Most SIEM deployments fail the same way: the platform works, the logs arrive, and the analysts stop reading the alerts. A queue of 800 alerts a day that nobody triages is worse than no SIEM at all, because it creates a written record of detections that were ignored. Tuning is not about silencing noise — it is about making the surviving alerts mean something.
Why the queue fills up
- Default rule packs. A vendor's starter pack is written to work across thousands of environments, so it assumes the worst about yours. Half of it will fire on activity you know is normal.
- Full log ingestion, no purpose. Collecting every log because you can doubles the noise without adding a single detection. Each source should exist to answer a question.
- Rules without context. A rule that says "five failed logins" knows nothing about the service account that always retries five times, or the scanner that logs in wrong by design.
- Duplicates. The same event arrives from the firewall, the proxy and the EDR, and each one raises its own alert.
Measure before you touch anything
Tuning without a baseline is guesswork. Pull these numbers for the last 30 days:
- alerts per day, and alerts per analyst per day;
- the top 20 rules by alert volume, with how many of those alerts closed as benign;
- alert-to-incident ratio overall and per rule;
- median time to acknowledge, and median time to close.
Any rule producing a large share of the volume with almost no confirmed incidents is a tuning candidate. A rule with low volume but a real hit is not — leave it alone even if an analyst finds it annoying.
Tune in this order
- Fix the log sources. A rule that fires because the timestamp is wrong, the hostname is missing or the parser dropped a field cannot be tuned. Correct parsing first.
- De-duplicate. Collapse alerts that describe the same event from different sensors into one case before anyone sees them.
- Baseline. Record what normal looks like for each rule: which accounts, which hosts, which hours. Most false positives become obvious.
- Allow-list narrowly. Name the account, the host and the reason, and set an expiry and an owner. A broad allow-list is a detection you decided to delete.
- Adjust severity, not existence. A noisy but real detection should become low severity and roll into a daily summary, not disappear.
- Change thresholds last. Raising a count from 5 to 50 hides the attacker who tries five times, waits, and tries again.
What good tuning looks like
A tuned rule has a written purpose, a narrow exception list with dates, a severity that matches the real-world consequence, and a runbook that tells the analyst what to check in what order. If an analyst has to guess what the alert means, the rule is not finished.
It also has an owner and a review date. Environments drift: a new application, a new subnet, a new service account. A rule that was clean six months ago can be the noisiest thing in the queue today.
Rules that should never be silenced
- Authentication from a new country or an unusual provider for a privileged account.
- New local administrator or new service account creation.
- Credential dumping activity and EDR alerts on security-sensitive processes.
- Security tooling stopped, disabled or reconfigured — including your own SIEM agent.
- Large outbound transfers from a server segment.
These are low-volume, high-consequence detections. If they are noisy, the environment has a problem worth investigating rather than an alert worth muting.
Overload is sometimes a staffing signal
If the queue is clean and still too large, the honest conclusion may be that one analyst cannot cover this estate alone. Automating enrichment — asset owner, business criticality, whether the source IP is internal, whether the account is normally seen — cuts triage time dramatically without weakening a single detection.
Related: SIEM and SOAR explained, and the first-hour incident response cheat sheet for what to do once an alert is real.