Marketing Attribution Models: One Sale, Different Answers

A customer discovers your website through organic search, later clicks a paid-search ad, returns through an email and finally visits directly to buy. Which channel gets credit?

An attribution model supplies the rule. First touch credits the earliest recorded interaction. Last touch credits the final one. Other models divide the credit across the journey. The sale stays the same; the report changes how it is allocated.

Consider this fictional CAD 1,200 order:

Organic search, 21 days before purchase → non-brand paid search, 14 days before → email, 7 days before → direct visit on purchase day.

Every interaction is eligible within the example’s 30-day window. Here is how six teaching rules allocate that one order.

Every row totals $1,200. None establishes that its credited channel caused the purchase or that removing another channel would leave the sale unchanged.

Download the model-selection worksheet and fictional journey calculations. The kit includes readable CSV examples, a selection worksheet and optional calculation code. The CSV results are static; changing an input CSV does not recalculate them. No account connection is required to read the files.

The six rules above are conceptual comparisons, not six current Google settings. Current platform choices are covered separately below. All customer journeys and monetary amounts in this guide are hypothetical teaching examples.

What the main attribution models emphasize

The models become easier to choose when you understand what each deliberately emphasizes and what it leaves out.

First touch: the beginning of the recorded path

First-touch attribution assigns all credit to the earliest eligible interaction. In the example, organic search receives the full $1,200.

This perspective helps you inspect where measurable journeys begin. You might use it to find early pages or campaigns worth investigating when a last-click report gives them little credit.

The word “recorded” matters. The customer could already know the business through a recommendation, an event or an earlier visit outside the available window. First touch identifies the beginning of the path you can see, rather than proving the moment demand was created.

It also gives later interactions no share. The email may have answered a buying question, and the paid-search visit may have helped the customer compare options. The rule does not assess those possibilities; it allocates the entire outcome to the first eligible point.

Last touch: the final recorded interaction

Literal last-touch attribution assigns all credit to the final eligible interaction. Our example includes direct visits, so the final direct return receives $1,200.

This gives a simple view of how measured paths end. It can be useful for inspecting completion behaviour or explaining why a reporting system concentrates value near the sale.

It says little about earlier influence. A direct return may follow weeks of consideration. Crediting that final interaction does not establish that the buyer would have returned without the earlier information.

Be careful with the label “last click” in a platform report. A product may exclude direct traffic or restrict which channels can receive credit. Our literal example should not be treated as a reproduction of every platform’s last-click calculation.

Last non-direct: the last identifiable non-direct interaction

This rule ignores the final direct return and credits the last eligible non-direct touch. Email therefore receives the $1,200 in the opening example. If a path consists only of direct interactions, the teaching rule credits the last direct one.

The approach prevents a direct return from taking all credit when another eligible interaction precedes it. It remains a single-touch allocation: earlier organic search and non-brand paid search receive nothing.

This can make it a useful baseline for understanding closing channels. It does not mean the final non-direct touch created all the demand. The email could have supplied useful information, delivered an offer or simply recorded a buyer who already intended to purchase.

Linear: equal shares for eligible interactions

Linear attribution divides credit equally among the eligible touches. Four interactions receive $300 each in the opening example.

This makes earlier and later contacts visible without choosing one endpoint as the winner. It can be useful as a transparent comparison when you want to understand how much a last-touch view concentrates credit.

Equal allocation is a reporting convention. It does not prove that each interaction had equal persuasive value. A brief visit and a substantive consultation would receive the same share if both were represented as eligible touches in the same linear calculation.

The unit also matters. This guide divides by touchpoint, then adds the shares by channel. A channel appearing twice can receive twice the credit of one appearing once. Dividing equally among distinct channels would be a different rule.

Position-based: emphasize discovery and completion

Our position-based example assigns 40% to the first interaction, 40% to the last and divides the remaining 20% across the middle. That gives organic search and direct $480 each, with $120 each for paid search and email.

This makes the endpoints prominent while retaining some credit for intervening contacts. It can help explore how a conclusion changes when discovery and completion receive a larger share.

The weights are a chosen convention, not findings about this customer’s behaviour. The 40/20/40 pattern should not be presented as a universally correct balance between early and late marketing.

Short paths need an explicit rule too. The supplied calculator gives a one-touch path 100% to that touch and a two-touch path 50% to each endpoint. Without that exception, a two-touch path would have no middle interaction to receive the remaining 20%.

