Blog

Notes from the data side of marketing

Practical write-ups on ad platforms, testing, SQL and putting AI to work on real business data — each one built around a mistake worth avoiding.

AI · Ad accounts

Your AI agent can now edit your ad account

(read is a report, write is a risk)

AI agents can now pause ads and move budgets. Why a wrong action costs far more than a wrong answer, and the permission setup that keeps an agent useful without letting it run up a bill.

7 min read AI analytics

What an AI analyst can't do yet

(an honest list, from someone who uses them daily)

Eight real limits of AI data analysts — why each one exists, which are shrinking, and four tests to run before you trust one with a board-deck number.

8 min read AI analytics

Seven questions to ask before you trust an AI's answer about your data

(before you paste it into the deck)

An AI-written query can run cleanly and still answer a different question. Seven quick checks — from the table it chose to timezones and refunds — that catch silent failures before a number reaches leadership.

7 min read AI analytics

Semantic layers vs. prompting: where business definitions should live

(definitions, not suggestions)

A metric definition in a prompt is a suggestion the model usually follows. In a semantic layer it is fixed SQL. Where each belongs, a simple test for deciding, and a one-week way to start.

7 min read AI analytics

How to keep an LLM from inventing a column name

(four layers, zero model upgrades)

LLMs invent column names because they can't see your schema. Four engineering layers — retrieval, constraint, dry-run validation and repair — fix it without a bigger model, plus the one failure they can't catch.

8 min read Ad testing

Your winning ad might be a coin flip

(check before you scale)

Two identical ads can show a 20% gap by pure chance. How much data a creative test really needs, why peeking manufactures winners, and a protocol for calling results honestly.

8 min read Google Ads

Where your PMax money actually goes

(read the channel report)

Performance Max now breaks spend and conversion value out by channel. How to read Search, YouTube, Display and Discover side by side, strip out view credit, and find the channels living off one blended ROAS.

7 min read SQL · AI

How to audit an AI-generated query in 60 seconds

(three places to look)

A one-minute routine for checking any AI-written SQL before you trust its number: the joins, the filters and the grain — in that order — plus a five-second cost and sanity check.

5 min read