Multi-Touch Attribution Models

Once dominant, now contested. Rule-based vs algorithmic models, the mechanics, signal degradation post-ATT, and where MTA still works.

Why MTA was dominant

From roughly 2010–2020, multi-touch attribution was the dominant cross-channel measurement methodology. It promised user-level granularity: see each user's journey across touchpoints, assign fractional credit, derive ROI per channel. Platforms (Google Analytics, Adobe Analytics) built MTA UIs. Vendors (Visual IQ, Convertro, Marketing Evolution, Neustar) sold sophisticated MTA platforms. CMOs used MTA dashboards to allocate budget.

Then iOS 14.5 happened. Then Safari ITP. Then Chrome cookie depreciation. The identity-based tracking that MTA relied on degraded sharply. MTA didn't become useless — but its dominance ended, and triangulation with MMM and incrementality became necessary.

Rule-based vs algorithmic MTA

Rule-based models

• Last click. 100% credit to last non-direct click.
• First click. 100% credit to first click.
• Linear. Equal credit across all touchpoints.
• Time decay. Exponentially more credit to recent touchpoints.
• Position-based (U-shaped). 40% to first, 40% to last, 20% distributed across middle touchpoints.

Rule-based models are simple, transparent, easy to explain. They're also arbitrary — no rule reflects causal reality. Their value is consistency and shared language across teams, not accuracy.

Algorithmic / data-driven models

• Markov chain. Models conversion paths as state transitions. Calculates the "removal effect" of each touchpoint — what happens to conversion probability if you remove this channel? Channels with high removal effect get more credit.
• Shapley value. Game-theoretic approach from cooperative game theory. Calculates each touchpoint's marginal contribution averaged across all possible orderings.
• Logistic regression. Models conversion as a function of touchpoint exposure; coefficients become weights.
• Survival analysis / hazard models. Models time-to-conversion as a function of touchpoint timing.
• Custom deep learning. Vendor-built neural-network models trained on user journey data.

Mechanics: credit assignment

How Markov chain MTA works

1. Build the conversion path graph: states = channels; transitions = sequential touches.
2. Compute transition probabilities from observed paths.
3. For each channel, simulate paths with that channel removed. Compute resulting conversion probability.
4. Credit = baseline conversion probability − with-channel-removed conversion probability.
5. Normalize so credits sum to total conversions.

How Shapley value MTA works

1. For each subset of channels, observe the conversion rate.
2. For each channel, compute its marginal contribution to every subset containing it.
3. Average those marginal contributions across all possible orderings to get the channel's Shapley value.
4. Normalize so values sum to total conversions.

Computationally expensive for many channels (factorial growth), but tractable for typical 5–15-channel mixes with approximation methods.

Data requirements

• User-level event data. Cookie ID, hashed email, device ID, or other persistent identifier.
• Touchpoint timestamps. Order of touches; not just presence.
• Channel/source/medium classification. Consistent attribution of touches to channels.
• Conversion events. Tied to user IDs in the same identity space.
• Sufficient volume. Hundreds of conversions per channel per modeling period.
• Path diversity. Variety of channel combinations; otherwise model can't learn cross-channel effects.
• Identity persistence. Across sessions, devices, channels. This is where ATT broke things.

Vendor landscape

• Tier 1 enterprise: Neustar, Nielsen Marketing Cloud, Marketing Evolution. Full-stack attribution with managed services.
• Mid-market: Northbeam, Rockerbox, AnyTrack, Triple Whale (DTC-focused). Modern MTA + MMM hybrid platforms.
• Marketing analytics platforms with MTA modules: Adverity, Funnel.io, Improvado.
• Native platform MTA: Google Analytics 4 data-driven attribution, Adobe Analytics, Heap.
• DIY: Build in BigQuery + dbt + Python (PyMC for Bayesian; ChannelAttribution package for R).

