Every quarter a brand asks us which attribution platform to buy, expecting a recommendation. The honest answer is that the question contains a false premise: no platform can observe what it needs to observe, because the underlying signal was removed by privacy changes and is not coming back.
What works instead is triangulation: three imperfect instruments whose errors point in different directions.
Tier one: clean capture, for the day-to-day
Server-side event collection through Conversions API and Enhanced Conversions, consent-aware identity resolution, and raw events stored in your own warehouse rather than in a vendor's interpretation of them. This tier answers operational questions: did something break overnight, is this campaign pacing, is match quality degrading.
What it cannot tell you is what was incremental. Treat it as telemetry, not truth.
Tier two: a model, for quarterly allocation
Media mix modelling reads aggregate spend against aggregate outcomes and estimates each channel's contribution, including the offline and upper-funnel activity no pixel ever sees. It is the right instrument for the question 'how should next quarter's budget be split'.
It earns its cost above roughly $500,000 per month in spend with two years of clean history. Below that, the confidence intervals are wide enough to drive a strategy through.
Tier three: experiments, for the arguments
Geo-holdouts, PSA lift tests and branded-search blackouts produce the closest thing to causal truth available in marketing. They are slower and more expensive per question, so reserve them for questions that matter: is this channel incremental at all, what is our true brand-search contribution, does upper-funnel video pay for itself.
And reconcile to finance
Whatever you build, tie it to bank-deposited revenue and real gross margin, after returns, shipping and payment fees. Marketing dashboards that cannot be reconciled to the P&L get quietly disbelieved by everyone who matters, and once that happens the reporting stops influencing decisions at all.