Skip to content
Primores AI wiki

Multi-Touch Attribution (MTA): What It Is and Why It's Now a Tactical Tool, Not Ground Truth

Multi-Touch Attribution (MTA)

TL;DR: MTA assigns fractional credit for a conversion across the multiple touchpoints in a customer journey — versus last-click, which hands 100% to the final touch. It uses user-level, path-level data, which is exactly what glossary/app-tracking-transparency and third-party-cookie loss took away. So in 2026 MTA is degraded and demoted to a tactical, in-platform optimization signal — best for relative channel/creative comparison inside a walled garden — not the cross-channel “source of truth” it was once sold as. The mature stack triangulates MTA (tactical) with glossary/marketing-mix-modeling (strategic, privacy-durable) and glossary/incrementality-testing (causal ground truth).

What it means

A customer rarely converts on first contact — they see an ad, click a search result, read an email, return days later. Last-click gives all the credit to the final touchpoint; first-click gives it all to the first. Both make every other touchpoint invisible. MTA spreads credit across the path, so each contributing channel gets a share.

The honest 2026 framing: MTA isn’t “dead” (that headline comes from vendors selling the replacement), but it is degraded — it remains the most-adopted method while no longer being trusted as cross-channel truth.

The model taxonomy

A. Rules-based — the analyst picks the credit-split rule in advance:

ModelCredit rule (typical default)
Last-click / first-click100% to the final / first touch
LinearEqual credit to every touch
Time-decayMore credit nearer the conversion (often a ~7-day half-life)
Position-based (U-shaped)~40% first, ~40% last, ~20% across the middle
W-shaped~30% each to first / lead-creation / opportunity touches, ~10% across the rest

(These splits are conventions, not laws — tools let you reconfigure them, and exact weights vary by source.)

B. Data-driven / algorithmic attribution (DDA) — the model learns the split from data instead of the analyst imposing it. It compares which touchpoint combinations preceded conversion vs non-conversion and credits each by how much it measurably moved conversion probability. Two dominant methods:

  • Shapley value (cooperative game theory): credit each touchpoint by its average marginal contribution across all touchpoint combinations — its “fair share.” Compute cost grows fast with the number of distinct touchpoint types.
  • Markov chains: model the journey as a graph of channel-to-channel transitions; credit each channel by its “removal effect” — how much total conversion probability drops if you remove it. (Shapley and Markov are different methods producing different splits — don’t conflate them.)

Concrete anchor — GA4: Google’s data-driven model is Shapley-based with a time-decay element (not Markov), considers up to 50 touchpoints over a 90-day lookback, and needs 400+ conversions per type to run. In late 2023 Google removed first-click, linear, time-decay, and position-based from GA4, leaving only last-click and DDA. (Google confirms the Shapley family but does not publish the full weighting internals — describe it as “Shapley-based per Google,” not an exact spec.)

Why MTA alone is no longer trusted

MTA needs user-level, cross-site/cross-device tracking to stitch a journey together. Third-party cookies were “the scaffolding most MTA systems were built on” — and glossary/app-tracking-transparency (2021) plus cookie deprecation removed it. The result: MTA now sees an incomplete journey and systematically mis-credits.

Reported signal loss (note: vendor-sourced ranges, not audited — the direction is uncontested, the precision isn’t): some advertisers reported losing visibility into 40–60% of iOS conversions; MTA coverage shrank to roughly 30–60% of its 2020 signal depending on channel mix. Every quantified figure here comes from a vendor with an interest in the “MTA is broken, buy MMM/incrementality” narrative — but the structural cause (Apple’s ATT design, Chrome’s cookie phase-out) is undisputed primary fact.

This is precisely why glossary/marketing-mix-modeling came back: MMM works on aggregate time-series data and needs no user IDs, cookies, or ATT opt-ins — structurally privacy-safe.

The 2026 positioning: triangulate, don’t pick one

The mature stack uses all three for what each does best — a consensus that holds even across competing measurement vendors:

  • MTA = tactical / in-platform. Relative channel & creative comparison, near-real-time optimization. Strongest inside a walled garden (a logged-in platform that still has user-level data). Weak as cross-channel truth.
  • MMM = strategic / privacy-durable. Long-term, aggregate budget allocation; no user tracking.
  • Incrementality = causal ground truth. Proves whether a channel caused conversions; when it disagrees with attribution, it wins.

This is the marketing/marketing-analytics-in-2026 “dual-model operating norm” (MMM strategic + MTA tactical + AI reconciliation), with incrementality as the causal third leg.

Key Takeaways

  • MTA fractionally splits conversion credit across the journey — better than last-click, but only if you can see the journey.
  • Two families: rules-based (last/first/linear/time-decay/U/W) and data-driven (Shapley, Markov).
  • ATT + cookie loss degraded it — reported 40–60% iOS signal loss (vendor ranges; direction is solid).
  • It’s tactical now, not ground truth — strongest inside walled gardens; “MTA is dead” is vendor framing.
  • Triangulate: MTA (tactical) + MMM (strategic) + incrementality (causal). No single source of truth.

Sources

Do-not-cite: the 40–60% / 30–60% signal-loss figures as audited fact (vendor self-reported ranges — attribute and hedge); exact rules-based splits as universal law (they’re configurable defaults); the absolutist “MTA is dead” framing (it’s degraded and demoted, not dead — still the most-adopted method).

By Andrej Ruckij