On Saturday 14 November, Morgan and I were at MeasureCamp Brussels. The train was late (it seems to have become a tradition for us), but we arrived just in time for the first sessions. Just enough time to grab a coffee and look at the session board.
For those who don't know the format: a MeasureCamp is an "unconference". There is no programme set in advance: attendees pitch their sessions on the morning itself, on a big board. The result is a very dense day, from tracking governance to YouTube ads, by way of AI and data quality.
Here is what we took away.
The sessions
Julien Coquet (N7 Data): Analytics As Code
First session, and already a serious topic. Julien Coquet starts from a simple observation. We manage our infrastructure as code, with Git. But the tracking plan still lives in Slack threads and outdated spreadsheets.
His proposal: write the tracking plan in YAML and version it in Git. For each event, you describe the required parameters, the owner and the governance rules (no personal data, consent required). Rules validate the spec automatically. The GTM implementation follows from it. A weekly check then verifies that the collected data still matches the plan.
He even showed Playwright tests generated from this plan, run on every pull request, with and without consent. His three takeaways, "apart from Belgian chocolate": the tracking plan is infrastructure, pull requests enable governance, and Git becomes the audit trail.
Aristide Riou: the GA4 interface, rebuilt on BigQuery
Aristide Riou had had enough of the GA4 interface. So he rebuilt it, directly on the BigQuery export: explorations, funnels, paths, segments. A nice personal project. What we mostly kept is a simple and effective idea: extract any URL parameter on the fly and turn it into an analysis dimension.
Juliana Jackson (Jellyfish): YouTube ad performance
Juliana Jackson shared what she learned from testing hundreds of creatives for 75 major brands. Her main message: we are leaving the age of optimisation, and now need to enter the age of exploration.
A few very practical tips. The first five seconds matter most. You do not reuse a TV ad as is on YouTube. And 95% of YouTube users have the sound on: say the call to action out loud, and prefer voice over music at the start of the video.
She also pointed to ABCD Detector, an open source Google tool that automatically checks videos against the ABCD creative framework.
Piotr Gruszecki: the semantic model, essential for conversational analytics
Why is everyone talking about semantic layers? Piotr Gruszecki gave a simple answer: AI agents cannot work without one. He showed an architecture for querying Google Search Console data in natural language. The data is loaded into Snowflake, a YAML semantic model describes the metrics and dimensions, and Claude accesses it through an MCP.
The key point: the AI does not guess the tables. It relies on validated business definitions. That is what makes its answers reliable.
Reports in the AI era: both, not either
Marilia Schuelter and Ivan Burbich led a discussion on what clients expect from reporting, backed by Coupler.io figures. 63% of 2026 MeasureCamp attendees still prefer traditional reports. Yet the share of AI tools in connector usage went from 3.9% to 28.4% in a year.
The session's conclusion: it will not be one or the other, but both. With one condition: only 7% of companies querying their data through an MCP have taken the time to document the context of their datasets. Without that written context, self-serve AI produces wrong numbers that look right. Once again, it all comes down to context.
Antoine Brion: offline conversions that really count
Antoine Brion tackled a problem many lead generation advertisers know well. Google Ads and Meta only see the submitted form. They know nothing about what comes next: qualified lead, site visit, signed quote, quote amount. As a result, bidding optimises for form volume, not value.
Using the example of an energy renovation company, he walked through a complete method: bring the CRM, the catalogue and the ad identifiers together in BigQuery, value each stage (expected value × close rate), then send the conversions back to the platforms every day.
One tip to remember: bid on a stage that happens within a week of the click, such as the site visit, rather than on the signed quote, which comes a month later.
Adrienne Vermorel: Green Checks, Wrong Data
The tag fires, the screenshot looks right… can we go home? Not so fast. Adrienne Vermorel proposed a way to read test results. She distinguishes where the data was observed (action, event, sent, received, usable data) and within what scope (environment, journey, cases tested).
She illustrated it with four traps, all from real projects. For example, an add_to_cart that also sends the products of a list viewed just before, because the ecommerce object was not cleared. Or a redirect to the site's language version that drops the gclid and UTMs. Or a debug_mode left on in production, which makes data disappear from reports while everything looks perfect in DebugView.
A very useful session, which will directly feed into our tracking QA.
Our session: expose your data fast, cheap and yours
Late in the afternoon, Morgan and I presented "Expose Your Data Fast, Cheap & Yours".
Our thesis: AI has changed the cost/benefit equation of coded dashboards. Data Studio has its limits: no versioning, no API, no automated creation, and strong dependence on the tool. Today development time is collapsing, and we build our dashboards directly in HTML, with Observable.
We then showed how to expose these dashboards securely, inside the client's Google Cloud project: Cloud Run, protected by Identity-Aware Proxy (IAP). Access is managed with Google identities, and the data never leaves the project. Behind the same door, you can also publish custom tools: a UTM generator, glossaries, specs and other small everyday tools. And we finished with our latest R&D topic, creating videos from data.
We also talked about the limits. This is not a self-serve BI tool. HTML is perfect for recurring, narrative reporting aimed at a non-analyst audience, while a BI tool remains better suited to ad hoc exploration. Our conclusion: "Stop building tools nobody uses, start shipping answers."
Thank you!
A big thank you to the MeasureCamp Brussels organising team, and to the sponsors, for a perfectly run day. The networking was really nice: between sessions, over coffee or at lunch, the conversations were as rich as the presentations.
What we take away
One common thread ran through the day: context. The versioned tracking plan, the semantic model, dataset documentation, the test reading grid: each time, it is about writing down what the data means and what has actually been checked. That is what then lets AI, ad platforms and business teams rely on that data.
At Woptimo, this is exactly the path we follow with our clients. If you would like to talk about it, we are here.