Rank vs detection

A high-ranking domain is not a safe domain

Open rank, page authority and popularity metrics measure how well known a site is. They do not measure whether a page on it is harvesting credentials right now — and the busiest hosting platforms on the internet prove it every single day.

0Verified phishing hosts tracked
0All-time domains tracked
0Concurrent DNS verify threads
<50msTypical lookup response
The category error

Two different questions, routinely confused

Popularity is a property of a registrable domain over months. Phishing is a property of a hostname over hours.

Rank and authority scores exist to answer a marketing question: among all the sites on the internet, how prominent is this one? They are computed from links, traffic estimates and reference patterns accumulated slowly, and they are attached to a registrable domain because that is the unit brands and search engines care about. Used for what they were built for, they are perfectly sound.

The damage starts when a rank score is treated as a security signal. A security control needs an answer about a specific hostname at a specific moment. Registrable domain versus full hostname, months versus hours — that mismatch is the entire disagreement, and it is visible in the sample data below.

What a rank score actually measures

Prominence, aggregated at the registrable domain, updated on a slow cycle. Every property below is a feature for its intended purpose and a defect when it is used to decide whether to let a user through.

  • It is computed for the registrable domain, so every subdomain inherits the parent's standing regardless of what is served there.
  • It moves slowly by design, because a metric that swung hourly would be useless for its actual audience.
  • It rewards age and accumulated references, which means a freshly stood-up campaign host either scores nothing or inherits a very high score.
  • It says nothing about what is being served at this moment, only about how much the wider web has referenced the name over time.
  • It cannot distinguish a hosting platform's own pages from the millions of user-created pages sitting under it.

What a phishing verdict has to measure

Whether this exact hostname is standing up live phishing infrastructure today. Different unit, different clock, different failure mode — which is why it needs a different dataset entirely.

  • The unit is the full hostname. 0-amazon.weebly.com is an entry; weebly.com is not, and never will be.
  • The clock is 24 hours. The database is rebuilt and republished at 04:30 UTC every day.
  • Presence requires proof of life: every candidate is resolved with 50 concurrent threads and only active A records are retained.
  • Absence of history is irrelevant. A host that first appeared yesterday is treated exactly like one that has been serving for a week.
  • Removal is automatic. When a host stops resolving it fails the next build and leaves the list.
Evidence

Four rows from the real feed that settle the argument

These are unmodified entries from the published export. Read each hostname, then read its parent.

A site builder with a global footprint

0-amazon.weebly.com

The parent is one of the best-known website builders on the internet, referenced by millions of pages and carrying whatever authority score you care to look up. The subdomain in front of it is a brand-lookalike phishing host.

Consequence: a control reasoning about the registrable domain sees weebly.com and allows it. The list sees the hostname and blocks it.

A major cloud application platform

0-amfc.firebaseapp.com

Firebase hosting is a first-class developer platform, and its parent domain's reputation is effectively unimpeachable. That reputation is exactly the asset a kit operator is renting when they deploy there.

Consequence: near-zero cost to stand up, near-maximal inherited standing — a combination no popularity metric can price correctly.

An edge platform, twice in a row

00-ede.pages.dev   001-ede.pages.dev

Two adjacent rows from the same export, differing only by a serial number. The naming is the fingerprint of bulk provisioning: dozens of hosts created in one session and rotated as each is taken down.

Consequence: a hit is rarely alone. When one comes back, search your logs for its siblings before you close the case.

Reassuring words on a trusted suffix

000-access-secure.web.app

A hostname engineered to read as legitimate in a browser address bar: security vocabulary in the label, a platform suffix that resolves, and a valid certificate as a matter of course because the platform issues one automatically.

Consequence: every heuristic a careful user is taught — check for HTTPS, check the domain looks plausible — passes here. Only a list of confirmed live hosts does not.
The correction

Three rules that fix the category error

None of them requires abandoning rank data. They require putting it back in its lane.

Enforce on the full hostname

Never collapse a lookup to the registrable domain. Doing so either misses every kit hosted on a large platform or blocks the platform itself, and both outcomes are unacceptable. Send the host exactly as it appeared.

Require proof of life, not reputation

Make presence on a verified list the signal, not absence from a popularity index. The database publishes only hosts that resolved to an active A record during the daily build, which is what makes it enforceable rather than merely indicative.

Keep rank for what it is good at

Prominence data is genuinely useful for prioritising a takedown, sizing a brand-abuse problem or triaging which incident to work first. Use it to rank your response; never use it to make the block decision.

Side by side

Rank data against a verified phishing list

Six properties that decide whether a signal can carry a blocking decision.

Property Popularity / rank data DNS-verified phishing list
Unit of measurement Registrable domain — every subdomain inherits the parent's score. Full hostname — each host is an independent entry.
Update cadence Slow by design, because prominence is a long-run measure. Rebuilt every 24 hours and published at 04:30 UTC.
Handles a brand-new host Either no data at all, or the parent's full inherited standing. Included from the moment it is reported and confirmed resolving.
Proof the host is live None — a score persists whether or not anything is served. Active A record required, checked with 50 concurrent threads.
Removes stale entries Not applicable; scores decay rather than expire. Pruned on failure to resolve, with a daily changelog of removals.
Safe to drive a block No — it answers a question nobody asked at the block point. Yes, with a consequence proportionate to reversal cost.
Do not use rank as a gate

Allowing traffic because a parent domain scores well is how a kit on a free hosting platform reaches your users with every heuristic satisfied. It is not a conservative default; it is an open door with a reassuring label on it.

Use a verified list as the gate

Ask about the hostname, require proof it is currently serving, and act proportionately on a hit. Keep prominence data for prioritising which of the resulting cases you work first — that is a job it does well.

Questions

Where teams get this wrong

Four patterns that show up repeatedly in architecture reviews.

Does the Phishing Detection API return a rank or authority score?

No, and deliberately so. The response carries domain, is_phishing, category, dns_status, last_checked, confidence and database_size — nothing about popularity. The full schema is in the response format reference.

Mixing a prominence score into a security verdict would invite exactly the confusion this page exists to dispel. Source prominence data separately and use it for prioritisation, never for the gate.
We already block newly registered domains. Is that not the same idea?

It is a different control with a different blind spot, and it is worth keeping. Registration-age heuristics catch the freshly bought lookalike; they do nothing at all about a subdomain created five minutes ago on a platform registered in 2007, which is where a large share of live kits sit.

Note what this database is not: a newly-registered-domain watch service. A name registered on Thursday and parked appears only once it is standing up and serving — the constraint that makes it enforceable, and the reason the two controls complement each other.
Can we allow-list a whole hosting platform to reduce credit spend?

Not one that hosts user-created content. Allow-listing pages.dev, web.app, firebaseapp.com or weebly.com would wave through four of the seven sample rows published in the feed. That is the single most expensive shortcut available here.

Allow-list your estate, your named suppliers and specific vendor hostnames you control — never a suffix under which anybody can provision. The credit packages page shows how to cut spend through deduplication instead.
Our filter categorises by domain. Do we need to change it?

You need whatever enforcement layer you use to be able to match a full hostname. Most DNS resolvers, response policy zones and next-generation firewall URL categories already can — the constraint is usually in the transform script that loads the list, not in the appliance.

The daily export is a plain three-column CSV precisely so that transform stays short. Recipes for resolvers, firewalls and SIEMs are on the daily threat feed page.

Ask about the hostname, not the brand behind it

Take a day of proxy logs, keep the full hostnames, and batch them against the list. The entries that come back on well-known platforms are the ones your current controls are already letting through.