Is your website ready for AI agents?

What is an agent-ready website?

AI agents can research products, compare suppliers and shortlist services. In some US services, they can also complete a purchase.

How much support your website needs depends on what sits between its front door and its money button. A retailer needs product data, live prices and an operable checkout. Most UK service businesses need something simpler: a site an agent can read, price and use to reach an enquiry form.

That is what agent readiness means in practice.

What changed for AI agents in 2026?

The large technology companies spent 2026 testing different places to put the transaction.

OpenAI introduced Instant Checkout inside ChatGPT, then retired it on 24th March 2026. ChatGPT now focuses on product discovery and sends the shopper to the merchant's checkout. Google launched checkout in AI Mode for US shoppers and plans wider UK shopping features, although it hasn't given a date.

Hey Savi and PayPal launched a standalone shopping app with in-app checkout on 2nd June 2026, with Debenhams Group as its first retailer. Microsoft and Amazon expanded their shopping tools, while Anthropic published an open-source commerce blueprint for retailers building their own agents.

For most UK businesses, the journey still begins with discovery and comparison in the conversation. The person then reaches your website, which must make sense to both of them.

Search results are becoming more interactive too. Google confirmed on 28th August 2026 that some AI Overviews can expand into longer AI Mode-style responses. Many important actions still finish on the website.

My earlier article explains why usage and referral are different things. Answering a question and completing a task are also different jobs. A website can support one brilliantly and make a mess of the other.

How does an AI agent read a website?

An agent can interpret a screenshot, inspect the raw HTML and use the accessibility tree. The accessibility tree is the structured version of the page that tells assistive technology what each element is, what it is called and whether it can be operated.

Agents usually combine these views. A button might look obvious in a screenshot but disappear as a control if it is really a styled <div>. A price might appear after JavaScript runs but remain absent from the raw HTML. A transparent cookie layer can cover an otherwise correct control.

Google's guidance on agent-ready site design is reassuringly practical:

• Make every action needed by a person or agent clear in the interface.
• Use real links and buttons instead of making generic elements behave like controls.
• Label form fields so an agent can match information to the correct input.
• Keep controls large enough to target reliably.
• Use the expected pointer cursor for clickable elements.
• Remove invisible overlays and other elements that block interaction.
• Keep layouts consistent between pages, so equivalent controls stay in equivalent places.

The Lighthouse and audit checks add accessible names and layout stability to that list. Together, these improvements support accessibility, content structure and dependable frontend behaviour. They also help hurried humans using a phone on a train.

Lighthouse's Agentic browsing category was turned on by default in version 13.3.0 on 7th May 2026. It checks areas including the accessibility tree, layout stability, llms.txt and experimental WebMCP tools and form annotations. The result is a pass ratio rather than a single performance score.

Does my website need to be agent-ready?

If a customer can complete the main transaction without speaking to you, your site is transactional. Retail, ticketing, accommodation and some appointment services belong here. The agent needs accurate data, current prices, availability and controls it can operate through to checkout.

If the sale needs a conversation, your website is a lead-generation site. Agencies, consultants and much of B2B belong here. The agent needs to understand the offer, compare it with alternatives and help the person make contact.

You can make the decision in two minutes:

1. Identify the action that directly produces revenue or a qualified lead.
2. Follow the shortest route from the homepage to that action.
3. Note every fact, choice, overlay, form and handover on the route.
4. Decide whether an agent must complete the action or prepare the human to complete it.

The more that sits between the front door and the money button, the more the agent must understand and operate.

What did 50 UK agency homepages reveal?

I tested the homepages of 50 UK agencies on 15th September 2026. They were the same agencies included in the August Best in Show study.

The rubric is my own. It draws on Google's web.dev guidance and the Lighthouse agentic-browsing checks. Each site received one desktop test on one day, so the grades are useful bands rather than permanent verdicts.

The mean score was 67.8 out of 100 and the median was 67. No agency received an A. There were 19 Bs, 30 Cs and one D, while 43 sites had at least one critical failure.

The strongest five were:

1. Wish Agency, 85
2. Blue Array, 84
3. Digital Next, 83
4. First Place SEO, 80
5. Builtvisible, 80

They deserve a straightforward well done.

Across the full group, agents could generally interpret the pages better than they could operate them. The average Perceive score was 73.8 and Extract was 71.7. Act fell to 60.1.

The most common critical problem was clickable <div> elements without a role or keyboard access, found on 20 sites. The next most common failures were sampled controls producing no change the test could observe on 18 sites and a missing <main> landmark on 17. The sandbox activated up to 12 sampled controls on each site.

Small targets were more widespread: 16 sites failed the check, 34 received a warning and none passed cleanly. Eight homepages used placeholder text instead of proper form labels. A person can often infer that a blank box wants a telephone number, while an agent using the accessibility tree may receive only an unnamed input.

These aren't unusual web problems. The WebAIM Million 2026 found missing labels on 51% of one million home pages, while only 46.1% had a main landmark.

Cookie consent overlays count as occlusion because an agent meets them on arrival. Two agencies lost most of their score to the consent layer alone. Another site set its custom cursor to none, removing an actionability signal that Google's guidance says agents use.

Content legibility didn't predict operability. For the Best in Show and agent-readiness scores, ranking the 50 agencies by each score gives two lists that barely overlap. (Spearman's rank correlation: rho 0.113, p 0.43, n 50, where a value near zero means no relationship, for the statistically inclined.)

