# Why “web agency” is getting too narrow > A website is becoming a knowledge and service interface for people and AI agents. Why the term “web agency” no longer covers that — with evidence. “Web agency” is getting too narrow because the term describes a trade — someone designs and builds pages that people look at in a browser — while what actually comes out of it is more and more often something else: a digital knowledge and service interface for people and for AI agents. That's my thesis, and I'm writing it down here because by now I find it again in almost every project. I'm deliberately framing this as an observation, not a prophecy. I have no year from which the term supposedly stops working, and I have no time for pieces about the future that are really sales pieces. What I do have are three developments that can be evidenced — and one consequence from them that I consider practically relevant. ## What I mean by “too narrow” “Too narrow” doesn't mean wrong, it means incomplete: the term cleanly describes a craft, but it says nothing about whether the result is also readable and usable for machines. It answers the question “who builds this?”, not the question “what is the thing being built, actually?”. This isn't a criticism of colleagues who do excellent work. Terms narrow all by themselves when the thing behind them expands. At some point “switchboard” no longer described what was going on in a company's communication — not because people had stopped making calls, but because so much had been added that the word only covered a fragment. “Web agency” is in exactly that phase right now. ## Where the term comes from — and why it fitted for a long time The term comes from a time when a website had exactly one audience: people with eyes. Search engines were merely the route there, a sorting office that ordered links. Anyone who made pages attractive, fast and findable had done everything that mattered — and for that, “web agency” was a perfectly precise word. That phase isn't over, it has only gained company. I traced this development as a whole in [Four phases every website is going through right now](https://btlabs.dev/en/posts/four-phases-website) — from the web for humans, through the hybrid web we are in today, to scenarios I explicitly flagged there as speculation. This article picks out exactly one spot from that: the point where language lags behind reality. ## Why AI agents are a real audience, not a footnote AI agents are a real audience because the user numbers behind them now run into the hundreds of millions: ChatGPT alone counts more than 900 million weekly active users — OpenAI's own figure, as of 27 February 2026. That is no longer a niche you serve on the side. The second figure is more uncomfortable and therefore more important. Cloudflare measures continuously how many pages an AI crawler fetches before it sends a single visitor back. In July 2025 the ratio stood at roughly 38,000:1 for Anthropic, roughly 1,091:1 for OpenAI and roughly 195:1 for Perplexity. That is a snapshot, the values swing widely depending on period and sector, and they will keep shifting. The direction is unambiguous all the same: your content gets read by machines far more often than machines send visitors back to you. For how a website is judged, that has an awkward consequence. Measure it by visitor numbers alone and you are measuring an ever smaller part of its effect — the recommendation an AI makes appears in no traffic statistic, yet it works. Why fewer clicks don't automatically mean less business, I have worked through in [Fewer Clicks, Better Customers](https://btlabs.dev/en/posts/fewer-clicks-better-customers) — with the arithmetic laid out. ## From reading to acting: the point where “building pages” stops being enough The real turning point isn't that machines read along — they have done that since the first search engines — but that they are starting to act, and that there is now a cross-vendor, open standard for it. The Model Context Protocol (MCP) is an open interface specification through which an AI system doesn't just read text but can call defined functions of a system. The path this standard has taken says more about the direction than any forecast: Anthropic published MCP in November 2024 and handed it over on 9 December 2025 to the newly founded Agentic AI Foundation under the umbrella of the Linux Foundation. Its platinum members include Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI. When companies that compete on almost everything have a protocol of all things managed jointly and neutrally, that's rarely a coincidence — it's a hint about where they expect the next bottleneck. And an honest qualification belongs here: an agent negotiating, comparing or booking on your behalf is not the normal state of things today, it's a build-out scenario. With us as with everyone else. I'm describing a direction, not a product promise. What already is everyday reality is the stage before that: machines read your facts and pass them on in their own words. ## What “digital