Metric naming conventions that end the attendance argument

The model had three measures called Attendance. Not similar names. The same word, three times, distinguished only by the folder they sat in, and two of the folders were called Core.

Their values for the same edition were 31,940, 27,510 and 24,300. Every one of them was correct. The first counted registrations created, all types, gross to date. The second counted everybody who entered the halls across show days including exhibitor staff and contractors. The third counted verified professional attendance across show days, which is the figure that had gone on the audit certificate.

Metric naming conventions exist to stop that, and the useful ones do it by making the filter and the window part of the name rather than part of a document.

Why does one word end up meaning three things?

Because each measure was created by a different person solving a real problem, and each of them called it the obvious thing.

Marketing needed a headline registration figure for a campaign report in March. Operations needed a halls figure for a capacity conversation in September. The audit team needed the professional attendance figure in October. Nobody was wrong at the moment of creation, and nobody looked at the field list first, because the field list had four hundred entries by then.

The damage arrives later and in public. The marketing deck quoted 31,940. The exhibitor prospectus quoted 27,510. The audit certificate said 24,300. An exhibitor who reads all three, which exhibitors do, is looking at a 7,640 gap between the largest and smallest figures for the same event, which is 24 per cent of the largest. From the outside that reads as one organiser reporting three attendances, and no explanation offered afterwards fully repairs it.

Kimball Group states the governing rule for this in its published technique on conformed facts, read on kimballgroup.com in August 2026. If the same measurement appears in more than one place, "care must be taken to make sure the technical definitions of the facts are identical if they are to be compared or computed together". Where they are consistent the facts "should be identically named", and where they are incompatible they "should be differently named to alert the business users and BI applications".

The second half of that is the part teams skip. Renaming an incompatible measure is treated as cosmetic. Kimball treats it as the alert mechanism, which is exactly right: the name is the only warning most readers will ever get.

A pattern with four slots

The convention I would publish has four slots, and two of them are optional.

Subject first, always. Registrations, Attendance, Exhibitors, Space, Revenue. Leading with the subject makes the field list sort into families, so a person looking for a registration measure sees every registration measure together and can compare them before choosing. This single rule does more work than the other three combined, because most bad selections happen when a reader picks the first plausible name they see.

Qualifier second. The thing that distinguishes this measure from its siblings. Gross, net, verified, professional, unique. If a measure has no sibling it needs no qualifier, and adding one anyway creates a false implication that a counterpart exists.

Window third. To date, show days, trailing twelve months, prior edition, at day minus sixty. Include it whenever more than one window exists for the same subject and qualifier, which for event data is nearly always.

Source or basis last, when it matters. Entry scans, registration records, audited. Use it where two measures of the same thing come from genuinely different instruments and the reader needs to know which.

dbt Labs, in the style guidance its documentation carried in August 2026, makes three rules that transfer directly to measure names. "Consistency is key! Use the same field names across models where possible." "Use names based on the business terminology, rather than the source terminology." And "Do not use abbreviations or aliases. Emphasize readability over brevity."

That last one is the one people fight. Attendance Verified Professional Show Days is 41 characters, which looks long in a field list and is entirely readable in a tooltip, a formula bar and a support conversation. Att Ver Prof SD is 16 characters and has to be decoded by everyone who meets it for the rest of the model's life. Length is paid once by the person naming it and brevity is paid repeatedly by everybody else.

Three attendances, renamed

Take the three measures above and run them through the pattern.

The 31,940 becomes Registrations Gross To Date. The subject changes, which is the important move: it was never an attendance measure at all, and calling it one was the original error. Somebody who wants a registration count now finds it next to Registrations Net To Date, and the two differ by cancellations.

The 27,510 becomes Attendance Total Show Days. Total is the qualifier that signals the inclusive population.

The 24,300 becomes Attendance Professional Show Days. Sitting directly beneath its sibling in the field list, the difference is visible without opening anything: 3,210 people, which is 13 per cent of the professional figure, being exhibitor staff and contractors.