Time decay: emphasize more recent interactions

Time-decay attribution gives greater weight to interactions closer to the outcome. Our teaching example uses a seven-day half-life: a touch seven days before purchase has half the raw weight of one on purchase day.

The four touches have raw weights of one-eighth, one-quarter, one-half and one. After those weights are normalized to the single order, they allocate $80, $160, $320 and $640.

This is useful for exploring a recency-weighted perspective. It does not establish that an older interaction was less influential. A detailed early consultation could matter greatly even though the formula gives it a small share.

The half-life is part of the model. Changing it changes the result. Seven days is an illustrative parameter in the downloadable calculator, not a recommended setting for every sales cycle.

How data-driven attribution differs

A data-driven model uses available data and an algorithm rather than fixed shares chosen in advance by the analyst. Its results depend on the system, eligible interactions, conversion definition and information available to that model.

Google Ads describes its data-driven attribution as using patterns in converting and non-converting paths for the particular advertiser and conversion action. That is different from assigning a predetermined 40% to the first and last touch.

There is no honest way to fill a “Google DDA” row for the four invented interactions by choosing percentages that look plausible. The supplied kit deliberately does not simulate Google’s proprietary allocation. A real platform export would need its own settings, date and source attached.

Data-driven attribution can produce fractional credit. If an interaction receives 0.2 conversions, that represents an allocated share of outcomes, not one-fifth of a customer. The channel’s attributed value and the number of actual orders answer related but different questions.

A sophisticated model also does not automatically repair inaccurate conversion values or missing business outcomes. If a system optimizes around form submissions while the business cares about completed paid work, the model’s sophistication does not settle that definition mismatch.

Use data-driven results in the context of the platform’s stated purpose and coverage. Keep the underlying business outcome counts visible, especially when the model informs a consequential decision.

Which models are currently available in Google?

Verify the current options and the affected conversion action before making a live change.

Google’s GA4 attribution overview explains the three report models. Direct visits are excluded from credit unless the path is entirely direct. Google paid channels last click prioritizes the last eligible Google Ads click and falls back to paid and organic last click where no Google Ads click exists.

Google’s Ads attribution documentation and GA4 documentation identify first-click, linear, time-decay and position-based models as retired product options. They remain useful concepts for understanding allocation, but an older tutorial showing them as current selectable Google settings is outdated.

Also check which channels can receive credit for the relevant conversion action. A model name alone does not reveal the scope of eligible interactions. Two reports labelled last click can describe different channel populations.

Changing a live Google Ads conversion action’s attribution setting can affect bidding strategies using that conversion data. Treat it as an operational change, with the existing setting and affected goals understood. An exported model comparison can help investigate the difference before any live setting is changed.

Choose a reporting view for a specific question

There is no single model that answers every marketing question. Choose a primary view whose purpose you can explain, then compare it with one alternative that tests an important assumption. The question matters most in paid media and media buying, where budget is split across channels and each model credits them differently.

For an everyday report about where observed journeys finish, a clearly defined last-click view can be an understandable baseline. For a supported advertising workflow, the relevant platform’s data-driven view may be appropriate. In either case, the outcome definition and channel coverage need to match the reporting purpose.

For an investigation into early discovery, inspect recorded paths and compare an early-touch perspective. For a question about concentration at the end of a journey, a linear comparison can show how much the result depends on a closing-touch rule.

These are ways to ask better questions, not categorical instructions that first touch is best for awareness or time decay is best for every short sales cycle. The same rule can be informative in one analysis and misleading in another.

Choose the outcome before the model

A submitted form, qualified enquiry, opportunity and completed sale can each be analyzed. Keep them distinct.

If one channel receives many form submissions but few qualified enquiries, the next task is to understand the gap. It could involve audience fit, duplicate submissions, intake delays or the qualification rule. Switching attribution models will not tell you which explanation is correct.

For a revenue analysis, specify whether value means an order before refunds, retained revenue, recognized revenue or a contribution measure after defined costs. A fixed estimated value assigned to a lead is an assumption, not collected customer revenue.

The ROAS vs ROI guide covers the separate revenue-and-cost calculation. Attribution decides how the chosen outcome is credited; it does not deduct the costs that determine what remains for the business.

