Insights · Trend · AI visibility
An AI assistant reads your website. But does it have anything to do there?
AI visibility starts with a model understanding what you actually do. We can already measure and improve that. But there is a further threshold few people talk about: an assistant can not only read your offer, it can perform a concrete action in your system. The standard for that is MCP. Here is what is changing and what to require to get it right.
What is changing right now
For twenty years we optimised websites for browsers and for search engines. Today a third audience is growing: the AI assistant that answers the customer instead of serving a list of links. The first step is an llms.txt file, a readable description of your business for models. The second step is different in kind: MCP is the standard an assistant uses to connect to a system and perform operations in it. The difference is that between a leaflet and a counter. A leaflet can be read. At a counter something gets done.
What it looks like in practice
An MCP server exposes a list of tools, that is, named actions with a description and a set of fields. Some are reads and usually work without any key: show the offer, check free slots, find a company in the directory. Some are writes and require a key generated by the account owner: post a listing, update a description, extend a record. The assistant does not guess how to fill in a web form. It receives the field definitions and the rules, and the system validates them exactly as it does for a form filled in by a person.
What to require from your vendor
Six questions that will save you rework
- Start with reads. Expose what customers ask about most: offer, prices, hours, free slots. The effect is immediate and the risk is nil.
- Require that writes through an assistant follow exactly the same rules as the form in the panel. Two parallel write paths mean two versions of the truth and certain trouble.
- Ask about the change log: who, when and through which path. Without it you cannot reconstruct what a person did and what an assistant did.
- Decide up front what the assistant must never do. Irreversible operations stay with a person, not a model.
- Ask about limits and how access keys are stored. It is the first question an auditor will ask anyway.
- Make sure user-entered content is treated as data, never as instructions for the model.
Proof from our work
We are not writing this from theory. We built such a server into the city portal we are launching for Otwock county. Forty-odd tools: reading the business directory, events and job offers without a key, and, for a listing owner, full management of their content using a key generated in the panel. The assistant posts a job, publishes an update, fixes a description, extends a record. Write rules are shared with the web forms, every change returns as a field-by-field diff, deletion is soft and confirmed, and the listing itself cannot be removed through the API at all. Every record carries its source, so you can see whether a change came from the panel or from an assistant. The call log itself we deliberately keep aggregated — tool, day, counter — with no request contents and no IP addresses. A conscious trade-off: less data in the log means less that can ever leak. We described the whole build in the Wirtualny Otwock case study. We are moving the same pattern into further systems, because there is only one hard dependency: the account and the data model you already have.
Frequently asked questions
How does an MCP server differ from the API I already have?
Technically it is still an API, only described in a way an assistant understands without developer documentation. The server itself advertises the available actions, the meaning of every field and which operations need a key. Connecting an assistant then means pasting an address and a key rather than running an integration project.
Is it safe if an assistant can write data?
Safety comes from the assistant holding no permissions beyond those of the account owner, with the same write rules as the web form. On top of that come key scopes, operation limits, a change log recording which path a change came through, and soft, confirmed deletion. Irreversible operations are simply not exposed.
I run a small business. Where do I start?
Start by tidying the data you already have: offer, prices, opening hours, availability. Then expose reads only, because that gives the largest effect for the smallest cost and carries no risk. Writes come later, once you see what assistants actually ask about. If you do not know how you currently appear in model answers, begin by measuring your AI visibility.
Check whether AI assistants have anything to do at your company
We start with a measurement: how models describe your company today and what they do not know about it. Then we decide what is worth exposing for reading, what for writing, and in what order.
