Multi-Touch Attribution: Read Credit First

A customer clicks a social ad, finds a buying guide through search, opens an email, and later returns through a branded search to buy. The paid search dashboard calls the final click the winner. The email team can point to its own interaction. A marketing manager trying to decide next month’s budget needs a better account of the journey than either channel report provides.

multi-touch attribution: a large centered tilted tablet showing an abstract journey chart, face-down phone, map pin, clock, coins, sealed envelope, pen, and potted plant

Multi-touch attribution offers one: it divides credit for a recorded conversion among several recorded interactions on the path. That can make earlier activity visible in a decision that would otherwise favor the last click. It also creates a tempting but costly misunderstanding. Credit is a way to describe and compare the paths the system can see; it is not automatically a measure of sales that would disappear if a channel stopped. Nielsen’s guide to multi-touch attribution describes the fractional-credit approach and the different model families. The useful question for a manager is therefore not simply which channel received the most credit. It is whether a particular definition of a conversion, path, and credit rule gives a dependable enough view to change a particular decision.

Decide which customer action deserves credit

Before assigning credit, decide what counts as a conversion. A form submission, a created contact, a purchase, and a closed sale are different outcomes. A model can divide credit for any of them, but an answer about contact creation should not be presented as an answer about revenue. A campaign might introduce many contacts while another helps them reach a purchase. Calling one campaign the winner without naming the outcome confuses those two jobs.

The distinction is especially consequential when sales happen in a CRM after marketing’s first interaction. HubSpot distinguishes contact, deal, and revenue reports, and says its revenue attribution depends on eligible closed-won deals with an associated contact and populated deal fields, including amount, in its attribution definitions. A closed-won deal with no associated contact cannot contribute its contact interactions to that report. The practical lesson travels beyond HubSpot: a revenue chart is built from records that meet its inclusion rules, not from every sale the business remembers making.

For a first decision, choose the outcome closest to the decision you actually control. If the question is which landing page brings people into a sales process, contact creation may be sufficient. If the question is whether a campaign deserves more of a revenue budget, a lead count is an incomplete substitute for sales. When deal records are patchy, present the lead result as a lead result and repair the link to the sale before giving it a revenue label. That choice may delay a polished revenue dashboard, but it avoids a stronger claim than the data can carry.

State the unit as well. Fractional credit can be allocated across one order, across its recorded value, or across many eligible conversions in a period. If a report says email received 24% of revenue credit, readers need to know which revenue entered the calculation and over what dates. A share of the eligible, linked deals is not necessarily a share of all company revenue. That denominator belongs beside every headline number because it determines how far the conclusion can travel.

Build a path from records that actually connect

The attractive diagram of a customer journey hides a practical task: each interaction must be recognizable as an event, placed before the conversion, and connected to the same person or record under a consistent rule. In Salesforce’s setup for its Multi-Touch Attribution app, the documented requirements include harmonized first-party touchpoint data and user-level identifiers across channels. The need for consistent identity is not a minor technical preference. Without it, an email interaction and a web purchase can sit in separate records even when a human considers them part of one journey.

Imagine that the customer in the opening example reads the buying guide on a work laptop and purchases on a phone. If the system cannot connect those records, the report may show a one-touch phone purchase instead of the four-touch path. The arithmetic may still be flawless for the recorded path. The business interpretation changes because the path is incomplete. A first-touch label then means first recorded eligible touch, not necessarily the first moment the customer learned about the company.

Define a touchpoint with similar care. A logged email click, a page visit, and an ad impression are not interchangeable observations. A company may record some of these events in detail, receive only summary information for others, and never see certain exposures at all. Combining them in one channel ranking without saying which events are eligible gives the most visible channels a structural advantage. Even a simple choice such as counting repeated email clicks separately can change a linear model’s allocation, because more eligible events create more places to assign credit.