Scale the conclusion to the evidence

A model comparison can justify examining a content group or a set of customer journeys. It provides less support for a major budget cut when the underlying data are sparse or the result changes sharply under another reasonable rule.

If a single large contract determines the channel ranking, show that concentration. You can add a clearly labelled sensitivity excluding it while retaining the full result. Do not quietly remove the order to create the ranking you prefer.

For small teams, a simple, consistent baseline with known gaps can be more useful than many elaborate views. The value comes from understanding the result and choosing an appropriate next investigation, rather than maximizing the number of model columns.

Watch the channel story change across four orders

The downloadable example extends the opening journey to four fictional orders worth CAD 3,000 in total. All use a 30-day inclusive window and the same value basis, with no tax, refunds, view-through interactions or modelled missing events in the example.

“Brand” and “non-brand” are custom categories in these examples. They should not be mistaken for automatic platform classifications.

Compare last non-direct with linear attribution while keeping those four orders unchanged.

Brand paid search falls from $1,500 to $600, a 60% decrease in attributed value. The business has not lost 60% of its sales. The same four orders still total $3,000.

Linear attribution gives more channels a share and spreads some credit away from the final non-direct interaction. That is a property of the rule. It does not establish that the linear column is a more accurate causal explanation of the purchases.

Single-touch and multi-touch describe the allocation, not the journey

First touch, last touch and last non-direct are single-touch models: each gives the entire outcome to one eligible interaction. They can still be applied to a journey containing many interactions. “Single-touch” does not mean the customer only encountered the business once.

Linear, position-based and time-decay are multi-touch rules because they can distribute one outcome among several interactions. A one-touch journey still receives all its credit at that one point. J03, the direct-only $300 order, therefore produces the same direct allocation under all six teaching rules.

J02 shows where the differences begin. Its paid-social interaction occurs seven days before a $600 purchase, followed by brand paid search on purchase day. First touch gives paid social $600. Last touch and last non-direct give brand paid search $600. Linear and the kit’s two-touch position rule split the value $300 each. With a seven-day half-life, the raw time-decay weights are one-half and one, producing $200 for paid social and $400 for brand paid search.

That short example helps explain why models can sometimes agree. Linear and position-based happen to produce the same result on this two-touch path because of their stated rules. Agreement here does not mean those models will agree on a longer path, or that two independent methods have verified the customer’s reasons for buying.

Conversion credit and value credit can tell different stories

Each order contributes one conversion before its credit is divided. J01 contributes one conversion worth $1,200; J03 contributes one conversion worth $300. An order-count comparison treats them as equal-sized outcomes, while a value comparison gives the first order four times the weight.

Under linear attribution, J01 gives each of its four interactions 0.25 conversion and $300. J03 gives direct one conversion and $300. Direct receives the same dollar amount from each of those two orders, but very different conversion shares.

Across all four journeys, the model must preserve four conversions and $3,000 of order value. That does not mean a channel’s percentage of conversion credit must equal its percentage of value credit. Different order values explain why the shares can differ without any arithmetic error.

Keep both measures labelled when the distinction matters. A campaign associated with fewer, larger orders can look different in a value report than in an order-count report. Whether larger orders also leave more contribution requires the separate cost analysis; attribution alone cannot answer that question.

Why email receives more than one share on one order

J04 has two email interactions followed by brand paid search. Linear attribution assigns $300 to each of the three touches, so email receives $600 for that order.

J01 contributes another $300 of email credit, producing the $900 total in the linear column. If the two email contacts had been collapsed into one before allocation, email would receive a different amount.

That is a reason to state whether your model allocates by interaction or distinct channel. It is also a reason to remove genuine tracking duplicates before modelling. An accidental duplicate can change the allocation even though the person did nothing new.

Repeated legitimate interactions should not automatically be removed. Two separate emails may represent two real contacts. The data preparation should distinguish a duplicated event from a customer returning through the same channel on another occasion.

Describe zero-based changes in dollars

Organic search moves from zero to $300 between the two model columns. A percentage increase from a zero base is undefined. Report the dollar change and the model definitions instead of displaying an impressive but meaningless growth percentage.

The whole comparison should be described as a change in credit. Words such as gained sales, lost revenue or improved performance imply a business change that the example has not demonstrated.

The lookback window can change the answer first

The lookback window determines which prior interactions remain eligible for credit. The model then allocates among the eligible interactions.

