Multi-Touch Attribution Explained: Models, Data Requirements, and Limitations

Multi-touch attribution (MTA) reconstructs the recorded interactions that precede a defined conversion and distributes that conversion’s credit or value across two or more eligible touchpoints. Rules-based models assign preset weights; algorithmic models learn weights from path data. The output explains how a model allocated observed outcomes. Unless the method adds credible counterfactual evidence, it does not prove that the credited touches caused incremental conversions.

Nielsen’s guide to multi-touch attribution defines the method around fractional credit: several journey touchpoints receive portions of the credit for a sale or another KPI event. First-touch and last-touch are different. Each gives 100% of the credit to one recorded interaction, so they are single-touch baselines even when they appear beside MTA models in the same tool.

The multi-touch attribution formula#

An ordinary normalized fractional-credit model shares a simple accounting constraint. For conversion c, let V_c be its value and let w_c,j be the weight assigned to eligible touch j:

Attributed value A_c,j = V_c × w_c,j

w_c,j ≥ 0 and Σ w_c,j = 1

Add the attributed values for all touches assigned to a channel to get that channel’s reported attributed value. The argument between models is not about this arithmetic. It is about how the weights are chosen.

• Linear attribution uses w_c,j = 1 ÷ n for n eligible touches.
• A common U-shaped rule gives 40% to the first touch, 40% to the last, and divides 20% among the middle touches.
• Adobe documents time-decay raw weights as r_c,j = 2^(-t_c,j ÷ h), where t is time before conversion and h is the chosen half-life; each final weight is r_c,j ÷ Σ r_c,j.

Here is a clearly illustrative example, not company data. Represent one conversion with 100 attribution points. Its eligible path is paid social, webinar, email, then branded search. Linear attribution assigns 25 points to each touch. A U-shaped model assigns 40 points to paid social, 10 to the webinar, 10 to email, and 40 to branded search.

The conversion still totals 100 points. The model changed the distribution, not the outcome. It also did not estimate how many points would remain if the marketing touches had been withheld. That is why a credit-allocation formula is not an incrementality formula.

Adobe’s attribution model documentation also shows an important implementation detail: one- and two-touch paths are normalized separately, and a lookback window decides which interactions are eligible before weights are applied. The name of a model is therefore not a complete specification. The container, conversion, window, exclusions, and edge-case rules are part of the model too.

Attribution, MMM, and incrementality answer different questions#

Multi-touch attribution starts with an observed conversion and asks how to divide its credit among recorded prior touches. Marketing mix modeling starts with aggregate variation in outcomes, media, and other business drivers over time. Incrementality starts with a treatment contrast and asks what changed because an intervention occurred.

The distinctions are operational, not semantic:

IAB’s guide to MMM and MTA treats the methods as complementary. MTA is generally more granular and tactical where interactions are addressable. MMM can cover aggregate effects from channels that do not produce person-level paths. The same guide is explicit that predetermined first-click, last-click, equal, or time-dependent weights do not reach true incrementality.

What each multi-touch model is actually assuming#

No model is neutral. A rules-based model makes its assumption visible in the weights. An algorithmic model moves the assumption into its features, training data, objective, and calibration. The right choice depends on the decision the report will support and on whether the data can support that level of inference.

Linear: every eligible touch counts equally#

Linear attribution is transparent and easy to reproduce. It is useful as a reference view when the team wants every observed step represented and does not yet have evidence for a stronger weighting rule.

Its limit is also transparent: repetition creates credit. Ten low-information page views collectively receive more credit than one high-information interaction unless the event definition or deduplication rule prevents that. HubSpot’s model documentation makes the problem concrete: in its linear view, every eligible interaction receives equal credit, so a highly targeted campaign interaction can be weighted the same as an ordinary page view.

Position-based: discovery and conversion get privileged#

U-shaped attribution emphasizes the first and last eligible interactions while preserving some credit for the middle. B2B systems may add milestone variants: HubSpot documents W-shaped and full-path models that reserve weight for contact creation, deal creation, or closed-won milestones.

