For most of my career, a website had two kinds of visitors: people and search engines. You designed for the first and optimized for the second, and the job was clear enough. Now there is a third visitor, and it behaves like neither. An AI agent arrives on behalf of a person, reads your page, decides what to click, fills in a form, and reports back. It never admires your hero image. It either completes the task or gives up and sends its user somewhere else.
This guide covers where sites tend to break for that visitor, how to test yours in an afternoon, and what I would fix first.
The visitor you never see in your analytics
Agents don’t all work the same way. Some read the structure underneath your page. Others look at a screenshot and guess where to click. Either way, they lean on the same things a screen reader does: clear labels, predictable layout and plain text. Products differ and change quickly, so treat everything below as a direction, not a specification.
That is why I tell people that readiness for AI agents is mostly old-fashioned craftsmanship. The sites agents handle well were usually already built with care: honest headings, real buttons, forms with proper labels. If you have been following our thinking on SEO vs GEO, this is the next layer of the same shift. Visibility used to mean being found. Increasingly it also means being usable by software acting for a customer.
Where agents get stuck
Most failures fall into five patterns, and you can probably name a page of yours for each.
- Unclear controls. Buttons labeled “Click here,” icon-only buttons with no text, and links that say “Read more” tell an agent nothing about what happens next.
- Walls in front of public information. CAPTCHAs, aggressive bot challenges and forced sign-ups before anyone can see a price or an address stop agents, and many humans, at the door.
- Interface tricks. Menus that open only on hover, infinite scroll with no fallback, pop-ups that cover the page, and carousels that hide the important slide all make a page hard to read in order.
- Poor forms. Fields with no labels, instructions that live only as placeholder text, and errors shown only by a color change leave an agent unsure what to type or why it failed.
- Key facts hidden in the wrong place. Prices, opening hours and shipping terms that sit inside images or PDFs, or load only after a click, may never be seen.
A two-hour audit you can run this week
You don’t need a special tool to start, although some new checkers have appeared this week that run around a hundred automated tests for agent-friendliness. A manual pass teaches you more, because you see the failure yourself.
Hour one: become the agent. Write down the three tasks a visitor most needs to complete on your site, such as finding a price, booking an appointment or buying a product. Then try each task three ways:
- Use only your keyboard. If you can’t reach a button with the Tab key, an agent may not reach it either.
- Turn on the screen reader built into your phone or computer and listen to the page. Wherever it says something meaningless, like “button” with no name, an agent faces the same fog.
- Ask an AI assistant that can browse the web to complete the task on your site. Watch where it hesitates, loops or gives up. Stop before any real payment or submission, and only test your own site.
Hour two: write the fix list. For each failure, record the page address, the task it blocked and a guess at what it costs you in leads or sales. Rank the list by money, not by how easy the fix is. I have seen plenty of teams spend a week polishing something that had no effect on revenue.
What I would fix first, and what I would leave alone
Start with names and labels. Every button, link and form field should say what it does in plain words. Then take down unnecessary walls from public information, and move key facts out of images and into real text. Finally, simplify your booking or checkout form so each step is obvious on its own. If you run a shop, our guide to AI for e-commerce sellers pairs well with this, because clear product descriptions help people and agents alike.
Here is what I would not chase. Don’t adopt every new protocol or markup standard that appears; many will fade. Structured data is worth adding when it matches what is visible on the page, but inflated or inaccurate markup does more harm than none. And use native HTML elements before reaching for ARIA attributes, because badly applied ARIA can make things worse. If your site came out of an AI website builder, test it like any other, since AI-built sites don’t guarantee clean structure.
Decide your policy before the agents decide for you
Being ready doesn’t mean letting everything in. I think of three stances. Welcome agents on public, informational pages such as products, prices, hours and guides. Limit them on actions that change something, with logins, rate limits and confirmation steps. Block private areas outright.
Two cautions. First, a robots.txt file only works on software that chooses to respect it, and an agent acting for a user may not behave like a traditional crawler, so read each vendor’s documentation. Second, check your security plugin and CDN settings. Bot protection tuned to block everything non-human can quietly block the helpful visitors too, including the ones that cite you. If you want to know whether you are being cited today, follow the steps in our guide to checking whether your site appears in AI Overviews.
Remember too that an agent working for a customer may handle personal data on your forms. Our privacy checklist for AI tools is a good companion read for deciding what you collect and keep.
What to expect, honestly
Agent-driven traffic is small for most sites today, and nobody can promise you a return on it. The case for acting now is that the same fixes help everyone else: visitors using assistive technology, people on slow phones, and search engines. That is why I treat this as low-risk work with compounding returns. For the broader playbook, see how to get featured in AI search results and the GEO strategy that increased my AI visibility.
Quick answers
Do I need special code for AI agents?
For the basics, no. Clean HTML, clear labels and accessible forms cover most of what agents need today.
Should I block AI agents?
For public information, usually not. For anything that changes data or handles payments, protect it with logins, rate limits and confirmations.
How will I know the fixes worked?
Repeat your three-task test once a month and keep a log. If a task that failed last month now completes, you have your answer.
Your next step
Pick the one task that matters most to your business and run the three tests on it this week. Write down every point of friction, fix the worst one, and test again. A website that works for a screen reader, a distracted human and an AI agent is simply a well-built website, and that is the kind that keeps winning whatever arrives next.
