Your acquisition report says paid search won the install. The journey says something else: a video introduced the app, an article answered a question, a referral brought the user back, and search captured the final tap. If the team funds only the reported winner, it may cut the channel that created the demand.
Multitouch attribution gives more than one interaction credit for a conversion. It can produce a fairer view of a multi-channel journey than last-click reporting, but it does not automatically reveal what caused the conversion. The useful outcome is not a prettier journey map. It is a budget decision that remains sensible when the model, window, and available signals change.
What is multitouch attribution?
Multitouch attribution is a marketing measurement method that distributes conversion credit across two or more touchpoints in a customer journey. Instead of assigning 100% of an install, purchase, or other outcome to one interaction, a multi-touch attribution model gives fractional credit according to a fixed rule or a data-driven calculation.
The touchpoints might include an ad impression, a social click, an organic visit, an email, a referral link, an app install, or a re-engagement campaign. Google describes an attribution model as a rule, set of rules, or data-driven algorithm that assigns credit along the path to a key event. Nielsen similarly distinguishes multi-touch attribution from first- and last-touch methods that give all credit to one endpoint.
That definition contains an important limit: attribution assigns credit. It does not observe an alternate world in which the user never saw the ad. A channel can appear in many converting paths because it caused demand, because high-intent users naturally encounter it, or both.
Multi-touch, multitouch, and MTA mean the same thing
You will see *multitouch attribution*, *multi-touch marketing attribution*, and *MTA model* used for the same family of methods. “Multi-channel attribution” is related but not identical: multi-channel describes credit across channels, while multi-touch can distinguish several interactions within the same channel or campaign.
How multitouch attribution differs from last-click attribution
Last-click attribution gives all conversion credit to the final eligible interaction. First-touch attribution gives it all to the first recorded interaction. Multitouch attribution shares the credit among several eligible interactions.
Consider an illustrative $100 subscription journey:
- The user sees a short-form video ad.
- They click an organic search result three days later.
- A friend sends a referral deep link.
- The user clicks a branded search ad and subscribes.
The same $100 of revenue can appear very different in a channel report:
| Model | Video | Organic search | Referral | Paid search | What the report encourages |
|---|---|---|---|---|---|
| First touch | $100 | $0 | $0 | $0 | Fund discovery |
| Last click | $0 | $0 | $0 | $100 | Fund closers |
| Linear | $25 | $25 | $25 | $25 | Maintain broad coverage |
| Position based, 40/20/40 | $40 | $10 | $10 | $40 | Protect opener and closer |
| Data driven | Learned from eligible data | Learned | Learned | Learned | Follow estimated contribution |
No customer behavior or revenue changed between the rows. Only the credit rule changed. This is why model comparisons should show both attributed conversion value and the unmodified business totals—spend, installs, purchases, revenue, and retention.
Last click remains useful when the journey is genuinely short, one channel dominates, or the team needs a simple and auditable operational rule. It also provides a stable baseline. Its failure mode is systematic: it cannot credit an earlier touch even when that touch introduced the product or changed the user's intent.
MTA can surface those assists, but it introduces more assumptions. The team must decide which interactions qualify, how identities are joined, how far back the model looks, whether impressions count, and how credit is weighted. Missing early touches can make a sophisticated MTA model behave like last click with extra decimals.
The main multi-touch attribution models
There is no universally correct multi-touch attribution model. Each one expresses a view about which part of the journey deserves credit. Pick the view that matches the decision, then test whether the result is stable enough to use.
Linear attribution
Linear attribution divides credit equally across all eligible touches. Four touches receive 25% each.
It is easy to calculate, explain, and replay from event data. It is a reasonable first diagnostic when the alternative is last click. Its weakness is also obvious: an incidental impression receives the same credit as a high-intent pricing-page visit unless the eligibility rules distinguish them.
Use linear attribution to reveal how much credit is hidden outside the final interaction, not to claim that every touch was equally influential.
Time-decay attribution
Time decay gives more credit to touches closer to the conversion. The weighting normally declines according to a chosen half-life or decay function.
This fits short promotions or journeys in which recent intent should matter more. It can still overvalue retargeting and branded search because those channels naturally appear near conversion. A short half-life can quietly recreate last-click bias.
