A tab opened in the background is not a slow page
The worst row on our load-time board said the home page took 7.4 seconds — on a page that paints in under half a second. The browser had deferred the paint because the tab was behind another one. Why a field measurement has to know when the page was hidden, and the one-line rule that keeps the board honest.
At six minutes past five on a Sunday morning the board recorded its worst number: the home page, 7,419 milliseconds. The home page is twelve kilobytes, its stylesheet travels inside it, and nothing in its head blocks the paint. In a lab on a slow connection it paints in about half a second. Seven seconds is not a slow page. It is a measurement of something else.
The something else is a tab. When a browser opens a page in a tab that is not in front — a middle-click, a link that opens behind, a phone that switched apps — it does not bother painting until the tab is shown. Nothing is wasted drawing pixels nobody can see. But the paint entry, when it finally comes, reports the time from the moment the page was requested, and that includes the whole time the tab sat hidden. Seconds. Minutes, sometimes. The number is real and it is not about the page.
The rule every field-metrics library applies
If the page was hidden at any point before its largest paint, the paint time is not a measurement of the page, and it is not sent. That is the rule; it is not ours, it is the one the tools that report these figures to the search engines all use, and our timer did not have it. It does now. The timer notes whether the document was hidden when it loaded, and the first moment it goes hidden afterwards; a paint later than that moment is dropped. Screens inside the app are unaffected — they are timed from a tap, in a tab that is by definition in front.
Proving it took a browser told to believe it was hidden from its first byte. Loaded that way, the page sent nothing. Loaded visibly, the same page reported a hundred and thirty-six milliseconds. The row that started this stays where it is — it is a real visitor's browser, not test data, and it ages out of the board's window in a week. Until then the "worst" column shows it, and the column is honest about what it is: the worst number received, which is not always the worst page.
Two lessons in two days
The day before, the board had been polluted by crawlers — machines that are not customers. This time by customers' own browsers doing something sensible. Both are the same shape: a measurement that includes the measurer. A board that ranks pages for a person to act on has to keep asking who else is sending the number, and what else the number contains.