The agent-readiness top five ranked 16th, 11th, 36th, 18th and 20th in Best in Show. None reached its top 10, and the third-placed site here ranked 36th there.

Half of the agencies had an llms.txt file, up from 23 in August. It remains the file many businesses build first and, as my earlier llms.txt test found, it is rarely the first problem they need to solve.

Dog on the Table scored 95. I wrote the rubric, so anything much lower would have required an uncomfortable meeting with myself.

Why do agents need you to publish prices?

A shopping agent can compare the price of men's Adidas Continental 80s in size 9 because retailers expose products, variants, stock and prices.

Ask what a local agency charges to rebrand a restaurant website and the evidence becomes much thinner. The agent must give a vague answer, rely on general market estimates or omit suppliers that don't publish enough information.

Of the 50 agencies, 10 published a price for their own services. Six offered only a general industry range, a calculator or another ballpark. The remaining 34, or 68%, published no price.

One site had a page called “Our Pricing” without any prices. Another /pricing/ address returned a JPEG.

This was a text-based scan followed by manual checks of flagged pages. It could miss prices stored only in images, PDFs or JavaScript content. These figures describe one run on 15th September 2026.

Keeping everything vague and secretive could lead to accusations of making up prices based on what you think you can get away with.

A service business doesn't need a fixed menu for every possible project. A useful article might say that rebuilding a 20-page brochure website, with a defined set of deliverables, usually costs £12,000 to £16,000. It avoids tyre kickers and time wasters, and feeds the robots that may well be sharing that information with interested parties.

I somewhat boldly (perhaps even foolishly) publish my own prices. “From” prices, day rates and sensible ranges can qualify enquiries while showing existing clients that the pricing has a basis.

Can a transactional site still stop an agent?

Being a shop doesn't automatically make a website agent-ready. Across the separate tests, 29 UK shop homepages averaged 59.7 and six clinic homepages averaged 66.7.

Gardening Express had the strongest shop homepage result, scoring 83 with no critical failures. Its product page exposed prices and stock in both the raw HTML and Product schema. The homepage carried Organization schema, while its server-rendered content matched the browser display.

The homepage and sampled product page used main and navigation landmarks. The homepage's interactive elements had accessible names. Its 8 of 11 sampled controls produced no observable change, however, which is why the audit praises real named buttons rather than claiming every button works.

On the sampled Acer product page, the add-to-basket control was a real named button inside a real form. The page scored 56 because 10 quantity steppers and product pagers were clickable spans without roles, one checkbox lacked a label and 4 of 10 sampled controls produced no observable change. The cookie dialog also covered the quantity box and basket button on arrival.

A private cosmetic clinic showed the opposite pattern. Its server-rendered pages had useful MedicalBusiness, MedicalProcedure and FAQ schema, so an agent could extract plenty of information.

It couldn't do much with it. A cookie layer blocked the navigation, clickable cards were absent from the accessibility tree and more than a dozen controls had no accessible name. The enquiry form relied on placeholders for fields including telephone, postcode and message.

The sampled treatment page gave no price, although some other pages on the site published “from” prices. A booking agent could read the treatment description but couldn't reliably navigate, price or complete the enquiry. Strong structured data didn't rescue the journey.

What should happen when bot protection blocks the form?

For a lead-generation site, the goal is to get the agent to the form with a well-informed enquiry.

Bot protection may stop the final submission, and sometimes it should. The agent can still gather the information, draft the message and open the correct form. The person checks the message and submits it.

This works when the route is clear. Use a direct link, visible service facts and properly labelled fields. Don't make the agent dismiss several overlays, decode unnamed inputs and guess whether a decorative box is secretly a button.

Marketing has called this work conversion rate optimisation for 20 years: reduce resistance between arrival and the intended action. We don't need another three-letter acronym beside SEO, GEO and AEO. ARO, AFO or Please Let the Robot Through can remain safely outside the strategy document.

What should you improve this quarter?

Transactional sites should start with the buying path. Put product names, variants, prices and availability in the HTML and appropriate schema. Use semantic controls, accessible names and a stable layout. Test cookie notices, product selectors, basket controls and checkout through the accessibility tree as well as the screen.

WebMCP is worth watching. It is a proposed browser API that lets a site expose functions such as search, add to basket or book a slot as named tools. It is currently a W3C community draft and Chrome origin trial, not a shipped standard. None of the 50 agencies used it.

If you run a complex application with important forms, an experiment may be useful. For most sites, repair the accessibility tree first.

Lead-generation sites should publish important facts and realistic price guidance in ordinary HTML. Add one clear <main> landmark, use real links and buttons and give every form field a persistent label. Add llms.txt if you find it useful, but don't let it jump ahead of controls an agent can recognise.

What can an agent do on Dog on the Table?

There are only two meaningful commercial actions on this site: send an enquiry or buy the £4,000 AI visibility audit. I don't expect anybody to hand a £4,000 decision entirely to a robot.

I want an agent to understand what I offer, why it matters, what the work includes, how the services fit together and what they cost. It can then help a person decide whether a conversation makes sense.

The free tools use bot protection. Limited access for agents working for real users is not out of the realms of possibility, but it isn't the first job.

The first job is to make the route readable, priced and reachable. You can use the free Agent-Readiness check to test your homepage and find where that route breaks. You can use the contact form to ask me to fix it for you.