Writing
Notes on tracking, clean data and reporting.
Short essays on how data gets collected, cleaned and reported, and on the awkward place where clean data has to meet the AI that everyone wants to put on top of it.
- Tracking·November 10, 2026·5 min read
Retargeting audiences are only as good as the events behind them
A retargeting audience is a list built from website events. If a checkout step is recorded wrong, the list is wrong, and the ad platform spends against it anyway. How to build the events and the audiences so they describe the same people.
Read → - Reporting·November 3, 2026·4 min read
When GA4 reports stop being enough
GA4's own reports are fine for a quick look. They run out when you need history, exact rows, or a dashboard that loads every time. The limits in the order teams hit them, and what moving the data to BigQuery does and does not fix.
Read → - Tracking·October 27, 2026·5 min read
Why GA4, Google Ads and Meta never agree on conversions
Google Ads, Meta and GA4 will each report a different number of conversions for the same week, and all three can be working correctly. Which differences are rules, which are mistakes, and which number to use for what.
Read → - Reporting·October 20, 2026·4 min read
The real reason your dashboard doesn't get used
It is not a training problem and it is not the design. A dashboard gets abandoned the first time someone senior finds a wrong number on it. What rebuilds that trust, and what does not.
Read → - Tracking·October 13, 2026·6 min read
What a GA4 and Google Tag Manager audit actually checks
A tag audit is an inventory of what your site records, compared with what you think it records. The GA4 and Google Tag Manager checks that find the most, in the order I run them.
Read → - Analytics Engineering·September 29, 2026·8 min read
Should I use dbt or write my own SQL pipelines?
A candid, practitioner take on when dbt is the right call, when hand-rolled SQL pipelines are, and what happens when a homegrown stack outgrows the one person who built it. Includes a decision framework for the actual fork in the road.
Read → - Agency Ops·September 22, 2026·6 min read
The three-day question
You emailed the account team a quick question on Monday. It is Wednesday. The three-day question is not a capacity problem. It is a structural one, and it is quietly the most expensive thing on the retainer.
Read → - AI Engineering·September 15, 2026·6 min read
Retrieval on dirty data is a distribution problem
Every RAG post is about chunking, embeddings, and rerankers. Almost none are about the corpus. If your warehouse or your docs disagree with themselves, retrieval just distributes the disagreement, and a smarter model gives you the same wrong answer with better prose.
Read → - AI Engineering·September 1, 2026·6 min read
The semantic layer nobody owns
Between your transformation layer and your dashboards there is supposed to be a place where 'active user' gets defined once. In most stacks it does not exist, and no one has been given the job of making it exist. AI on top of that gap just distributes the disagreement faster.
Read → - AI Engineering·May 4, 2026·5 min read
"We're using AI for our data" is not a plan
Four different jobs get hidden inside that sentence, and most of them end in wrong answers or wrong data. Map the stack first, then give the model a specific job with a review gate.
Read → - AI Engineering·April 6, 2026·5 min read
What is data hygiene, and why does it matter for AI
Data hygiene means your data is accurate, consistent, and documented. Most data stacks are one of those three things. Here's where the gaps live and why they hit harder once AI is in the loop.
Read → - AI Engineering·March 30, 2026·4 min read
Natural language vs the transformation layer: do you still need SQL?
The transformation layer isn't going anywhere. The question is who gets to interact with it. Right now the answer is whoever knows SQL. That's a small group.
Read → - AI Engineering·March 23, 2026·4 min read
Why your AI chatbot keeps giving wrong answers
The chatbot isn't broken. It's working exactly as designed on bad inputs. Here's what's actually happening and where the real fix lives.
Read →