Alert fatigue in event reporting starts the week you subscribe everybody
Alert fatigue in reporting happens when alerts fire on a schedule instead of on an exception, so readers filter them unread within weeks. The repair is measurement. Count how many times each rule led to a recorded action over 90 days, state a minimum alert to action rate, and delete every rule below it.
Somebody on the audience team asks why nobody spotted that the pacing had gone flat for nine days. The answer turns up in the show director's mailbox, where a rule created eight months earlier files everything from the reporting platform into a folder called Reports, unread, 412 messages deep.
Alert fatigue in reporting is usually diagnosed as a threshold problem, and the fix proposed is a better threshold. That is part of it. The larger part is that somebody subscribed eleven people to a rule that fires every morning whether or not anything happened, and a message that arrives every morning stops being an alert by about the third week.
What makes an alert worth sending?
The clearest test I know comes from outside the events industry. Rob Ewaschuk wrote the monitoring chapter of the Google Site Reliability Engineering book, edited by Beyer and colleagues and published by O'Reilly in 2016, and he set out four questions to ask of any rule that pages a human. The first is whether the rule detects "an otherwise undetected condition that is urgent, actionable, and actively or imminently user-visible". The others follow from it: will I ever be able to ignore this alert knowing it is benign, does it definitely indicate that users are being affected, and can I take action in response.
His statement of the underlying constraint is the part worth pinning above the desk of whoever administers your reporting workspace. "Every time the pager goes off, I should be able to react with a sense of urgency. I can only react with a sense of urgency a few times a day before I become fatigued." And on what happens when that budget is exceeded, the same chapter is blunt: "when pages occur too frequently, employees second-guess, skim, or even ignore incoming alerts, sometimes even ignoring a 'real' page that's masked by the noise."
Nothing about that reasoning is specific to distributed systems. A show director has the same finite budget of attention and the same behaviour when it runs out. The difference is that a production engineer has usually been forced to think about the budget, and a reporting team almost never has, because sending an email costs nothing at the point of sending.
The alert to action rate, and how to work it out
The measurement is simple enough to do in a spreadsheet, and it is more persuasive than any argument about principles.
Take every alert rule and every report subscription pointed at your show teams. For each one, count the messages it sent over the last 90 days. Then, for each one, count the occasions where somebody did something because of it: made a call, changed a spend, moved a deadline, opened an investigation. The second number comes from asking the recipients, which takes an afternoon and is the only honest source for it.
Here is what that exercise looks like on a portfolio I would consider typical in shape. Twenty-two rules, 90 days, 412 messages delivered. Recipients between them could name 19 occasions where a message caused an action. That is an alert to action rate of 19 divided by 412, or 4.6 per cent.
Then split it. Five of the rules accounted for 63 of those 412 messages and 17 of the 19 actions, a rate of 27 per cent. The other seventeen rules sent 349 messages and produced 2 actions, a rate of 0.6 per cent. The distribution is always this lopsided, and it is why an average rate is nearly useless on its own and a per-rule rate settles the argument immediately.
Now state a floor before you look at which rules fail it, because stating the floor afterwards turns the exercise into a defence of what you already have. Ten per cent is a defensible place to start. On these numbers it deletes seventeen rules and cuts the volume from 4.6 messages a day to 0.7, which is about five a week.
The choice of 90 days matters more in an event business than it would elsewhere. A rule written for the eight weeks before doors will look worthless if you measure it in June and indispensable if you measure it in September, so the window has to span at least one full pre-show ramp for every show whose rules you are judging. On a portfolio with shows in March, June and October, a 90 day window measured from any starting point covers the ramp of one show and the quiet period of the other two. The usable version is a rolling 90 days per show, anchored to that show's own calendar, which is more work and produces a number you can defend when the owner of a deleted rule objects.
Why is subscribing everybody the actual cause?
Two mechanics in the tooling push in the wrong direction, and both are worth knowing precisely.
The first is that alerts are personal in some platforms and shared in others, and teams mix the two without noticing. Microsoft's documentation for data alerts in the Power BI service, current in 2026, states plainly that "Only you can see the alerts you set, even if you share your dashboard", that alerts can be placed only on gauges, KPIs and cards pinned to a dashboard, and that a user can create up to 250 alerts across all their models. A personal alert cannot become the team's early warning system, however carefully it is configured, because nobody else can see it or maintain it. The moment its owner goes on leave, the warning stops.
