datascale
DE EN
Contact

Stack

The tools we use. Honest and transparent.

We recommend what we run ourselves or build in production for clients, and name the tradeoffs honestly. Where a partnership exists, it is declared openly, commission model included.

EU-first für DACH

Where your data lives.

Default stack for DACH clients, seven layers from consent to activation. Filter by trust category to see where your own tools land.

collect

  1. CMP

    Usercentrics

    Munich · EU

    EU-native

    ?

  2. Tagging

    sGTM · stape.io

    Frankfurt · EU

    EU-native

    ?

  3. Analytics

    GA4

    via sGTM · EU

    EU-configurable

    ?

visualize

  1. BI

    Power BI (EU tenant)

    Frankfurt · EU

    EU-configurable

    ?

activate

  1. Activation

    Hightouch

    US · managed

    US with SCCs

    ?

Foundation · under every stage

  1. Warehouse

    BigQuery

    europe-west3

    EU-native

    ?

  2. Modeling

    dbt

    runs on the warehouse

    EU-configurable

    ?

Also supported:

  • Cookiebot
    ?
  • OneTrust
    ?
  • Plausible CE
    ?
  • Matomo
    ?
  • Snowflake
    ?
  • Data (Looker) Studio
    ?
  • Census
    ?
  • RudderStack
    ?
  • funnel.io
    ?

Diagram: Datascale default stack for DACH clients. We adapt for international projects. As of Jun 2026.

Reference architectures

What a real stack looks like.

Three typical setups we build, with the actual tools at every layer. No tool stands on its own, the value is in a clean data flow from source to activation.

Shopify store, paid-heavy. Conversion tracking has to survive ITP and adblockers, otherwise Meta and Google optimise against gaps.

  1. Source

    Shopify

    Webshop

  2. Collection

    sGTM via stape.io

    Frankfurt · EU

  3. Storage

    BigQuery

    europe-west3

  4. Modeling

    dbt

  5. Reporting

    Data (Looker) Studio

    Power BI

  6. Activation

    Meta CAPI

    Google Ads

Example architectures, not a mandatory setup. We adapt per project. As of Jun 2026.

Filter
57 of 57 tools
Tool We use it. Integration We build it in production. Partner Certified · disclosed openly.

Collection & Tracking

Server-Side & Tag Management

We build this Measurement & Privacy Engineering →

Event Collection & CDI

We build this Modern Data Stack & Composable CDP →Measurement & Privacy Engineering →

Analytics

Web & Marketing Analytics

We build this Measurement & Privacy Engineering →

Product Analytics

Qualitative & Behavior

Ingestion

We build this Modern Data Stack & Composable CDP →

Storage & Modeling

Data Warehouse

We build this Modern Data Stack & Composable CDP →

Transformation

Orchestration & Scheduling

We build this Modern Data Stack & Composable CDP →

Reliability & Governance

Data Observability & Quality

We build this Data Reliability & Governance →

Data Catalog & Metadata Governance

We build this Data Reliability & Governance →

BI & Dashboards

Measurement

Privacy broke attribution, not measurement.

Privacy broke multi-touch attribution, not measurement. Safari ITP, iOS ATT and consent flows pushed MTA's identity coverage from over 90% to roughly 30 to 60%. The 2026 answer is not a single tool but a layered measurement stack: MMM as the strategic backbone, validated by incrementality tests on the largest channels, with attribution left as a directional signal for ongoing optimisation. Beneath it sit first-party event collection and a warehouse as the golden record, on which a composable CDP activates. That is marketing engineering. It matches our warehouse-centric, self-hosted default exactly.

We build this Measurement & Privacy Engineering →

Experimentation

We build this Experimentation & Conversion Intelligence →

Activation & Lifecycle

Activation & Composable CDP

We build this Modern Data Stack & Composable CDP →

Marketing Automation & Lifecycle

We build this Revenue Intelligence & Executive BI →

AI & Intelligence

Semantic layer first, then agent.

An AI agent is only as good as the governance layer beneath it. Raw schema access for an LLM reproduces the Shadow-BI problem at the agent level: no metric definitions, no lineage, no row-level security. So the order is: semantic layer first, then agent, then observability. Not the other way round. Snowflake measures roughly 20 percent higher text-to-SQL accuracy with a semantic model than with schema alone. The market is catching up: in January 2026 Usercentrics acquired MCP Manager, a consent and governance layer for the Model Context Protocol.

We build this AI Strategy & Data Readiness →Modern Data Stack & Composable CDP →

What this site runs on

This very site runs on exactly this stack: self-hosted in the EU.

Hosting & Infrastructure

CMS

Email

  • Q01
    Do I need a semantic layer before deploying an analytics agent?

    Yes. Without a semantic model the agent guesses from the raw schema and reproduces Shadow BI at the agent level. Snowflake measures roughly 20 percent higher text-to-SQL accuracy with a semantic layer than with schema alone, plus consistent metric definitions and lineage.

  • Q02
    Is MCP GDPR-compliant?

    That depends on the governance layer, not the protocol. Raw MCP schema access often uses a service account instead of user identity and bypasses row- and column-level security. It becomes compliant when MCP runs on top of the semantic layer with real access control.

  • Q03
    Self-hosted LLM observability in the EU, what should I use?

    Langfuse as the default, MIT-licensed and self-hostable in the EU. Arize Phoenix if you've standardised on OpenTelemetry. Both keep the trace data in your own infrastructure.

  • Q04
    Cortex or BigQuery ML?

    Follow the warehouse, not the trend. If the lakehouse runs on BigQuery in the EU, BigQuery ML is the coherent default. In a Snowflake-committed shop, Cortex is the more direct path.

  • Q05
    MMM or attribution in 2026?

    Both, but layered. MMM carries strategic budget allocation, incrementality tests validate the largest channels, attribution stays the directional signal for ongoing optimisation. Privacy ended multi-touch attribution as the single source of truth, not measurement itself.

  • Q06
    Is open-source MMM (Meridian, Robyn) worth it?

    Yes, but count honestly: the licence is free, the total cost of ownership is data-science headcount. An unmaintained MMM decays, and when the data scientist leaves, the model breaks. Running and maintaining it is exactly our job.

  • Q07
    How do I measure without third-party cookies?

    With the layered stack: first-party event collection as the foundation, MMM for strategic contribution, incrementality tests for causality. None of these layers needs third-party cookies.

  • Q08
    Composable CDP or packaged?

    It depends on your data-engineering maturity. If a cleanly modelled warehouse is your golden record, a composable CDP activates directly on it, with no data copy. Without that foundation, a packaged solution can be the faster start.