In the opening example, a 30-day window includes organic search 21 days before purchase. If the teaching window is shortened to 14 days, that organic interaction drops out while the paid-search interaction exactly 14 days before remains eligible under the kit’s inclusive rule.

First-touch credit would then move from organic search to non-brand paid search. The model still uses the same first-touch rule, and the customer still made the same purchase. The eligible path changed.

That distinction helps explain why two reports can disagree despite displaying the same model name. One may include an early interaction that the other excludes. Longer windows can capture more recorded history, but they also change which contacts receive credit; more history does not automatically establish more causal influence.

Google’s lookback-window guidance describes the eligibility role of the window and states that changes apply going forward. Record the actual setting associated with an analysis rather than assuming an older export used the current configuration.

Decide which interactions belong in the comparison

A click, a page visit, an impression and a completed sales conversation are different observations. Before comparing models, decide which types of interaction the available dataset represents. Do not silently combine one platform’s impression-inclusive paths with another source’s click-only paths and attribute the entire difference to the model.

As a separate hypothetical variation, suppose the opening four-touch journey gained one additional eligible interaction. A five-touch linear allocation would give each touch $240 instead of $300. The $1,200 outcome is unchanged, but the population of credited interactions has expanded.

That is a coverage change, even though the linear formula stays the same. It could represent useful additional information, or it could be a duplicated tracking event that should be removed. Inspect what was added before judging the new allocation.

For a fair teaching comparison, this guide keeps the same eligible interactions across all six rules. Real reports may require a narrower shared scope before their model results can be compared meaningfully.

A reporting month and a lookback window are different

A report about purchases in October can allocate credit to September interactions if those interactions remain eligible. A report about October interactions can include outcomes that occur later, depending on its date basis.

Google’s attribution-model comparison documentation distinguishes event time from ad interaction time. Check which basis you selected before comparing the report with a sales ledger or another platform.

A similar calendar range does not prove that two reports follow the same outcomes. Align the chosen outcome, population and date basis before interpreting differences as a disagreement between models.

Acquisition dimensions are not outcome attribution

In GA4, a first-user source, a session source and an event-scoped attribution dimension have different jobs. One describes user acquisition, another describes sessions and another allocates credit for the selected outcomes.

A first-user source attached to a revenue report should not automatically be described as a first-touch allocation of that purchase journey. You need to know the dimension and metric together.

Google’s reporting-attribution setting documentation says the change affects relevant event-scoped reports and explorations, while user- and session-scoped traffic dimensions remain unaffected. It also says the reporting model applies to historical and future data.

That means a historical chart can change its channel allocation after a reporting-model change without the business’s historical sales changing. Preserve the setting and extraction date with important exports so the team can explain the difference later.

A plain label prevents an avoidable argument

Instead of calling a chart “marketing revenue,” label it with the outcome and view: purchase revenue credit, named model, eligible channel scope, date basis and extraction date.

The note can be short. It simply needs to survive when the chart is copied into a presentation. A colleague should not have to reopen an analytics account to discover whether the figure covers purchases, estimated leads or an earlier version of the reporting rules.

If the outcome definition changes, show that change separately from the model change. A new rule for what counts as a qualified opportunity affects the business population; a new attribution model affects how credit is divided within the chosen population. Combining both in one before-and-after comparison makes the result difficult to explain.

Identify the missing evidence that matters

Attribution relies on the outcome and the path being represented well enough for the intended analysis. Missing information can affect different parts of that picture.

If an order has no linked path, the business outcome may still be real. Keep it visible as unmatched or unallocated in your own analysis rather than quietly dropping it. Otherwise the matched journeys can appear to represent the entire business.

If campaign labels are missing, preserve the raw information and explain any repair. Assigning every unknown source to direct just to complete a chart invents a classification the evidence does not support.

If identity is fragmented across devices or systems, the available path may be incomplete. Work within the information the business is permitted to use and state the limitation. More intrusive collection is not the automatic solution to every measurement problem.

Offline conversations, recommendations and prior brand familiarity may be absent entirely. A digital attribution chart should not imply that its visible channels capture every influence on the buyer.

Direct and missing paths are not always the same label

The teaching calculator treats an explicit direct-only path differently from an empty eligible path. The former receives direct credit; the latter stays Unallocated.

