Your next customer will never visit your site.
Their AI agent will.
The internet is shifting from humans browsing websites to AI agents that search, compare, book and buy on their behalf. Businesses that are unreadable to those agents will vanish from the shortlist — without their traffic ever dropping.
Written by Manuel Cilurzo · AI consultant, Belgium · updated August 2026
The agentic web is the phase of the internet in which autonomous AI agents take over most of the searching, comparing and transacting from humans. The user states a goal — “book a table for six on Friday, Italian, within ten minutes’ drive” — and an agent works through the sources, weighs the options and completes the task. The website stops being a shop window and becomes a data layer.
On this page
What is the agentic web?
For thirty years the internet was built on one assumption: at the other end of the connection sits a human with a screen. Everything follows from that — navigation, cookie banners, beautiful photography, A/B tests on button colour, search optimisation to pull that human onto your page.
That assumption is dissolving. More and more digital actions are performed not by a person clicking, but by an AI agent: a model that receives a goal, consults sources itself, takes intermediate steps, uses tools and only returns with a result. Anyone who has asked ChatGPT, Claude, Gemini or Perplexity a question and received a composed answer with three recommendations has already lived through that shift. The ten blue links were skipped.
The next step is already under way: agents that do not just answer but act. They fill in forms, compare prices, place orders, book appointments and pay. At that moment, the visitor to your website is literally software.
Why it is tipping now
This is no longer a forecast. It is the sum of four developments that reinforce each other.
- Models can reason across multiple steps. A task like “find, compare, book” used to fall apart. Now it holds together.
- There is a standard for offering tools. With the Model Context Protocol (MCP), any company can expose its own systems to any agent, without building a separate integration per assistant.
- The interface is moving into conversation. Answer engines return one composed answer instead of a list. There is no page two where you can still be found.
- Payment by software is being normalised. Once an agent can transact safely, the last reason a human had to make the final click disappears.
What changes for your business
Today you optimise for
- A human who scrolls and clicks
- Position in a list of ten results
- Persuasion through imagery and design
- Counting visitors in analytics
- A contact form as the endpoint
- Beautifully written, vaguely stated
Soon you optimise for
- An agent that parses and summarises your page
- Being named in the answer, or not existing
- Persuasion through facts, figures and structure
- Counting mentions in ChatGPT, Claude, Perplexity
- A machine path an agent can book through
- Explicit, unambiguous, quotable
The painful part: you will not notice this in your traffic. An agent that reads your prices and then does not recommend you leaves almost no trace in your statistics. The first signal is usually that the phone rings less, with nothing visibly broken.
The agent ladder: four layers of agent readiness
Each layer rests on the one below. Building layer three without layers one and two means building tools nobody ever finds.
Findability
Are you retrieved at all when an agent looks for your kind of solution? This is the layer of answer engine optimization: being present in the sources models draw on, in language that matches the questions people actually ask.
Readability
Can a machine take your offer apart without error? Prices, opening hours, terms, what you do and explicitly do not do. This is the layer of structured data, schema.org, llms.txt and clean semantics — the agent-ready website.
Usability
Can an agent actually do something with you — check availability, reserve, order, get a quote? That requires a machine path into your systems, usually through your own MCP server or a well-documented API.
Trustworthiness
Why would a model recommend you over someone else? Consistent data across every source, a recognisable entity, demonstrable expertise and reviews that hold up. This layer takes longest and is hardest to copy.
The agent ladder is the model I use in every AI readiness audit. You are free to reuse it, with attribution to this page.
Test: is your business agent-ready?
Work through these ten questions. Every “no” is a point where an agent drops you and picks a competitor.
- Does your current price exist as a machine-readable fact, not only as text inside an image or PDF?
- Could a model explain in one paragraph what you do, for whom, and where?
- Are your opening hours, address and phone number identical across your site, Google, social media and directories?
- Does your site carry structured data (schema.org) for your services, prices and frequently asked questions?