knowledge and service interface” actually means A digital knowledge and service interface is a website that does two things at once: it provides reliable facts about a business in a way that people and machines understand identically (knowledge) — and it offers defined, controlled entry points through which something can be triggered (service). The term sounds clunkier than “website”, but it captures more precisely what is actually being built. The knowledge part is the unspectacular but decisive one: opening hours, services, responsibilities, locations, conditions — maintained in a single place and additionally published in machine-readable form, for instance as structured data following Schema.org or as a short profile in an llms.txt. For people, nothing visible changes. For a machine, everything changes: instead of guessing, it can quote. The service part is the one where most people immediately think of a chatbot — and that is exactly the wrong turn. It isn't about a little window in the bottom right that answers questions, it's about clean, permission-limited interfaces through which an authorised system may carry out one specific task. Why those are two completely different things is set out in detail in [“Agent-Ready” Doesn't Mean “Install a Chatbot”](https://btlabs.dev/en/posts/agent-ready-doesnt-mean-chatbot). ## Why people don't become an afterthought in all this People stay on equal footing because both audiences are served from the same source: there is no “AI version” of the truth sitting next to the human one. When an opening time changes, it changes for the person on the page and for the language model in the answer at the same moment — otherwise you have two truths, and the second one is always the outdated one. That's why I think the opposition “beautiful for humans or readable for machines” is the wrong framing. Both are the same work, done once visibly and once invisibly. Which is exactly why, in btlabs Core, the machine-readable side isn't an add-on module you book later but the baseline state: the same maintained facts feed the presentation for people and the interfaces for machines. Not because that's a feat, but because anything else drifts apart sooner or later. What incidentally doesn't become less important in all this is design. A person still decides within seconds whether they trust a business to do good work — and that decision runs through images, language and a calm layout, not through structured data. The second audience is added; it doesn't replace the first. ## An example that makes the difference tangible The difference becomes most tangible in a simple process almost every business knows — an enquiry about a service you offer. The following example is made up, but typical of what we see in projects. Two businesses in the same area, both with a well-maintained, well-designed website. Business A describes its offering in carefully written prose: history, values, services, all nicely phrased. The opening hours sit in the footer as a graphic, the responsibilities are tucked into a sentence in the middle of the text. For people, that works well. Business B has the same information stored in addition as clearly delimited facts: service, location, hours, contact person — each as a self-contained, machine-readable entry, plus paragraphs that answer a typical customer question in a single, self-contained sentence. To people, the page looks almost identical. Now if someone asks an AI, “Who does this in my area, and are they open tomorrow?”, here is what happens. With Business B the system can lift the information straight out, reproduce it correctly and name the business. With Business A it has to interpret — it can't read the graphic with the opening hours, and the responsibility is buried in a subclause. At best the answer stays vague; at worst Business A drops out of the recommendation entirely. Not because the work is worse, but because it isn't graspable for a machine. The point of this example isn't the technology. The point is that in both cases the same “web agency service” was delivered — designed, fast, well-maintained pages — and still only one of the two businesses shows up in the answer. That gap is exactly what the term doesn't cover. ## Why this matters earlier in South Tyrol and the DACH region than elsewhere In South Tyrol and the DACH region this shift hits small, specialised businesses particularly early, because a lot of value here rests with providers who have barely any marketing budget but genuine professional substance. Someone who has built a good reputation in their own valley over decades gets recommended by people — but by a machine only if the facts about them are written down somewhere in machine-readable form. Then there is multilingualism. A question to an AI can be asked in German, Italian or English, and the answer draws on whatever is available and comprehensible in that language. A business whose facts are in good shape in all three languages can be cited in three language regions — one with a translation pushed through afterwards