Cloudflare Just Entered This Category From the Network Layer

On 6 August 2026 Cloudflare announced an AEO Visibility Dashboard, and the sentence that matters is its stated differentiator: it uses network-layer signals rather than test sampling. Cloudflare sits between AI platforms and the websites they fetch, so instead of asking an engine a question and reading the answer, it can watch the retrieval happen.

Every tool in our comparison, ours included, works the other way. We ask, we read, we count. That method has a sampling error we have worked through the arithmetic on: a hundred runs puts a 30% rate somewhere between 21% and 39%. A network observer does not sample. It sees the requests.

That is a real advantage on one of the two things this category measures, and it cannot exist for the other. Here is why.

What this is based on: Cloudflare’s own announcement, read on 6 August 2026, the day it was published. We have no account and have run no test of the product. Everything below about what it does is quoted from that page; everything about what a network layer can observe is our own reasoning, and you can check it.

What was announced

Four metrics, in Cloudflare’s words:

• Citation Rate, “which AI platforms use a brand’s site as a trusted source”
• Mention Rate, “whether AI assistants name a brand even without citing its site”
• Prominence, “how much of an AI answer is attributed to the brand”
• Share of Voice, competitive positioning on specific customer questions

It is “available in early access today”, requested from the Overview tab of the Cloudflare dashboard, and the announcement says it will eventually be generally available to current and potential customers. Which AI assistants are covered is not stated. The price is not stated. The timeline is not stated.

Why the network layer is genuinely better at one thing

When an AI assistant answers a question by searching, something fetches your page. That fetch is a real, countable event. If you are behind Cloudflare, Cloudflare sees it: which crawler, which URL, when, how often.

Compare that with how the rest of us establish the same fact. We run a prompt, read the cited sources, and record whether your domain appears. If the answer varies between runs, and it does, we need many runs before the number means anything. We are estimating a rate from a sample.

Cloudflare is not estimating. For crawler traffic to sites it fronts, it has the population, not a sample of it. No confidence interval is needed for a thing you counted completely.

That is the strongest methodological position anyone in this category has, and it comes from infrastructure rather than from cleverness. It is not something a prompt-sampling tool can catch up to by running more prompts.

And why it cannot do that for the second metric

Now read the second metric again: “whether AI assistants name a brand even without citing its site.”

An assistant naming your brand without citing your site produces no request to your site. There is nothing for a network layer to observe. The event happens entirely inside a model and a user’s screen, on infrastructure Cloudflare is not between.

We measured exactly this today, on three questions in our own category. Two of the three ChatGPT answers cited nothing at all, and they still named brands: Profound, Semrush, Peec AI, Otterly and others. Not one of those namings generated a fetch of anyone’s website. A network observer watching that day would have recorded silence, and the silence would have been wrong.

So Mention Rate cannot come from network signals. Either Cloudflare samples prompts for it, like everyone else, or it infers it from something the announcement does not describe. The same question applies to Prominence, which requires reading the answer text, and to Share of Voice, which requires knowing which competitors were named.

None of that makes the product weak. It means the differentiator in the headline covers one of its four metrics, and the announcement does not say what covers the other three. That is the question to ask in the early-access call.

The two things this category calls one thing

This is worth separating properly, because the whole market blurs it.

Retrieval is an assistant fetching your page. It is observable at the network layer, countable, and it tells you your content was considered. It does not tell you the answer used it.

Naming is an assistant putting your brand in the text a person reads. It is only observable by reading answers, which means sampling, which means uncertainty. It is also the thing that actually sells anything.

A tool built on network signals is strong on the first and structurally blind to the second. A tool built on prompt sampling is weak on the first, because it only sees the citations an answer happens to show, and it is the only way to see the second at all.

The honest conclusion is that these are complementary instruments, and anyone serious about measurement will end up wanting both. That is an inconvenient thing for us to write, and it is what the evidence says.

How we found out, which is its own finding

We did not read this in a newsletter. We found it by running a probe.

