A semantic layer for event metrics so verified attendance means one thing
A semantic layer holds each event metric as a single definition: the fact table it reads, the joins it needs, the filters it applies and the aggregation it performs. Every dashboard and export is served from that definition, so verified attendance returns the same value regardless of which tool asked for it.
The exhibitor relations lead says the show did 11,689. The audience team's deck says 11,204. Finance has 10,927 in the profitability model. All three numbers came out of the same badge scan fact table, in the same week, from people who know what they are doing.
A semantic layer for event metrics exists to make that impossible. Define the measure once, with the joins and filters that constitute it, and serve every consumer from that one definition. What follows is what the definition needs to contain, and what it does not fix.
Three answers to one question
Start with the arithmetic, because the numbers are not mysterious once you see them.
The edition produced 38,900 badge scans across 11,689 distinct badge holders. Marketing counted the badge holders and reported 11,689. The audience team excluded 485 staff badges and reported 11,204. Finance excluded staff and a further 277 contractor badges issued to stand builders, and reported 10,927.
The spread is 762 badge holders, which is 6.5 per cent of the largest figure. Every one of the three queries is defensible. Nobody made an error. The warehouse contains no record of which rule is the official one, so the rule lives in three separate SQL files owned by three separate people, and it changes whenever one of them edits their file.
This is a different failure from two reports disagreeing because they joined tables at mismatched grain. There the SQL is wrong. Here every query is right and the disagreement is about what the words mean.
What a semantic layer holds
The Cube documentation, current in 2026, describes the pattern as centralising "metric definitions, joins, access rules, and caching upstream of every BI tool, application, and AI agent that queries the data." The same documentation splits the layer into four parts: data modelling, access control, caching and APIs.
The data modelling part is where the argument above gets settled. Cube's model is built from cubes and views, and it defines the two ingredients of any measure in terms an event team already uses: "Measures are referred to as quantitative data, such as number of units sold, number of unique visits, profit, and so on. Dimensions are referred to as categorical data, such as state, gender, product name, or units of time (e.g., day, week, month)."
Views then control what consumers see. The documentation describes them as sitting "on top of cubes" to "create a facade of your whole data model, with which data consumers can interact." For an event portfolio that facade is where you decide that the commercial team gets verified attendance and does not get to reach past it into the raw scan table.
The access control part earns more attention in an event business than the tooling usually gets, because exhibitor facing reports and internal reports are served from the same measures. Cube's documentation places policy in the layer so that "the same policies apply to every consumer". That property is what you want when an exhibitor portal and a board dashboard both ask for lead counts and only one of them may see other exhibitors' figures. Implementing that rule separately in each consumer is the mechanism by which an exhibitor eventually sees a competitor's scan totals, and it happens in the sixth consumer, built by somebody who never saw the first five.
The five parts of a measure definition
A definition that resolves the 762 badge holder argument has to say five things, and event metrics tend to fail on the last two.
- The fact table and its grain. One row per badge scan, from the scan fact.
- The aggregation. A distinct count of badge identifier, which is what makes attendance different from scan volume.
- The joins required. To the badge dimension for badge type, to the edition dimension for the show and year, to the date dimension for show day.
- The filters that constitute the measure. Badge type is excluded for staff and contractor. Scan timestamp falls inside published show hours. Scan is not flagged as a test read.
- The entity being counted. Distinct badge holders, where a badge holder is a person, so two badges issued to the same person under different registrations count once.
The fourth and fifth are where the three teams diverged, and they are exactly the parts that a column name cannot express. A warehouse column called verified_attendance in an aggregate table carries a number with no statement of the rules that produced it, which is why every consumer eventually rebuilds it from the atomic fact and gets a slightly different answer.
Why does the definition belong outside the BI tool?
Every BI tool has somewhere to put a calculated measure, and the reflex is to define attendance there. It works, for one tool.
An event portfolio has more consumers than one tool. There is the internal dashboard, the exhibitor facing post-show report, the sales workspace where renewal conversations happen, an annual PDF the marketing team builds, a spreadsheet the finance lead maintains, and increasingly a language model answering questions over the warehouse. Define the measure in the BI tool and the other five consumers cannot reach it. They query the fact table directly and reinvent the filters, which is precisely how you get to 11,689 and 10,927.
