---
title: "AI agents: what business websites let them do"
description: "Across 98,981 local business websites in 34 countries, 99.6% let AI in, but only 2.1% offer AI agents an action in a structured format."
url: https://www.weftd.com/resources/insights/business-websites-ai-agents
locale: en
alternates:
  en: https://www.weftd.com/resources/insights/business-websites-ai-agents
  fr: https://www.weftd.com/fr/resources/insights/sites-entreprises-agents-ia
---

# AI agents: what business websites let them do

Across 98,981 local business websites in 34 countries, 99.6% let AI in, but only 2.1% offer AI agents an action in a structured format.

_Charles Cousyn · 2026-10-24 · updated 2026-10-24 · AI agents, Agentic web, Structured data, AI presence, Local businesses_

For a long time, we asked AI to answer a question. Several products launched since August 2026 offer something else: handing a mission to an agent that searches, compares, checks and can go as far as booking or ordering. To do that, the agent goes through business websites.

Two surveys published in late September and early October 2026 measure both sides: how much room these agents now take on the web, and what local business websites give them to work with.

## The short answer

Almost every local business website lets AI in. A minority gives it enough to understand the business and very few let it act. Across 34 countries, 16.0% of sites state their name, address and phone number in a structured format, 7.5% an offer or a price and 2.1% a booking, an order or a contact request ([Metatron Intelligence, October 2026](https://intelligence.metatron.marketing/local-business-index/)). For everything else, the agent has to infer.

## Summary

- On Cloudflare's network, which sits in front of about 20% of the web, daily requests from AI agents grew more than 1,700% in one year.
- Across 98,981 local business websites in 34 countries, 99.6% do not block AI, but only 2.1% offer it an action in a structured format.
- When information is not structured, the agent infers it from sources that may contradict each other, and its decision depends on them.

## AI agents have become an audience of the web

On 30 September 2026, Cloudflare published an analysis titled "The Internet has a second audience". The company reports that more than half of the traffic it observes is now automated, and that daily requests from AI agents on its network grew more than 1,700% in one year ([Cloudflare, 30 September 2026](https://blog.cloudflare.com/agentic-web/)). Cloudflare sits in front of about 20% of the web: its figures describe a large share of the internet, not all of it.

These bots do not all do the same thing. Telling them apart matters, because they do not call for the same response from a business.

| Visitor | Example | What it does | What Cloudflare measures |
|---|---|---|---|
| Search crawler | Googlebot | Indexes pages for search engines | Fewer than 1% of sites block it |
| Training crawler | GPTBot, from OpenAI | Collects content to train a model | 17% of sites block it |
| AI agent | Dots, Muse, Instinct | Acts on behalf of a person | Requests up more than 1,700% in one year |

The share of crawler requests meant for AI training rose from 22% in spring 2025 to 52% in June 2026. It is the category sites block most. Agents follow a different logic: they show up because a person gave them a specific request, such as finding an available tradesperson, comparing two quotes or booking a table.

These agents now come from the largest AI companies. Meta launched Muse on 8 September: an agent with its own online computer, able to browse, fill in forms and buy, available in the United States ([Meta, 8 September 2026](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/)). OpenAI introduced Dots on 29 September: agents that each have a cloud computer and a browser. They connect to more than 4,000 apps ([OpenAI, 29 September 2026](https://openai.com/index/introducing-dots/)). The Pro plan does not cover the European Economic Area, Switzerland or the United Kingdom, where Dots requires Business Premium ([IT-Connect, 30 September 2026](https://www.it-connect.fr/openai-dots-agents-ia-toujours-actifs/)). We tested two other agents, Instinct and Grok Bot, in September ([see our test](/resources/use-cases/same-mission-three-ais-chatgpt-instinct-grok-bot)).

Delegating to agents is still new for most people. Using AI to search is not: in France, for example, 48% of people used generative AI in June 2025, up from 20% in 2023. Among them, 73% used it to look for information ([Crédoc for Arcep, Baromètre du numérique 2026](https://www.arcep.fr/uploads/tx_gspublication/barometre-du-numerique-edition-2026_INFOGRAPHIE.pdf)).