These models are readable, which makes them useful when a team deliberately wants a funnel policy reflected in reporting. Their weights are judgments, not measured effects. They also depend on clean milestone order. HubSpot documents cases where a W-shaped or full-path result is null when deal and contact dates do not fit the required sequence.

Time decay: recency gets privileged#

Time decay gives more credit to touches nearer the conversion. A half-life makes the policy explicit: after one half-life, a touch’s raw weight is half that of an otherwise comparable touch at conversion time.

This model can be useful for a time-boxed promotion or an acceleration question. It systematically discounts early discovery, however, and its answer changes when the half-life or lookback window changes. A long B2B sales cycle does not make a seven-day default inherently correct.

Custom: the organization writes the policy#

A custom model can encode known milestones, exclude administrative events, or reflect a carefully argued operating hypothesis. It is the most honest rules-based option when the default models plainly do not match the journey.

Custom does not mean learned or validated. If the team writes 50%, 30%, and 20% into a model, the output will reflect that policy with mathematical precision. It will not tell the team whether those percentages were true.

Algorithmic: the data helps choose the weights#

Algorithmic or data-driven attribution estimates weights from observed paths rather than fixing them in advance. Google Analytics says its current data-driven model evaluates converting and non-converting paths and considers timing, device, interaction count, sequence, and creative type. Adobe documents an algorithmic model based on a cooperative-game allocation method. The label therefore describes a family, not one portable algorithm.

An algorithmic model needs more than converted journeys. It needs enough representative non-converting paths to distinguish common exposure from conversion-associated exposure, stable event definitions, and enough observations across the channel combinations it is expected to score. Even then, prediction is not automatically causal identification.

Model availability is also a product fact, not a timeless taxonomy. Google Analytics currently lists data-driven attribution, paid and organic last click, and Google paid channels last click in its attribution reports; it no longer offers first click, linear, time decay, or position-based models there. A warehouse model, Adobe, HubSpot, Salesforce, and another analytics product can expose different choices under different names.

The minimum data contract for MTA#

MTA is a path-building system before it is a weighting system. If events cannot be linked to the right subject and outcome in the right order, a more sophisticated model only assigns more sophisticated weights to a broken path.

A minimum event-level contract needs these records:

This table is a practical synthesis, not a universal vendor schema. Salesforce’s MTA setup requirements illustrate the foundation: first-party data must be ingested and harmonized, and user-level identifiers must work across multiple channels before touchpoints and conversions can be configured. Adobe is equally direct about the join requirement: summary-level data without a visitor identifier cannot be used for its attribution processing.

Revenue attribution needs CRM evidence, not only web events#

For a B2B revenue report, a cookie or analytics identifier is not enough. The path must reach the person, account, opportunity, closed status, and approved value used by the business.

HubSpot provides a bounded product example. Its revenue attribution requires a closed-won deal, at least one associated contact, and populated amount, create-date, and close-date properties. Sales activities must be associated with both the relevant contact and deal. Interactions outside HubSpot need supported campaign tracking and a page carrying its tracking code or they are absent from that report.

The general lesson is not to copy HubSpot’s objects. It is to test the full join. Select a small sample of closed outcomes and reconstruct every link from source event to campaign, identity, opportunity, status, and value. Then start from a touch and trace forward to the outcome. A report total can look plausible while both directions contain silent omissions.

B2B identity is an account problem as well as a person problem#

One buyer can switch devices, reject tracking, use multiple email addresses, or interact anonymously before identifying. One account can have several stakeholders. A person-level model may over-credit the individual who submitted the form while missing the colleagues who attended a meeting or influenced the purchase offline.

Define the unit of analysis explicitly: person, account, opportunity, or buying group. Document how anonymous activity becomes known, how contacts associate with accounts and opportunities, how duplicates are resolved, and whether a later identity merge rewrites historical paths. Do not silently treat a person ID as a complete account journey.

