Methodology · Open formula
Hosting rating methodology v1.1
Six pillars, published weights, and an explicit rule for everything we could not measure.
1. What is measured — and what is not the provider’s fault
The provider’s own landing page is not its hosting, and its own status page is a self-report. What we measure is the behaviour of customer sites hosted there, attributed to the provider by the ASN that announces the address.
A site being down does not mean the hosting is down. So only infrastructure-level failures count against a provider:
- the provider’s DNS does not resolve;
- the TCP connection is refused or times out;
- the TLS handshake fails;
- the node itself answers 5xx.
4xx and application-level failures do not count — those belong to the site owner, not the host. That separation is what makes the number defensible in an argument.
1a. Test sites and how availability and response are normalised (v1.1)
Since v1.1 the availability and response pillars come from our own test sites: a tariff we buy from the provider with a page of ours on it, watched by our monitor every minute. The purchase is the attribution — we know whom we bought from — and the announcing ASN is recorded with every measurement so a tariff fronted by a third-party shield is visible as such. Only infrastructure failures per §1 count; expected-code and keyword mismatches are our page, not the host, and leave the denominator.
- Window and floors. 30 days; a number appears only after at least 28 observed days and at least 80% of the expected checks — otherwise the pillar is unmeasured, not zero. Expected checks are counted from each test site’s own age, so a newly added site never hides an established number.
- Availability = up / (up + infrastructure failures). Normalised linearly: 99.0% → 0, 99.95% → 1 (so an SLA-grade 99.9% reads 0.95).
- Response is the TTFB of successful checks, p50 and p75 by nearest rank — never the mean. p50: 200 ms or less → 1, 1000 ms or more → 0; p75: 400 ms → 1, 1500 ms → 0; linear between.
- Several test sites at one provider are pooled: one number for “all our sites there”, not an average of averages. The raw percentage and milliseconds are printed in the table and in the dataset next to the normalised share.
2. Pillars and weights
| Pillar | Weight | What it reads |
|---|---|---|
| Availability (infrastructure only) | 30 | our own test sites at each provider, failures per §1 |
| Response time (TTFB p50/p75, never the mean) | 20 | our own test sites at each provider (Moscow probe in v1.1) |
| Network quality (IPv6, HTTP/2–3, TLS 1.3, DNSSEC) | 15 | our protocol tools |
| Reviews from verified customers | 15 | §4 — shrunk, decayed, group-capped |
| Secure defaults (panel headers, TLS grade, RFC 2142 abuse contact) | 10 | our security tools |
| Transparency (public status page, SLA with compensation, incident history, RIPE data completeness) | 10 | checkable facts |
Where the network and security probes are pointed. At the provider's storefront when it is served from the provider's own network (attribution by ASN). When the storefront sits behind a third-party shield (a CDN or DDoS filter), the probes go to the first service host of the provider — panel, cabinet, billing, mail — that answers over HTTPS from the provider's own network, and the basis of every measurement says so. No such host, no measurement: the pillar stays unmeasured, never zero.
Nothing here is a self-reported marketing figure: a claimed «99.9 % SLA» earns points under transparency only for the published document, never as an availability measurement.
3. An unmeasured pillar leaves both sides of the ratio
Score = 100 · Σ(wi · pi) / Σ(wi)
The sums run over measured pillars only. An unmeasured pillar is not scored zero (which would punish a provider for our missing data) and not scored fifty (which would invent a measurement). It is excluded, and the denominator shrinks with it. The rating always prints how many pillars stood behind the number — 72 / 100 · 4 of 6 measured.
Below the publication threshold — fewer than the required number of independent sites or days of observation for a provider — the figure is not published at all, rather than published with a caveat. A caveat does not survive a screenshot.
4. Reviews are not an arithmetic mean
Rw = n/(n+m) · R + m/(n+m) · C
Bayesian shrinkage toward the category mean. R is the provider’s own mean, n the number of counted reviews, C the mean across all providers, m the confidence threshold. One five-star review does not produce a five-star provider.
- Recency decay
2^(−Δmonths/12)— a host that was good in 2023 is not evidence about today. - Author weight by verification level: self-reported reviews are published but carry weight 0; a review with technical proof of hosting carries 0,7; proof plus confirmed payment carries 1,0.
- Group cap against astroturfing. Reviews arriving from one /24 subnet, one proven hosting domain or one account carry the combined weight of one — on every axis at once. Twenty five-star reviews in a week from one network weigh as much as one; one domain reviewed through twelve accounts is one observation.
- Trust gate. While
n < m/2the review pillar counts as unmeasured and leaves the ratio per §3 — it does not drag the score toward the middle.
Pre-moderation is a spam classifier, not a judge of opinion: it labels spam, abuse, third-party personal data, off-topic and astroturfing signals, and it never edits the text. An honest negative review passes. If the classifier is unavailable, the review stays in the queue — there is no auto-publish on failure, and the rating never moves on an unmoderated review.
5. Where our data is thin — said plainly
The sites we can observe are the sites our own users monitor. That sample is biased (it is whoever uses Enterno) and per-provider it can be small. So our existing measurements are a seed and a control sample — not a census — and the publication threshold in §3 exists precisely to stop a thin sample from being dressed up as a verdict.
6. What the rating does not claim
It is a measurement of publicly observable technical behaviour plus verified customer experience, on a stated date range. It is not an audit of the provider’s internal infrastructure, not a legal or financial assessment of the company, and not advice that a given provider fits your workload. Positions are not for sale and cannot be edited on request: a provider can dispute a measurement with evidence, and a corrected measurement changes the score the same way any other measurement does. Rating pages carry no advertising: the companies being rated buy the same ad inventory, and a banner from one of them under its own row would make that promise worthless.
This page is published before any score appears in the UI — the same order every methodology here follows.