Breach Counts Compared: The Lowercase Letter That Decided Who Got Notified
Archi3 min read
Contents
The number that opens every breach file#
The first real decision in a breach response is arithmetic. Someone exports a customer table, counts the rows, and that total travels into the regulator filing, the notification letters, the press statement and, eventually, the complaint.
Three counting methods show up again and again. They disagree by large margins. Picture a mid-size insurer that lost a backup through an unpatched web form, with 1.2 million rows sitting in a marketing database. Which number goes on the filing depends entirely on which method the team picks under deadline pressure.
Option one: count the rows#
Fastest. Export, count, report. The number matches the file, which is exactly what an auditor or a regulator will ask for later. It also overcounts. In a marketing table that has been through four re-imports, 10 to 20 percent of addresses are duplicates, internal test accounts, or records that bounced years ago and were never cleaned.
Option two: count people#
Normalize, dedupe, then match against a customer master. Two to five days of work. The total usually drops 15 to 30 percent, which matters when state notification thresholds begin at 500 or 1,000 residents. The catch is that the customer master is often stale, so the dedupe introduces a fresh set of errors.
Option three: count confirmed exposure#
Only count the records the attacker actually read. This needs row-level access logs. Most databases do not keep them, and the ones that do roll them off within days. It is the most accurate answer and the least available one at the moment a deadline is running.
The detail that moved the number#
Deduplication is where things get strange. Email normalization looks trivial until you write it down. Lowercase the address, trim spaces, strip the plus tag, and on some providers strip dots from the local part. Every step is a judgment call, and every step changes who gets a letter.
Case sensitivity was the one that bit. The script hashed addresses as-is. J.Smith@example.com and j.smith@example.com hashed differently, so 38,000 policyholders were counted twice. That inflated total crossed a state threshold the accurate count did not, which meant an extra regulator filing, an extra print run, and a public number that was about 4 percent too high.
Plus tags did the same thing in the other direction. A customer who signed up as jane+news@example.com and later as jane@example.com is one person. Strip the tag and they collapse correctly. Leave it and they are two people. Neither choice is wrong in the abstract. Both change the mailing list.
Why the first number sticks#
The count in the initial filing is hard to revise. Regulators build timelines from it, plaintiffs quote it, reporters archive it. Correcting it downward a week later reads as a walk-back even when the correction is right. Correcting it upward reads worse.
That is the argument for spending the extra two days on option two, and for writing the normalization rules before the export happens rather than after.
Comparing the three#
Rough tradeoffs, based on how these plans usually run: , Row count: same-day answer, 10 to 20 percent overcount, easiest to defend as a literal reading of the file., People count: two to five days, 15 to 30 percent lower, needs a written rule set and a customer master that is not stale., Confirmed exposure: most accurate, rarely possible, and usually finished after the filing deadline has passed.
Cost follows the same order. A row count is an analyst's afternoon. A people count is an analyst, a data engineer and a privacy counsel review. Confirmed exposure is a logging project that should have started a year before the incident.
The rule worth writing down now#
Decide the normalization rules in advance, in writing, and test them against a sample export. Lowercase or not. Strip plus tags or not. Strip dots or not. Count internal test accounts or not. Each rule is one line in a document that takes an afternoon to produce.
Then keep a second number beside the first: the count of people you can actually reach with a working mailing address. That figure decides whether the notification lands, and it is always smaller than the headline total.
The gap between the two is the part almost nobody plans for. A 1.2 million record breach with 300,000 dead addresses is a 900,000 person notification with a 25 percent bounce rate. The call center gets staffed against the first number. The complaints arrive against the second.
More articles