7 attribution models
Most attribution arguments start because teams compare numbers without naming the credit rule behind them. First-click answers a discovery question. Last-click answers a closing question. Linear and position-based models create simple multi-touch views. Time-decay favors recency. Data-driven attribution tries to infer contribution from observed patterns.
| Model | How it works | Best for | Weakness | Platform support |
|---|---|---|---|---|
| First-click | Gives 100% of conversion credit to the first known touchpoint. | Discovery programs, SEO, creator campaigns, and top-of-funnel reporting. | Ignores all later persuasion, retargeting, and closing activity. | GA4 exploration, warehouse SQL, BI tools. |
| Last-click | Gives 100% of credit to the final eligible click before conversion. | Simple purchase paths and short-cycle direct response. | Overvalues branded search, affiliates, email, and retargeting. | Most ad platforms and analytics tools. |
| Last non-direct click | Ignores direct visits when a prior campaign source exists, then credits the latest non-direct source. | Analytics views where direct traffic is often returning demand. | Can hide the role of direct intent and still collapses the journey to one source. | Universal Analytics legacy, GA-style warehouse logic. |
| Linear (equal weight) | Splits credit equally across every eligible touchpoint. | Early multi-channel audits and simple stakeholder education. | Pretends every touchpoint was equally useful. | Warehouse SQL, BI tools, some attribution vendors. |
| Time-decay (half-life) | Gives more credit to recent touches, with older touches losing weight by a fixed half-life. | Longer consideration cycles where recent engagement matters most. | Requires an arbitrary decay curve and can undervalue discovery. | Analytics tools, attribution vendors, warehouse models. |
| Position-based (40/20/40) | Gives 40% to the first touch, 40% to the last touch, and splits the remaining 20% across middle touches. | Campaigns where opening and closing roles are both important. | The 40/20/40 split is a convention, not evidence from your own customers. | Attribution vendors, spreadsheets, warehouse logic. |
| Data-driven (ML) | Uses modeled contribution patterns to estimate which touchpoints changed conversion probability. | High-volume accounts with enough paths, conversions, and stable tracking. | Opaque, platform-specific, and difficult to compare across destinations. | Google Ads, GA4, Meta modeled reporting, enterprise MTA tools. |
Example: a $200 purchase through 4 touchpoints
Imagine one shopper finds you on Day 1 through organic search, returns on Day 3 after a Meta ad click, compares again on Day 7 after a Google Ads click, then types your URL directly on Day 10 and purchases for $200. The business outcome is one purchase, but each model tells a different operational story.
Initial discovery from a non-paid query.
Paid social brings the shopper back.
High-intent search captures comparison demand.
The shopper returns directly and buys.
| Model | Organic | Meta | Google Ads | Direct | Readout |
|---|---|---|---|---|---|
| First-click | $200 | $0 | $0 | $0 | Discovery gets all credit. |
| Last-click | $0 | $0 | $0 | $200 | Direct closes the path. |
| Last non-direct click | $0 | $0 | $200 | $0 | Direct is skipped, Google Ads wins. |
| Linear | $50 | $50 | $50 | $50 | Every touchpoint gets 25%. |
| Time-decay | $20 | $35 | $60 | $85 | Recent touches get more weight. |
| Position-based | $80 | $20 | $20 | $80 | First and last get 40% each. |
| Data-driven | $38 | $44 | $72 | $46 | Modeled from conversion paths. |
None of those rows changes the customer journey. They only change the accounting rule. That is why attribution should be treated as a measurement lens, then compared with experiments, contribution margin, incrementality tests, and actual customer behavior.
Multi-touch is mostly a lie
The phrase multi-touch attribution sounds like one neutral system is watching every ad, email, organic visit, referral, affiliate click, and direct session. That is rarely what marketers are looking at. In practice, most platform reports are platform-side attribution. Meta attributes conversions to Meta interactions. Google Ads attributes conversions to Google Ads interactions. TikTok, Pinterest, Reddit, and affiliate networks do the same inside their own boundaries.
That means platform totals can overlap. If a customer clicked a Meta ad and a Google ad before buying, both platforms may report the purchase because each platform sees an eligible interaction inside its own window. This is not always a bug. It is a symptom of separate measurement systems answering separate questions.
True MTA requires raw touchpoint data in a warehouse, an identity layer that can connect anonymous sessions to known customers, a consent-aware event model, and rules for channel taxonomy, sessionization, deduplication, and lookback windows. Without that, multi-touch reporting is usually a dashboard label wrapped around partial platform views.
What TrackLayer ships
TrackLayer does not force one attribution model on every team. That would be the wrong abstraction because the ad platforms, the warehouse, and the finance team often need different views of the same conversion. Instead, TrackLayer focuses on preserving the underlying event facts: touchpoints, click IDs, browser IDs, timestamps, session context, consent state, and conversion payloads.
For destinations, TrackLayer sends clean conversion events with the identifiers those platforms need to attribute in their native models. Meta can use its own reporting windows. Google Ads can use data-driven attribution. TikTok, Pinterest, and Reddit can apply their configured rules. TrackLayer's job is to make those events complete, deduplicated, fresh, and eligible.