Platform definitions can differ. GA4’s attribution-report documentation includes cases where Direct appears when no path data are available. Preserve the platform’s actual label and consult its definition rather than interpreting every Direct row as proof of someone typing the website address.

This is one reason a locally calculated teaching model should not be described as a replica of a platform model, even where the model names sound similar. The data preparation and exception handling can differ before the allocation is applied.

Modelled outcomes need an accurate description

GA4 can combine observed and modelled key-event information in reporting, as explained in Google’s modelled key-event documentation. A modelled aggregate estimate is different from recovering one identifiable customer’s exact missing journey.

Neither observed nor modelled automatically means useful or useless. The distinction affects how you describe the evidence. Keep the source’s qualifications attached and avoid treating a report as a complete event-by-event customer history when it is not.

Recent results can also develop as outcomes arrive and reporting updates. Compare periods with an appropriate observation allowance for the question, and record the extraction date. A fresh change in attributed totals does not by itself diagnose broken tracking or a genuine marketing improvement.

Use model disagreement to decide what to investigate

In the four-order example, organic search receives no last-non-direct credit but does receive linear credit. A reasonable next investigation is to inspect the early organic visits: which pages appear, what questions they answer and how those paths relate to the selected business outcomes.

That can inform content research. It does not establish how many sales would vanish if those pages disappeared. First-touch and linear models can make early interactions visible without proving their incremental contribution.

Brand search receives substantial last-non-direct credit and much less linear credit. That suggests investigating its role in completing recorded journeys and how much demand was already present. The comparison alone does not justify pausing brand advertising or treating it as unnecessary.

Email receives repeated-touch credit on J04. Investigate whether those are meaningful separate contacts, repeated navigation or a tracking duplication. The useful action depends on the actual events, not simply on email’s total allocation.

Hold the main view steady while investigating

Choose the primary reporting view before looking for the most favourable channel result. Keep one comparison that tests a real concern, such as whether a closing-touch view hides early interactions.

Changing the official model whenever a preferred channel looks weak makes trends harder to interpret. A model can legitimately change as the business or measurement system changes, but the reason and effective date should remain visible.

If multiple reasonable models lead to the same broad conclusion, that gives the conclusion some robustness to those particular allocation rules. It still does not prove causal accuracy. If the conclusion reverses, that sensitivity is itself useful information and should narrow the confidence of the recommendation.

Separate attribution from the decision about more spending

A channel’s share of attributed revenue is not automatically the budget share it should receive. Additional spend can reach different buyers, face different prices or require more creative and sales capacity.

Historical attributed averages describe the measured activity. A decision about the next spending increment needs evidence about the additional outcomes and costs. Use the attribution comparison to identify questions, then evaluate the proposed change on the terms that actually matter to the business.

When the question is causal, use different evidence

Attribution asks how to distribute credit among eligible interactions. A causal question asks what changed because an activity happened compared with what would otherwise have happened.

A returning buyer may click an ad shortly before a purchase they already intended to make. Another buyer may have been influenced by an earlier interaction that receives little closing credit. The path alone does not show the alternative outcome without either interaction.

Google’s Conversion Lift documentation describes controlled treatment-and-control comparisons for estimating incremental outcomes and notes that the feature is not available to every account. A suitable experiment is a different evidence exercise from reallocating the same set of observed sales.

Where a credible test is not feasible, use a narrower conclusion. You can report the observed business outcomes, how their credit changes across models and what commercial constraints apply. State that the incremental effect remains unresolved.

This guide does not provide an experiment design or promise a causal answer from the downloadable journeys. A material spending decision may warrant specialist analysis of the available comparison, timing and uncertainty.

Use the worksheet without making it a project

The first useful task in the kit is the blank model-selection worksheet. It asks what decision the analysis could change, which outcome is being counted and what data are available.

Begin with those questions, then choose the primary and comparison models. Explain what each emphasizes in one sentence. If the comparison cannot change an investigation or decision, adding another column may create work without adding insight.

Complete the fields that establish a fair comparison: the population, eligible interactions, lookback window, direct-traffic rule, date basis and value definition. Record important missing evidence and its consequence. For example, a missing link between qualified opportunities and marketing paths may prevent a channel-level qualified-opportunity analysis even though form submissions are available.

Keep internal evidence references meaningful to the reviewer. A saved export name and date can be enough. The worksheet does not need customer names or other unnecessary personal information to explain how the analysis was prepared.

