Why do AI agents fail to buy from most websites?
The main reason is bot protection, not missing technology. When Axidort sent a declared AI agent to 50 leading commerce sites, 31 of them (62%) served a bot challenge before the agent could finish a purchase journey. The median Axidort Score was 11 out of 100 (scanned October 1, 2026).
By the Axidort team · Updated
What we measured
For the Axidort Index we scanned 50 well-known B2B and B2C commerce sites with the same agents and scoring as a public Axidort scan. A browser agent identified itself as AxidortAgentCheck, searched for a product, read its price and stock and tried to reach checkout. A second agent read the same page without a browser. Nothing was ordered. A site that scored under 70 was scanned twice and the better score kept. All scans ran from one data-centre location, which matters: bot protection is stricter towards data-centre traffic than towards a home connection.
The results
| Finding | Result |
|---|---|
| Sites scanned | 50 |
| Median Axidort Score | 11 out of 100 |
| Median, B2C sites | 15 |
| Median, B2B sites | 8 |
| Served a bot challenge to the agent | 31 (62%) |
| Agent reached a cart or buying flow | 21 |
| No blocker recorded | 16 |
| Grade A or B | 7 |
| Grade F | 36 (72%) |
| Highest score | IKEA (97) |
The scores are lopsided. A small group of sites let the agent through and score well. Most stop it at the door and score near zero, because an agent that is challenged on the first page never gets to show whether the product data or checkout would have worked.
The five things that stop agents
1. Bot challenges
By far the most common. A CAPTCHA, a JavaScript challenge or a block page is served to automated browsers, including ones that identify themselves honestly. The agent has no way through and should not try to find one. Fix: keep protection for abusive traffic, and add rules that let verified or declared agents reach product and cart pages. See robots.txt for AI agents for how agents identify themselves.
2. Facts that only exist after JavaScript runs
A reader agent fetches the HTML and reads it. If price and stock are injected later by a script, it sees a product with no price. Fix: render name, price, currency, unit and availability in the server response and mirror them in Product structured data.
3. Add-to-cart requests that are refused
The button is there and the click happens, but the request behind it is rejected for automated sessions and the page shows no error. Fix: test the cart endpoint with an automated browser, and return a clear message when a request is refused.
4. Login and quote walls
Requiring an account before checkout, or selling by quote only, ends the journey for an agent acting without credentials. This is common and legitimate in B2B. Fix: offer guest checkout where you can. Where you cannot, make the sign-in or quote step standard and clearly labelled, and publish what an agent may see without an account.
5. Search and navigation an agent cannot operate
Search boxes hidden behind icons with no label, filters built from unlabelled div elements, and overlays that cover the page. Fix: use real form controls, labels and links. See how to make a website agent friendly.
What the high scorers do
- They serve the page to a declared agent without a challenge.
- Product name, price and availability are in the HTML and in structured data, and the two agree.
- Search works with a plain query and returns product links.
- Add to cart works in a clean, logged-out session, and checkout can be reached as a guest.
None of this needs a new protocol. It is ordinary, well-built commerce that does not treat every automated visitor as an attacker.
Limits of this data
- One location, one declared agent and one point in time. A site may behave differently for another agent, from another region or next week.
- Pages behind a login were not scanned, which understates what B2B portals offer signed-in buyers.
- A low score describes what an agent experienced. It is not a judgement of a business, its products or its service.
The per-site results, including each site’s seven component scores, are public in the Axidort Index and as JSON.
Frequently asked questions
01What percentage of websites block AI agents?
In the Axidort Index, 31 of 50 leading commerce sites (62%) served a bot challenge to a declared AI agent. The sample is large, well-known commerce sites scanned from a data centre on October 1, 2026, so it is not a figure for the whole web.
02Should I let AI agents through my bot protection?
Let through the ones that act for your customers and identify themselves, and keep blocking abusive traffic. Most bot-management products let you write rules by verified bot, user agent or signed request rather than blocking all automation.
03Do AI agents solve CAPTCHAs?
Well-behaved agents do not attempt to defeat a CAPTCHA. They stop and report that they were blocked, which for the buyer means your site drops off the shortlist.