What should an Ad Manager drop alert compare against?

A percentage drop alert on Ad Manager impressions, ad requests or revenue should compare a day with the same weekday a week earlier, or with the average of that weekday over the last four weeks. Yesterday is the wrong baseline for anything with a weekly shape. Rate metrics such as CPM and fill rate can use a trailing seven-day average.

The baseline decides whether the alert is useful. A 20 percent threshold against the wrong baseline either fires every Monday or stays quiet through a real drop, and after three weeks of that nobody reads it.

Why yesterday is the wrong baseline for most alerts

Most sites have a weekly rhythm. Traffic on a Saturday is not traffic on a Tuesday, and the ads follow the traffic. If your alert compares Monday with Sunday, it measures your weekly pattern, not a problem. The same alert against Saturday on Sunday may hide a real 15 percent drop inside the normal weekend dip.

You can see your own shape in two minutes. Run an Ad Manager report for the last eight full weeks, split by day, with impressions, ad requests and revenue. Put the days in a column by weekday. If the Sunday column sits well below the Wednesday column week after week, yesterday is not a fair comparison for you.

Some metrics barely care about the weekday. Rates such as CPM, fill rate and viewability usually move less from day to day than volume does, though seasonal demand still shows up in CPM. The same eight-week report tells you which of your metrics are flat across the week and which are not.

What Ad Manager's own compare option does

Ad Manager's reports have a Compare setting, off by default, with three choices: previous period, same period last year, and a custom period. Google's help page spells out what happens when you also split the report by day:

• Previous period: each day is compared with the day before it.
• Same period last year: each day is compared with the same day in the previous year.
• Custom period: each day is compared with the period you specify.

Previous period with a day split is the yesterday baseline from the section above. Same period last year matches the calendar date, so Tuesday 27 October 2026 is set against Monday 27 October 2025. For a metric with a weekly shape, that is a weekday mismatch built into the comparison. If you want a year-on-year view, set a custom period that starts 364 days earlier, which lands on the same weekday.

For an alert, the custom period is the one to use: seven days back for a same-weekday comparison.

The baselines side by side

Baseline	Good at	Fails on	Use it for
Yesterday	Sudden breaks on a flat metric	Every weekday change	Rates, zero checks
Same weekday last week	Weekly shape	A bad or unusual week last week	Impressions, requests, revenue
Average of same weekday, last 4 weeks	Weekly shape, one odd week	Slow to follow a real trend	Revenue alerts you want quiet
Trailing 7-day daily average	Smooths noise	A single day against a weekly average still sees the weekday	Rates such as CPM, fill rate
Last 7 days against the 7 days before	Weekday-neutral	Notices a break days late	Slow declines, weekly reviews
Same date last year	Seasonality	Weekday mismatch, a year of site changes	Context, not alerts

Two rows need a note. The trailing seven-day average is the default in many tools because it is easy to compute. Against a single day, it still carries the weekday problem: if your Sundays always run 20 percent under the weekly average, a 15 percent threshold fires every Sunday. Comparing the last seven days with the seven days before removes the weekday entirely, because each window holds every weekday once. The price is speed: a break on Tuesday moves a seven-day sum by a seventh, so the alert needs a lower threshold and will still trail the event.

A zero-activity check needs no baseline. If an ad unit that served yesterday has no ad requests today, that is the alert.

Which baseline for which metric

A workable default set, which you can adjust once you have your own eight weeks of data:

1. Revenue, impressions and ad requests per ad unit or site: same weekday last week, or the four-week same-weekday average if last week was unusual.
2. CPM and fill rate: trailing seven-day average, with a wider threshold than the volume metrics.
3. A slow decline you want to catch in the weekly review: last seven days against the seven before.
4. Zero activity on ad requests or impressions: no comparison, just the zero.

Keep the dimension narrow enough to see a break and wide enough to have volume. An alert on one small ad unit swings with every quiet hour. An alert on the whole network hides a section that went to zero.

Weeks that skew the baseline, including this one

A baseline assumes last week was a normal week. This week and next have three reasons it may not be.

Start with daylight saving time. Ad Manager reports use the network time zone by default, and Google's help page says your personal time zone setting does not change that. Clocks in the EU went back on Sunday 25 October, the last Sunday of October. In the United States they go back on Sunday 1 November, the first Sunday of November. On that day the 1 a.m. hour repeats, so the day has 25 hours. For a network on New York time, a same-weekday alert compares the 25-hour Sunday 1 November with the 24-hour Sunday 25 October, which looks about 4 percent up with nothing changed on the site. A week later, Sunday 8 November is compared with that 25-hour day and looks about 4 percent down. A network on a European time zone had its 25-hour day on 25 October, so this coming Sunday is the one that looks low. A 4 percent swing will not trip a 20 percent threshold on its own, but it stacks with the normal noise.

Then there is Prebid. Prebid.js 11.44.0 came out on 24 October. Its notes list a fix for a bug where a higher-desirability bid did not result in targeting, a change to when GPT targeting is reset after a bid wins, and an opt-in for priority bids. If your wrapper picks up that release, the bids that reach Ad Manager after the upgrade come from a different stack than the week before. Mark the upgrade date in your change log and expect the first same-weekday comparison after it to say more about the release than about demand.

Ad Manager is changing too. The 19 October release notes say rows with a zero-value ratio but other data, such as impressions, will no longer be left out of reports, so report totals may show more activity. From 2 November, overdelivery reconciliation extends to Programmatic Guaranteed deals with DV360, with overdelivered impressions identified at month end and not paid. Further out, from February 2027, web display impressions count only when the creative begins to render. Each one changes the numbers without anything changing on the page.

The fix is the same for all three. Keep a short change log with the date and what changed, and when an alert fires in the week after an entry, read the log before you open an incident. For a known shift such as the clock change, you can also skip the affected day as a baseline and compare with the four-week same-weekday average instead.

Setting the threshold once the baseline is right

Use the same eight-week report. For each day, work out the percentage change against the baseline you chose. Sort the drops and look at the biggest ones that turned out to be nothing. Set the threshold a little past them. That is your normal noise, measured, instead of a round number someone picked.

Do this per metric, because a CPM threshold and an impressions threshold are not the same number. If the largest harmless drop on impressions is close to the size of a real problem, the dimension is too narrow or the baseline is wrong; fix that before you lower the threshold.

One more trap: comparing today so far with a full day. At 10:00 the day is less than half over, so any comparison with a complete day reads as a drop. Compare the same hours, or wait for the day to close. The newest hours can also be short because the report itself is late, and telling a late Ad Manager report from a real revenue drop covers that case.

What to check tomorrow

Open your alert rules and write next to each one what it compares against. Any rule on impressions, ad requests or revenue that compares with yesterday is the first to change, to the same weekday last week. Then look at your network time zone in Ad Manager's settings and put the next clock change in the change log.

In PubPilot, a percentage drop rule compares today, yesterday or the last few days with the daily average of an earlier window you set, from 1 to 60 days, so a 7-day window against the previous 7 days is the weekday-neutral option from the table above.

If an alert does fire and the baseline holds up, the checklist in why did ad revenue drop in Google Ad Manager is where to go next.