For internal analysis, users can export raw events to their warehouse and build their own MTA. That is where teams define their channel taxonomy, identity rules, lookback windows, and model choice. The same raw stream can support last non-direct, time-decay, position-based, Markov, or a custom model without mutating the source event history.
Platform-specific attribution
Platform defaults matter because they shape the numbers buyers see every day. Treat them as operating settings, not universal measurement laws. Always verify the conversion action, event type, attribution setting, and reporting column before comparing one platform against another.
| Platform | Typical setting | Important caveat |
|---|---|---|
| Meta | Purchase commonly reports on 1-day click plus 1-day view by default, with configurable windows such as 7-day click. | Meta attributes inside Meta's own delivery and identity system, not across your whole media plan. |
| Google Ads | Data-driven attribution has been the default for most conversion actions since 2023. | Google Ads credits Google Ads interactions and uses Google modeling, consent signals, and click history. |
| TikTok | Common default reporting uses last-click with a 7-day click window. | View-through and click windows depend on account configuration and campaign objective. |
| Often uses 30-day click and 1-day view windows. | Pinterest conversion reporting is heavily affected by tag, CAPI, and enhanced match quality. | |
| Common default reporting uses a 28-day click window. | Reddit attribution depends on Reddit click identifiers, Pixel or CAPI coverage, and event matchability. |
The iOS 14.5 / ATT impact
Apple's App Tracking Transparency framework changed mobile attribution by requiring apps to ask permission before tracking users across apps and websites owned by other companies. When users deny permission, platforms lose access to device-level ad identifiers such as IDFA, and view-through attribution becomes harder to support with user-level confidence.
The practical effect is shorter, noisier attribution windows and more modeled reporting. Clicks remain easier to reason about than views because a click can carry explicit campaign context into a browser session. View-through attribution depends more heavily on platform identity, modeled conversion estimates, and aggregated reporting. That is why many teams saw paid social reporting shift after iOS 14.5 even when their stores kept selling.
Server-side tracking helps by preserving first-party events, consent state, click IDs, user agent, IP context, and hashed customer information when eligible. It does not bypass ATT or consent rules. A good measurement stack respects the permission boundary while reducing technical loss inside the data that is still allowed to be measured.
Markov chains for MTA
Markov attribution is a common warehouse-side method for teams that want something more analytical than fixed rules but more transparent than a platform black box. The basic idea is to turn customer journeys into paths, estimate transition probabilities between channels, then calculate what happens to conversion probability when a channel is removed from the path graph.
A simplified path might be Start → Organic search → Meta → Google Ads → Direct → Conversion. A Markov model learns how often users move between those states and how often they convert or exit. Then it removes Meta, Google Ads, or another channel and measures the drop in conversion probability. That drop is the channel's removal effect.
In practice, sophisticated teams build this in Python or R from raw event tables. They sessionize events, normalize campaign sources, exclude ineligible consent states, cap lookback windows, and compare the resulting weights against holdout tests. Markov is useful, but it is only as credible as the journey data and identity resolution that feed it.
Which model should you use?
Pick the model that matches the decision in front of you. Do not ask one attribution view to answer every budget, creative, merchandising, and finance question. Mature teams often keep multiple readouts side by side and use experiments to decide where the model is misleading them.
Single-channel acquisition: Last-click
When almost every paid interaction happens in one channel, a simple close-of-path model is usually enough for operating decisions.
Multi-channel conversion paths: Data-driven
When customers touch search, paid social, email, affiliates, and direct before purchase, modeled contribution beats a fixed rule.
SaaS with long cycles: Time-decay
A recent demo request or nurture email usually deserves more credit than an old awareness click, while still preserving the early journey.
BFCM-type spikes: Position-based
Launches, seasonal peaks, and promotion windows often have a clear discovery touch and a clear closing touch.
FAQ
Is last-click attribution wrong?
No. It is a narrow answer to a narrow question: what was the last eligible interaction before conversion? It is useful for short-cycle performance marketing, but it should not be treated as a full explanation of demand creation.
Why do Meta and Google both claim the same purchase?
Each platform sees its own eligible interaction and applies its own attribution window. If a shopper clicked Meta and Google before buying, both platforms may have a legitimate platform-side claim even though your business only received one order.
Can server-side tracking fix attribution disagreements?
It can improve event delivery, match quality, deduplication, and raw data completeness. It cannot force every platform to use the same identity graph, lookback window, or credit rule.
Do we need data-driven attribution?
Use it when you have enough conversion volume and multi-touch paths for a model to learn from. If volume is low, fixed rules plus incrementality tests are often easier to explain and harder to overfit.
What is the difference between MTA and MMM?
Multi-touch attribution works at the event and user journey level. Marketing mix modeling works at the aggregate time-series level. Mature teams often use both, then reconcile them with experiments.
Where does Markov attribution fit?
Markov is a warehouse-side MTA method. It estimates each channel's removal effect by comparing conversion probability with and without that channel in observed journey paths.