The company believed their strongest conversion channel was paid social. The numbers said so. The dashboard confirmed it. Budget followed — growing quarter by quarter, because the ROAS figures from Meta looked so convincing. The CMO presented them proudly to the board.
But the numbers were wrong. Not slightly wrong. Fundamentally wrong. A misconfigured GA4 setup — combined with a pixel firing double under specific conditions — was double-counting conversions from one specific campaign type and creating an illusion of performance that didn't exist in reality.
I discovered it by chance. We were in the middle of another project — a general review of their analytics setup — when I compared conversion data from GA4 against actual orders from their Shopify backend. The discrepancy wasn't marginal. For the relevant campaign type, GA4 was recording 2.7 conversions per actual order. A channel that appeared to have a ROAS of 4.2 actually had a ROAS of under 1.6.
The next question was: how far back does this go? We reviewed historical data. The error was present for 14 months. For 14 months, the company had been optimising against a signal that didn't mean what they thought it meant. Budget had been allocated, campaigns scaled, strategic decisions made — all based on data that was systematically wrong.
The hardest conversation in the entire project wasn't the technical fix. It was the conversation with the CMO about what we had found. He had presented these figures to the board four times. He had defended budget increases with them. His credibility as a leader was tied to these numbers.
We chose an approach that prioritised constructiveness over confrontation. We didn't present findings as "you've been making mistakes." We presented them as "here's an opportunity to make significantly better decisions going forward." That's not spin — it's the true perspective. An error you discover is an opportunity to learn. An error you don't discover is a permanent handicap.
The technical solution took two intensive weeks. We went systematically through the entire tracking architecture: GA4, Google Tag Manager, Meta Pixel, server-side setup via Google Cloud, and finally a dedicated data pipeline pulling order data directly from Shopify's backend and matching it against campaign data in BigQuery.
Server-side tracking was the key to the long-term solution. Browser-based tracking is vulnerable — ad-blockers, ITP in Safari, third-party cookie restrictions — all of this erodes data quality over time. A server-side implementation via Google Cloud Tag Manager sends data directly from the company's server, not from the user's browser. It gives significantly better data quality and is far more robust against future browser restrictions.
We also implemented a dedicated reconciliation system: an automated daily check that compares conversion data from analytics with actual orders from the backend. If the discrepancy exceeds 5%, an automatic alert is generated. The company now discovers tracking errors within 24 hours instead of 14 months.
While working on the technical side, we simultaneously conducted a deep analysis of what the historical figures actually meant when corrected. Which channels were actually performing? Which were overvalued? The answer was both surprising and actionable: organic search and email were significantly undervalued. Paid social was still positive — but at a far lower scale than previously assumed.
Budget was reallocated based on the corrected figures. The paid social budget was reduced by 35%. Email and SEO received significant investment. The result three quarters later: total marketing ROI up 28%, with a lower overall budget. Those are the opportunities you discover when data is trustworthy.
From a leadership perspective, the project created an important cultural shift: data quality moved from being a technical matter to being a leadership question. The CMO started asking different types of questions in meetings — not "what do the numbers say?", but "are we confident about what the numbers actually measure?" That's a subtle but critical difference.
The most important lesson from this project isn't technical. It's human. All organisations have a tendency to trust data that confirms what they want to believe. Confirmation bias is real, and it's dangerous when combined with unreliable tracking. The best insurance isn't better algorithms — it's a culture that constantly questions what the numbers actually measure.