AI agents now book for customers. Does your website let them in?

In September 2026, OpenAI and Meta each launched AI agents that act for people on the web. Ask OpenAI's dots or Meta's Muse to book a table or buy a gift and it opens a browser and tries to finish the job. Many attempts end at the business's website. TechCrunch reported on 6 October that Walmart, which had signed up as a Muse partner, was still turning agents away at a button that asks the visitor to prove they are human, and that Yelp only admits agents that pay for its data. Walmart wanted the agents in; its own check sent them away. Neither agent is fully open in Europe yet, so a business in Europe has a short window to test whether an AI agent can get through its booking or enquiry form, and to decide what the agent should be allowed to do.

TL;DR: AI agents from OpenAI and Meta now book and buy for their users, and many business websites stop them by accident, usually at a "verify you're human" check. Walk your own booking or enquiry path, list every step only a person can pass, and decide which should stay. Cloudflare's new default rarely touches a site without ads.

What are AI agents doing on business websites?

A personal AI agent is an assistant that completes a task on a website for someone: it opens pages, fills in forms and, with the user's approval, pays. Two arrived in September 2026.

• Muse, Meta's personal agent, is available in the US and Canada. On 29 September Meta added Muse for Small Business, which connects to tools such as Shopify, Stripe and QuickBooks.
• Dots, launched by OpenAI at DevDay on 29 September, each run on their own cloud computer and browser inside ChatGPT. They are rolling out to Pro and Business Premium subscribers. OpenAI said Pro users in the European Economic Area, Switzerland and the UK will not get dots for now, while Business Premium covers every region ChatGPT supports.

That split decides who reaches a European booking page first. A guesthouse in Porto or a tour operator in Lisbon whose bookings come from American and Canadian travellers can meet these agents now, because the agent works wherever its owner lives, not where the business is. A dentist in Lyon whose patients all live nearby will meet them later, when OpenAI or Meta opens Europe to consumers, and can use the time to fix the form first.

Why do AI agents fail on business websites?

Picture a Canadian couple planning a week in Lisbon who ask their agent to book a dinner table and a tour. The agent finds both sites, picks a time and reaches the form. Then it meets a checkbox that asks it to prove it is human, or a request to create an account, or a code sent by SMS to a phone it doesn't hold. The task fails, and the couple are told the restaurant "couldn't be booked".

TechCrunch's reporting separates two kinds of failure:

• Deliberate blocks. Amazon began blocking Muse from its shop. Yelp told TechCrunch it doesn't permit non-human traffic unless the agent has paid for access to its data. Delta said it takes "steps to protect customers from unauthorized automated activity".
• Accidental blocks. Walmart wants agents in, and still loses them at its human-verification button. If that check is interrupted, it fails, and the agent is sent away.

The person behind the agent can't tell which kind they hit. In TechCrunch's words, "what's not clear to the end user is whether or not the block is accidental", so the business and the agent both look broken. Meta's own line on it: "Turning away a personal agent means turning away the customer behind it." Meta sells agents, so read that with its interest in mind; a guesthouse that loses the booking still loses it.

On a small business site, the usual blockers are a CAPTCHA or "I'm not a robot" box on the form, a required account before booking, an SMS code, a booking widget run by a platform with its own rules, and pages that only say "call to book".

Does Cloudflare's new default block AI agents on your site?

Cloudflare's agent block covers pages that show ads, so a small business site without ads is usually untouched. You may be on Cloudflare without knowing it, because whoever set up your domain may have added it. Ask them, or look for "cloudflare" in the "server" line when you run curl -sI https://yoursite.com. On 15 September 2026, Cloudflare split its AI controls into three categories: Search, Training and Agent. Agent covers "user-directed agents visiting a page on behalf of a human", which is what dots and Muse are.

For new domains, the recommended setting depends on whether the site earns money from ads:

• No ads: agents allowed.
• Ads: agents blocked on the pages that show them.

Existing domains kept their settings. A domain that had the old "Block AI bots" switch on was moved to "Block on pages with ads" for agents, which still leaves an ad-free booking page open.

The bigger risk is tightening it by hand. Since the same date, setting Training to Block also stops Googlebot, Bingbot and Applebot, the crawlers behind Google, Bing and Apple search. To keep AI companies from training on your content and stay in search, the setting is "Disallow AI Training". If your site runs on another host or a security plugin, its bot settings sit in its own dashboard; note what each is set to before changing anything.

How do you test whether an AI agent can book with you?

Two checks cover it: one for your server and one for the path a customer takes from homepage to confirmed booking.

1. Ask your server to load a page for an AI fetcher. This one needs a terminal, so hand it to whoever maintains your site if that isn't you. Request your booking or contact page with an AI agent's name as the user agent:

curl -s -o /dev/null -w "%{http_code}" -A "ChatGPT-User" https://yoursite.com/contact

A 200 means the page was served. A 403, or a challenge page instead of your content, means a rule turns that agent away. We ran the same command against sharphaw.com/contact on 28 October 2026 with the ChatGPT-User and Claude-User names, and both got a 200. This only covers agents that announce themselves; a dot runs a real browser, which is what the second check is for.

2. Walk the path and count the human-only steps. On your phone, go from your homepage to a finished booking or enquiry and write down every step a person passes and software can't:

• a checkbox, puzzle or "press and hold" challenge
• an account you must create before booking
• a code sent by SMS or email before the form submits
• a payment or card guarantee
• a phone number as the only way to book

If the booking runs through a third-party widget, read that platform's terms on automated booking too. Its rules apply on your page, whatever you decide.

Should you let AI agents book for your customers?

Let them in up to the point where money or a slot is committed, and keep a person there. Bot checks exist for real reasons: spam enquiries, fake bookings that hold a table nobody turns up for, scripts testing stolen cards. An agent sent by a real customer is a different visitor, and the same check treats both alike.

A workable split for most service businesses:

• Let agents through to read prices, opening hours and availability, and to send an enquiry or a booking request.
• Keep a person-only step where money or a scarce slot is committed: a deposit, a card guarantee, a confirmation link sent to the customer.

Both companies already put the person back in at that point. Meta says Muse does nothing that "publishes, sends, or spends without your approval", and OpenAI tells dots users to "always review consequential work".

The enquiry form on sharphaw.com uses a hidden field that spam bots fill in and people never see, in place of a visible "I'm not a robot" box. Anyone who can type a message can send one, including an agent working for someone. The cost is that a determined spammer gets through more often than with a challenge, and we accept that for a form whose whole job is to be sent.

When you do block on purpose, say so on the page. A plain line such as "Groups of eight or more: please book by phone" gives the agent something to report back, and the customer a next step. Meta, Walmart, Stripe and others have begun work on an open standard for how agents and businesses talk to each other; until one exists, the text on your page is the only instruction the agent gets.

What to do this week

Run the curl check on your booking and contact pages, then walk the path on your phone and list each human-only step. Decide which steps protect something real and remove or replace the rest. If you use Cloudflare, open Security settings and note what Search, Training and Agent are set to, and swap any Training "Block" for "Disallow AI Training".

If the walk shows that the booking path itself needs rebuilding, that is the work our Conversion-First Websites service covers: the form, the steps before it and its upkeep. Our concept builds show booking flows built for invented businesses. To have us look at yours, tell us about your site, or read the plans and fees first.