How CrawlWarden decides a Search Console number left its normal range

The comparison set is the same weekday on each of the previous eight weeks. Weekdays are matched because a business site routinely halves at weekends: a plain eight week average would call every Saturday a collapse and miss a real Tuesday drop.

Two kinds of day are then removed. Days Google itself lists as a data anomaly, because comparing against a day Google says it got wrong compares against nothing. And days when an incident for the same check was already open, because a drop that is allowed to become its own baseline stops being reported after a fortnight, which is the exact moment the customer most needs it.

If that leaves fewer than eight usable days, the window reaches back one more week at a time. Fewer than five usable days even then, and the day is not judged.

The normal level is the MEDIAN of the usable days, not the mean. One viral Tuesday would otherwise raise the bar for eight weeks and silence every ordinary Tuesday after it.

A day is below its normal range when it falls under 70 percent of that median for a site's clicks or impressions, and under 60 percent for a single watched page's clicks. The page threshold is looser because one page moves far more than a whole site does.

A day above its normal range is not an alert and never will be. Traffic going up is not a problem to be told about at seven in the morning.

Below these, nothing is recorded and no alert can fire, however large the percentage fall looks.

A single day below its range is noise: a bank holiday, a Google reporting wobble, a slow Monday. So a site level check needs three consecutive finalised days below, and a page level check four.

That is the same rule as every other check in the product, and it sets the earliest an alert can arrive. Google finalises a day two to three days late, so a drop beginning on a Monday is reported on the Friday or the Saturday. When the cause was a change on the page itself, the page checks will usually have said so days earlier, which is the point of having both.

The incident closes on its own when the numbers come back inside the range.

Search Console reporting breaks occasionally, and when it does it breaks for everybody. If an unusual share of the properties CrawlWarden watches fall below their range on the same day, that is far more likely to be Google's reporting than every customer losing traffic at once.

So the date is HELD rather than folded in, and the operator checks it against Google's own data anomalies page and the Search Status Dashboard before anything is sent. A date confirmed as real is released and alerts normally; a date confirmed as Google's is dropped, and the days around it are excluded from future comparisons.

The gate stays silent while CrawlWarden watches fewer than 20 eligible properties, because below that a shared pattern is as likely to be coincidence as a fault.

A slow decline. The normal level slides down with the traffic, so a site losing two percent a week for twelve weeks stays inside its range the whole way. This catches drops, not erosion, and nothing in the method pretends otherwise.

A drop on a site below the floors. A page going from eight clicks a day to three is a 60 percent fall and is not reported, because at those numbers it is indistinguishable from a quiet week.

Why anything happened. CrawlWarden reports that a number left its normal range, on which dates, and what it recorded changing on those pages beforehand. It does not state a cause, and a test fails the build if any generated text implies one.

There is no average position check. Position moves when a site starts ranking for new and weaker queries, which is good news that reads as bad. It appears as evidence beside an incident and is never judged.

Every evidence block shows the day, the value, the normal level it was compared against, the percentage, and which dates were used or left out. Export the same date range from Search Console, take the medians by weekday, and the answer should match. If it does not, that is a bug worth an email.

• the four checks this method drives, and what each one means
• what Google reports about each page, which needs no traffic at all
• what Search Console can and cannot tell you when clicks fall
• every check CrawlWarden runs, by group