## What local business websites give them

The Local Business Index, published by the UK company Metatron Marketing, covers the websites of 98,981 local businesses in 34 countries. It was carried out on 6 and 7 October 2026 ([Metatron Intelligence, October 2026](https://intelligence.metatron.marketing/local-business-index/)). Addresses were selected at random from OpenStreetMap, excluding schools, associations and public administrations. For each site, the homepage, the robots.txt file and the llms.txt file were analysed.

| What the survey measures | 34 countries | United States | United Kingdom | Germany | France |
|---|---|---|---|---|---|
| The site blocks no AI that provides answers | 99.6% | 99.7% | 99.5% | 99.8% | 99.6% |
| The homepage has a description | 66.7% | 69.0% | 68.6% | 63.6% | 74.6% |
| Name, address and phone in a structured format | 16.0% | 25.2% | 18.0% | 11.6% | 18.2% |
| Offer or price in a structured format | 7.5% | 11.9% | 7.1% | 5.4% | 9.9% |
| Booking, order or contact request in a structured format | 2.1% | 3.6% | 2.8% | 1.1% | 3.5% |

A structured format means information encoded to a standard, such as schema.org, that search engines and AI can interpret directly, without inferring it from the page text.

The report is titled "Found, but not understood". The figures drop at each step. Access is almost never a problem. Understanding concerns fewer than one site in five and action about two in a hundred. Among the twelve countries the report details, the United States and France are ahead on action, at 3.6% and 3.5%. Germany, the largest sample with 24,833 sites, stands at 1.1%.

By sector, across all 34 countries, restaurants are furthest ahead on action: 8.0% of their sites offer a booking or an order in a structured format. They are followed by accommodation (1.0%), health (0.5%), retail (0.4%), crafts and local services (0.3%). This figure is a minimum: only structured entry points are counted, not every form.

The tool used to build the site weighs heavily. Depending on the system, the share of sites that give their name, address and phone in a structured format ranges from 0.4% to 52.9%. The report does not name these systems. It also notes that 8.4% of sites show almost no text without JavaScript, which makes them hard to use for a bot that does not run that code.

## When information is not structured, the agent infers

An agent that finds no structured information does not stop there. It reads the pages, checks other surfaces such as the Google Business Profile, booking platforms or directories, then decides between them.

In September, we gave the same mission to several agents: find three restaurants near Lyon Part-Dieu station with a table free on a specific date ([see the full test](/resources/use-cases/same-mission-three-ais-chatgpt-instinct-grok-bot)). For one of the restaurants suggested, the official website, a second website, Google and the booking platform showed four different sets of opening hours. Two agents carried out the same operation and reached opposite conclusions. The one that trusted the official website, the most rigorous method, ruled out a restaurant that did have a table free.

A study presented in July 2026 by two researchers, one of them at Université Paris Cité, measures what structure changes under controlled conditions ([Elnaffar and Rashidi, July 2026](https://arxiv.org/abs/2607.12056)). The authors built two versions of the same online store, with the same products and the same prices, the second one prepared for agents. Over 150 tasks per version, run with three different models, agents succeed on 89.3% of tasks on the prepared version against 49.3% on the standard one, with 30% fewer steps. It is a prototype rather than real websites, as the authors point out themselves.

Both observations point the same way: an agent's decision depends on the form in which information reaches it.

## Three levels of an agent-ready website: legible, understandable, operable

The surveys make it possible to describe how ready a website is in three levels.

| Level | What the agent can do | Share of sites, 34 countries |
|---|---|---|
| Legible | Access the site and read its text | 99.6% do not block AI |
| Understandable | Know without ambiguity who the business is, where it is, what it offers and at what price | 16.0% for identity, 7.5% for the offer |
| Operable | Trigger a booking, an order or a request | 2.1% |

Almost every site reaches the first level. A minority reaches the next two. What separates a legible website from an operable one is covered in [why your website is a black box to AI](/resources/academy/website-black-box-to-ai).

To make a business operable, two paths exist today. The first goes through the website itself: complete structured data and booking or contact journeys described explicitly. The second goes through a direct connection between the agent and the business, in the form of an MCP server or an app connected in ChatGPT, Claude or other AI. Dots and Muse use both: their browser for websites, connected apps for the services that offer them. With a direct connection, the business provides the information instead of leaving the agent to infer it ([see how an AI presence works](/resources/academy/how-does-an-ai-presence-work)).

This second path is the one [Weftd](/solutions) works on: building, from the business's own information, an official source that agents query directly.

## What you can check on your website

Five points can be checked on your own or with whoever manages your website.

