The scale
Every scan produces a score from 0 to 100, where 0 means nothing suspicious was found and 100 means multiple independent sources consider the target dangerous. The score maps onto five bands.
| Score | Verdict | What it means in practice |
|---|---|---|
| 0–4 | Safe | No signal fired. Nothing known against it. |
| 5–19 | Low | Something minor, usually a redirect or a moderately young domain. |
| 20–39 | Medium | Mixed signals. Read the breakdown before trusting it. |
| 40–69 | High | At least one reputable source flagged it. Avoid. |
| 70–100 | Critical | Multiple sources agree. Do not visit, download or log in. |
The signals and what each is worth
The score is additive. Each signal that fires contributes points, and the result page lists every contribution so the arithmetic is visible. Signals that do not fire are not shown, which is why a clean scan has an empty breakdown table.
VirusTotal detection ratio, up to 30 points
The proportion of the 95+ engines that flagged the target, multiplied by 30. One engine out of ninety-five is worth well under a point. Forty out of ninety-five is worth around thirteen. This is deliberately gradual, because a single engine disagreeing with ninety-four others is usually a false positive rather than a discovery.
VirusTotal flagged at all, 20 points
A flat addition the moment any engine flags the target, on top of the ratio above. The ratio alone under-weights the case where two or three engines have found something real that the rest have not caught up with yet.
Google Safe Browsing listing, 15 points
Google's blocklist is authoritative but narrow: it covers what Google has confirmed, which is a fraction of what exists at any moment. A listing is close to conclusive, so it combines with other signals rather than needing to carry a verdict alone.
urlscan.io sandbox score, up to 15 points
The urlscan verdict score, scaled by 0.15. If urlscan marks the page outright malicious, its score is floored at 70 before scaling, so a malicious verdict always contributes at least ten and a half points.
Domain age, 5 to 15 points
Read from the registry record. Under 7 days adds 15, under 30 days adds 10, under 90 days adds 5. This is the highest-value cheap signal in the entire model. Disposable scam infrastructure is measured in days. Real businesses are measured in years.
Brand impersonation in the hostname, 20 points
A known brand name appearing anywhere in the hostname while the registrable domain does not belong to that brand. This catches both the subdomain trick and the more common case of a scammer owning a lookalike domain outright. An allow-list prevents the brands' own infrastructure domains from being flagged against themselves, so s3.amazonaws.com and login.microsoftonline.com are not treated as impersonating Amazon and Microsoft.
Redirect destination impersonation, 20 points
Applied when a link redirects across domains and the destination impersonates a brand that the original address did not. This is the signal that catches a shortened link resolving to a fake login page, a case that scored nothing at all in earlier versions of the model.
SSL certificate mismatch, 15 points
The certificate does not cover the hostname being visited. Rare and serious.
Very new certificate on a young domain, 10 to 15 points
Only applied when both the certificate is under 30 days old and the domain is under 90 days old. A new certificate alone means nothing, because major sites rotate certificates every ninety days by design. Treating certificate age on its own as suspicious was a real false-positive source and was removed.
Redirect chain shape, capped at 25 points total
Structural properties of the chain rather than its destination. An HTTPS to HTTP downgrade adds 12. Landing on a different registrable domain adds 5, reduced to 3 when the link started on a known shortener, because that is the shortener working as designed. Crossing three or more domains adds 5, and a chain of four or more hops adds 5. The whole group is capped so a long chain can never carry a verdict by itself.
Rules that override the arithmetic
Fail closed
If VirusTotal returns no engine verdicts, Safe Browsing is unavailable and urlscan produced nothing, the scan raises an error instead of returning a score. A zero built from three failed lookups is indistinguishable from a genuinely clean result, and presenting one as the other is the worst failure a safety tool can have.
A missing source is never a clean source
When one source fails but others answer, the weighting is applied only across the sources that responded. A failed lookup is never blended in as a zero, which would drag the score down and make an unscanned target look safer than it is.
Scores are capped and rounded
The total is capped at 100 and rounded to one decimal place.
What the free tools use instead
The QR checker, link expander, fake website checker, scam email checker and bulk checker run a lighter model that excludes VirusTotal and urlscan, because those have per-scan costs and the free tools have no account or quota. The lite model weights domain age, Safe Browsing, certificate state, redirect shape and hostname heuristics more heavily to compensate, and every result is labelled preliminary with a link to escalate to the full scan. It is good at catching disposable scam infrastructure and blind to malware hosted on an otherwise clean domain.
Which reports get published
A public safety report is only published when the scan completed with at least three independent data sources, produced no upstream errors, and is under 120 days old. Anything failing that gate stays reachable but carries a noindex tag, because a page of boilerplate with no real data underneath it is worthless to a reader and to a search engine alike.
Known limitations
- Zero-day phishing. A page live for an hour can pass every blocklist. Domain age is the main defence here and it is not sufficient.
- Compromised legitimate sites. An old domain with a clean history that was hacked this morning scores well. Age cuts both ways.
- Aged domains bought to defeat age checks. Expired domains with history are purchased specifically to bypass the strongest cheap signal.
- Cloaking. A page can serve different content to a scanner than to a person.
- Content and conduct. Nothing here judges whether a real business ships orders or honours refunds. That is a different question from whether a link is technically dangerous.
Corrections
If a verdict looks wrong, use the report button on any result. Reports are reviewed by hand. Weights change when evidence says they should, and this page changes with them. Three scoring bugs that produced wrong verdicts were found and fixed in 2026 by exactly this route.