Is your shop agent-ready? · Article 3 · 2026
Every website gets two readers
Between 5 October and 2 November 2026, 65 per cent of all page requests on the Cloudflare network came from bots. For requests from Germany the figure was 74 per cent. Every website therefore has two readers: the people it was built for, and the programs that read on behalf of search engines, AI assistants and agents. This article covers what is measured about the second reader, what is not, and what it needs from your shop.
What has been measured
On 2 November 2026 we queried the Cloudflare Radar API. Cloudflare says it serves more than 20 per cent of the web, and it classifies each request as human or bot. For page requests (HTML pages, without images and scripts) in the four weeks from 5 October to 2 November 2026, it returned:
- 65 per cent bots worldwide,
- 74 per cent bots for requests from Germany.
"From Germany" means the requesting address is in Germany, data centres included. In early May the weekly figure was still 55 per cent worldwide and 59 per cent for Germany. In June, Cloudflare's chief executive Matthew Prince reported that bots now request more pages on his network than humans do. In March he had expected that to happen in 2027.
Other networks count differently and report similar figures. Fastly measured 49 per cent bots across all requests in January 2026, and Thales 53 per cent of all traffic for 2025. Counting all content, images and scripts included, Cloudflare reports 39 per cent, up from 28 per cent a year earlier.
Retail is the most heavily read target. In October 2026, 30 per cent of all AI bot requests on the Cloudflare network went to retail sites. A shop's web analytics shows little of this, because it usually filters bots out. In the 2026 online retailer survey by FHNW, a Swiss university of applied sciences, half of the retailers could not put a figure on their bot share. Those who did gave a median estimate of 0.8 per cent.
Who the second reader is
Most bots are not agents. They include search engines, link preview services, monitoring tools, scrapers and attackers. For AI bots, Cloudflare also reports the purpose their operators declare. In the four weeks to 2 November 2026 the split was:
- training language models, 43 per cent,
- mixed purpose, 39 per cent (Googlebot, Bingbot and Applebot, whose requests serve several purposes),
- search for AI answers, 12 per cent,
- agents acting for a user, 4 per cent,
- undeclared, 2 per cent.
The agent that searches and buys for a customer is a small part of the second reader today. On retail sites its share was even smaller in October 2026, at 1.5 per cent. What grew there was search: from 1.4 per cent of AI bot requests in October 2025 to 11.6 per cent in October 2026. ChatGPT, Claude and Perplexity need these search requests for their answers to shopping questions (article 1). A shop that blocks their search crawlers drops out of those answers.
Retail responds mainly by throttling. On retail sites in October 2026, one in five AI bot requests received the answer "too many requests". A year earlier it was around one in two hundred.
Where the 90 per cent comes from
The figure "90 per cent bot traffic" comes up often in conversations about agentic commerce. We found no measurement with a stated method behind it. The figure appears in at least eight contexts, and almost all of them mean something different:
- synthetically generated content online (Europol, 2022, as an expectation for 2026),
- B2B purchasing volume (Gartner, forecast for 2028),
- the amount of text that AI models process (Huawei, forecast for 2035),
- individual websites (the chief executive of Webflow in an interview, 2025),
- advertising campaigns (a click fraud expert, 2017).
Measurements today show 49 to 65 per cent, depending on network and counting method. As a direction, the figure is still useful. If the rise from May to October 2026 continues, bots on the Cloudflare network will request nine out of ten pages between late 2027 and mid-2028. That is our own extrapolation from six months of data, which include a dip in September. Matthew Prince expects a thousand times as much bot traffic as human traffic by around 2031. He adds that all his predictions on this have been wrong so far.
For your shop, what counts is your own server or CDN log. It shows who requests pages, how often and under which identifier.
Software first, then shops
For developer tools and business software, a second interface has been standard since 2025: one for agents, next to the one for humans. Netlify coined the term "Agent Experience" for this in January 2025. The shared connector is the Model Context Protocol, MCP for short. Through it, AI assistants reach tools and data. In October 2025 OpenAI built the apps in ChatGPT on MCP. In December 2025 the Linux Foundation counted more than 10,000 published MCP servers and named Claude, Cursor, Microsoft Copilot, Gemini, VS Code and ChatGPT as adopters.
Downloads show how widely MCP is used. Our query of 2 November 2026 returned, for the last 30 days, 258 million downloads of the TypeScript package from npm and 235 million of the Python package from PyPI. A download is not a user. The scale does show that software is now built with this second interface.
The vendors speak of two audiences. In January 2026 SAP wrote about its Commerce Cloud: "Optimization must now serve two audiences: humans and machines." The interface for humans stays. Agents use it as well.
