Articles

How Website Safety Verdicts Work: Blocklists, Scores and AI

Blocklists, trust scores and AI each judge a website differently. How each verdict is made, what a 0–100 score means, and what no checker can see.

verified.fyi
11 min read
On this page

Several distinct evidence shapes pass through a sorting mechanism and emerge as one balanced indicator.

A red warning page in Chrome, a "Dangerous" label from a site checker and a clean VirusTotal report all look like answers to the same question: is this website safe? They aren't. Each is built a different way, from different evidence, and each answers a narrower question than it appears to.

Knowing how a verdict is made tells you how much weight it can carry. This guide covers the two main kinds, blocklist verdicts and trust scores, how a score like verified.fyi's is put together, where a language model fits in, and what no automated check can see.

Blocklist verdicts: has anyone reported this site?

Most of the warnings people meet come from blocklists. The biggest is Google Safe Browsing, which Google says helps protect over five billion devices a day. It lists sites that serve malware or unwanted software, and social engineering pages such as phishing. Chrome and other browsers, Gmail and Google Search all use it to warn you before you open one.

For years Chrome checked each site you visited against a copy of that list kept on your device and refreshed every 30 to 60 minutes. In March 2024 Google moved Chrome's standard protection to real-time checks against its servers, because, it said, the majority of unsafe sites now "exist for less than 10 minutes". Even real-time checks in standard mode only protect you from sites Safe Browsing has already confirmed as unsafe. Catching sites too new to be confirmed is left to Enhanced protection, which adds machine-learning classification of pages.

Machine learning isn't new here. In 2010 Google engineers described the classifier that kept its phishing blocklist up to date, checking the address and content of millions of pages a day to decide which were phishing.

Other services work the same way:

A blocklist hit is strong evidence: someone looked at the site and found it doing harm. But a blocklist answers only one question, whether anyone has found and reported this site yet. Silence means "not yet", not "safe".

Why "not on a blocklist" isn't a pass

Scammers know about the lag and build around it. Interisle Consulting's Phishing Landscape 2025 found that 77% of the domain names used for phishing in the year to April 2025 were registered by the criminals themselves. A domain bought this morning has no past for any list to have caught.

Our own data shows the same gap. In our Born Yesterday study of 21,433 sites, 85% of those registered less than a week before we checked them scored Suspicious or Dangerous, yet only 4.9% were on Google's Web Risk threat list at the time.

There is a gap in scope too. Safe Browsing's categories are malware, unwanted software and social engineering. A shop that takes your money and sends nothing, or sends a counterfeit, doesn't obviously fit any of them, and may never be listed at all.

Trust scores: what do all the facts add up to?

A trust score asks a different question. Instead of looking a site up on one list, it collects everything it can find out about the site and weighs it together.

At verified.fyi that means 200+ facts, collected live from public sources at the moment you check, and scored in six areas:

  • Security: is the connection encrypted, and is the site clean on malware and phishing scanners?
  • Identity: who owns the site, how long has it existed, and are they hiding?
  • Reputation: what do blocklists, web history and reviews say about it?
  • Transparency: can you find out who runs it and how to reach them?
  • Compliance: are the legal basics in place, and is the company real?
  • Infrastructure: is it built on proper hosting, DNS and email?

Areas that protect your money and identity count the most.

The facts come from sources most people never check by hand. Malware and phishing flags come from Google Web Risk, Google's enterprise service for checking addresses against its lists of unsafe sites, and from VirusTotal and abuse.ch. Age and ownership come from WHOIS records, certificate issuance from the public Certificate Transparency logs, and the site's history from the Internet Archive. Popularity comes from the Tranco ranking, a research list that averages several top-site rankings over 30 days so it is harder to manipulate. Company and VAT numbers are checked against Companies House, the EU's VIES service and other national registers, and financial licences against the FCA and ASIC.

Some sources look past the site to the network it sits on. Spamhaus's DROP list names address ranges "leased or stolen by professional spam or cyber-crime operations", and a site hosted in one is marked bad however tidy it looks.

Blocklists are still part of this. A blocklist hit is one of the heaviest findings there is. But it is one input among many rather than the whole verdict.

