Skip to main content
Answer Engineered

ENTRY 010 · GEO · By Answer Engineered Research

Cloudflare Shipped an AI Visibility Score. It Asks Two Models.

Cloudflare's AEO dashboard pairs real crawl and referral logs with a Citation Rate scored against Claude and GPT only, computed once per category.

· 6 MIN

What actually shipped on 6 August

The launch bundles two things, one of which is not new.

Agent Readiness — the diagnostics half — shipped on 17 April 2026, originally as a standalone tool. What is new on 6 August is the AEO Visibility Dashboard and its integration into the main Cloudflare dashboard.

The new dashboard reports five named scores:

MetricWhat it is scored on
Citation RateAssistant answers to a prompt panel
Mention RateAssistant answers to a prompt panel
ProminenceAssistant answers to a prompt panel
Share of VoiceAssistant answers to a prompt panel
Industry Fit scoreAssistant answers to a prompt panel

Alongside them sits AI Operator Activity, which Cloudflare describes as showing “the real crawl and referral traffic on your site, per operator (OpenAI, Google, and so on): who reads your content, who sends visitors back, and the errors they hit on the way (403 blocked, 404 dead link).”

That table above is the whole argument of this post. Five of the six surfaces come from the same sampled source.

Which models, and how often

Cloudflare states plainly which assistants generate the scores: “the leading assistants (today, Anthropic’s Claude and OpenAI’s GPT).”

Two. Not Gemini, not Perplexity, not Copilot, at least not on day one. The word “today” is doing honest work there — it reads as a roadmap statement, and it would be unfair to treat a launch-day model list as permanent.

The cadence is the more consequential detail, and again Cloudflare says it directly: “Rather than re-querying models every time a site owner runs a scan, we run this panel once per category and reuse the baseline across all accounts in that domain.”

Read that twice, because it changes what the score can mean. The prompt panel is not run for your site. It is run for your category, and the result is reused across every account in that category. That is a sensible engineering decision — re-querying frontier models per customer per scan would be enormously expensive — and it is a real constraint on interpretation. A movement in your Citation Rate can come from your site changing, or from the category baseline being recomputed, and from the outside those two look identical.

No refresh interval is published. We looked for one and did not find it.

The sentence in the press release

Cloudflare’s press release for the same launch describes how competing tools work:

Most tools attempt to fill that gap by only sending test prompts to AI chatbots and sampling the responses — a method limited in scale and prone to inconsistency if not paired with other data signals.

The critique is accurate. Prompt sampling is noisy, and a category of vendors has been selling sampled scores as though they were measurements for two years.

The qualifier at the end is not decorative, and it is where a lazy version of this post would cheat. Cloudflare does not claim it has stopped sampling. It claims sampling is a problem when it is not paired with other data signals — and Cloudflare genuinely has the other data signal, described in the release as operating “at the network layer between AI platforms and the websites they access, observing actual crawl and referral activity directly, across millions of sites.” That is a real answer to its own objection, and no competitor can make it.

So the fair statement is narrower than “Cloudflare does the thing it criticises,” and more useful:

Cloudflare pairs a sampled score with network data, exactly as its own standard requires. But the pairing happens in the interface, not in the metric. Citation Rate is still a sampled number. The crawl logs do not make it less sampled; they sit beside it. And Citation Rate is the number that will end up in a slide deck, because it is a single figure with a percentage sign on it, and “your 403 rate to GPTBot” is not.

The traffic claim, and the number that isn’t there

The post opens with a striking line: “fewer than half of all HTML page requests now come from a human.”

We tried to pin that to a figure and could not. No percentage is published in the post. The linked Radar traffic dashboard renders its charts in JavaScript, and fetching it returns interface scaffolding rather than values. So the honest position is that the claim is directional and unquantified in the source, and anyone citing a specific percentage for it is supplying a number Cloudflare did not.

Cloudflare also qualifies it in the same breath, noting that not all of those machines are agents acting for a person. Non-human is not a synonym for AI assistant. Uptime monitors, security scanners, feed readers and scrapers are all in that bucket, and they have been for years.

One more caveat on the same data: Cloudflare’s network is large, but it is a footprint, not the web. Every figure here describes traffic that passes through Cloudflare.

What to do with it

The dashboard is worth opening, with a clear split in how much weight each half carries.

Treat AI Operator Activity as real data about your site. Crawl counts, referral counts and error codes per operator are first-party facts, and the error codes are the most immediately actionable thing in the entire launch. A 403 to an AI crawler you meant to allow is a bug with a fix. A 404 an operator keeps hitting is a link you can repair this afternoon. Neither requires you to believe anything about scoring methodology.

Treat the five scores as a category-level sample with a two-model panel. Useful for orientation, poor for attribution. Specifically: do not report a Citation Rate change as the result of a content change, because the panel refreshes on a schedule you cannot see and describes your category rather than your site.

And note the one thing no tool in this space can currently give you, Cloudflare included: a per-site, per-query, repeatable citation measurement across all the major assistants. That product does not exist. What is being sold as it, everywhere, is a sample.

We have made this point before about llms.txt and about citation freshness, and it holds here: the useful question about any GEO number is not what it says, but what it was measured on.

Sources

All three were fetched and read in full on 7 August 2026. Quotations are verbatim from those pages.