Somebody asked an assistant about your product this week. It answered. You were not consulted, you did not see the question, and you have no idea whether what it said was true.
This is the part of the AI shift that most companies have not priced in yet. The concern everyone talks about is AI writing content. The more consequential change is AI reading it — and then speaking on your behalf to someone who will never visit your site.
What an agent does with your website today
An assistant asked "does product X support single sign-on?" does roughly what you would do: finds your site, reads what is there, and forms an answer.
What is there is marketing copy. Marketing copy is written to persuade, which means it is written in the register of implication. "Enterprise-ready security" implies SSO without asserting it. "Works with the tools you already use" implies an integration that might be on the roadmap. A human reader discounts this automatically — they have spent their whole life reading advertising. A language model reads it as a description of the world and reports it back with the confidence of a spec sheet.
So the failure mode is not that the AI lies about you. It is that it takes your marketing literally, and then a prospect makes a decision based on a capability you do not have.
The gap between what your site implies and what your product does used to be absorbed by human scepticism. Machines do not have any.
The Agent Guide
The Agent Guide is a structured, machine-readable statement of what you actually offer and what you actually claim, published from the same CMS that runs your site and served to agents over MCP.
It has three parts, and the shape of them is the whole idea.
Products. What you sell or provide, described in structured fields rather than prose. Not a paragraph an agent has to parse, but a record it can read.
Claims. Specific assertions you stand behind — that a thing does X, supports Y, is certified for Z. Each is a discrete, addressable item rather than an adjective buried in a sentence.
Evidence. What backs each claim. A documentation link, a certification, a benchmark, a source. A claim without evidence is an opinion, and the structure makes that visible rather than letting confident phrasing paper over it.
An agent connected to a Birkly instance can list your published products, fetch one with its resolved source links and linked claims, and retrieve a specific claim by ID. It gets a citable statement, attributable to you, instead of an inference from a hero headline.
Why this is a CMS feature and not a file
The obvious objection: this is just a well-known JSON file. Why does it need to live in a content management system?
Because the alternative rots. Anything maintained separately from the content it describes drifts out of date the first week someone is busy, and the further it drifts the more actively harmful it becomes — a stale machine-readable claim is worse than none, since it is stated with structure and therefore trusted more.
In Birkly, the Agent Guide is built from the same collections your site renders from. Products in the guide are the products in your catalog. Claims link to documentation entries that exist. When you update the docs, the evidence link still points at the thing it described, because it points at an entry rather than a URL you remembered to update. The guide is a view of your content, not a copy of it.
There is a second reason, which is the reason it is in the core rather than a paid plugin: this only works if it is normal. A structured claims format that a handful of companies pay extra for is a curiosity. One that ships with the CMS and costs nothing is infrastructure.
What it does not do
Some honest limits, because this is an area where it is easy to promise a great deal.
The Agent Guide does not make an AI believe you. There is no mechanism by which publishing a claim compels a model to repeat it, and any vendor implying otherwise is selling something. What it does is remove the excuse for guessing: an agent that has your structured claims available and still invents a capability is making an error it did not have to make.
It does not verify anything either. If you publish a claim your product does not support, the guide will faithfully transmit your falsehood in a machine-readable format, which is a considerably worse outcome than saying nothing. The evidence field exists to make that uncomfortable, not to prevent it.
And it is not a substitute for documentation. It is a layer above documentation that says what your docs assert and where to find it.
The bet underneath
We think a meaningful share of the questions people ask about products in five years will be answered by an intermediary rather than by a page. Whether that share is thirty percent or eighty, the direction is not seriously in doubt.
If that is right, then every company faces a choice it has not consciously made yet: let the intermediary infer what you offer from copy that was written to persuade humans, or state it in a form the intermediary can read directly. The first option is the status quo and requires no work. The second requires a small amount of structure and gives you back the ability to be quoted accurately.
That is the Agent Guide. It ships in the core, it costs nothing, and it will be open source at release. See how it fits into the product, or join the waitlist.