Position-based attribution
A position-based or U-shaped model reserves substantial credit for the first and last touches, then splits the remainder across the middle. A familiar rule is 40% to the first touch, 40% to the last, and 20% shared by all assists.
The model protects discovery and conversion without pretending that every middle interaction is equal. The percentages are a policy choice, not a measured law. If the first recorded touch is merely the first touch your systems could see, the model can reward a data boundary rather than true discovery.
Milestone or custom attribution
Some teams assign weights to milestones such as first touch, qualified visit, install, trial, and purchase. This can be clearer than a generic W-shaped or full-path model for app funnels because it maps credit to events the product team already governs.
Custom rules are valuable when they remain documented and replayable. They become dangerous when every stakeholder negotiates a weight that protects their channel.
Data-driven attribution
Data-driven attribution learns weights from observed paths rather than applying one fixed split. Google Analytics says its current data-driven model evaluates converting and non-converting paths and estimates how the presence of an ad interaction changes key-event probability. Google Analytics currently offers data-driven attribution plus two last-click options; its first-click, linear, time-decay, and position-based report models were deprecated in November 2023.
A data-driven model can detect patterns that a fixed rule misses, but “algorithmic” does not mean assumption-free. Results depend on identity coverage, event quality, channel access, sample size, conversion definition, and the model's treatment of missing paths. If the weighting cannot be explained at all, it may be hard to turn a surprising score into a defensible budget change.
When multitouch attribution actually changes budget decisions
MTA earns its cost when it changes a decision that a simpler model would get wrong. Four situations are especially useful.
It exposes channels that create demand but rarely close
Video, creator campaigns, display, content, and referrals often appear earlier than branded search or retargeting. Last click can make the closer look efficient while giving the opener no credit. Multi-touch reporting keeps discovery visible long enough to test whether it deserves investment.
It separates a channel's roles
One channel can introduce users in one campaign and close them in another. Roll both into a single platform total and the distinction disappears. Journey-level data lets the team compare campaign roles instead of labeling an entire network “upper funnel” or “lower funnel.”
It reveals assisted paths with stronger downstream value
The install is not always the right optimization target. Compare retention, purchase rate, subscription value, or another downstream event for single-touch and assisted journeys. If assisted users repeatedly create more value, the team has a reason to investigate the sequence—without claiming the sequence alone caused the difference.
It makes hidden assumptions visible
Running first touch, last click, linear, and a chosen multi-touch model over the same cohort shows how much the channel ranking depends on the rule. Large movement is not proof that MTA is correct. It is evidence that the budget decision is model-sensitive and needs stronger validation.
A practical output is a sensitivity table:
| Channel | Last-click ROAS | Linear ROAS | Position-based ROAS | Rank range | Decision |
|---|---|---|---|---|---|
| Channel A | 2.8 | 2.1 | 2.4 | 1–2 | Hold while testing |
| Channel B | 1.2 | 2.3 | 2.0 | 2–4 | Test incremental lift |
| Channel C | 1.9 | 1.8 | 1.9 | 2–3 | Stable; optimize within channel |
These are illustrative values, not benchmarks. The useful column is the rank range: if a channel is first under one reasonable rule and fourth under another, a large reallocation is premature.
The relationship is operational, not merely theoretical. A published large-scale advertising study compared last-touch and multi-touch inputs in a deployed budget-allocation system, illustrating why the credit method changes the performance signal used to distribute spend.
Which attribution model should you use?
Choose based on journey complexity, data quality, and the decision horizon—not on which model awards the most conversions.
| Situation | Starting model | Why | Upgrade trigger |
|---|---|---|---|
| One or two channels, short install path | Last click plus first-touch comparison | Simple, traceable baseline | Assists become common or channel roles diverge |
| Multiple digital channels, modest data | Linear or position based | Exposes assists with auditable logic | Results stay stable and conversion volume supports modeling |
| Short promotion or re-engagement cycle | Time decay | Reflects recency while retaining assists | Decay choice materially changes rankings |
| High-volume, well-instrumented journey | Data driven plus a rules-based benchmark | Learns interaction patterns | Model passes holdout and sensitivity checks |
| Fragmented mobile identity or large offline journey | Last click for operations; MMM and experiments for planning | User-level paths are too incomplete for strong MTA claims | Deterministic coverage improves |
The best starting point for many app teams is not an algorithmic model. It is a clean last-click baseline, a first-touch view, and one auditable multi-touch rule applied to the same mature cohort. If those views recommend the same action, the decision is robust. If they conflict, investigate coverage and run an experiment before moving a large share of spend.
