Cloudflare Radar launches Domain Rankings to fill the Alexa gap

With Alexa's domain ranking service now deprecated, Cloudflare is stepping in with a new dataset built on traffic from its 1.1.1.1 resolver. The new Radar Domain Rankings (Beta) aims to measure which domains people actually use, not just which ones generate the most traffic.

The dataset draws on aggregated 1.1.1.1 resolver data that is anonymized in line with Cloudflare's published privacy commitments. The idea is to identify popular domains based on global Internet usage patterns without tracking individual users.

What counts as a domain

Domain names are hierarchical, dot-separated labels that map to network resources via DNS. But defining what should be ranked is trickier than it sounds. Many countries use second-level registrations — the UK's .co.uk, for instance — while others, like Japan, offer registrations at the third level under regional hierarchies.

The Public Suffix List provides a useful reference for these variations. Radar Domain Rankings applies a ruleset informed by such nuances:

  • Domains are reduced to second-level registrations (e.g., cloudflare.com, google.com).
  • If the second level is .edu, .com, .org, .gov, .net, .co, or .mil, the third level is used instead.
  • No distinction is made between website domains and infrastructure systems — a domain is treated as any Internet-available resource.
  • A curated, semi-automated list will map multiple domains that belong to the same platform (e.g., fb.audio, fb.com, and fb.watch all representing Facebook) in future updates.

Popularity as a population measure

Raw traffic volume does not reliably reflect what people consider popular, according to Cloudflare's research. So Domain Rankings defines popularity differently: the estimated relative size of the user population that accesses a domain over a given time period. The more people interested in a domain, the higher its rank.

Sensing unique users without violating privacy is the challenge. Cloudflare strips data that could identify individuals, but it can still recognize automated traffic patterns. Using data points stored only briefly — such as ASN, high-level geolocation, and the Cloudflare data center that served a request — the team built a proxy for unique-user counts that does not require cookies, IP addresses, or other tracking artifacts.

Building the model

The 1.1.1.1 resolver, launched in 2018, now sits among the world's top DNS services. Its data is aggregated and privacy-scrubbed, and the setup has passed an independent privacy examination by a Big Four accounting firm. But the resolver's diverse user base presents a challenge: humans, servers, IoT devices, home routers, and bots all query it, making it hard to isolate human traffic.

Cloudflare spent six months developing machine learning models to predict domain ranks. The team experimented with regression techniques for interpretability, then moved to more complex algorithms for accuracy, testing various feature combinations and synthetic datasets.

A key finding was that no single model handled both the top-ranked domains and the long tail well. This aligns with academic work, including a Cloudflare Research collaboration paper for IMC 2022 on the accuracy of popular website lists: the highest-ranked domains (like google.com or facebook.com) have feature values disproportionately larger than lower-ranked domains, while the long tail's ranking is based on smaller, less reliable numbers.

The solution uses two models:

  • A model trained on the population with the highest feature values, using more features, produces the ordered top 100 domain list.
  • A second model, trained on a broader group of domains with fewer features, produces the top one million most popular domains.

The top 100 list is fully ordered. The broader one-million list is divided into ranked buckets — 200, 500, 1,000, 2,000, 5,000, 10,000, 20,000, 50,000, 100,000, 200,000, 500,000, and 1,000,000 — where domains within each bucket are intentionally unordered. This reflects the natural instability of long-tail estimates while still conveying how close a domain is to the top.

What's available now

Domain Rankings recognizes that no global view is perfectly accurate. Data collection context, geographic distribution of users, product characteristics, and behavioral diversity all introduce bias. The results represent Cloudflare's view based on its aggregated signals.

The dataset launched on the Radar Domains page with two offerings:

  • An ordered list of the top 100 most popular domains globally and per country, based on the first model, covering the last 24 hours and updated daily.
  • Unordered global datasets in ranked buckets up to one million domains, based on the second model, covering the last 7 days and updated weekly.

Both are accessible on Cloudflare Radar, with CSV downloads available. The service will continue to be refined based on monitoring and user feedback.