05 · What I do # Analytics Data strategy · 5 min read Analytics is often described as the work of producing numbers. It is closer to the work of deciding what the numbers mean: what each concept refers to, how it is measured, how it is grouped, which movements deserve attention and which are noise. That agreement is the semantic layer, and it has quietly become the most valuable asset a data function owns, because everything downstream now reads from it, including the models. In short - A metric is a definition before it is a number, and those definitions belong where the computation reads them rather than in a document. - Not everything measurable deserves watching. A tracked metric needs an expected range and a decision it would actually change. - The semantic layer is the machine readable version of those agreements, and the context every AI system consumes. ## A metric is a definition before it is a number Every metric carries a set of decisions that rarely get written down. What population counts. At what grain. Measured when, in whose timezone. What happens to refunds, trials, internal accounts, cancelled orders, partial deliveries, currency conversion. Two teams can compute revenue with identical SQL skill and disagree by a fifth, purely because those decisions were taken silently and differently. So the first deliverable is not a dashboard, it is a definition with an owner: what this metric means, what it excludes, when it changes, and who decides. Written once, stored where the computation actually reads it rather than in a document nobody opens. A definition that exists only inside one report is not a definition; it is that report's private opinion. ## Dimensions decide which questions can exist Measures get the attention, but grouping is what determines the range of answerable questions. If customer segment is defined one way in marketing and another in finance, no amount of correct arithmetic will make their funnels agree. If a product hierarchy has no stable level between item and category, nobody can ask the question that lives in between. Conformed dimensions are the unglamorous fix: shared definitions of customer, product, region, channel and time, used identically everywhere, with hierarchies that match how the business actually reasons rather than how a source system happens to store things. Get this right and questions compose. Get it wrong and every cross-domain question becomes a reconciliation project. ## Signal, noise, and the cost of watching Not everything measurable deserves to be measured, and almost nothing deserves to be watched continuously. Every standing metric is a claim on human attention: someone will look at it, wonder about a movement, and spend an afternoon explaining a fluctuation that was always within normal variation. Which is why a metric worth tracking needs more than a definition. It needs an expected range, a sense of what a meaningful change looks like against its usual noise, and an answer to the question of what anyone would do differently if it moved. A number that changes nobody's behaviour is decoration, however precisely it is computed. > If a definition lives only inside a dashboard, you do not have a metric. You have an opinion. ## One time answers, and standing instruments Two very different products get confused with each other. One is the answer to a question somebody has now: is this campaign working, why did churn spike in March, would this pricing change have helped last year. Those deserve a fast, rigorous, deliberately disposable analysis. Writing it as production code, wiring it into a scheduler and maintaining it for three years is waste dressed up as diligence. The other is an instrument: a measurement the business will consult repeatedly to steer something. That earns automation, an owner, tests, monitoring and a place in the semantic layer, because it will be trusted without being re-examined every time. The triage question is simple and worth asking out loud before any work starts. Is this a decision that recurs, or a question that will be answered once? Teams drown when everything ad hoc is promoted to permanent, and they burn out when the same answer is rebuilt by hand every month. ## The semantic layer is the machine readable version All of the above can live in people's heads, and in most organisations it does. The semantic layer is what happens when it is written down in a form that systems can read: entities and their keys, metrics with their formulas and grain, dimensions and hierarchies, valid joins, time semantics, ownership and freshness expectations. The immediate payoff is consistency. The same definition serves the BI tool, the notebook, the embedded chart in the product and the export to finance, so the number is the number regardless of the door it came through. The second payoff is speed, because analysts compose from agreed building blocks instead of re-deriving joins and filters that somebody already got right. ## Context is what AI systems actually consume This is where the semantic layer stopped being a nice to have. A language model asked a business question has no way of knowing that active customer excludes trials, that revenue is net of refunds, that these two tables must not be joined directly, or that last quarter means the fiscal one. Given the raw schema it will produce something fluent, plausible and wrong. Given the semantic layer as context, it is answering with the organisation's own definitions. Machine learning has the same dependency in a different shape. Features built on inconsistent definitions produce models that disagree with the reports the business trusts, and training sets assembled without point in time correctness quietly leak the future into the past. A semantic layer that records grain, effective dates and lineage is what makes a feature reproducible rather than merely available. The pattern is the same for both: the model is only as coherent as the metadata it is handed. Investing in definitions used to pay off in agreement between humans. It now also decides whether the automated systems built on top of the data are worth listening to. ## Treating definitions as a product Semantic layers rot in the same way documentation rots, unless they are maintained like software. Versioned, so a change to a definition is visible and dated. Reviewed, so a new metric has to justify its existence against the ones that already exist. Deprecated on purpose, so the graveyard of abandoned measures does not become part of the context every consumer reads. Done that way, it becomes the thing that outlives tools. Dashboards get replaced, warehouses get migrated, models get retrained. The agreement about what the business means by its own words is what carries forward, and it is worth more than any of them. Written by Gonzalo Rodriguez Pardo, Chief Data Officer