1. **Access.** Your robots.txt file does not block AI search crawlers. Blocking training crawlers is a separate decision. At OpenAI, for example, the training crawler (GPTBot), the search crawler (OAI-SearchBot) and the one that acts at a user's request (ChatGPT-User) are distinct. Each setting is independent of the others ([OpenAI, crawler documentation](https://developers.openai.com/api/docs/bots)). An accidental block is one of the [six reasons ChatGPT does not recommend a business](/resources/academy/why-chatgpt-doesnt-recommend-your-business).
2. **Text.** Your homepage displays its content without depending on JavaScript.
3. **Identity.** Your name, address, phone number and opening hours are declared in structured data and identical across all your surfaces: your website, your Google Business Profile, the platforms you appear on.
4. **Offer.** Your services and prices appear as text on the website and in structured data, not only in a downloadable document.
5. **Action.** The main action of your business, whether booking, ordering, requesting a quote or making an appointment, is available through a structured entry point or a direct connection.

## The limits of these measurements

The Metatron survey only covers the homepage. The access, identity and offer indicators rest on 77,187 of the 98,981 sites, those measured after an extension added during the survey. Blocking figures are minimums, since Google's crawler was not measured. The action indicator only counts structured entry points. The report details twelve countries and does not publish groups under 300 sites.

Cloudflare's figures cover its own network and the company does not detail its method. The study on the two versions of an online store is a controlled experiment on a prototype. Our September test covers one mission, one city and one date.

**[See what this looks like in your line of work →](/solutions)**

## FAQ

**What is an AI agent?**

An AI agent receives a mission rather than a question. It searches, compares, browses websites and can act on behalf of its user: fill in a form, send a request, make a booking. OpenAI's Dots, Meta's Muse, Instinct and Grok Bot are recent examples.

**What is the difference between a search crawler, a training crawler and an AI agent?**

A search crawler indexes pages for search engines. A training crawler collects content to train a model. An agent steps in when a person gives it a request and acts for them. According to Cloudflare, fewer than 1% of sites block the first and 17% the second.

**Should you block AI crawlers in robots.txt?**

The question has to be answered crawler by crawler. Blocking a training crawler is an editorial choice that does not keep you out of answers. Blocking a search crawler, on the other hand, removes your site from that AI's answers: OpenAI states this for its OAI-SearchBot crawler. Across 98,981 local business websites, 99.6% block no AI that provides answers.

**What is a structured format?**

It is information encoded to a standard, such as schema.org, that search engines and AI can interpret directly. An address written in a paragraph has to be inferred. The same address declared in structured data is identified without ambiguity.

**Can an agent book on a website with no structured format?**

Agents with a browser can fill in a form the way a person would. They then have to interpret the fields and steps. In the controlled experiment mentioned above, agents succeeded on 49.3% of tasks on the standard version of the site, against 89.3% on the version prepared for them.

**Which countries are furthest ahead?**

Among the twelve countries the report details, the United States (3.6%) and France (3.5%) have the highest share of sites offering an action in a structured format, followed by the United Kingdom (2.8%). Germany stands at 1.1%. The gap between what agents look for and what websites give them exists in every country measured.

**What makes a website agent-ready?**

Three questions are a good place to start. Can an agent access your website and read its text? Does it understand without ambiguity who you are, what you offer and at what price? Can it trigger the main action of your business? The five checks above go through each of these questions.

## Sources

- [Cloudflare, "The Internet has a second audience", 30 September 2026](https://blog.cloudflare.com/agentic-web/)
- [Metatron Intelligence, Local Business Index, survey of 6 and 7 October 2026](https://intelligence.metatron.marketing/local-business-index/)
- [OpenAI, introducing Dots, 29 September 2026](https://openai.com/index/introducing-dots/)
- [IT-Connect, Dots availability, 30 September 2026](https://www.it-connect.fr/openai-dots-agents-ia-toujours-actifs/)
- [Meta Newsroom, Muse launch, 8 September 2026](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/)
- [OpenAI, overview of OpenAI crawlers, accessed 23 October 2026](https://developers.openai.com/api/docs/bots)
- [Crédoc for Arcep, Baromètre du numérique, 2026 edition](https://www.arcep.fr/uploads/tx_gspublication/barometre-du-numerique-edition-2026_INFOGRAPHIE.pdf)
- [S. Elnaffar and F. Rashidi, "Designing Agent-Ready Websites for AI Web Agents", ICEME 2026, July 2026](https://arxiv.org/abs/2607.12056)
- [Weftd, "ChatGPT, Instinct, Grok Bot: Our Test", 23 September 2026](/resources/use-cases/same-mission-three-ais-chatgpt-instinct-grok-bot)