The point is not to assemble a perfect history of every person. That is usually unavailable. The point is to know the observed population. Count the conversions that are eligible for the chosen outcome, the conversions with at least one linked touch, and the share with multiple linked touches. Then look at which channels are reliably recorded and which are sparse. If very few purchases have more than one visible interaction, a more elaborate multi-touch model will spend most of its time assigning full credit to single recorded touches. It cannot recover an earlier interaction simply by using a cleverer weighting formula.

Privacy choices and device boundaries are among the reasons paths can break. Google’s explanation of modeled conversions describes lost observation arising from privacy and technical limits, including movement between devices. A missing interaction is therefore not proof that a channel did nothing. Equally, the possibility of missing data is not permission to assume that every unobserved channel mattered. The honest report says which part of the journey was observed and confines its credit claims to that view.

Choose a credit rule for the question at hand

Consider an illustrative $100 order with four recorded touches in this order: a paid social click, an organic visit to a buying guide, an email click, and a branded paid search click. Last-touch attribution assigns the entire $100 to branded search. First-touch assigns it to paid social. Both make a useful comparison because they show how much the answer depends on an endpoint, but neither describes the middle of the observed path.

A linear multi-touch model divides the same $100 equally: $25 to each recorded interaction. That is an intelligible starting point when the team wants to see which channels appear along linked journeys without claiming to know that one interaction was more influential than another. Its cost is built into its fairness. A brief, repeated interaction can accumulate credit alongside a substantial buying-guide visit because the rule counts eligible touches, not the customer’s reason for buying. The report describes how often channels appear in the defined paths under that rule; it does not discover equal causal importance.

A position-based rule deliberately values the endpoints more. For example, an illustrative rule could assign 40% to the first touch, 40% to the last, and divide the remaining 20% between the two middle touches. On the $100 order, paid social and branded search would each receive $40, while the guide and email would each receive $10. Choose such a rule when the team has a clear reason to emphasize introduction and final return. Do not present the resulting $40 as a measured effect of either endpoint. The weights express the chosen policy for allocating credit. Adobe’s model descriptions show how linear and position-based configurations distribute conversion value differently.

Time decay makes recency the organizing assumption. If the four touches occurred 30, 15, two, and zero days before the order, respectively, the later interactions receive more weight under a time-decay rule. That can be a sensible descriptive view of a short promotion whose immediate path is the question. It is a poor default for judging an earlier buying guide if customers commonly return much later to purchase. In that case the model would reduce the guide’s credit because of elapsed time, regardless of whether the guide helped the customer make sense of the product. The dates and the chosen decay rate matter as much as the model name.

Algorithmic attribution estimates weights from observed data rather than fixing the same proportions for every path. One reason to consider it is that it can use information beyond the converting journeys. Google Analytics’ account of data-driven attribution says its method evaluates converting and non-converting paths. That is a meaningful difference from merely splitting each completed sale by a fixed rule. It does not turn an output labeled “data-driven” into a universal truth about a channel. The model still depends on the events, identities, conversion definition, and available comparison data in the product using it.

There is no ranking in which the most complicated rule is automatically the best. Start with linear credit if the immediate job is to reveal the distribution of recorded touches across a path and make the arithmetic easy to challenge. Put first- and last-touch beside it as endpoints. Use position-based or time decay when the business question makes their weighting choice explicit. Consider an algorithmic report when its inputs and method are clear enough that a change in credit can be explained to the people who must act on it. Whichever model is chosen, label the result as credit under that model, not as a discovered percentage of causal contribution.

Set the path boundary before comparing channels

Changing the model is only one way to change the result. An attribution configuration also decides how far back to look and which events are considered part of the same path. Adobe describes three components: a model, a container that sets path scope, and a lookback window. The same linear rule can produce different channel shares if its path includes only one visit instead of a longer sequence of visits.

