280 Sites Measured From Inside China: A Third Never Finished Loading

The question people search for is whether a site is blocked in China. It is the wrong question, and this is the measurement that shows why.

We requested 280 enterprise software and infrastructure sites from a probe inside mainland China, three times each, with a 60-second ceiling on every attempt. 840 requests. A run counts as complete only if the page finished loading; a run still going at 60 seconds is recorded as a failure rather than dropped.

The result

Across every site that answered at all, the median time to first byte was 742ms (90th percentile 1,093ms). Among the sites that did finish, the median largest-contentful-paint was 2.1 seconds (90th percentile 4.7s).

Then the number that matters more than any of those:

Of the 96 sites that never finished, 66 answered. A first byte arrived, in most cases a paint was recorded, and the page still never completed. Only 30 produced no response at all.

Two different failures wearing the same result

Both groups look identical to a user — the page does not work — and they are entirely different problems.

The 30 that produced nothing are the ones people mean by “blocked”. The list is dominated by two companies: ten Google properties (Maps, Docs, reCAPTCHA, Tag Manager, Analytics, Cloud Functions, Hosted Libraries, the Fonts browsing site, YouTube, and support pages) and four from Meta and X (Facebook, Instagram, Twitter, X). Snapchat is there too. This is the well-understood part of operating in China, and it is not a delivery problem — a request to a destination that does not answer is not made to answer by moving where it comes from.

A useful contrast sits in the same dataset: bing.com/maps returned the fastest first byte of all 280 sites, at 66ms. Maps as a category is not the issue. The specific host is.

The 66 that answered and never finished are the ones nobody plans for, and they are twice as numerous. A connectivity check passes. A speed test reports a time. Monitoring that pings for a 200 stays green. And the page a visitor opens never becomes usable.

Some examples of how far that gap can stretch:

Authentisign answered faster than any other site in the census and completed none of its runs. Contentstack painted after twenty-eight seconds — a number that exists only because a measurement tool kept waiting, not because anyone would.

Time to first byte does not tell you whether a page works. That is the single most useful thing in this dataset, and it is the reason a speed test taken from outside China, or a first-byte figure taken from anywhere, cannot answer the question a buyer is actually asking.

What we learned that no single site could show

Public asset CDNs: two of five worked

All five are normally referenced from the document head, where the browser will not paint until they resolve. That position is what makes them disproportionately damaging — a page that references a failing script CDN does not load slowly, it stops.

The rule people reach for — “public CDNs don’t work in China” — is not supported. Three of five failed and two were fine. The workable advice is narrower and more useful: measure the host you actually depend on, or bundle the library into your own build and remove the question. Bundling is usually a smaller change than any delivery work.

A shared failure signature across one vendor — but not all of it

Four Adobe properties failed with ERR_HTTP2_PROTOCOL_ERROR rather than a timeout: Adobe Commerce, Magento, Marketo and Adobe Dynamic Tag Management. A protocol-level failure is a different diagnosis from congestion — it indicates the connection being interrupted in transit, and it is not improved by adding bandwidth or points of presence.

Adobe Fonts, measured in the same run, completed every attempt. Vendor-level generalisations do not survive measurement, which is why this census tests hosts rather than brands.

The same product category, opposite outcomes

Of nine podcast hosts, five produced no usable load and four completed every attempt. Of two loyalty apps for the same storefront platform, one completed every run and the other completed none. Learning platforms split the same way: one finished all three attempts at a 1.6-second paint, another finished none.

There is no category-level answer available here. That is not a rhetorical point — it means advice of the form “use a headless CMS” or “avoid embedded video” is not actionable, and the only thing that resolves it is measuring the specific host.

Even one vendor’s own hosts can disagree

Google Fonts is the cleanest example. fonts.googleapis.com, the stylesheet endpoint a page calls, answered 73 out of 73 times in a separate 12-hour sampling run at a median 111ms. fonts.google.com, the site a designer browses to choose a typeface, completed none of three attempts. Visitors receive the fonts; the designers cannot go shopping for new ones.

What we excluded, and why

Five hosts are not in the 280.

Three of them — fauna.com, stepzen.com and thin.dev — do not resolve from outside China either. Those services no longer exist. The fourth was a page in our own knowledge base carrying a wrong vendor URL; the domain it pointed at is not registered anywhere.

The fifth is the one worth dwelling on, because it nearly got published. edg.io failed from inside China with ERR_CERT_COMMON_NAME_INVALID — the only certificate failure in the run, and a genuinely diagnostic-looking signature. We had already written it up as a distinct finding when a check from outside China returned the same error. Edgio, which had acquired Edgecast, filed for Chapter 11 in September 2024 and its content delivery network went offline on 15 January 2025; Akamai bought the customer contracts but not the network. The certificate on that host is unmaintained because nobody is maintaining it, anywhere.

A dead product is not a China finding. Any failure with an unusually clean signature — a name that will not resolve, a connection closed immediately, a certificate that does not match — is worth verifying from outside the country before it is written down. We caught five this way, including one that would have had us publish that Google Fonts does not work in China while our own sampling data said the opposite.

Limits

• One probe, one city. A commercial cloud line in cn-zhangjiakou. International transit from a datacentre is materially better than from a corporate broadband connection, so every figure here is a floor rather than a worst case. A real office would see worse.
• Three attempts per site. Enough to separate “arrives” from “does not” reliably; not enough to be a precise latency figure. Where we sampled across a full day — twelve hours of third-party requests — one host’s time to first byte swung from a median of 493ms in its quietest hour to 1,086ms at the evening peak, a spread a three-run median cannot see.
• Vendor sites, not customer deployments. How a vendor’s own site behaves is not how a site built on it behaves. It is a signal about the host, not a verdict on the product.
• A 60-second ceiling is a choice. Some of these pages might complete at ninety seconds. No visitor waits that long, so we did not either.

Why the answer is not only about delivery

It would be convenient to end here, because everything above is a reachability problem and reachability is the part we can fix by changing where bytes are served from. But moving delivery inside mainland China is not only an engineering decision. Serving from inside the mainland is precisely what triggers the ICP filing requirement — the filing is issued to a mainland legal entity, it binds to your hosting and access arrangement, and access providers are prohibited from carrying an unfiled site.

So the two questions arrive together, and they are decided once, at architecture time: can the page arrive, and are you permitted to serve it from where it would arrive from. We measured what each of the available arrangements actually buys on the first question in four ways to deliver a website into China — including the arrangement that is inside China, correctly filed, and still failed to finish loading.

What we are measuring next

The same census from a Chinese office broadband connection rather than a datacentre line, which we expect to make every number here worse and the comparison less flattering to publish. We will publish it anyway. The method, including the eight ways we got it wrong first, is written up separately.

Measurements were run by the 21YunBox China engineering team on 28–29 August 2026 from a probe inside mainland China. Every site’s result — completion rate, median time to first byte, median paint and failure type — is recorded per host.