GT DIGITAL / Gregor Trela

Experimental Media Lab — Next

I don’t just talk about where media is going.I build things to find out.

The Lab is where my media experience meets tools I build myself — small, working prototypes that test how intelligence, automation and agents could change the way media is planned and bought. Each one is published with its status and its limits.

Lab rules

Four rules for every experiment.

  1. Real constraints

    Real audiences, real costs, real trade-offs. If it only works in a demo, it doesn’t count.

  2. Show the working

    Every assumption is visible, sourced where possible and editable. Nothing is a black box.

  3. Human sign-off by design

    The tools recommend and challenge. A person decides — and the tool is built around that moment.

  4. Honest status

    Live, in development or an open question — and what is not true yet, said out loud.

Experiment log

What’s on the bench.

Experiment 01 Live October 2026

Media planning that shows its working.

Can a planning tool be fast — and still show every assumption behind its numbers?

What it does

Takes a campaign brief, allocates the budget across twelve channels, estimates net reach and frequency against the real Swiss audience universe and compares three strategic modes — then challenges its own plan before a human signs it off.

What I learned

Credibility comes from disclosure. The planner earns trust because its benchmarks, formulas and limits are on the page — and anyone can overwrite them with their own rate card.

Limits

Planning-grade Swiss benchmarks, not a booked rate card. Deterministic arithmetic in the browser — no language model, no live media data, nothing leaves the page.

Open the planner
Experiment 02 In development Not public yet

A media agent that reads the brief — and leaves the maths alone.

What should a language model do in media planning — and what should it never do?

What it does

A brief goes in as prose; a calculated media plan comes out. The model reads the brief, derives the parameters and names contradictions and assumptions. The engine allocates the budget and calculates reach, frequency and waste. The model then writes the hand-over from the calculated numbers.

What I learned

The model must never decide the channel shares. My first scaffold let it guess the percentages — which can’t be reproduced twice and can’t produce reach or frequency. Now the arithmetic sits in the engine: same brief, same plan.

Next

Replace the benchmarks with a real rate card before it meets a client, and add LinkedIn as a channel for B2B briefs.

Open questions

What I want to find out next.

Not experiments yet — the questions the next ones are meant to answer.

  • Open question

    Live research

    What changes when a planner reads current cost, audience and inventory data instead of static benchmarks — and a model reads the brief instead of matching keywords?

  • Open question

    A partner graph

    Can a structured map of DSPs, SSPs, CTV platforms, retail media networks and measurement partners become part of planning itself?

  • Open question

    Controlled actions

    Where can an agent act — with write access to buying platforms behind approval gates — and where must it only recommend?

  • Open question

    Machine visibility

    When people ask AI for recommendations, what makes a brand understood, trusted and recommended — and how would you measure it?

Next — 04 PerspectiveHow I read the industry — and why these experiments matter.

Contact

Let’s talk.

If one of these sounds like you, write to me. A few lines are enough.

  • A leadership roleIn media, marketing or AdTech.
  • AdvisoryA media, investment or growth question that needs an independent, senior view.
  • A partnershipMedia, AdTech and technology companies looking for strategic collaboration.
  • An exchangeAbout where media is going — or about something in the Lab.
hello@gt-digital.ch LinkedIn Zürich, Switzerland