The second is that report subscriptions look like alerts and behave like newsletters. The same documentation notes that when an alert threshold is met, the service checks the time since the last alert and sends only if the one hour or 24 hour window you chose has passed. That is a rate limit on a genuine alert. A subscription has no such logic. It goes out on the schedule regardless, which is correct behaviour for a scheduled report and fatal if the team has come to treat it as the thing that will tell them when something is wrong.
So the first cut is not about thresholds at all. It is separating the rules that fire on an exception from the messages that arrive on a calendar, and giving the second group a different name, a different subject line and a different expectation.
What survives the cut
Five rules is a reasonable landing point for a single show, and the five tend to be the same ones.
- Registration pacing outside its band for three consecutive days. One recipient, the audience acquisition lead. Three days rather than one because a single day is noise.
- A registration source stopping entirely. Zero rows from a channel that produced rows yesterday is a broken integration far more often than it is a market signal.
- Exhibitor contract signature falling behind the same point in the prior edition by more than the band. Sales director only.
- A refresh failing twice in a row, which belongs to whoever owns the pipeline and not to the show team at all.
- Space sold crossing the point where the overflow decision has to be made. This one fires once per edition if it fires at all, which is exactly what you want.
Everything on that list has a named owner, a defined action and a plausible reason to be unexpected. The rules that got deleted mostly failed the last of those: they told somebody that registrations had gone up, on a day when registrations go up.
Two design details do most of the remaining work. Put the action in the subject line, so the message reads as an instruction and can be triaged without opening it, and carry the operands with the claim so the recipient can check the arithmetic in the preview pane. A subject reading "Pacing 9 days below band, check paid search spend" gets a decision in about two seconds. A subject reading "Your daily report is ready" gets a folder rule.
The second detail is the review date. Every rule gets an owner and a date, six months out, when somebody has to look at its alert to action rate and either renew it or let it lapse. Rules accumulate because creating one is a two minute job and deleting one feels like a decision that needs a meeting. A standing expiry turns the default from keep to cut, and the whole cleanup stops being a project that happens once every few years after somebody misses nine days of flat pacing.
Where this stops
Deleting seventeen rules will eventually delete one that would have caught something. That is a real cost and it should be stated rather than argued away.
The defence is that the deleted rule was not catching anything anyway, because it was arriving in a folder nobody opened. A rule with a 0.6 per cent action rate is not providing 0.6 per cent of the protection it was designed for. It is providing close to none of it, while also degrading the rules around it. But the honest version of that argument admits that both states carry risk, and the one you are moving to concentrates the remaining risk in five rules that had better be correct.
The other limit is that this measurement cannot be automated, and I have seen people try. You can count messages from the platform. You cannot count actions without asking humans, because the record of the action lives in a call that was made or a budget line that was changed, and neither is joined to the alert. Anyone who offers you an automated alert to action rate is counting clicks in the email, which measures curiosity.
There is also a limit on what alerting can fix. A rule fires on a number crossing a line, and it can only be as good as the line. Deriving the line from prior editions instead of picking a round one is the other half of this work and V32 covers it. Deciding whether a single unusual point deserves a mark on the chart at all is a question about the chart rather than about the mailbox, which is V30's. And an alert that fires because a refresh silently failed is a data quality event wearing a business costume, which is why the refresh state belongs on the page itself as V36 argues.
This week, export the list of every subscription and alert rule pointed at your show teams, with the recipient count on each. Sort by recipient count descending. Anything with more than three recipients is either a report that should be scheduled and named as one, or an alert that has no owner, and you will be able to tell which within about ten minutes of reading the list. That list is also the cheapest thing to hand whoever is being asked to make the BI and reporting surface trustworthy again.
Questions people ask about alert fatigue in reporting
- How do you measure whether an alert is worth keeping?
- Count two things per rule over a fixed window, usually 90 days: how many messages it sent, and how many of those led to somebody doing something they would not otherwise have done. Divide the second by the first. A rule that sent 90 messages and produced no action has an alert to action rate of zero and should go.
- Should a daily summary email count as an alert?
- No. A message that arrives on a schedule is a report with an email delivery mechanism, and it needs a different test: does anyone open it. An alert exists to interrupt somebody, so it earns its place only when the condition it detects is both unexpected and worth an interruption at the moment it fires.
- Who should receive a registration pacing alert?
- One named person who can act on it, with a second named as cover. Broad distribution lists are the usual cause of fatigue, because a message addressed to eleven people is acted on by nobody and filtered by everybody. If two teams both need to know, send one of them a weekly digest instead of the alert.