- Is there an
llms.txtpointing language models to your most important pages? - Are you blocking AI crawlers in
robots.txtwithout ever having decided to? - Is there any way for software — not a human — to check availability or make a reservation?
- Is there a page that literally answers the questions customers ask, in their words?
- Is there a clear, repeated link between your company and a recognisable person with demonstrable expertise?
- Do you know today whether ChatGPT, Claude or Perplexity names you when someone asks for your service?
Anyone scoring below seven is effectively invisible the moment the request arrives through an agent. An AI readiness audit measures this point by point and sets a concrete action plan beside it.
Five expensive misconceptions
1. “Our SEO is good, so we’re fine”
Classic SEO optimises for ranking. Answer engines do not necessarily cite the highest-ranked page, but the most clearly stated, best structured and most consistent source. Good SEO helps, but it covers at most layer one of the agent ladder.
2. “We’ll add a chatbot”
A chatbot helps the visitor who is already on your site. The agentic web is precisely about the customer who never opens it. Different problem, different solution.
3. “We block AI crawlers, they steal our content”
A defensible choice — but understand what you are buying. You trade the risk of copied text for the certainty of never being recommended. For publishers that can add up. For a service business or a restaurant it is almost always self-harm. Decide this deliberately, not by accident through a plugin default.
4. “We’ll wait until it matures”
The layers that take longest — consistent data, entity building, reputation — are exactly the ones you cannot buy your way through later. This is a head-start game.
5. “Our sector is too personal for AI”
The agent does not replace your craft. It replaces the search for your craft. That is precisely the part where you win or lose customers today.
Frequently asked questions about AI agents and the agentic web
What exactly is the agentic web?
The agentic web is the phase of the internet in which AI agents search, compare and transact on behalf of people, instead of people visiting websites themselves. The user states a goal; the agent works through the sources and completes the task. For businesses it means their website is no longer only a shop window for humans, but also a data source for machines.
What is the difference between SEO and AEO?
SEO (search engine optimisation) gets your page high in a list of results. AEO (answer engine optimization) gets you named inside the answer an AI assistant composes itself. SEO wins a click; AEO wins a recommendation. AEO leans much harder on clear phrasing, structured data, consistency across sources and demonstrable expertise.
Should I allow or block AI crawlers?
For most service businesses, shops and restaurants, allowing them is the right call: being blocked means never being recommended. For publishers whose product is the text itself, the trade-off is different. What matters is that it is a deliberate decision — many sites block AI crawlers today without anyone ever having chosen to, through a plugin default.
What is llms.txt and do I need one?
An llms.txt is a file in your website’s root that tells language models, in plain language, what your business does and which pages carry the important facts. It complements — it does not replace — structured data. It is cheap to add and costs nothing if it goes unread, so the trade-off is usually easy.
What is MCP (Model Context Protocol)?
MCP is an open standard that lets an AI agent use tools and data sources. Instead of building a separate integration for every assistant, you publish one MCP server that any compatible agent can connect to. For a business it is the way to be not only read by agents but also used: checking availability, reserving, ordering.
Is this only relevant for tech companies?
No — it hits hardest in businesses where the customer journey starts with searching and comparing: hospitality, construction and trades, clinics and practices, retail, professional services. Precisely the sectors that invested in being found by humans, and precisely where the channel is now shifting beneath them.
How do you measure whether it works?
Three kinds of measurement: how often you are named in the answers of the major assistants to a fixed set of questions, how much traffic and how many enquiries demonstrably come from AI sources, and how complete and consistent your structured data is. I always take the first measurement before we start, so improvement is provable.
How long does an engagement take?
An audit produces a report and action plan within two weeks. The technical layers — readability and usability — are typically a matter of weeks. The layer that takes longest is trustworthiness: consistent data and entity building work over months. Which is exactly why starting early is the whole strategy.
Want to know what AI says about your business today?
I test your business in the major AI assistants and send you back what they answer — including what is missing. Free, no obligation.