Good, neutral, warning and bad signals

Each fact becomes one of four kinds of signal:

  • Good: in the site's favour, such as a valid certificate or a 14-year history.
  • Neutral: useful context that doesn't change the score, such as a popularity rank or a review count.
  • Warning: worth knowing, but common on honest sites, such as missing email protection.
  • Bad: strongly linked to scams and fraud, such as a listing on a scam blocklist or a copycat name.

Which kind a fact becomes depends on its detail. A domain that imitates a top website is marked bad when it uses lookalike letters, writes the real domain in front of another, or puts the brand next to words like "login" or "payout". The brand name alone, or a one-letter typo, is only a warning.

Only warnings and bad signals cost points. Good signals add nothing, because trust is the default: a site doesn't have to earn its way up from zero.

That rule matters because many reassuring things are cheap to fake. Let's Encrypt, which issues certificates for free, validates control of the domain and offers no organisation-validated certificates at all. In Born Yesterday, 93% of the Dangerous sites under a month old still showed a valid padlock. If padlocks earned points, scams would collect them.

The rule also runs the other way: some gaps cost nothing, because honest sites have them too.

  • Most small sites have too few visitors to appear in Google's Chrome UX Report, whose traffic threshold isn't disclosed, so having no data there costs nothing.
  • A registration hidden behind a privacy service costs nothing, because redacted WHOIS records have been the norm since GDPR took effect in May 2018.
  • Not being blocked by a protective DNS resolver earns nothing, since new scams often aren't listed yet.

How points come off

Every site starts at 100 and loses points according to how serious each problem is:

Severity What it covers Points
Cosmetic Tidiness issues with no real risk: a missing favicon, no social links −1 to −3
Minor Sloppy but survivable: weak security headers, no cookie notice −4 to −9
Major Genuine trust problems: no identifiable owner on a store, a very young domain, no contact details −10 to −24
Severe Strong scam indicators: blocklist hits, malware detections, brand impersonation −25 to −60

One severe finding is enough to sink a score on its own. And because nothing a site does well adds points back, a tidy About page can't make up for a malware detection or a blocklist hit.

Why a language model does the weighing

A fixed table of deductions soon runs into a problem: the same fact can be harmless on one site and alarming on another.

Take a missing privacy policy. On a personal blog it's not great, but a hobby blog isn't asking for your card number, so it counts as a cosmetic issue worth a few points at most. On a shop taking payments, the same gap is a serious problem and a major deduction.

The newer checks are full of the same kind of judgement. A homepage that asks for a password or card details isn't marked bad for that alone; it tells the scoring how much an anonymous owner matters. Only a request for a crypto wallet's recovery phrase is bad outright. A founding year from long before the domain was registered can simply be an old firm that moved to a newer domain, so it only becomes a warning when the domain is under a year old and nothing else (a company record, archive history, real traffic) shows an older business.

Writing a rule for every combination of site and finding isn't practical, so verified.fyi hands that judgement to a language model. It reads each signal in the context of what the site actually is, then weighs it and writes the verdict.

Language models have two well-known weaknesses, and both shape how one should be used for this.

The first is making things up. NIST's Generative AI Profile calls it confabulation: "the production of confidently stated but erroneous or false content". A safety verdict resting on a fact the model invented would be worse than none. That's why, at verified.fyi, the AI never invents facts. It only weighs the signals the checks actually found, and every report shows you those signals, so you can see the working for yourself.

The second is that a model can be steered by what it reads. OWASP puts prompt injection first in its Top 10 risks for language model applications, and describes indirect prompt injection, where the instructions arrive inside a website or file and can work even when a person can't see them. A scam page is written by someone who wants you fooled, so any checker that shows pages to a model has to assume the page will try to talk to it. The same problem at a larger scale drives the agentic AI security risks businesses face when AI agents act on the pages and documents they read.

Being able to trace a verdict back to its evidence is also what good AI practice asks for. NIST's AI Risk Management Framework lists "explainable and interpretable" among the characteristics of trustworthy AI, and explainability is a recurring demand in the compliance-aligned governance frameworks organisations are building around AI. For a safety check, it comes down to this: if you can't see why a site scored 34, you have no way to tell whether the reason applies to you.