IAB identifies identity resolution, shared devices, cookie rejection, offline matching, walled gardens, and unavailable user-level impression data as material MTA constraints. These are coverage facts, not minor implementation defects.

An algorithmic model needs a volume statement, not a borrowed threshold#

There is no universal number of conversions that makes a learned attribution model reliable. Required volume changes with conversion frequency, path diversity, number of channels and features, class imbalance, model design, and how quickly behavior changes.

Published platform limits show why thresholds do not travel. Search Ads 360 currently requires 15,000 clicks and 600 Floodlight conversions in the preceding 30 days for its own data-driven model. If volume falls below its product requirements, it can retain a prior model or fall back to linear allocation under documented conditions. Those numbers describe one product’s inputs and fallback—not a general definition of adequate MTA data.

When volume is limited, compare transparent rules-based views and label the result as policy-based. An unstable learned weight is not more informative merely because it came from an algorithm.

The limitations that survive clean implementation#

Perfect event plumbing would remove many errors. It would not solve the identification problem.

Exposure is selected, not random#

People who see an ad, search a brand, attend a webinar, or open an email can already differ from people who do not. Intent affects both the path and the probability of conversion. An observational model can learn that branded search is highly predictive without establishing how much demand branded search created.

A peer-reviewed comparison of observational methods with 15 randomized Facebook experiments found that the observational estimates often failed to recover the randomized effects even after the models included extensive demographic and behavioral controls. The study is historical and platform-specific, so it is not a current universal error estimate. It demonstrates why more granular observations do not automatically reveal the missing counterfactual.

Missing paths are not missing at random#

Consent choices, browser restrictions, cross-device movement, offline interactions, and platform boundaries remove some touches and conversions from direct observation. The missing population can behave differently from the observed population.

Google’s modeled-conversion documentation describes how its own systems estimate some unobserved outcomes with aggregate signals and withhold modeled reporting when there is insufficient information. That is a specific platform capability. An in-house MTA table cannot repair the same gaps merely by adding a modeled label.

A precise allocation can still be wrong for the decision#

Rules-based models are deterministic: the same path and rule always produce the same weights. That precision says nothing about whether the rule represents influence. Algorithmic models can be precisely predictive while still carrying selection bias or omitted-channel bias.

Amazon’s 2025 MTA methodology paper states the trade-off plainly: machine-learning models trained purely on observational data scale well and can be precise, but may yield biased estimates of ad effects; randomized trials provide less biased causal evidence but can be noisy and too coarse for touch-level allocation. Amazon describes combining experiments, machine learning, and event histories in its own system.

Model choice can reverse the story#

If linear says content leads, U-shaped says paid social leads, and time decay says branded search leads, the correct conclusion is not to select the most convenient winner. The conclusion is that the ranking is sensitive to defensible policy choices.

Run at least one simple baseline and two plausible multi-touch models on the same frozen path table. Compare channel rank, attributed value, and the decision each view would trigger. Investigate reversals. Sensitivity is itself a result.

Six receipts to require before acting on MTA#

Use MTA as a governed reporting model, not as a number generator. Before a channel decision, require six short receipts:

1. Outcome receipt: the conversion, value, status, and reversal policy are written and reconciled to the system of record.
2. Path receipt: included channels and touch types are listed, and known blind spots are quantified where possible.
3. Identity receipt: the person, account, or opportunity unit and its merge, split, and association rules are explicit.
4. Window receipt: the lookback, timezone, late-arrival, deduplication, and direct-traffic rules are versioned.
5. Sensitivity receipt: the result has been compared with a simple baseline and another plausible model on the same frozen data.
6. Calibration receipt: any causal claim points to an experiment or defensible quasi-experiment; otherwise the output is labeled attributed, associated, or directional.

There is no broadly accepted benchmark for a good credit share, a minimum journey length, or the amount of agreement models should show. IAB notes that MTA benchmarks are scarce and results vary with category, media mix, investment, data availability, freshness, and provider methods. Internal receipts are more useful than a borrowed target because they reveal whether this particular result is fit for this particular decision.