How AI agents experience your business: a manifesto
In short
Ten principles for leaders, written on 4 November 2026. Each states one belief about how AI agents meet a business and carries the number behind it, with its source, its date and a label where the number is a vendor finding. The principles are opinions; the numbers are the evidence.
1. Agents are already your audience
Machines read more of the web than people do, and the machines that act for a person arrive every time someone asks about you. Treat them as a customer segment with needs, not as an annoyance in the logs.
- 30 Sep 2026
Cloudflare (30 Sep 2026): Daily requests from AI agents on Cloudflare's network grew by more than 1,700 percent over the past year, and more than half of Internet traffic this year was not human.
2. Most sites have not stated a position
A robots.txt written for search engines says nothing about agents. Few sites declare what AI may do with their content, and almost none publish a way for an agent to act.
- 17 Apr 2026
Cloudflare Radar (17 Apr 2026): Cloudflare Radar's scan of top domains found 78 percent have a robots.txt, 4 percent declare AI usage preferences with Content Signals, and 3.9 percent support markdown content negotiation.
3. The cheapest fixes are the ones that matter
In a controlled test, a long list of popular measures changed nothing. Two things broke retrieval: an answer hidden in script and a refusal returned to agent user agents.
- 19 Oct 2026
Ora's lab found (19 Oct 2026): One docs site served under 19 configurations and probed on 5 topics by a plain-fetch client and a JavaScript-running client (190 probes): only hiding the answer in JavaScript and returning 403 to agents stopped retrieval; llms.txt, JSON-LD, sitemaps and markdown mirrors changed nothing.
4. Agents use the front door
They go to docs and the home page, and only rarely to machine files. When they do open a machine file, they follow a link to it rather than guess its path.
- 19 Oct 2026
Ora's lab found (AgentReady guide, Oct 2026): Agents reached docs pages in 83 percent of traced runs and the home page in 69 percent; .well-known files 23 percent, openapi.json 21 percent where shipped, and sitemap.xml 4 percent.
5. Agent readiness changes what is said about you
A site an agent can read gets recommended more often, and a site it cannot read gets described from other people's pages.
- 8 Oct 2026
Ora's lab found (8 Oct 2026): Agents clearly recommended agent-ready businesses in 20 percent of answers against 11 percent for businesses that were not agent ready (1.9 times), with two blind judges both giving the top grade.
6. Missing facts are the cost
When the agent cannot read you, it does not stop. It answers from somewhere else, and the buyer's question goes unanswered or answered wrongly.
- 8 Oct 2026
Ora's lab found (8 Oct 2026): Answers built without reading the business's site were 3.7 times as likely to contain none of the facts the buyer asked for: 25 percent versus 7 percent, across 2,499 graded answers from 131 businesses.
7. Search is not the thing to block
Most sites keep search crawlers and a minority refuse training. The choice is between purposes, and it can be made one purpose at a time.
- 30 Sep 2026
Cloudflare (30 Sep 2026): Fewer than 1 percent of sites on Cloudflare block search crawlers, while 17 percent block training.
8. The scoreboard is public and low
Scanning has become cheap and common. The averages are low, which means the sites that do the work stand out for some time.
- 5 Nov 2026
Ora, vendor-published (5 Nov 2026): Ora's home page reported 145,922 sites scanned, an average score of 44 and 7 percent at B or above.
9. Measure the work
A number is only worth reporting if the method is. A study of tens of thousands of agent journeys is one way to know what happens; your own task test is another.
- 8 Oct 2026
Ora's lab, method: The study ran 37,927 agent journeys on 1,056 business sites across four agent setups on Claude and GPT models; sites were scored on 24 August 2026 and journeys ran in late August and September 2026.
10. Write the position down
A default is a decision somebody else made. Write down who may enter, what may be asked of an agent and who owns the answer, and review it each quarter.
- 17 Apr 2026
Cloudflare Radar (17 Apr 2026): Cloudflare Radar's scan of top domains found 78 percent have a robots.txt, 4 percent declare AI usage preferences with Content Signals, and 3.9 percent support markdown content negotiation.
Sources
- Cloudflare: The Internet has a second audience (30 Sep 2026) (blog.cloudflare.com/agentic-web), read 4 Nov 2026.
- Cloudflare: Introducing the Agent Readiness score (17 Apr 2026) (blog.cloudflare.com/agent-readiness), read 4 Nov 2026.
- Introducing AgentReady.org (19 Oct 2026) (ora.ai/blog/agentready-launch), read 4 Nov 2026. Vendor claim.
- AgentReady.org guide (agentready.org), read 4 Nov 2026. Vendor claim.
- Ora research: AX is the new AEO (8 Oct 2026) (ora.ai/blog/ax-is-the-new-aeo), read 4 Nov 2026. Vendor claim.
- Ora home page (ora.ai), read 5 Nov 2026. Vendor claim.
Last reviewed