Shops are next
For shops, the second interface arrives through the platforms first. The Universal Commerce Protocol (UCP) from Google and Shopify describes how an agent finds a shop's catalogue, prices and basket. According to UCPChecker, more than 19,000 shops had a valid UCP profile on 30 October 2026, around 10,950 of them on Shopify and around 6,700 on Wix. Around 150 shops worldwide served one from their own platform. Around 98 per cent of the profiles use MCP as their connector.
In Germany, Austria and Switzerland we retrieved the domains of 45 retailers on 2 November 2026. Only one served a UCP profile from its own platform: Hornbach, with catalogue search and without a checkout.
Retailers are just starting. In an EHI Retail Institute survey of just under 220 people in German-speaking retail, published in September 2026, around 60 per cent are looking into autonomous AI systems. Just over 11 per cent already use concrete applications.
The evidence points to a sequence: developer tools first, then payments and business software, then platform shops, and retailers with their own platform last. Shopify switches the interface on for its shops by default. Retailers on SAP, Salesforce, Intershop or their own platform have to build it.
Until then, agents use the interface built for humans. According to the security vendors HUMAN (April 2026) and Forter (October 2026), around 70 per cent of agent activity runs through browsers that operate pages the way a person does (article 2).
What the second reader needs
The second reader needs three things, in this order.
Readable pages. An agent must be able to retrieve the product page without failing at a security check. Price, availability and GTIN must be in the page the server delivers. Article 5 shows how many shops turn a new reader away: in our product page check of 3 November 2026, 19 of 30 shops did not release a single product page to our checker.
Structured data. The same details belong on the page in machine-readable form, using the schema.org vocabulary, and in the feeds for Google and the price comparison sites. Price and availability in the data must match the text. Google does not ask for new files made specially for AI. It continues to read products from Merchant Center feeds.
An interface. Then comes a separate entrance for agents: search, product, price and stock as an interface, described in UCP and fed from the same data as your feeds. Through it, agents get fresh prices, and you see who is asking. A policy per credential decides which agent gets what (article 6).
The order has a reason. There is no evidence that any major assistant retrieves the UCP profiles of third-party shops. They already use product pages and feeds.
What we make of it
We call the goal an Agent Channel: a separate entrance for agents into your shop, which recognises who is asking and gives each agent the data you intend for it. The first level, Agent-Ready, starts with the readable pages from this article.
Sources
- Cloudflare Radar, our own API queries on 2 November 2026: share of bots in HTML requests worldwide and by location of the requesting address (5 October to 2 November 2026, weekly values since May 2026), share of all requests with the prior year, purpose of AI bot requests. On 1 November 2026 for the retail industry (4 October to 1 November 2026, same days a year earlier): share of all AI bot requests, purpose, response codes.
- Cloudflare, blog post of 30 September 2026 (reach of the network). Matthew Prince: SXSW, 19 March 2026 (TechCrunch); post on X, 3 June 2026 (reported by hellopartner.com and ppc.land); Decoder podcast, 26 September 2026 (reported by ppc.land).
- Fastly, Threat Insights Report 26Q1 (16 April 2026); Thales, Bad Bot Report 2026 (29 April 2026). Both vendor reports.
- FHNW, Onlinehändlerbefragung 2026 (online retailer survey, September 2026), chapter 6.2.
- Origins of the "90 per cent": Europol Innovation Lab (2022, via Futurism); Gartner (November 2025, via Digital Commerce 360); Huawei, "Intelligent World 2035" (16 September 2026); Linda Tong, Webflow, interviewed by Turing Post (13 July 2025); HORIZONT (7 August 2017). Extrapolation: our own calculation of 2 November 2026.
- npm API and pypistats.org, our own queries on 2 November 2026: downloads of
@modelcontextprotocol/sdk(30 September to 29 October 2026) andmcp(last month). - Mathias Biilmann, Netlify, blog post on "Agent Experience" (28 January 2025); Linux Foundation, formation of the Agentic AI Foundation (9 December 2025); OpenAI Apps SDK (6 October 2025, TechCrunch); SAP News, Commerce Cloud (21 January 2026).
- UCPChecker, State of Agentic Commerce, October 2026 (as of 30 October 2026); our own retrievals of
/.well-known/ucpat 45 retailer domains in DACH, 2 November 2026. - EHI Retail Institute, survey on agentic commerce commissioned by Softserve (July and August 2026), reported by E-Commerce Magazin on 14 September 2026.
- HUMAN Security, State of Agentic Traffic (April 2026); Forter, network data on agents (October 2026). Both vendor studies.
- Google Search Central: no new machine-readable files are needed for the AI features in Search.
- Own product page check, 3 November 2026: up to 100 product pages per shop from the sitemap, 30 shops that could be checked (details in article 5).
What does an agent see on your product pages?
We look at it together with you, on your own pages.