Are you agent ready?

Don Norman has said he coined the term user experience at Apple in 1993 because the words then in use, human interface and usability, were too narrow for the work. The discipline that grew from it is now one of the best resourced functions in digital business: research teams, usability labs, design systems, experimentation platforms, and a head of experience on most leadership teams. Any serious organisation can tell you its conversion rate, where the journey leaks and what its customers said afterwards.

Ask the same leadership team what an agent experiences when it arrives at the site on a customer's behalf, and the room goes quiet. Not from complacency. Nobody has looked. There is no team, no metric and no dashboard, and the phrase agent experience is less than two years old.

The human experience has been designed on purpose for three decades. The agent experience is whatever is left over.

Left over is not nothing. It is the document the server sends, with the pictures, the scripts and the styling stripped away. Every website already has an agent experience. It has simply never been seen by the people who own the site, which makes it the one experience nobody designed.

It helps to be literal about this, because most of the mystique around AI search and agentic commerce dissolves once you watch what the software actually does.

An agent working for a prospective customer does not open a browser. It fetches the page with a plain HTTP client. No JavaScript runs, no images load, no stylesheet applies. What arrives is the document the server sent, and if the words on the page are drawn later by a script, as far as this reader is concerned they were never there.

Before it reads anything it checks whether it is allowed in, and whether there is a map. Then it reads, and it pays for every token, so navigation, banners and boilerplate are a cost before they are a nuisance. It looks for structured data to work out who you are. Then it tries to answer the questions the person asked: what does this organisation do, who is it for, what does it cost, how do I buy or get in touch. And if the person wants to proceed, it looks for something it can act on: an address it can write to, a form it can submit, a link it can follow, an interface it can call.

A form drawn by a third-party script is, to this reader, a blank space. A price that lives only in a PDF is a price that does not exist. A beautiful page with the substance in an image is an empty page.

None of this is hidden. All of it is unfamiliar, because nobody on the human side of the screen has ever had to read a website this way.

So we built something that does, and writes down what it finds. agentready takes a web address and a kind of site. It fetches what an agent would fetch, in the order an agent would fetch it: the robots file, the sitemap, the index, the home page and a handful of linked pages, probing as it goes for markdown versions, structured data, forms and any documented interface. A model then reads the home page text alone and writes down what it takes the business to be, and what a prospective customer would still want to know. That paragraph sits at the top of the report. If it is wrong, every answer an agent gives about you is wrong in the same way, and that is worth knowing before anything else.

The rest of the report is twenty checks in five areas. Each check is a page of plain English that tells the assessor what to look for in the evidence and what a score from zero to four means. The model judges. The arithmetic is done in code. Choosing the kind of site changes the weights and switches off the checks that make no sense for it: a consulting firm is not marked down for lacking an API description, and a shop is not expected to publish its catalogue as a single text file.

Each area scores out of a hundred, and the areas combine by weight into one number. Seventy-five and above is strong, fifty is fair, twenty-five is weak, and below that is poor. The number is the least interesting thing on the page, for reasons I will come to.

It would be odd to publish a measuring stick without standing against it. On 20 October we scanned theoverstorygroup.com as a services site. The scan read the home page and three linked pages, made eighteen requests to the site and took twenty-four seconds.

Seventy-five overall, which is strong by exactly one point. Four of the five areas sat between seventy-eight and eighty-five. We built this site with the second audience in mind, with a markdown twin for every page and an index for agents from the first day, so that much was expected.

The paragraph at the top of the report is the part worth reading twice.

It got us right, from the text of one page, with nothing else to go on. Where it marked us down is more useful than where it did not.

The gap between building for the second audience and being found by it is made of small things nobody owns.

Three items went on a list that afternoon. None of them needed a workshop. All of them needed someone to look.

Treat the report the way you would treat a first usability test. It is not a verdict. It is the first time anyone has watched this customer try to use your site, and the useful output is a list, not a score.

Some honesty about the instrument. The assessor is a language model, and if you scan the same site twice a few checks will move by a point. It reads a handful of pages, not the estate. It measures whether you are legible to agents, not whether agents are in fact citing or choosing you, which needs measurement most organisations do not yet have. It is a reading, not an audit, and it is free because the conversation it starts is the point.