AI marketing claims. The evidence, the limits, and what I'd test next.

AI × Marketing. Measured.

What to do with AI in marketing.
And how to check it works.

Claims about AI in marketing, read against the evidence: what holds up, what doesn’t yet and what I’d try next.

Free. One-click unsubscribe.

Read the research →
  • Claim.
  • Evidence.
  • Limits.
  • Next step.

The latest

Ideas you can put to the test.
What I’d do with these findings.

Field notesOpinion

AI analytics

When everyone has the same models, the advantage is the data they don't have

The top AI models are now within 25 points of each other and getting 10x cheaper a year. In an AI-native go-to-market, the edge moves to proprietary ground truth.

Chart from “When everyone has the same models, the advantage is the data they don't have”
Field notesOpinion

AI analytics

What made Google's AI understand users better wasn't more data

Before shipping an LLM that reads billions of watch histories, Google tested what to feed it. Cleaner fields, grouping by meaning and curated examples beat volume and size.

Field notesOpinion

AI advertising

The agency fee is about to become ad spend

A top SEO told me AI made him more expensive, not cheaper. Then I built Meta campaigns by talking to Claude. Here's what that means for agencies, and the two jobs left.

Field notesOpinion

Measurement

Google's recommender showed new content 41% less. It performed better

In a live test on a platform with billions of users, the new interests Google's LLM suggested were shown far less and watched more. Your ad budget runs on the same loop.

Field notesOpinion

Measurement

Why I paid $1,000 for one word

Unmodeled data is data nobody filled in with guesses. Here's why that word became the name of this newsletter, and why it matters more now that AI runs on our data.

View all research →

New pieces reach subscribers first. Get them by email →

Unmodeled LAB

Less talking about AI.
More building it, together.

A small group of marketers in one room, building with AI on real marketing data and sharing what they find.

LAB #Pilot is the first one. I’ll document what works and what doesn’t, and publish it here as research. Not in Barcelona? You’ll be able to read what came out of it.

LAB #PilotResults being written up

Build Your Own AI Growth Analyst

· SB Hotels Plaza Europa · Free

Can 20 marketers build a functional AI Growth Analyst in one morning?

One morning in Barcelona. 20 marketers try to build an AI growth analyst on real marketing data, in the room, together. I document what happens and publish it as research: what got built, what broke and what worked.

Explore LAB #Pilot →

The Unmodeled method

I don’t predict the future.
I check it.

  1. 01

    Hypothesis

    What am I trying to find out?

  2. 02

    Experiment

    How can it be tested?

  3. 03

    Evidence

    What actually happened?

  4. 04

    Verdict

    What should you do next?

A test without a proper control doesn’t prove cause and effect, and I say so. Every piece keeps results, observations and interpretation apart, and carries one label: Experiment Observation Opinion

Read the methodology →

Written by one person.
Checked against the data.

Unmodeled: what was observed, before estimates fill the gaps.

An independent publication powered by Sealmetrics. I pick the topics and the verdicts; Sealmetrics doesn’t.

I’m not here to prove AI always works. I want to know when it does, when it doesn’t and what to do about it.

Written by Rafa, founder of Sealmetrics.

More about Unmodeled →