How to attribute conversions when last-click is broken (2026 playbook)

Last-click attribution stopped being honest around 2021. iOS 14.5 broke pixel fidelity. The Privacy Sandbox shipped. Browser cookies stopped persisting reliably. And starting in 2024, a meaningful share of consumer research moved into LLM conversations that carry no UTM at all. The result: a typical 2026 marketing dashboard credits 30-50% of conversions to a channel that did not actually cause them.

You can't go back to deterministic attribution. You also can't sit and wait for last-click to fix itself. The teams that ship in 2026 use a layered approach — four different attribution lenses, each catching what the others miss.

Layer 1 — platform-reported attribution (where you start)

Meta's pixel + CAPI, Google's Enhanced Conversions, TikTok Events API, Snap CAPI, X Conversion API. Each platform claims credit using its own model — usually 7-day click + 1-day view. Sum them up and you'll exceed 100% of your actual conversions, often by 40-80%, because each platform credits the touches it can see.

Use platform-reported attribution for in-platform optimization (bid signals, audience signals). Don't use it for cross-channel budget allocation; it's structurally inflated.

Layer 2 — server-side first-party tracking (your truth)

Server-side conversion APIs send the conversion to each platform with first-party data — your CRM record, your Stripe payment ID, your hashed email. This is more iOS-resilient than the pixel and is the bare minimum in 2026.

Even better: a composable customer data platform (Segment, Rudderstack, custom) as a single source of truth — every conversion logged once, distributed to every platform. Cometly does this end-to-end; Northbeam has its own pixel; Floowzy reads from Stripe directly without requiring CDP plumbing.

Layer 3 — MTA model overlay (the cross-channel allocator)

Run a multi-touch attribution model (data-driven if you have volume, position-based if you don't) across your server-side conversion log. The full MTA pillar guide covers when each model is honest and when it isn't — short version: linear and time-decay are reasonable defaults for $100k-$1M monthly spend; data-driven needs $1M+ monthly to be statistically credible.

MTA isn't a source of truth either. It's a perspective. Run it alongside platform-reported and call out where they disagree — that's where the interesting questions live.

Layer 4 — incrementality + MMM (the sanity check)

Incrementality tests (Meta's Conversion Lift, Google's Conversion Lift, Snap brand lift, geo holdouts) and media-mix modeling are how you catch the systematic biases in layers 1-3. They're expensive and infrequent — most teams run incrementality once a quarter and refresh MMM yearly — but they're the only honest answer to 'is this channel actually driving incremental revenue.'

Northbeam and Measured are the category leaders here; Meta's geo-based testing has caught up significantly in 2025-2026.

The new layer — LLM-mediated dark conversions

None of the four layers above catches the consumer who asked ChatGPT 'what's the best ROAS tracking tool for a $50k/mo agency', got three named recommendations, then searched the brand directly the next day. That conversion lands on direct or organic.

Three compensations: (1) Post-purchase surveys with an explicit AI-assistant option in the 'where did you first hear about us' field. (2) Watch branded search lift as a leading indicator of LLM visibility. (3) Stretch attribution windows to 30-60 days for considered purchases; the LLM conversation often precedes the direct search by 7-21 days.