Read the example outputs safely

The four-journey CSV repeats an order’s value on each touch row so the record is readable. Do not sum that input-value column across all touch rows. The unique order values are $1,200, $600, $300 and $900, totalling $3,000.

The credit-by-touch and credit-by-channel files contain results for all six teaching models. Filter to one model before adding its values. Adding every model together counts the same four outcomes six times.

The reconciliation worksheet in the pack uses a separate fictional dataset. It is an optional aid for understanding report coverage and should not be joined to the four journeys as though it describes the same orders.

Recalculation is an optional technical route

The included JavaScript calculator reads the JSON journey input and produces static CSV outputs. Editing the readable CSV input does not trigger it. The companion instructions explain the local Node.js workflow and the exact treatment of single-touch, repeated-channel, direct-only and missing paths.

Use a separate working folder so a recalculation does not overwrite the bundled reference outputs. The script requires no account connection or API key. Readers who only want to understand or choose a model can use the worksheet and supplied results without running code.

The arithmetic checks cover the stated teaching rules. They do not validate a live GA4 property, advertising account, CRM integration or spreadsheet application. If you introduce real data, its definitions, permissions and completeness require their own review.

Present the finding without implying the sales changed

A useful summary of the four-order comparison would be:

“Across four fictional orders worth CAD 3,000, brand paid search receives $1,500 under last non-direct attribution and $600 under linear attribution. The total order value is unchanged. The model choice materially changes the credit assigned to closing interactions, so the comparison supports investigating those paths before making a budget decision.”

That says what was measured and what follows. It does not claim the channel lost sales, that the linear result is causal truth or that a budget cut has been justified.

For your own report, keep the chosen outcome, model, window and source coverage beside the result. Use one alternative to expose the assumption that matters most. The aim is to understand why a channel receives credit and what evidence would make the next decision more dependable.

Frequently asked questions

What are marketing attribution models in plain terms?

A marketing attribution model is the rule a report uses to split credit for a sale across the interactions that came before it. First touch credits the earliest recorded interaction, last touch the final one, and other models share credit along the journey. The sale itself does not change; only the way the report allocates it does.

What are the main types of marketing attribution models?

The common rule-based types are first touch, last touch, last non-direct, linear, position-based and time decay. Single-touch models give one interaction all the credit, while multi-touch models divide it. Data-driven attribution is different again, because an algorithm sets the shares from the available paths instead of fixed weights chosen in advance.

Which marketing attribution model should a business choose?

No single attribution model answers every marketing question, so start with the question you need answered. Pick a primary view whose purpose you can explain, then compare it with one alternative that tests an important assumption. Decide the outcome being credited, such as completed paid work rather than form submissions, before choosing the model.

What is attribution analysis in marketing used for?

Attribution analysis compares how different credit rules describe the same set of sales, which shows where conclusions depend on the rule chosen. When models disagree sharply about a channel, that disagreement points to what deserves a closer look. It describes how credit is allocated and does not show that a channel caused a purchase.

How does a time decay attribution model split credit?

Time decay gives more weight to interactions closer to the sale. In the post's fictional example with a seven-day half-life, touches 21, 14, seven and zero days before a CAD 1,200 order receive $80, $160, $320 and $640. The half-life is part of the model, so changing it changes the result, and a detailed early conversation can still matter more than its small share suggests.

Which attribution models can I choose in GA4 today?

GA4 attribution reports offer data-driven, paid and organic last click, and Google paid channels last click. Google has retired first-click, linear, time-decay and position-based as selectable options, so older tutorials that show them are outdated. Direct visits receive credit only when the whole path is direct, and it is worth verifying the current options before changing a live setting.

How is data-driven attribution different from linear or position-based models?

Linear and position-based models apply fixed shares, such as equal parts or 40 per cent to the first and last touch. Data-driven attribution uses an algorithm and the converting and non-converting paths for that advertiser and conversion action. Its credit can be fractional, and it cannot repair inaccurate conversion values or a missing business outcome.

Can marketing attribution show that a channel caused a sale?

No, every attribution model only allocates credit for sales that already happened. None of them establishes that the credited channel caused the purchase, or that the sale would have been lost without another channel. When the real question is whether more spending would produce more sales, you need different evidence than an attribution report.