often only in one. Why that's more than a translation question is covered in [Three Languages, One Citability](https://btlabs.dev/en/posts/three-languages-one-citability). Both together create a situation I consider unusually favourable: the advantage here does not lie with the biggest advertising budget but with whoever maintains their facts most cleanly. That is a discipline in which a small business can genuinely be better than a large one — and one of the few points where digitalisation favours the small players for once instead of putting them at a disadvantage. ## What this changes about choosing a service provider What mainly changes in the selection is the list of questions — regardless of whether the business card says “web agency”, “digital agency” or something else entirely. I find four questions more telling than any portfolio: That the starting question itself is often the wrong one, [I have set out elsewhere](https://btlabs.dev/en/posts/getting-a-website-made-wrong-question). - **In how many places do I have to enter a change?** If a changed opening time has to be updated in five places, sooner or later it will be wrong in three of them — and it's precisely the wrong version that ends up in an AI answer. - **Can I see what a machine reads from my site?** Structured data and a machine-readable profile are verifiable, not a matter of taste. Anyone who can't show you them probably doesn't have them. - **Who owns the content, the data and the access?** Data sovereignty isn't a contractual formality, it's the precondition for being able to control your own facts at all. - **How is success measured when answers happen without a click?** Visitor numbers alone no longer answer that question. A classic web agency can answer these questions just as well as a freelance collective like ours. The point isn't the label, it's whether the questions get asked at all. Why we don't describe ourselves as a classic web agency is on our page [Web agency Bolzano — the alternative](https://btlabs.dev/en/web-agency-bolzano). If you want to know where your own site stands on these four questions today, [get in touch](https://btlabs.dev/en/contact) — that's a topic for a short, concrete conversation, not for an off-the-shelf quote. ## What I am explicitly not claiming I'm not claiming the term “web agency” will disappear. Job titles are stubborn, often outlasting what they originally meant, and that's fine. All I'm saying is: it describes an ever smaller part of what's needed. I'm also deliberately not naming a year. “Over the next few years” is as precise as I can honestly be. Anything more exact would be made up. And I'm not claiming that acting agents are standard today — they are not. Anyone who tells you otherwise is selling you a picture of the future as a finished product. ## How I'd know I'm wrong A thesis that can't fail isn't a thesis — so let me say what would change my mind. First: if AI answer engines fell back to mainly handing out links instead of formulating answers. Then the old division of roles would be back, and “web agency” would be the right word again. Second: if the open-standard approach fell apart — if every large provider went back to its own closed interface after all. Then the idea of a general service interface would be dead, and what remained would be a patchwork that smaller businesses would sensibly ignore. I'm watching both. At the moment I see the opposite: more shared standards, not fewer. Until that turns, the thesis stands. That is why I prefer to call what we build a digital foundation rather than a website. Not out of fondness for labels, but because the word already accounts for the second audience, which has been here for a while. “Web agency” isn't becoming wrong. It is just becoming too small for what is actually being built. ## Sources & evidence - [OpenAI — Scaling AI for everyone (more than 900 million weekly active users)](https://openai.com/index/scaling-ai-for-everyone/) (OpenAI's official figure, 27 February 2026). - [Cloudflare — The crawl-to-click gap: Cloudflare data on AI bots, training, and referrals](https://blog.cloudflare.com/crawlers-click-ai-bots-training/) (crawl-to-visitor ratios, monthly figures for July 2025). - [Linux Foundation — Formation of the Agentic AI Foundation](https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation) (founded on 9 December 2025, MCP as a founding project, member list). Note: the crawler ratios above are a snapshot from the summer of 2025 and vary considerably between providers, sectors and measurement periods. I quote them as an order of magnitude for an imbalance, not as a stable metric. --- Source: https://btlabs.dev/en/posts/why-web-agency-is-too-narrow Last-Modified: 2026-09-01T18:36:12.995Z Languages: [de](https://btlabs.dev/llms/de/posts/warum-webagentur-zu-eng-wird) · [it](https://btlabs.dev/llms/it/posts/agenzia-web-termine-troppo-stretto) See also: [llms.txt](https://btlabs.dev/llms.txt) · [ai.txt (Policy)](https://btlabs.dev/ai.txt) · [identity.json](https://btlabs.dev/identity.json)