Three renames, and the audit trail improves immediately, because the gaps between the three numbers now have names too. Whether professional attendance is the right headline figure for your show, and what counts as professional, is a definitional question that the post-show reporting cluster owns and this convention deliberately does not settle. The convention's job is to ensure that once the definition is settled, exactly one measure carries it.

The qualifiers that sound precise and are not

Some words look like they narrow a measure and do nothing at all. They are worth banning explicitly in the convention, because each one gets used by somebody who believes they have disambiguated.

Total is the worst offender. Total registrations tells a reader that something has been summed and nothing about what was included. In the example above, Attendance Total Show Days works only because Total is doing real work against a named sibling, Professional. Standing alone, with no sibling, Total means whatever the formula happens to do.

Actual is the second. Actual attendance implies a contrast with a forecast that may not exist as a measure anywhere, and it invites the reader to assume the number has been verified when it may be a raw count.

Final belongs to the same family and is worse, because it makes a claim about time that ages badly. A measure named Registrations Final gets built in November and is still called Final in February when late corrections have moved it by 300.

New is ambiguous in event data specifically, since new can mean first-time at this show, first-time in the portfolio, or created since the last refresh. Three different populations, one word, and the reader has no way to tell which.

Real, Clean and Adjusted round out the list. Every one of them describes what somebody did to the number rather than what the number counts, and none of them survives being read by a person who was not in the room when the adjustment was agreed.

How do you rename without breaking every report?

The honest answer is that you break some reports, and the choice is whether you break them at a time you chose.

Take the renames in one pass rather than gradually. A model half-converted to a convention is worse than one that never had a convention, because readers now cannot tell whether a name lacking a window means the window is implied or means nobody has got to it yet.

Sequence that works. First, list every measure and its current formula, and group by what they actually compute rather than by what they are called. That grouping alone usually finds two or three genuine duplicates that can be deleted rather than renamed. Second, agree the new names with the people who use them, in one meeting, with the numbers on screen: agreement is much faster when somebody can see that their attendance and the audit team's attendance differ by 3,210. Third, rename, and keep the old name as a deprecated alias only if your tool supports it and only for one reporting cycle with a fixed removal date.

Fourth, and this is the step teams skip, fix the reports. A rename that leaves twelve reports pointing at a deprecated alias has moved the problem rather than solved it. Budget the report fixes as part of the rename rather than as follow-up work, because follow-up work does not happen in the month before a show.

One thing to decide up front is what happens to the names in a report that goes outside the building. An exhibitor scorecard reading Attendance Professional Show Days is precise and slightly cold. Most teams want a friendlier label on the tile itself, which is fine as long as the friendly version is a display label over an unambiguous measure rather than a second measure with a friendly name. Getting the label under the tile right is its own decision and it depends on this one being made first.

Where this stops

A naming convention makes two different measures distinguishable. It does not make either of them correct, and it cannot detect that the formula behind Attendance Professional Show Days quietly counts contractors because the badge type mapping has been wrong since 2024.

Names carry intent. Only tests carry correctness, and the two get conflated by teams that have just finished a renaming project and feel like the problem is solved. If you rename without adding a check that each measure returns what its name claims, you have improved communication and left the underlying defect in place with a more confident label on it.

The second limit is that conventions decay unless somebody owns the field list. Every new measure is created by somebody in a hurry, and a convention with no gatekeeper lasts about eighteen months. Practical mitigations are a naming check in whatever review your model changes go through, and a periodic listing of measures whose names do not parse against the pattern. Attaching the definition to the measure itself so it travels into every report tooltip closes the remaining gap, since a name can only carry so much.

Export your measure list this week and sort it alphabetically. Read the first fifty names and mark every one where you cannot tell the population, the filter and the window from the name alone. That count is your backlog, and it will be larger than you expect. Doing this before the next report request arrives is the cheapest governance available to an event business intelligence team, and it starts with a conversation about what the next report is actually for.