Signal degradation and implications

MTA quality depends on signal quality. Post-2021, the major degradation vectors:

• iOS ATT. ~70% of iOS users don't opt in to tracking. Their conversion paths are partial or missing.
• Safari ITP. 7-day cookie cap means cross-session attribution >7 days back is broken.
• Ad blockers. 20–40% of users have ad blockers in many demographics.
• VPN and proxy use. Geographic and IP-based deduplication degraded.
• Consent declines in EEA. 30–50% of EEA users decline cookies; MTA can't observe their paths.

The implication: MTA underweights upper-funnel and view-through; overweights direct and self-attribution; misses cross-device journeys; biases toward identity-rich channels (logged-in platforms) and against identity-poor (CTV, display).

Where MTA still works

• Logged-in ecosystems with persistent identity (apps with strong login, B2B SaaS with account-based tracking).
• Short-cycle conversion paths where signal degradation has minimal time to compound.
• Direction (which channels are punching above their weight), not magnitude.
• Optimization signals for platforms that need cross-channel context (e.g., Google Ads with offline conversions from MTA).

Where MTA doesn't work

• iOS-heavy audiences with ATT signal loss.
• Long sales cycles (B2B enterprise with 6–18 month cycles).
• Cross-device journeys without unified user identity.
• Strategic budget allocation across major channels — use MMM and incrementality instead.
• Brand-building channels (CTV, OOH) whose contribution doesn't show up as identifiable clicks.

Advanced playbook

• Hybrid attribution stack. Use MTA for daily optimization, MMM for strategic allocation, incrementality for calibration of both.
• First-party identity investment. Email-based identity, account-based identity, hashed cross-device matching. The more first-party signal, the more durable MTA.
• Server-side data collection. Reduces ad-blocker impact; enriches with server-known user attributes; sends to all destinations from one source.
• Conversion API everywhere. Meta CAPI, TikTok Events API, LinkedIn CAPI, Google Enhanced Conversions. Server-side conversion data improves identity matching by 15–40%.
• Modeled conversions awareness. Platforms model conversions for cases they can't track. Understand what's observed vs modeled in each platform's attribution.
• Acknowledge MTA limits in stakeholder communication. Don't represent MTA as truth post-ATT. Show ranges; show triangulation; show incrementality calibration.
• DIY MTA in BigQuery for transparency. Build Markov or Shapley models in SQL/Python on your raw data. Avoids vendor lock-in; gives full methodology transparency.
• Channel role classification. Brand channels and conversion channels have different attribution properties. Build the classification into MTA outputs.
• Test MTA against incrementality. Quarterly: compare MTA estimates against incrementality test results. Calibrate.
• Communicate confidence levels. Don't report MTA point estimates without confidence intervals or qualitative confidence indicators.

Common mistakes

• Trusting MTA as truth post-ATT without calibration.
• Using MTA for strategic budget allocation when MMM is more appropriate.
• Choosing rule-based models (last-click) for sophisticated programs because they're simple.
• Choosing algorithmic models you can't explain to stakeholders.
• Treating data-driven attribution as a black box; not auditing methodology.
• Implementing MTA without CAPI; observed-conversion data is so degraded that MTA misleads.
• No identity persistence; cross-device journeys missed entirely.
• Underweighting brand and view-through channels because MTA can't see them.
• No periodic comparison of MTA estimates against incrementality.
• Building MTA on platform-reported data only (no warehouse foundation).

Operating checklist

• Server-side data collection for all major platforms (CAPI/Events API)
• First-party identity strategy: email-based, account-based, cross-device hashing
• MTA model chosen with documented rationale (rule-based vs algorithmic)
• BigQuery-based MTA or vendor solution with methodology transparency
• Quarterly comparison of MTA vs incrementality test results
• Channel role classification (brand vs conversion) in MTA outputs
• Confidence intervals or qualitative confidence in MTA reporting
• Stakeholder communication acknowledges MTA limits post-ATT
• MTA used for tactical optimization; MMM used for strategic allocation
• Periodic model audit for drift or methodology decay

---

Part of the Attribution & Measurement series.