Reading the final score

  • 80–100, Trusted: established and clean across the checks.
  • 60–79, Mostly safe: minor issues, nothing alarming.
  • 40–59, Use caution: a mixed picture. Double-check before you commit.
  • 20–39, Suspicious: multiple red flags. Avoid sharing anything.
  • 0–19, Dangerous: strong scam signs. Stay away.

The number is a summary; the signals beneath it are the useful part. Two sites can land in Use caution for very different reasons, a hobby blog with weak security headers or a shop that hides who runs it, so read what pushed the score down before you decide. Our list of website reputation indicators explains what each kind of signal means.

What no score can tell you

A score measures signals, not intentions. A high score means low risk; it isn't a promise, and it isn't legal or financial advice. A few things can fool any automated check:

  • Careful new scams. A domain registered an hour ago has no reviews, history or listings. A score still counts the domain's age against it, but a patient scammer can make a new site look tidy in every other way.
  • Hacked sites. An old, legitimate site can be compromised. Spamhaus's Domain Blocklist has a separate category for hostnames on legitimate domains that are being abused, and in Born Yesterday 14% of the sites rated Dangerous were more than five years old. Some hacks give themselves away: verified.fyi also opens the homepage as a visitor arriving from a Google search, and a site that sends only those visitors somewhere else is marked bad.
  • Change. A score is a snapshot of the day it was checked. Re-scan later: if the owner has fixed the problems, the score recovers, and if the site has been hacked since, it falls.

When tools disagree, a specific finding beats a general impression: a current listing on a reputable blocklist outweighs a clean score elsewhere. Google's Transparency Report site status shows whether Safe Browsing flags an address right now. If you run a site that has been flagged, Search Console's Security issues report shows what Google found and lets you request a review once it's fixed. For a VirusTotal detection, contact the vendor that made it.

No checker replaces the rest of your defences. Browser protection, an independent lookup and your own read of the page each catch different things, the same layering argument made in independent comparisons of privacy and security tools. And when more is at stake than one purchase, such as a supplier that will hold your customers' data, a score is where vetting starts, not where it ends. A cybersecurity assessment checklist covers your own business's side; our third-party website vetting checklist covers theirs.

See the working on any site

Paste an address into verified.fyi and you get the score, the verdict and every signal behind it, so you can check the reasoning rather than take it on trust. Nobody can pay for a better score, and badges show the live score, whatever it is. The methodology lists the sources and what each of the newer checks looks for, and recent checks show the scoring on real sites.

For the checks you can do by eye, see how to tell if a website is legit. If the link reached you in an email or a text, start with phishing.

Wondering about a site right now?

Paste the address — we'll run 200+ checks and give you a plain-English verdict in seconds.

Frequently asked questions

How is a website safety score calculated?

On verified.fyi every site starts at 100 and loses points for each warning or bad signal, from 1 to 3 points for cosmetic issues up to 25 to 60 for severe ones such as a blocklist hit or brand impersonation. Good signals add nothing, because trust is the default, so one severe finding can sink a score on its own.

What do the verdicts mean?

80–100 is Trusted, 60–79 Mostly safe, 40–59 Use caution, 20–39 Suspicious and 0–19 Dangerous. Below 40, don't share anything with the site; between 40 and 59, check further before you commit.

If a site isn't on a blocklist, is it safe?

Not necessarily. A blocklist only knows about sites someone has already found and reported. Google says most unsafe sites now exist for less than 10 minutes, so a brand-new scam is often on no list at all.

Why use AI to score a website?

Because the same fact means different things on different sites: a missing privacy policy is minor on a hobby blog and serious on a shop taking payments. On verified.fyi the AI only weighs the signals the checks found, never invents its own, and every report shows those signals.

Can a site with a high score still be a scam?

Yes. A score measures signals, not intentions: a carefully built new scam or a hacked legitimate site can look clean for a while. A high score means low risk, not a guarantee.

More from the blog

View all posts →

Check before you trust

Free, instant, no account needed — paste any site and get a plain-English verdict.

Check a site →