Asking Google’s AI Mode, in Spanish, from Spain, “¿qué herramientas miden la visibilidad de una marca en respuestas de IA?”, the answer named Cloudflare’s AEO Visibility Dashboard and cited the investor-relations page announcing it. That page is dated 6 August 2026. We ran the probe on 6 August 2026.

A corporate announcement was retrieved, read and used in a generated answer, in another language, on the day it was published. Whatever the indexing latency for this kind of source is, it is not measured in weeks.

The same probe is worth reading for a second reason. Of the twenty-two domains cited across the two Spanish surfaces, eight were Spanish-language publishers: blog.hubspot.es, latevaweb.com, ethinking.es, animorstudio.com, cyberclick.es, inboundcycle.com, aibrandpulse.ai and es.semrush.com. The most-cited domain of all was YouTube, with five citations across both surfaces.

Run the equivalent question in English from the United States and you get a different set, dominated by vendor blogs and with YouTube nowhere near the top. That is the argument for measuring per market rather than averaging, made with our own numbers instead of asserted. Both probes are single runs, so both are draws rather than rates, and the presence or absence of a domain in one draw is weaker evidence than the shape of the two sets together.

What it means that Cloudflare is here at all

Two things, and the second matters more.

The category just acquired a distribution problem. Cloudflare fronts a large share of the web. A visibility dashboard that appears in an existing tab, for customers who already pay for something else, does not need to win a comparison to get used. Most of the tools in our comparison, including us, have to be found and chosen first.

Network-layer measurement will become table stakes, and that is good. Crawler analytics are the part of this problem with a correct answer. Once a big infrastructure provider gives them away, the argument moves to the part that is genuinely hard, which is what the models say when they are not fetching anything. That is a better argument to be having.

It also validates something we have been unable to prove and have said so. Search Console shows impressions on long, conversational queries that look like prompts an assistant turned into a search, and it shows no way of telling who caused them. A network layer can.

What we could not verify

• Which AI assistants it covers. Not stated in the announcement.
• The price, or whether it is bundled with an existing plan.
• How Mention Rate, Prominence and Share of Voice are produced, given the network layer cannot observe an uncited naming.
• Whether it works for sites not behind Cloudflare, which is the population question that decides how general the numbers are.
• Anything about accuracy. We have no account and have run no test. This article is a reading of an announcement, not a review.

What we are doing about it

Adding it to the catalogue as an active tool with no published price, which is what it is today, and asking for early access. When we have run it, it gets a review with prices, arithmetic and a section on what we could not verify, like everything else here.

If it turns out that Cloudflare has solved crawler-side measurement and is honest about the limits of the other three metrics, that is a better product than most of this market and we will say so. Our comparison of every tool we verified now lists 81.

Common Questions About Network-Layer AI Visibility

What is network-layer AI visibility measurement?

It counts requests from AI crawlers to your website, observed by infrastructure that sits in front of your site. Because it observes every request rather than estimating from a sample of prompts, it produces exact counts of retrieval, with no confidence interval needed.

Is it better than prompt sampling?

For measuring retrieval, yes, decisively. For measuring whether an assistant names your brand, it cannot work at all, because an assistant that names you without linking to you never touches your server. The two methods answer different questions and a complete picture needs both.

Does Cloudflare’s dashboard cost anything?

The announcement does not say. It says the dashboard is in early access as of 6 August 2026, requested from the Overview tab of the Cloudflare dashboard, and that it will become generally available to current and potential customers. No price, no timeline and no assistant list are stated.

Can it measure my site if I am not a Cloudflare customer?

The announcement does not address it, and structurally the network-layer signal requires traffic to pass through Cloudflare. That is the question that decides whether its numbers describe the web or describe the part of the web it fronts.

Does this make AI visibility tools obsolete?

No, and it makes the useful half of them more clearly useful. Crawler analytics is the tractable part of the problem. What a model says when it answers from memory, with no retrieval and no citations, is the hard part, and we measured two answers out of three doing exactly that today.

What should I ask Cloudflare in the early-access call?

Which assistants are covered, what the population is, and above all how Mention Rate is produced. Naming without citing generates no request, so that number cannot come from the network. Whatever produces it is the part of the method that carries the uncertainty.