Attribution Is Broken. Here Is What We Do Instead.
Last-click lies, platform-reported ROAS double-counts, and multi-touch models mostly launder assumptions. A practical measurement stack for companies without a data science team.
Key takeaways
- If you add up every platform-reported conversion you will usually exceed your actual order count. That is the whole problem in one sentence.
- Geo holdout tests are the cheapest credible incrementality measure available to a mid-market company.
- Branded search is the most commonly over-credited line item in almost every account we audit.
- You do not need a perfect model. You need a consistent one your finance team accepts.
Add up the conversions reported by Google, Meta, your affiliate platform and your email tool. If that total is larger than the number of orders you actually shipped, you already understand the problem.
Why every model is wrong in its own way
- Last-click credits the final touch, which is usually branded search or direct. It systematically over-credits the bottom of the funnel.
- Platform-reported conversions use view-through windows and modelled conversions, and every platform claims the same order.
- Multi-touch models distribute credit using weights somebody chose. The output looks scientific and is largely an assumption made visible.
- Media mix models need years of data and meaningful spend variation. Most mid-market companies have neither.
The stack we actually run
Four layers, in this order. None of them is perfect and together they are good enough to make budget decisions with confidence.
- Clean server-side event collection with deduplication, so the raw numbers are at least internally consistent.
- A blended CAC number: total sales and marketing spend divided by new customers. Crude, unfakeable, and the number to show the board.
- Self-reported attribution on the order or lead form. A single "how did you hear about us" field catches word of mouth and podcasts that no pixel will ever see.
- Geo holdout tests, run quarterly on the largest line items, to establish what is genuinely incremental.
How to run a geo holdout without a data team
- Pick eight to twelve matched market pairs by historical revenue and similar seasonality.
- Pause the channel under test in one side of each pair for four weeks. Change nothing else.
- Compare revenue in test versus control against the pre-period baseline.
- The difference is your incremental lift. Compare it to what the platform claimed and adjust your budget accordingly.
What good looks like
You do not need a perfect attribution model. You need one number your finance team trusts, one test that proves incrementality, and the discipline to stop optimising against a metric you know is inflated. In practice that is blended CAC, a quarterly geo holdout, and the honesty to say "we are not sure" when you are not sure.
Precision you cannot defend is worse than an honest range. Pick the metric you would still believe if it made you look bad.
Want this applied to your account?
We will run the same analysis on your channels and send back what we find — free, and yours to keep whether or not you hire us.
Request an auditMore from the blog
How Much Should Digital Marketing Actually Cost in 2026?
Real benchmark ranges for SEO, paid media, content and social retainers, what drives the number up or down, and the pricing models that quietly transfer risk onto you.
Read articleSEO vs PPC: A Framework for Splitting Your Budget
A decision framework based on sales cycle, margin, urgency and competitive position — instead of the usual answer that you should obviously do both.
Read articleCore Web Vitals and Conversion: What the Data Actually Shows
Site speed is a ranking factor, but the conversion impact is the bigger prize. Here is what moved on real sites, and the fixes that produced most of it.
Read article