Attribution After Cookies: What Actually Works in 2026

The cookie era ended not with one dramatic deprecation but with a thousand cuts — browser ITP, app tracking transparency, consent banners nobody accepts, and finally the slow browser phase-outs. What remains is a measurement landscape where every ad platform reports a conversion number, the numbers sum to more revenue than you earned, and finance stopped believing all of them. Correctly.

Accept the epistemology, then build

User-level cross-site tracking is not coming back; measurement now rests on three legs, each honest about what it knows:

1. Server-side conversion APIs — feed the machines. Meta CAPI, Google Enhanced Conversions, TikTok Events API: first-party purchase events sent server-to-server, matched probabilistically. This does not fix attribution truth — it fixes ad delivery, because the platforms' optimisation is only as good as the signal you return. Brands that implemented server-side events properly saw paid efficiency recover meaningfully; those that did not are training algorithms on noise. This is table-stakes engineering, part of the commerce data layer, and it is the one place where better plumbing directly buys cheaper customers.

2. Media mix modelling — the budget allocator. MMM regressed spend against outcomes long before cookies; it is now the grown-up answer again, and modern open-source tooling plus weekly-grain data made it accessible below enterprise scale. MMM tells you channel-level marginal return without touching user data — GDPR-serene — but it needs spend variation to learn from (steady-state budgets teach it nothing) and honest priors. Run it quarterly; let it argue with the platform dashboards; believe it more.

3. Incrementality tests — the truth serum. Geo holdouts, audience splits, PSA tests: deliberately not spending somewhere measurable is the only method that measures rather than models causation. Every brand can afford one clean test per quarter. Typical findings, humbling as ever: branded search largely harvesting, retargeting less incremental than its ROAS claims by a wide margin, upper-funnel undervalued by click-based numbers. Each finding reallocates real money — which is the point.

What to stop doing

Stop reconciling platform dashboards against each other; they measure different fictions. Stop last-click as a decision rule; it is a participation trophy for the bottom of the funnel. Stop buying "AI attribution" tools that promise user-level truth from probabilistic dust — they re-launder the fiction with better UX. And stop treating consent rates as someone else's problem: consent UX quality now directly determines the volume of signal your entire stack learns from.

The operating cadence

Weekly: platform numbers for directional in-flight decisions (they are relatively consistent with themselves). Quarterly: MMM refresh reallocates budgets; one incrementality test settles the loudest argument. Annually: the model's marginal-return curves set the plan. Wire the whole thing to contribution margin, not revenue, and marketing measurement finally speaks the same language as the P&L — which was the goal the cookies always pretended to serve.