What an AI agent is and how it differs from a search engine
Google's crawler reads your site to show it in search results. An AI assistant such as ChatGPT reads your site to answer a question about your company — I covered that in “What AI says about your company”. An agent goes further: it doesn't describe your site, it acts on it on a person's behalf. It opens the catalogue, picks a size, adds a product to the basket, types a name and phone number into a booking form.
This is no longer an experiment. In January 2026 Google launched auto browse in Chrome: Gemini works through multi-step tasks by itself — it finds products within a budget, fills the basket and goes all the way to checkout, then stops and asks the person to confirm the purchase. For now it is available to paying subscribers in the US, but the direction is clear. OpenAI, Perplexity, Anthropic and others have similar agents.
One important point: an agent is not a separate “robot shopper” you can block and forget about. Behind it is a real person with real money who simply delegated the routine. Cut off the agent and you cut off that customer.
How an agent sees your site
A person looks at the picture: colours, photos, a big button. An agent sees the page differently — more like a blind user with a screen reader. It relies on text, labels on buttons and fields, headings and page structure. Some agents also take screenshots, but even they do better when the meaning is in the code itself, not only in the design.
Hence the main takeaway of this article: a site that works well for an agent is almost always a site that works well for people. Especially for those using it on a phone on the go, with poor eyesight, or simply in a hurry. In most cases there is nothing to build “specially for robots”.
Seven places where agents trip up
1. Price “on request”
The customer asks the agent to “find where it's cheaper”, and instead of a price you have “ask our sales team”. A person might call. An agent simply drops you from the comparison: there is nothing to compare. If an exact price isn't possible, even a range helps — “from 1,200 PLN”, “from 90 PLN per hour”.
2. Key information only in images
Opening hours on a banner, a price list as a photo, delivery terms inside a graphic. A person will read it; an agent won't always. Everything that affects the customer's decision should be plain text on the page.
3. Buttons without labels
A basket icon, an arrow, a little plus — a person guesses from context. In the code such a button often has no name at all. The agent sees ten identical nameless buttons and can't tell which one is “Add to basket”. The fix is a label for screen readers — the same change accessibility law requires; more in the article on website accessibility.
4. Puzzle forms
A field with no label, just grey placeholder text inside. A phone number accepted in only one format, with the error shown as a red border and no explanation. Dropdowns drawn by a script instead of a normal list. A person manages on the second try; an agent may not. Simple rules: every field has a clear label, errors are written in words (“enter the number as +48…”), required fields are marked.
5. CAPTCHA at every step
CAPTCHA exists to keep bots out — and to it, an agent is a bot. Well-behaved agents won't solve it: they'll stop and ask the person. If there's a CAPTCHA on the order form, every such order needs the customer to come back to the screen. Spam protection is necessary, but it's better to choose options that don't get in the way at every step: invisible checks, rate limits on submissions, server-side verification.
6. Pop-ups on top of everything
A newsletter window after three seconds, a chatbot covering half the screen, a discount banner with no close button. A person irritably hunts for the cross. An agent wastes steps fighting the windows or can't reach the button underneath at all. Fewer pop-ups make life easier for people and agents alike.
7. A slow site
An agent is in no hurry, but it does have a time and step limit per task. A page that takes eight seconds to load risks the task being cut off halfway. How we measure speed is described in the article on testing with PageSpeed Insights.
What helps an agent
- Products and services described in markup. Schema.org structured data is the page's “passport” in machine-readable form: name, price, availability, rating, address, opening hours. Google has read it for years, and now AI services do too. Customers don't see it, but the agent doesn't have to guess.
- Clear addresses and structure. A “Delivery and payment” page at
/delivery, not a PDF tucked away in the footer.
- Up-to-date data in one place. If the price on your site says one thing, your Google profile another and the PDF price list a third, the agent will pick any of them — not necessarily the right one.
- Checkout without registration. A mandatory account means extra steps for a person and a dead end for an agent, which doesn't create accounts on the customer's behalf.
What about the new protocols?
There is a lot of noise right now around standards “for agents”. Google and partners are pushing agentic commerce protocols, OpenAI has its own, and Chrome is testing WebMCP — a way for a site to tell an agent directly: “here is the product search function, here is the booking function”. It's a good idea, but as of autumn 2026 WebMCP is an experimental Chrome feature, not an adopted standard, and among popular agents only Gemini is starting to use it.
My advice to business owners: keep an eye on it, but don't rush. If you run a large online shop on Shopify or a similar platform, the platform itself will most likely ship the necessary modules. If you have a services site or a small shop, tidying up the list above will do far more: it works with every agent at once and improves the site for people and for search along the way.
Block or let in?
Some site owners, seeing a rise in “robot” traffic, block everything in sight in their hosting or Cloudflare protection settings. It's easy to overdo: blocking bots that harvest data for training models is one decision; blocking agents acting for a live customer is quite another. Before switching on aggressive protection, check exactly who it cuts off. And if you yourself give an AI assistant access to your site or computer, I covered security in a separate article.
A 15-minute check
The simplest test is to give an agent a task and see if it copes. If you don't have one to hand, you can walk through the site “through an agent's eyes” yourself:
- Find the price of your main service or product on the home page. Is it written as text, or missing?
- Open the site on your phone and try to order or book without registering. How many steps and pop-ups along the way?
- Fill in a form with a mistake. Is it clear in words what went wrong?
- Turn on your phone's screen reader (VoiceOver or TalkBack) and go through the home page buttons. Does it name them meaningfully, or say “button, button, button”?
- Compare opening hours, address and prices on your site and in your Google profile. Do they match?
If the answer is “no” to at least two of these, agents are already getting lost on your site — and most likely some people are too.
In short
- AI agents already visit websites and act for customers: they search, compare, fill in forms and build baskets.
- An agent sees a site as text and structure, not as a picture.
- The main obstacles are hidden prices, information in images, nameless buttons, complicated forms, CAPTCHA and pop-ups.
- Almost everything that helps an agent also helps people and search.
- New protocols: watch them, but start by putting your site in order.
If you'd like to know how your site looks to AI agents and get a list of concrete fixes, see what we do in AI — or just write to us: we'll check your site and tell you honestly what's worth changing and what isn't.
