, whether the first response is a bot wall, what your robots.txt says to AI agents, and whether product pages carry Product/Offer data. A real product from your sitemap makes the task concrete.
- 2Live run
Two agents shop side by side. Every step shows the screenshot the agent saw and what it was thinking.
- 3Report
Stage scores from finding the product to checkout, a friction map (steps, backtracks, retries), stuck points with screenshots and fixes, the flows that work, and confidence based on how far the agents agree.
FREE RUN AND FULL REPORT
See it free; get the full picture with your work email
Full report
WORK EMAIL- Three scenarios × three agents, plus a computer-use agent that works from screenshots only
- Your agent-readiness audit as a PDF
- A personal project panel where every finding becomes a task
- Sent to your inbox, usually within 30 minutes
Getting stuck is normal: shopping agents are still young and stall on many stores. The report separates what is on your side from what is on ours.
WHY NOW
AI assistants are becoming a place where shopping starts
Figures published by the organizations named, mostly for the US market. Read them as a leading indicator.
+138%
Growth in AI-referred visits to US retail sites in May 2026 versus a year earlier
Source: Adobe Analytics (Digital Commerce 360) · Jun 17, 2026
+54%
Higher conversion rate for AI-referred visits than for visits from other sources
Source: Adobe Analytics (Digital Commerce 360) · Jun 17, 2026
51%
Share of US consumers who used at least one AI tool to shop last month; the first time above 50%
Source: NIQ Agentic Commerce Tracker (The Next Web) · Sep 24, 2026
The figures are published by the organizations named, not by Specoria.
Questions
Will the agents place an order or create anything in my store?
No. They stop at checkout and never press the order or payment button. A product may sit in a cart session, the same as when a visitor leaves without buying.
Does my store see this as ChatGPT?
No. Our browser runs on Cloudflare and is identified as Cloudflare’s signed bot. We don’t imitate any assistant, solve CAPTCHAs or use residential proxies.
My bot protection blocked the agents. Is the test broken?
No, that is a finding. Real shopping agents can be stopped the same way. If you want verified agents in, allow them in your WAF or bot settings and run the test again.
Why two agents, and why do they disagree?
Different models find different paths. If every agent gets stuck at the same stage, we report a repeated block and show the technical evidence: each step and screenshot up to where they stopped. If only some do, it’s friction. One run is not a final verdict, so fix the cause and run it again.
Does it work for a hotel or services site?
Yes. The pre-check recognises from your home page whether it is an e-commerce, accommodation or services site. On a hotel site the agent chooses dates, the number of guests and a room and goes as far as the guest details form; on a services site it finds the service and reaches the contact or quote form, reading which details the form asks for. In both cases nothing is typed into the form and no form is sent. If your booking opens in an engine on another domain, the agent doesn’t follow it there; the report marks that part as “not measured”.
Which stores can I test?
Your own store’s domain. Marketplaces and a short list of large sites whose terms forbid automated agents are excluded. Free runs are limited to 3 a day per visitor.
What happens to the data?
We keep the run and its screenshots for 30 days so the shared link works, then delete them unless you ask for the full report. We store a salted hash of your IP address for the daily limit, never the address itself.
Fix what stops the agents
Starter re-measures every month with real agent simulations and turns each finding into a task with ready-made fixes. First month $1, then $149/month.