# How an AI agent reads your business: the four interfaces behind the answer > How AI agents read a business: machine-readable profile, structured data, text twin and controlled query — with first-party measurements and the honest limit. When someone asks an assistant about a business today, that assistant rarely opens the website the way a person sees it. It fetches what is machine-readable — and what it finds there decides whether your business appears in the answer or not. This is not speculation. On this website, AI and search systems triggered **23,520 requests** over the past 30 days. Among them, systems that clearly belong to assistants: ClaudeBot with 1,171 requests, GPTBot with 815, OAI-SearchBot with 476. ## What does an AI agent actually fetch? An AI agent does not fetch one page — it fetches four different things, and each answers a different question. The machine-readable profile answers "who is this and what is here?". The structured data answers "which facts apply?". The text twin answers "what does it actually say?". The controlled query answers "what is current right now?". Together, these four make up a company's discovery surface. If one is missing, the agent has to guess — and guessing is exactly where wrong recommendations come from. ## Interface 1: the machine-readable profile A machine-readable profile is a simple text file in the root directory that tells a language model, in plain form, what the site is about and where the important content sits. The proposal came from Jeremy Howard and was published in September 2024. The honest framing matters here: an Ahrefs analysis from 2026 found the file on 137,210 domains — and established that **97 per cent of them were never requested**. The file on its own therefore achieves little. It only works when what it points to is cleanly readable as well. ## Interface 2: structured data Structured data are machine-readable annotations in the page source that separate facts from their presentation: opening hours, address, services, authorship, publication date. A person reads "we're here for you Monday to Friday". A machine only derives a reliable time span from that if it is annotated separately as a date. That is the difference between a statement a model has to interpret and one it can simply read. Interpreting produces errors; reading does not. ## Interface 3: the text twin A text twin is the same page in plain text — no navigation, no layout, no scripts. For an agent, that is the difference between a document and a building site: it doesn't have to work its way through menus, consent dialogues and deferred loading, but gets the content directly. That saves it not just effort but time — and time is a scarce resource for agents. Anything fetching on a limited time budget gives up on slow or deeply nested pages and sees only a fraction. ## Interface 4: the controlled query The first three interfaces are static: they deliver what was true at the moment of publication. The fourth is a controlled query option — a defined access point through which an authorised system can ask specifically for current details, instead of inferring them from a page. The difference shows in everything that changes: availability, current services, contacts. Facts fetched statically age between two visits. A query is always current. The decisive part is the control. Such an access point is not an open gate but a door with a key: whoever uses it is known, and what they may do is defined. ## Why four interfaces and not one? Because no system uses all four the same way. Some assistants take only the text. Others follow the profile. Others read nothing but the structured data. Serving only one interface makes you invisible to precisely those systems that prefer another. It is the same reasoning as with multilingual sites: you don't build for the language you hope will be asked, but for all of those that occur. ## Not everyone claiming to be an assistant is one A detail from the same measurement that gets underestimated in practice: of 5,728 requests from a system presenting itself as a major provider's agent, **six** could be verified. With systems that genuinely belong to assistants, the ratio is the other way round — for GPTBot, 684 out of 815. Anyone opening their discovery surface should therefore know who they are opening it to. Identity is verifiable, and verifying it is the difference between openness and loss of control. ## The honest limit: being read is not the same as being recommended This website was fetched by machines 23,520 times in 30 days. In the same period, that produced **four** counted citations and **four** referrals. That gap is the most honest sentence in this article. A clean discovery surface makes sure an assistant has your facts right when it needs them. It does not make the assistant recommend you. That is decided by what you offer and what others say about you. Anyone promising you otherwise is selling you a number they cannot prove. ## How to tell whether it works for you Three checks you can run without tools. Open the text version of one of your most important pages and read it: is your offer there in full, or is half of it missing because it loads later? Ask an assistant about your business and check whether the facts it cites are correct. And look at whether your opening hours and address exist as annotated data — or only as running text. If you find something that doesn't hold up, it is not a detail: a machine repeats a wrong value until it is corrected. If you want to know what these four interfaces look like for your business, [talk to us](https://btlabs.dev/en/contact) — a short plain-talk conversation is enough for an honest assessment. ## Sources - Jeremy Howard (Answer.AI), proposal for machine-readable website profiles, September 2024 - Ahrefs, 2026 analysis: 137,210 domains with a machine-readable profile, 97 % without a single request - Own measurement, btlabs.dev, 30-day window: 23,520 machine requests, 18,179 of them verified Which of these interfaces you actually encounter on the open web is something we count: 33 routes are in the running measurement, and 16 of them returned not a single response across the entire panel. It is all in the [ai-discovery radar](https://btlabs.dev/en/ai-discovery-radar). --- Source: https://btlabs.dev/en/posts/how-ai-agents-read-your-business Last-Modified: 2026-08-15T09:42:54.116Z Languages: [de](https://btlabs.dev/llms/de/posts/wie-ki-agenten-deinen-betrieb-lesen) · [it](https://btlabs.dev/llms/it/posts/come-gli-agenti-ia-leggono-la-tua-azienda) See also: [llms.txt](https://btlabs.dev/llms.txt) · [ai.txt (Policy)](https://btlabs.dev/ai.txt) · [identity.json](https://btlabs.dev/identity.json)