MihaelTurkalj

Most people own one end of this. I have worked across all of it, which is usually why the numbers disagree in the first place.

1. 01

   The campaign that looked fine

   Performance MaxGoogle Business ProfileLocation assetsGA4

   An employer brand Performance Max campaign was bidding on completed job applications and reporting solid results. Nobody had asked me to look at it. I went in anyway and found budget going to Google Maps, where the campaign was optimising toward local actions rather than applications.

   The cause was the location asset, inherited automatically the moment a Business Profile is linked to the account. It can be excluded at campaign level. I excluded it and followed the behaviour through GA4.

2. 02

   A hundred dashboards instead of a hundred slides

   Looker StudioSupermetricsMulti-market data modelling

   A multi-brand FMCG client was drowning in its own reporting: spreadsheets consolidated by hand, then rebuilt as slides, across eight to ten markets and a long list of brands. There were more reports than anyone could act on.

   I built and maintained the reporting layer that replaced it — over a hundred Looker Studio dashboards a year, consumed directly by the client rather than mediated through a deck.

3. 03

   Inheriting someone else's account

   Account auditGTMConsent ModeConversion-based bidding

   A recurring part of agency work is taking over an account from the agency before you. The first job is finding out what is actually in there.

   I audited the structure, connected the data sources, implemented tracking and Consent Mode where they were missing, and only then moved the account onto conversion-based bidding — because conversion bidding on broken measurement is worse than no conversion bidding at all.

4. 04

   Brand safety against the brand's own rules

   PythonYouTube Data API v3Claude Agent SDKPlacement reports

   YouTube placement reports run to thousands of rows, and the usual answer is a keyword list that catches the obvious and misses the rest. I built a tool that takes the raw placement export, filters it down to the placements that actually matter, resolves each one to a YouTube channel and enriches it through the YouTube Data API.

   Then a language model reads every channel against the brand's own written safety guidelines — the actual PDFs the brand publishes, not a list of banned words — and returns a paste-ready exclusion list with a written reason per channel. Fully cached, so a re-run costs nothing. Built as a personal project; no client data on this page.

5. 05

   The agent that ran in the team, not in the pitch

   PythonGoogle Ads APIGemini

   I built the first AI agent inside the agency, on Google's adspace_agent framework: campaign structures stood up through the Google Ads API, optimisation recommendations generated with Gemini. It went into daily use in my team. Then I worked out the per-account cost, established it was uneconomic across the whole portfolio, and narrowed it to the cases where it genuinely beat a person on both speed and cost — which is where it kept running.

   It never made it into the pitch it was originally built for: it needed a terminal to start, and the people presenting could not run it. A packaging failure, not a capability one, and the reason I now care as much about how a tool is handed over as about whether it works.