A growing share of the traffic to any website is not a person. It is a machine fetching pages on someone’s behalf. This is what that machine actually receives, what it is graded on, and what changes when you fix it.
SuperSchema | Measured 2026-08-16 | Reference build: superschema.ai 100/100
Disclosure: the seven days of traffic data below come from anahana.com, which is founder-owned. It is used because it is the only site whose logs we can publish in full, not because it is an independent third party.
When you open a website, your browser downloads the page, runs its JavaScript, assembles the result, and paints it. That last part is doing an enormous amount of work. On a modern site, most of what you see did not exist in the file the server sent.
Most AI crawlers skip that step. GPTBot and ClaudeBot fetch HTML and read it. They are not browsers. Whatever the server sends is the entire site as far as they are concerned.
Which means a site can look finished to every human who visits and be substantially blank to every machine that does. We know because it happened to us. Here is what our own homepage returned on 2026-08-16, before we fixed it.
Headline, copy, product explanation, a call to action. The page did its job.
`` $ curl -sL https://superschema.ai/ | grep -oE '<(h1|h2|p)[^>]*>[^<]{5,}' (no output) $ curl -sL https://superschema.ai/ | wc -c 8352 $ curl -sL https://superschema.ai/zzz-not-real | wc -c 8352 (identical) ``
$ curl -sL https://superschema.ai/ | grep -oE '<(h1|h2|p)[^>]*>[^<]{5,}' (no output) $ curl -sL https://superschema.ai/ | wc -c 8352 $ curl -sL https://superschema.ai/zzz-not-real | wc -c 8352 (identical)
No headline. No copy. Byte-identical to the 404 page.
The same was true of /about, /blog, /contact and every registry page. /pricing, /docs and /features were fine, so this was a per-route setting, not a broken site. It is fixed now. The rest of this page is the record of what that took and what it was worth.
This would be an interesting curiosity if it were only a forecast. Two things in the last year make it concrete instead.
On 6 August 2026 Cloudflare released Kitesurf, described in its own changelog as a stateless, highly scalable browser that runs entirely on Workers and is designed for AI agents. It throws away tabs, extensions and pixel-perfect rendering, the things a browser has for a person, and uses 3 to 7 times less CPU and memory than Chromium on common agent tasks.
Worth noticing who that is. The same company that publishes the scanner grading your site is building the browser the agents will arrive in. Those are not two products. They are one bet, placed by a company that can see the traffic.
Its top strategic prediction for 2026: AI agents will intermediate more than $15 trillion in B2B spending by 2028, with 90% of B2B buying agent-intermediated by that point.
Read that precisely, because the precision is the whole point. Intermediated, not autonomous. Agents source, compare, negotiate and execute, and humans stay in the loop for the complex and the sensitive. And it is B2B spending specifically, the buying process your own customers are already running.
The forecast may prove wrong. Forecasts do. The browser is already shipped.
Every technical detail on this page rolls up to one of four questions. They are worth learning in this order, because each one only matters if the previous one is answered.
AI recommends what it can verify. Pass all four and you are what the answer recommends. Fail one, and it recommends someone it can verify.
Agent ready means yes to all four. They are a checklist, not a sequence. You do not work through them in order, you either pass each one or you do not.
Cloudflare publishes a free scanner at isitagentready.com. Free, no account, one page. It is the most widely used public grade, which means any prospect can run you through it in thirty seconds without asking your permission. It is worth understanding precisely.
The two marked rows are ours. superschema.ai scored 21 on this same scanner in August, three checks out of fourteen, Level 1. Section 07 is what closing that gap actually involved. n/a means the scanner reported a level without a published check count.
Most of the internet scores badly, including companies you would expect to score well. That is the opportunity and the risk in the same fact.
Cloudflare’s is not the only spec. Vercel publishes its own Agent-Readability Spec with a separate 0 to 100 rubric, and third-party scanners grade against it. One runs 70 checks including protocols Cloudflare does not score at all, such as commerce and signed bot authentication.
Which means 100 on one grader is not 100 on all of them. A site can be perfect on Cloudflare’s fourteen and still have gaps on someone else’s seventy. Treat any single score as one instrument’s reading, not a finished state, and expect the specs to keep moving for a while yet.
The scanner measures protocol, whether the right files exist at the right paths. It is silent on three things that matter at least as much.
anahana.com was taken to a perfect 100. Every file, every header, DNS records, a live MCP server. Then we measured seven days of traffic before announcing any of it anywhere. This is the honest result.
Roughly 6,000 to 1. Every file was live and correct. Almost nothing found them. Being fully conformant generated no measurable agent traffic on its own.
There is a second finding inside the same dataset, and it changes how the traffic should be read. Crawlers are not one thing.
Live agent fetches, the class that means a person is asking a question in real time, totalled 3,745 across the whole window, of which ChatGPT-User is the bulk. That is 2.3% of verified AI bot traffic. The other 97.7% is training and indexing crawl. Meta and Apple alone are 61% of the load and neither one ever returns a person. Never total these numbers. Segment them, or you will report a win that contains no customers.
This is why the work does not end at a green scan. The files have to be announced, through registry submissions, link headers, an agents.md, or they sit there correct and unread.
This is our own record, not a recommendation written from the outside. Fourteen checks, and they are not evenly distributed across who can do them. That is the part that surprises people: the work does not belong to one team. Roughly half of it is files, and the rest splits between the application and the DNS zone.
Marketing owns content, IT owns DNS, engineering owns the app. Handing all three a single undifferentiated checklist is why this work stalls. It queues behind whoever is busiest. Three lists, three owners, run in parallel.
One rule overrides the score. Never publish an artifact that asserts something untrue. A file announcing an API that does not exist scores a point and is a lie, and it fails the moment anyone tries to use it.
Every check carries a prerequisite. If the prerequisite is not met, the check is not shipped, the score stays lower, and that is the correct outcome. This matters more here than elsewhere: a company selling verified truth cannot publish hollow conformance artifacts.
The honest value is trust and sales-cycle compression, not sessions. Being cited produces a reference most readers never click. If the goal is traffic volume, this is not the lever, and anyone who tells you otherwise is selling against measured evidence, including our own.
The reason to do the work is narrower and more concrete: the grade is public, it takes thirty seconds to check, and it is increasingly the first thing a technical buyer runs. Scoring badly is a live liability regardless of what the traffic does.
Run a free scan and find out what a machine actually receives from your pages, where you stand on all four questions, and the first things to fix.
Run a free Scan - Results in seconds, no signup.
All figures first-party and verified 2026-08-16
---
Human-friendly version: https://superschema.ai/what-ai-agents-see Generated by SuperSchema - https://superschema.ai