You Need a Data Warehouse Later Than You Think

Under roughly €1M in annual revenue, you do not need a data warehouse, and buying one anyway burns runway on infrastructure your questions cannot fill. Past a certain point, you cannot afford to keep operating without one. Both statements are true, and the threshold between them has little to do with revenue as such. Revenue is a decent proxy at the extremes; what actually moves the needle is the complexity of the questions your business needs answered. Analytics needs evolve with stage, and the expensive mistakes sit at both ends: the seed-stage brand paying for infrastructure it cannot use, and the €8M brand steering ad budget with numbers it knows are wrong.

A data warehouse, for this discussion, is a database built for analysis rather than operations: a place like BigQuery where copies of your data from every tool land in one system, so they can be joined and queried together. The joining is the entire point. Every signal below is some form of needing two sources in one query.

Signs you do not need one yet

If you are spending on one channel, the channel's own dashboard is honestly fine. Meta's reporting on Meta performance is flawed in known ways, but with nothing to join it against, a warehouse would hold one lonely table that answers nothing the dashboard could not.

If every question your team actually asks this quarter can be answered inside a platform UI, the gap a warehouse fills does not exist yet. Be honest about the questions you ask, as opposed to the ones you imagine asking after the next fundraise.

And if the team is still just the founders, the constraint on decisions is time and attention, never data infrastructure. A warehouse nobody queries is a monthly reminder of an optimistic afternoon.

There is no shame in any of this. Skipping the warehouse at this stage is the correct engineering decision, the same way not building for a million users on launch day is.

One cheap move is worth making early, though: switch on the free GA4 BigQuery export now, even if nothing queries it for a year. The export only collects data from the day you enable it, so a brand that flips it on at €800K crosses the warehouse threshold later with two years of event history waiting, while a brand that waits starts from zero. Ten minutes of setup, and storage at this volume costs cents.

The four signals you do

The threshold announces itself. In practice it looks like one or more of these.

Cross-source questions start mattering. Blended CAC across Meta and Google. Lifetime value cut by first-touch channel. Whether email is retaining the customers paid social acquires. No single dashboard can answer these, because each platform sees only itself, and tab-switching cannot multiply two numbers that live in different tools. The moment real budget decisions hang on cross-source questions, dashboards have hit their limit.

Two tools report different numbers for the same metric, and money rides on which is right. Shopify says the weekend did €80,000; GA4 says €64,000; Meta claims credit for most of both. The discrepancies have knowable causes, but adjudicating them requires putting the sources side by side at order level, which is warehouse work. Until then, every channel argument in the company is two people quoting different tools at each other.

Someone rebuilds the same spreadsheet every Monday. Export from four platforms, paste, fix the columns, VLOOKUP, send. It costs a morning a week, breaks without warning when a platform renames a field, and the person doing it is usually your best analyst. Manual, repeated, cross-source assembly is a warehouse pipeline being executed by hand.

You want AI on your data. Asking an assistant "which channel drove our best customers last quarter" only works if channels, orders, and customers are already joined, governed, and consistently defined somewhere the AI can query. Pointing AI at five disconnected dashboards produces five confident, disconnected answers. Joined, governed data underneath is the prerequisite, and it is a warehouse.

The second signal deserves the most weight. Question complexity, in the end, is what separates the stages: a brand asking "did this campaign pay off" does not need infrastructure, and a brand asking "which acquisition channel produces customers still buying at month twelve" cannot answer without it, whatever either brand's revenue is.

What getting one actually means in 2026

The dated mental model is a six-month platform build with consultants and a steering committee. That era is over at this end of the market. The current version: a BigQuery project (yours, from day one), managed pipelines landing your ad, analytics, and store data in it, transformation models that turn raw exports into clean orders and sessions and spend tables, and a semantic layer holding one agreed definition of each metric. Live in weeks, not quarters.

The sequencing matters less than people expect, but a typical first month looks like this: pipelines and the GA4 export land in week one, core models for orders, sessions, and spend take shape in week two, and by week three there is a blended CAC number both the founder and the performance lead accept, usually for the first time. The early wins are mundane on purpose: one agreed revenue figure, one Monday spreadsheet retired.

Two fears keep brands stuck past the threshold, and both are smaller than they look. Warehouse cost fear dissolves once you run the arithmetic, since marketing-sized data usually lands near a euro a month on BigQuery. Build cost is the realer concern, and self-assembly does carry a heavy people bill, which is precisely why managed models exist for teams without data engineers.

Maturity by stage

Rules of thumb, with the usual caveat that the questions matter more than the revenue band:

Most brands cross the warehouse line somewhere in the second row, and almost all of them describe the same regret: waiting a year past the signals, because the Monday spreadsheet still technically worked.

Where Verity fits

Verity is the managed version of the 2026 model: pipelines into a BigQuery project you own, SQLMesh transforms with tests and lineage, a semantic layer with agreed definitions, and Data Chat for plain-language questions on top. It exists for teams that have hit the four signals without hiring data engineers, with plans from €500 per month on the pricing page.