There is a second reason that matters more over five years. A definition inside a BI tool is not reviewable. It has no diff, usually no author, and no history, so a change to the staff badge filter in March appears as a movement in a trend line in April with nothing connecting the two. Keeping the definitions in version control with owners and a changelog is a discipline in itself, argued in metric definitions as code.
Naming the measures that genuinely differ
A semantic layer is sometimes sold as a way to force everyone onto one number. That is the wrong goal, and the Kimball Group made the better argument in 2013 in its technique for conformed facts:
If the same measurement appears in separate fact tables, care must be taken to make sure the technical definitions of the facts are identical if they are to be compared or computed together. If the separate fact definitions are consistent, the conformed facts should be identically named; but if they are incompatible, they should be differently named to alert the business users and BI applications.
Apply that to the three attendance numbers. Finance genuinely needs a figure that excludes contractors, because contractors do not buy anything and the profitability model treats attendance as a proxy for commercial audience. Marketing genuinely wants the badge holder count, because that is the number of people who were in the building.
So the semantic layer holds three measures with three names. Verified attendance is 10,927 and excludes staff and contractors. Badges scanned is 11,689 and excludes nothing. Trade attendance is 11,204 and excludes staff only. Each has one definition, each has an owner, and any report using one of them can be traced to the rule that produced it. The disagreement has not been suppressed. It has been given names, which is what makes it survivable.
What does this cost to run?
Two costs, and they are not the ones people expect.
The compute cost is real and manageable. Every dashboard request now resolves against a definition and generates SQL, so the warehouse does the work on each load. Whether to precompute those results into an aggregate table or pay the query each time is a genuine engineering trade, and it depends on how many measures and editions you have. Forty editions by nine measures is small enough that both options are open, which is the argument made in transform time versus query time metrics.
The organisational cost is the one that sinks projects. Somebody has to convince finance, marketing and exhibitor relations to agree on names, and to accept that their number is now visible to everybody with its rules attached. That conversation takes weeks and produces the actual value of the exercise, because the definitions written down at the end of it are the first honest description of what your organisation measures. Buying a tool and skipping the conversation gives you a layer that nobody's numbers come from.
Where this stops
A semantic layer makes definitions consistent. It has nothing to say about whether they are right.
If your badge scan coverage is 62 per cent because two halls had no readers at the entrance, then verified attendance defined perfectly, served identically to every consumer and traceable to the row, is still 38 per cent low. The layer will deliver that wrong number with total consistency to six systems, and the consistency will make it more believable than it was when three teams disagreed. The disagreement was at least a signal that somebody should look.
It also does nothing for the metrics that are not computable from what you have. Repeat visitor rate needs identity resolution across editions before any definition of it means anything. Dwell time needs scans at exit as well as entry. A definition of a measure your data cannot support produces either a null or a confident approximation, and the second is worse.
Start with one measure and one meeting. Take verified attendance, get the three people who publish a version of it into a room, and write down the filters each of them applies. If the answers differ, you have found the gap and you have the beginning of a definition. If they agree, write the definition down anyway and put it in the data platform repository, because agreement that exists only in three people's heads has a way of expiring quietly.
Questions people ask about semantic layer for event metrics
- What is a semantic layer in an event data platform?
- A layer between the warehouse and its consumers that holds metric definitions as code. Instead of each dashboard writing its own SQL against the scan fact, each one requests a named measure, and the layer generates the query from a definition that specifies the joins, the filters and the aggregation.
- Why do two teams get different attendance numbers from the same table?
- Because attendance is a definition rather than a column. One team counts every badge that scanned, another excludes staff badges, a third also excludes contractors and scans outside show hours. All three queries are correct against their own rules, and nothing in the warehouse records which rule is the official one.
- Does a semantic layer replace the data warehouse?
- No. The warehouse still holds the facts and dimensions and still does the heavy work. The semantic layer sits above it and holds the definitions, translating a request for a named measure into SQL. A badly modelled warehouse produces the same wrong answers through a semantic layer, only more consistently.
Related reading
- Keeping metric definitions as code under review like any other change
- Transform time versus query time metrics for a portfolio reporting stack
- A fact table grain mismatch is why your two reports disagree