AI Agent Compatibility Fixes: Let Agents Complete the Task

How AI agents use a website

Several AI products now act on websites for their users. OpenAI's ChatGPT agent, Anthropic's Claude in Chrome, Google's Gemini agent features and browsers with a built-in assistant such as Perplexity's Comet can all be asked to find a product, compare prices, book a table or fill in an enquiry. The list changes quickly, but they work in broadly one of two ways.

• Driving a real browser. The agent loads the page in a browser, looks at a screenshot and at the page's accessibility tree, the same structure screen readers use, and then clicks, types and scrolls. These agents run JavaScript, but they depend on buttons, links and fields being clearly named and predictable.
• Fetching the page. Lighter tools download the HTML and read it without running scripts. If your prices, stock levels or opening hours only appear after JavaScript runs, these tools see an empty page.

Both kinds work on behalf of a real person who asked for something specific, which is what makes them different from search engine crawlers. An agent that cannot finish is a lost enquiry or sale, not a lost ranking.

Agent compatibility and agent readiness are different audits

We run two AI audits because they test two different things, and a site can pass one and fail the other.

If your problem is not being mentioned in AI answers at all, start with the Agent Readiness Guide. If AI tools find you but cannot get anything done once they arrive, this page is the right place.

The fix guides

Each guide takes one finding from the audit and walks through the cause, the fix, how to test it and the questions people ask most.

Getting in the door

Fixing agent blocks covers firewalls and bot protection that treat every automated visitor as a scraper, how to recognise agents that identify themselves, and how to adjust rules in Cloudflare, Akamai or AWS WAF without opening the door to abuse. Fixing slow rendering covers pages that take so long to become usable that an agent moves on, and how server rendering and lighter scripts fix it.

Reading what matters

Fixing JavaScript-only content covers putting prices, stock, dates and key text in the first HTML the server sends. Fixing structured data for agents covers the Product, Offer, Event and LocalBusiness details agents lean on to answer "is it in stock, how much, when is it open". Fixing contact details covers phone numbers, email addresses and opening hours that exist only as an image or inside a script.

Getting the task done

Fixing navigation covers menus that only open on hover and links that are really scripted buttons. Fixing forms covers fields with no label, IDs that change on every load and multi-step forms that lose their place. Fixing checkouts covers session timeouts, CAPTCHAs on every step and the hand-over to the person when payment is needed.

By platform

The platform guides cover where agent problems usually come from: Shopify, including apps that add pop-ups and scripts to the checkout; WordPress, including booking and form plugins; and single-page apps built with React, Vue or similar, where links and page changes often happen only in script.

Letting good agents in without letting everyone in

Most sites never decide to block AI agents. It happens because bot protection sees automated traffic from a data centre and challenges it, the same way it challenges scrapers. The fix is not to switch protection off but to recognise agents that say who they are.

• User agent names. Several providers publish the names their user-triggered fetchers send, such as ChatGPT-User, Claude-User and Perplexity-User. These are separate from their training crawlers, so you can allow one and block the other. User agent strings are easy to fake, so treat them as a hint rather than proof.
• Published IP ranges. Some providers publish the address ranges their agents use. Matching the user agent and the IP range together is much harder to fake.
• Signed requests. Web Bot Auth is a proposal at the IETF that lets an agent sign each request with a cryptographic key the site can check. Cloudflare supports it, and some agent providers already sign their traffic. It is the most reliable way to tell a genuine agent from an impostor.

Keep rate limits in place for everyone. A genuine agent working for one person makes a handful of requests, not thousands, so sensible limits do not get in its way.

Where agents should stop, and how to make the hand-over work

Agents are designed to stop and hand control back to the person at sensitive moments: signing in, entering payment details, confirming a purchase or solving a CAPTCHA. That is a safety feature, not a failure. What breaks the task is when the hand-over goes badly.

1. Keep the basket. If the session expires while the person is being asked to confirm, everything the agent did is lost. Allow enough time and keep the basket when a session is refreshed.
2. Challenge on suspicion, not on every step. A CAPTCHA on the product page, the basket and the address form stops the task three times. Use risk-based checks that only challenge unusual behaviour.
3. Make the confirmation step obvious. A clear summary of item, price, delivery and total lets the person check what the agent chose in seconds.
4. Support guest checkout. Forcing an account to be created is the step most likely to stop an agent completely.

Payment standards for agents are starting to appear, such as the Agentic Commerce Protocol from OpenAI and Stripe and Google's Agent Payments Protocol. They are early and mostly reach shops through their payment provider or platform, so check what yours supports before building anything yourself.

Which findings to fix first

1. Blocks at the door. If bot protection stops agents from loading the site, no other fix matters. Check this first.
2. Your most valuable journey. For a shop that is search, product, basket and checkout. For a service business it is the contact or booking form. Fix the path that earns money before anything else.
3. Unlabelled controls. Buttons with no name, icon-only links and fields without labels are the most common reason an agent clicks the wrong thing or stops.
4. Key facts only in JavaScript. Prices, stock, opening hours and contact details need to be in the page HTML and in structured data.
5. Speed and stability. Pages that shift around while loading or take many seconds to become usable make agents mis-click or give up.

Many of these fixes are the same as accessibility fixes. A site that works well with a keyboard and a screen reader is usually one an agent can use too, so the accessibility fixes are worth reading alongside this page.

A sensible order of work

1. Run the AI Agent Compatibility audit on your home page and on one page from each main journey.
2. Check your firewall and bot protection logs for challenges and blocks against known agent user agents.
3. Walk through your main journey with only a keyboard. Anything you cannot reach or operate this way will also trouble agents.
4. Name every button, link and form field in the code, not just visually.
5. Move key facts into the first HTML response and add matching structured data.
6. Remove extra steps: forced accounts, repeated CAPTCHAs and short session timeouts.
7. Try the journey with an AI agent you have access to and note where it hesitates or asks for help.
8. Run the audit again after any theme change, new plugin or checkout app.

For the underlying ideas about how AI systems read sites, see the Agent Readiness Guide, the llms.txt Guide and the AI Schema Guide.