Why AI shopping agents fail before checkout: 13 obstacles and how to fix each one

How an AI shopping agent reads your store

An AI shopping agent is software that takes a task, such as “running shoes, size 42, under $150”, and tries to complete it in a web browser on its own. Browser agents work in a loop: load the page, read it, choose an action, click or type, and repeat until the task is done or they get stuck.

What they read is not what a person sees. An agent builds its picture of a page from four sources:

• The HTML the server sends: links, product names, prices and button text.
• The accessibility tree: the names and roles of buttons, links and form fields, the same information a screen reader uses.
• Structured data: Product JSON-LD with the price, availability, brand and identifiers.
• Screenshots: many agents also look at the page as an image. Reading text from an image is slower and less certain than reading it from the code.

That is why a store can look perfect to people and still stop an agent. A size shown as a picture of “42”, a close button drawn as an icon, a product tile that is not a link: a person never notices, an agent has nothing to read.

The five gates of a purchase path

Every purchase passes five gates. The agent has to get into the store, find and understand a product, pick the right variant, get from the cart to checkout, and reach the checkout form. One closed gate is enough to lose the sale.

We built each obstacle below into its own copy of the same demo store, next to a control store where everything is done right (S01). Each copy differs from the control store in one thing only, so when an agent fails, the cause is clear. Every obstacle is listed in our obstacle registry with its number.

Entry: can the agent use the store at all?

<button type="button" class="modal-close" aria-label="Close">
  <svg aria-hidden="true" viewBox="0 0 16 16">…</svg>
</button>

curl -s https://yourstore.com/products/aero-run-2 | grep -i "add to cart"

Product: can the agent find and understand it?

<a class="product-tile" href="/products/aero-run-2">
  <img src="/img/aero-run-2.jpg" alt="Aero Run 2 running shoe, blue">
  <span class="name">Aero Run 2</span>
  <span class="price">$139.00</span>
</a>

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Aero Run 2",
  "sku": "AR2-BLU-42",
  "gtin13": "5901234123457",
  "brand": { "@type": "Brand", "name": "NORDLYS" },
  "image": "https://yourstore.com/img/aero-run-2.jpg",
  "offers": {
    "@type": "Offer",
    "url": "https://yourstore.com/products/aero-run-2",
    "price": "139.00",
    "priceCurrency": "USD",
    "availability": "https://schema.org/InStock"
  }
}
</script>

Variant: can the agent pick the right size and colour?

<fieldset>
  <legend>Size</legend>
  <button type="button" aria-pressed="false">42</button>
  <button type="button" aria-pressed="false" aria-label="Size 43">
    <img src="/img/size-43.svg" alt="">
  </button>
  <button type="button" disabled>44 – out of stock</button>
</fieldset>

<label for="size">Size</label>
<select id="size" name="size">
  <option value="42">42</option>
  <option value="43">43</option>
  <option value="44" disabled>44 – out of stock</option>
</select>

Cart: can the agent get from the cart to checkout?

Checkout: can the agent reach the checkout form?

Check your own store in 10 minutes

You can find most of these obstacles yourself with a browser and its developer tools:

• Fetch a product page without a browser (or turn JavaScript off). The name, price and “Add to cart” are in the HTML.
• Open the accessibility tree in the developer tools. Every size, colour and close button has a name.
• View the page source and find Product JSON-LD with the price, currency and availability. It passes validator.schema.org.
• Open the store in a private window. Every banner and pop-up can be closed with a named button, and the cookie banner offers “Only necessary” on its first screen.
• Unavailable sizes are disabled and say so in text.
• Every product in a listing is a real link to its page.
• Walk to checkout as a guest, without logging in. Every place where you have to guess will stop an agent too.
• Read your /robots.txt rules for OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot and Claude-User.