Take another illustrative order. A customer clicks an introductory ad 45 days before buying, reads a product page ten days before, and clicks an email on the day of purchase. A 30-day lookback excludes the first ad; a 60-day lookback can include it if the records connect. Under a linear rule, the 30-day path gives the product page and email half the credit each. The 60-day path gives each of the three eligible touches one third. Neither calculation is mathematically wrong. They answer different questions because the set of included touches changed.

The choice is consequential for a longer buying process. A short window can quietly erase the channel that introduced a customer. A very long window can keep assigning credit to an old interaction that may be less relevant to the current purchase. The correct boundary depends on the buying behavior the business wants to study and on whether the data can connect interactions over that period. It should be stated before anyone reviews which channel “won.” Adobe’s attribution FAQ also notes that conversions can have no eligible touch within a selected window, making the absence of credit itself a configuration outcome.

Keep the reporting period distinct from the lookback period. A September purchase report asks which conversions occurred in September; its lookback may reach into August or earlier to find eligible touches. When a team changes both dates at once, a difference in channel credit has no clean interpretation. Hold the conversion population, path scope, channel definitions, and window steady while comparing models. If the team then wants to test another window, change that one setting and explain the resulting difference.

Read fractional credit without mistaking it for new sales

Multi-touch totals can look unfamiliar because a single conversion is split into fractions. In the illustrative four-touch order, linear attribution could show a quarter of an order for each touch if the metric is order count, or $25 for each if the metric is order value. The customer did not place four quarter-orders. The fractions are accounting shares assigned by the model. Across many orders, those shares are summed to describe where the selected rule placed credit.

This also explains why a channel’s assigned revenue should not be added to the same revenue as though it were extra business. If a $100 order is fully allocated among four touches, the total allocated value remains $100. A different rule moves the $100 around; it does not create a second sale. If two dashboards display different numbers, first check whether they use the same conversion event, eligible records, attribution window, and identity rules. A model comparison only begins after those pieces align.

Coverage determines whether even a clean allocation is representative. Consider a purely illustrative report with 1,000 purchases in a month. Suppose 740 purchases can be linked to at least one eligible touch and 260 cannot. If the report says email received 22% of fractional order credit among the 740 linked purchases, it has not established that email drove 22% of all 1,000 purchases. The unlinked 260 might differ from the linked group in ways the report cannot show. The 740 denominator belongs in the sentence, along with the model and window. Without those details, a precise percentage invites an imprecise decision.

Comparing channel shares across periods needs the same discipline. A rise in email credit might reflect more useful email activity, but it might also reflect a new email integration, a change in lookback, or a loss of visibility into another channel. A report can show a change in allocation while the customer behavior of interest remains uncertain. Check whether the number of eligible conversions and linked paths changed before treating a share movement as a performance movement. The cost is a slower readout; the benefit is that a tracking change is less likely to masquerade as a marketing success.

Credit shows a route to conversion, not what caused it

The causal question is different: how many of these purchases would have happened without the advertising? Fixed attribution weights cannot answer that from a completed path alone. The IAB’s guide to marketing mix modeling and multi-touch attribution places that distinction among the limits of using path-level credit to judge marketing. A rule that assigns 40% of a sale to a first ad is a rule about the observed sale, not a demonstration that removing the ad would have removed 40% of its value.

The distinction matters most at budget time. Someone already intent on buying may see a branded search ad immediately before purchase. Last-touch assigns the ad all credit, yet the person might have purchased anyway. The reverse can also occur in the recorded path: an early exposure might matter, but later recorded clicks claim most of the credit. The data in one completed journey cannot tell us which alternative would have occurred without a particular ad. Adding non-converting paths to a model gives it more information; whether a specific estimate identifies a causal effect depends on its design and data, not on the word “algorithmic.”

This is a practical concern, not a semantic objection. In a comparison with randomized Facebook field experiments, observational estimates often failed to match the experimental estimates even when the researchers had extensive observed data. The study does not give a universal correction factor for an attribution report. It shows why rich records of who saw ads and who bought can still be a shaky basis for saying what the ads caused.

