What a marketing data warehouse is, and when an agency needs one.

A data warehouse is a database built for reading rather than writing. A client's ad platforms, CRM and store each hold their own slice of the truth, and each answers questions only about itself. A warehouse is one place all of it lands, for every client on the roster, so a question that spans two of them can be answered at all.

For an agency that means a cloud data warehouse in an account the agency controls, holding every client's data in one set of tables. The important part of that sentence is not which warehouse product. It is who controls the account.

If every question a client asks is about one platform, you do not need a warehouse. Google Ads answers questions about Google Ads perfectly well, and moving that data somewhere else to ask the same question is work with no payoff.

You need one when the questions start spanning sources. What did the client spend to acquire the customers who actually paid. Which channel produced the deals that closed. For an agency there is a third kind: which clients are underspending this month, asked once across the whole roster. None of those can be answered inside any single platform, and no amount of dashboard building inside one will produce them.

A warehouse has layers, and skipping the middle one is the most common way to end up with an expensive database nobody trusts.

• Raw

  Exactly what each platform returned, unmodified, one dataset per source. Never reported on directly. Its job is to be re-runnable.

• Modelled

  The raw tables reconciled into consistent ones: one name per metric, one type, one grain, one currency. This is the layer reports and AI should read.

• Reporting

  Optional. Pre-aggregated tables shaped for specific dashboards, when a dashboard is slow enough to need them.

It is not a product category, it is a property of the data. A metric is AI-ready when it has one name, one type, one grain and one currency across every source it arrives from.

The reason it matters is that a language model cannot ask you what you meant. Given two columns that are secretly the same measure at different grains, it will average them, and it will present the result with the same confidence as a correct answer. Data that is merely centralised is not AI-ready. Data that is modelled is.

Three, in the order people make them. Loading raw exports and reporting straight off them, which produces reports that disagree. Modelling inside the BI tool instead of the warehouse, which puts the definitions somewhere only that tool can see. And having no owner, so nothing is documented and the first schema change upstream breaks a dashboard nobody knows how to fix.

All three come from the same place: treating the warehouse as storage rather than as the layer where meaning gets decided.

If you take one thing from this: whoever owns the cloud project owns the data. A tool that stores your clients' history on its own infrastructure and shows it to you is a tool you cannot leave without losing that history.

SightShark ingests every client's sources into your warehouse and models them in place, in your cloud project, in a region you pick, with each row scoped to its client. If you stop paying, the history is already where it lives.