Mobile privacy limits what multitouch attribution can observe
User-level MTA needs a way to connect touches to a person, device, account, or privacy-preserving report. Mobile platforms deliberately restrict that connection.
On Apple platforms, apps need permission through AppTrackingTransparency to link user or device data across companies for advertising measurement. Apple explicitly says developers may not use fingerprinting to identify a device, and third-party deep-linking or deferred-deep-linking services that create a shared cross-company identity for ad measurement require permission. The full rules are in Apple's user privacy and data use guidance.
Apple's AdAttributionKit provides signed campaign postbacks without tracking individual users across companies. That protects privacy, but it also means a marketer cannot expect a complete person-level path through every impression, click, install, and post-install event.
Google's proposed Attribution Reporting API for Android is likewise designed to support app and web conversion measurement without relying on cross-party advertising identifiers. Aggregation, reporting rules, walled gardens, opt-outs, cross-device behavior, and dark social all leave gaps.
Treat measurement coverage as a metric. For each conversion cohort, report the share with:
- a deterministic campaign click;
- a connected first and last touch;
- an authenticated or consented cross-session identity;
- only a privacy-preserving aggregate or modeled signal;
- no supported attribution signal.
Do not fill the last category with fingerprinted certainty. An explicit “unattributed” bucket is more useful than a complete-looking journey graph built on prohibited or unstable matching.
How to implement multitouch attribution for mobile campaigns
Treat implementation as a measurement policy and data-quality project. The model comes after the event definitions.
1. Define the decision and conversion
Write the decision first: “Should we move 15% of prospecting spend from paid social to paid search?” Then name the conversion and value being allocated: install, activated user, trial, purchase, or 30-day revenue.
One dashboard should not silently use install credit to answer a revenue question. Preserve the install source, then measure downstream cohorts separately.
2. Create a canonical touchpoint schema
For each eligible touch, capture a stable event name, timestamp, channel, source, campaign, creative, interaction type, destination, consent state, and available first-party identity. Store the raw parameters as well as normalized channel labels so taxonomy fixes do not destroy history.
Document exclusions. Internal test clicks, duplicate events, bot traffic, self-referrals, and direct sessions after a known campaign touch all need consistent treatment.
3. Preserve campaign context through the install
Mobile journeys often break at the App Store or Play Store. A deep link should route an existing user directly to the intended screen, while a deferred deep link should restore eligible campaign context after a first install.
Deeplinkly combines deep linking, deferred deep linking, install attribution, and raw exports through CSV, API, and webhooks, giving app teams one deterministic event trail to replay across attribution models. When no supported signal exists, the install stays unattributed rather than being assigned through device fingerprinting.
4. Set eligibility and attribution windows
Decide which clicks, views, and engagements can enter a path and for how long. Keep view-through and click-through results separate because an impression and an intentional tap provide different evidence.
Use the same mature cohort when comparing windows. The attribution-window guide shows how to replay conversion lag and audit the conversions added by a wider rule.
5. Replay several models before choosing one
Run last click, first touch, and one multi-touch attribution model against the same deduplicated paths. Compare channel credit, ROAS, cost per attributed conversion, and rank.
Then change one assumption at a time:
- shorten and lengthen the window;
- remove impression-only touches;
- separate new from returning users;
- compare install credit with downstream revenue credit;
- exclude paths with weak identity stitching.
If the recommended budget changes every time, the model is not ready to drive automation.
6. Reconcile systems before debating the model
An ad network, analytics platform, and MMP may disagree because they use different windows, time zones, identity scopes, view rules, event definitions, or deduplication priorities. Those are data-contract differences, not necessarily missing events.
Run an attribution discrepancy audit and establish one source for cross-channel allocation. Platform reports can still optimize campaigns inside their own inventory, but adding self-reported conversions across networks will double count shared journeys.