When the decision is whether to increase or remove a significant amount of spend, ask a question that can change the action: what happens to purchases when otherwise comparable people or markets do and do not receive that activity? A randomized holdout can address that question if it is feasible and designed around the outcome that matters. Attribution remains useful for describing linked journeys and finding plausible places to investigate. It should not carry a major causal budget claim on the strength of fractional credit alone.

Use the report to make a smaller, better decision

The best immediate use of multi-touch attribution is often a specific operational decision. Suppose branded search takes nearly all the last-touch credit, while paid social repeatedly appears early in linked converting paths. Under a broader credit rule, paid social may receive more credit. That does not justify moving a large budget to paid social by itself. It does justify examining whether paid social is introducing people who later return, whether its audience and creative match that role, and whether a controlled spend test is worth running. Meanwhile, cutting the early channel solely because last-touch gives it little credit would discard information the fuller path provides.

The same logic can protect a valuable middle touch. If the buying guide repeatedly appears between initial discovery and purchase, a first- or last-touch report gives the guide no credit by design. A linear report brings it into view. The next decision might be to improve the guide, make it easier to find, or check whether visitors who use it progress differently. The report alone cannot tell which edit would create more purchases. It can tell the team that the guide belongs in the journey being considered and that an endpoint-only story omitted it.

Prioritize comparisons that survive reasonable changes in configuration. If a channel looks strong only under a short window that excludes earlier activity, its ranking is fragile. If a channel appears on many linked paths under several sensible windows and receives meaningful credit under different rules, there is a stronger descriptive case that it participates in the recorded journey. That still leaves the incremental-value question for a test. A stable credit pattern is a reason to investigate a channel carefully, not a license to turn an assigned revenue share into a spend target.

The trade-off is time. A single model produces a neat ranked list quickly. Defining the population, looking at alternate windows, and connecting a test to the decision takes longer. The extra work is most valuable when the proposed budget change is large or when the channels differ sharply in how observable they are. For a small creative or page improvement, a clear descriptive signal may be enough to act. For a large channel cut, demand evidence about what happens without the spend.

A workable first readout

Begin with one conversion outcome and one period. Record the number of eligible conversions, the number with a linked path, and the number with more than one recorded touch. Specify which events are eligible as touches and how they connect to the conversion. For a CRM revenue report, reconcile the closed-won deals and associated contacts before debating weighting. For a web purchase report, check that the purchase and its preceding interactions use the same identity and dates. This modest starting point makes a missing-data problem visible while it is still fixable.

Then publish a transparent baseline: linear credit, a stated path scope, and a lookback period tied to the customer process being studied. Show first- and last-touch views next to it, because the gap between them tells the reader where an endpoint rule changes the story. If a position-based, time-decay, or algorithmic result is added, explain what that result values and which settings differ. Preserve the same conversion population while comparing rules. Otherwise, the dashboard may attribute a difference to the model that actually came from a changed set of orders.

The final readout should name the decision it informs in ordinary language. For example: “Among linked September purchases under a 30-day window, email received a larger share of credit under linear allocation than under last-touch. That suggests email often appears before the final recorded interaction; it does not establish that email generated the difference in sales.” The numbers in that sentence would come from the company’s own report, not from a generic benchmark. The next line should say whether the team will improve an email step, inspect missing paths, or run a spend test. Credit becomes useful when it clarifies that next choice and stays within what the recorded journey can support.

Multi-touch attribution is worth using when more than one recorded interaction matters to the decision and the underlying paths are connected well enough to read. Its strongest contribution is to replace a reflexive “last click wins” story with a clearer account of where a chosen rule places credit. The boundary is equally important: a credit share belongs to a defined outcome, eligible population, path scope, and time window. Keep those definitions visible, and use a separate causal test when the real question is whether spending created sales that otherwise would not have happened.

One person. A whole marketing team.

Invite only