7. Validate budget moves with incrementality
Attribution asks how observed conversions should be credited. Incrementality asks how many additional conversions occurred because the campaign ran. Google explains that its Conversion Lift compares outcomes for a treatment group that saw ads with a control group that did not, deliberately ignoring normal attribution rules for that causal estimate.
Use audience holdouts, geo tests, or platform lift studies for decisions with meaningful spend. A good operating loop is:
- Use MTA to find a channel that may be undervalued or overvalued.
- Form a specific budget hypothesis.
- Test the change with an eligible control.
- Compare lift with the model's predicted direction.
- Adjust the model policy or its decision weight.
For broader guidance, use the incrementality testing guide. The IAB and MMA also describe MTA and marketing mix modeling as complementary rather than mutually exclusive: user-level attribution provides granularity, while aggregate methods can capture effects that individual paths miss.
8. Version the policy and monitor drift
Record the model, formula, eligible touch types, windows, identity rules, conversion definition, lookback period, exclusions, and effective date. Recompute historical comparisons from fixed raw data where possible.
Monitor unattributed share, duplicate rate, path length, channel mix, consent coverage, delayed postbacks, and the percentage of revenue whose first and last touches are connected. A sudden shift in any of these can move MTA credit even when campaign performance is unchanged.
Common multitouch attribution mistakes
- Treating assigned credit as causal lift. Presence in a path is not proof that a channel created the outcome.
- Starting with the most complex model. Clean events and stable identities matter more than an impressive algorithm.
- Letting impressions overwhelm clicks. High-volume, low-intent exposures can absorb credit unless eligibility is controlled.
- Ignoring the unattributed population. A model trained only on visible journeys may not represent all converters.
- Mixing platform and cross-channel truth. Every ad network has a partial view and an incentive to report its contribution.
- Changing model and window together. You will not know which assumption moved the result.
- Using one model for every decision. Campaign optimization, financial reporting, journey analysis, and long-term planning have different needs.
- Automating budget changes without guardrails. Require minimum data, a stable rank across models, and an experiment for large reallocations.
Frequently asked questions
What is the difference between single-touch and multitouch attribution?
Single-touch attribution gives 100% of conversion credit to one interaction, usually the first or last eligible touch. Multitouch attribution distributes credit across two or more interactions using a rule or data-driven model.
What is the difference between multi-channel and multi-touch attribution?
Multi-channel attribution focuses on how credit is shared across channels such as search, social, email, and referrals. Multi-touch attribution can go deeper by assigning credit to several interactions, campaigns, or creatives, including multiple touches within one channel.
How do you calculate multi-touch attribution?
First define the conversion value and eligible path. Then apply weights whose total equals 100%: a linear model uses 1 ÷ number of touches, while position-based and time-decay models use chosen rules. A data-driven model estimates weights from observed converting and non-converting paths.
How much data do you need for multitouch attribution?
Rules-based MTA has no universal minimum, but it still needs enough complete journeys to produce stable channel comparisons. Data-driven requirements vary by platform and model. Instead of using one arbitrary threshold, check whether results hold across time periods, cohorts, windows, and reasonable model choices.
Can a small marketing team use multitouch attribution?
Yes. Start with clean campaign tagging and compare first touch, last click, and a linear model in a spreadsheet or warehouse. Do not buy complex modeling infrastructure until the comparison reveals a recurring budget decision that simpler reporting cannot answer.
Is first-touch attribution a multi-touch model?
No. First-touch attribution is a single-touch model because one interaction receives all the credit. It is still useful beside MTA because it shows which recorded channel introduced users and helps expose the opposite bias from last click.
Does multitouch attribution work without cookies or device IDs?
It can work on the subset of journeys connected by consented first-party identifiers, deterministic campaign signals, authenticated sessions, or privacy-preserving platform reports. It cannot reconstruct every person-level touch when the underlying identity signal is unavailable, and it should not use fingerprinting to pretend otherwise.
Choose a model your budget can survive
Use last click when the journey is short and auditability matters. Add an auditable multi-touch model when assists are common and early channels disappear from the budget conversation. Use data-driven attribution only when the event, identity, and conversion data are strong enough to support it.
Before reallocating meaningful spend, test whether the conclusion survives a different reasonable model and whether the credited channel creates incremental lift. If it does, move budget with confidence. If it does not, improve the measurement policy before asking the model for a more precise answer.