Marketing attribution models assign credit for a recorded outcome to the marketing interactions that qualify under a set of rules. They help you interpret a report and choose what to investigate. They do not reconstruct every influence on a customer, establish what would have happened without marketing, or tell you exactly where to spend the next dollar.
The useful starting point is a decision: explain a reporting change, examine how channels participate in recorded journeys, or support campaign optimisation. Choose a model whose assumptions fit that decision, then document the outcome, eligible interactions, time window and missing evidence. Use a second model to test whether the conclusion depends on the credit rule.
This guide is for Canadian business owners, marketing managers and analysts who need to explain channel performance without overstating it. It includes a model-selection worksheet and identical fictional journeys calculated under six transparent rules. Every business scenario and numerical example below is fictional. Dollar amounts are Canadian dollars unless stated otherwise.
Start with the decision you need to make
“Which attribution model is best?” is too broad to produce a useful answer. A finance review, an advertising bid strategy and a content planning meeting need different evidence. Start by writing the decision in one sentence, including who will act on it and what could change.
For example, a fictional team might ask: “Should we investigate the role of non-brand search before reducing its budget next quarter?” That is a defensible question for an attribution comparison. “How much revenue did non-brand search cause?” requires a different design. The first asks how credit changes across observed paths; the second asks about a counterfactual outcome.
| Decision | Useful evidence | Limit to state |
|---|---|---|
| Explain a channel credit shift | The same outcomes under two models, with identical filters | Reallocated credit does not mean sales changed |
| Find early interactions worth investigating | Observed paths and a first-touch sensitivity view | First recorded contact may follow earlier unseen influences |
| Support advertising optimisation | A supported platform model and reliable conversion definitions | Platform credit depends on eligible channels and available data |
| Reconcile marketing and sales reports | Outcome records, identifiers, dates and a documented bridge | A different model cannot repair missing or duplicated outcomes |
| Estimate additional sales from advertising | A suitable incrementality study | Attribution alone cannot answer the causal question |
We plan and run B2B marketing across search, ads, content and email for Canadian companies.
Next, identify the consequence of being wrong. A small content investigation can proceed with incomplete evidence if the uncertainty is visible. A large budget cut needs stronger support. Avoid giving an exploratory report the authority of a tested investment recommendation.
Record the available alternatives as well. A team choosing between continuing a campaign, reducing it or improving measurement needs to compare those actions. It does not need to settle an abstract debate about the fairest allocation of every sale.
Choose one outcome before comparing models
Separate a form submission, a qualified enquiry, an opportunity and a completed sale. Each can be useful, but they answer different questions. A channel credited with many submissions may have few qualified opportunities. That could reflect audience fit, duplicate records, slow follow-up or the qualification definition. Changing attribution cannot decide which explanation is correct.
For a sale-based analysis, specify whether value means order value, revenue after refunds, recognised revenue or contribution after defined costs. For a lead-based analysis, label any assigned value as an assumption. A fictional $200 lead value is not $200 collected from a customer.
If the broader strategy or buying process is still unclear, establish that context first with the B2B marketing operating framework. The attribution exercise then becomes one part of a defined management decision.
Write a measurement contract that a colleague can reproduce
A model name is only one part of the specification. Two reports can both say “last click” while allocating credit across different interactions, periods and outcomes. Write down the full measurement contract before exporting numbers.
- Outcome and counting unit: the exact event or business stage, counted by order, person, enquiry or opportunity.
- Population: the markets, products, customer types and reporting dates included.
- Eligible interactions: clicks, sessions, views or other recorded contacts, including the treatment of direct traffic.
- Identity and joins: how interactions are connected to an outcome, and where those connections are unavailable.
- Time rules: the lookback window, time zone, date basis and date the report was extracted.
- Value basis: currency, refunds, tax treatment and any estimated values.
- Allocation rule: model name, parameters and handling of short or empty paths.
These fields prevent several avoidable arguments. If one report counts every enquiry and another counts unique qualified opportunities, their totals should differ. If one includes existing customers and another only new customers, the channel mix may differ for a valid reason. Resolve the population before debating attribution.
Use a reporting label that travels with the chart. “Purchase credit, paid and organic channels, event date, CAD, extracted 8 October” communicates more than “marketing revenue”. Add the exact model and window in the chart note or adjacent table. A screenshot separated from its filters should still be interpretable.
Keep acquisition dimensions separate from outcome credit
In GA4, a first-user acquisition dimension, a session acquisition dimension and an event-scoped attribution dimension describe different things. Google states that changing the reporting attribution model affects relevant event-scoped reports and explorations; user- and session-scoped traffic dimensions are unaffected. It also says the reporting model change applies to historical and future data. See the GA4 reporting attribution model documentation.
That distinction matters when a colleague compares two exports. A first-user source is not automatically a first-touch allocation of the revenue in the other report. Label the dimension scope and metric together, then ask whether they describe the same population. Do not rename incompatible columns until they appear to agree.
Keep a dated copy of the reporting specification. Otherwise, a later change to definitions can make an old chart look inconsistent even when the original calculation was correct.
Understand the assumption inside each model
A rules-based model makes an allocation choice explicit. That transparency is useful, even when the choice is deliberately simplistic. The six rules below form this guide’s teaching calculator. They are conceptual comparison tools, not a list of current settings in Google products.
First touch, last touch and last non-direct touch
First touch assigns all credit to the earliest eligible recorded interaction. It can help you examine which sources begin the paths you can observe. Its central weakness is the word “recorded”: the person may already know the business, have received a recommendation or have interacted outside the available window.
Last touch assigns all credit to the final eligible recorded interaction. In this guide’s calculator, it includes direct visits. It is easy to explain and useful for inspecting the end of a measured path. It gives no allocation to earlier interactions, even when those interactions were important to the buyer.
Last non-direct touch assigns all credit to the final recorded interaction that is not direct. Our rule falls back to the last direct interaction when the entire path is direct. It prevents a returning direct visit from taking the whole allocation, but it still concentrates credit on a single point. Direct-only does not mean marketing had no influence.
Linear and position-based models
Linear attribution divides credit equally among eligible interactions. In our calculator, three interactions receive one-third each. Two email interactions count twice: we allocate by touchpoint first, then aggregate by channel. Equal shares per channel would be a different model.
The assumption is not that all interactions had equal persuasive power. It is that equal allocation is the comparison rule. A repeated contact can receive more total channel credit simply because it appears more often. A duplicate tracking event could therefore change the result unless it is removed before modelling.
Position-based attribution gives special weight to the first and last eligible interactions. Our version assigns 40% to each endpoint and divides the remaining 20% across the middle. For two interactions, it uses 50% each; for one, 100%. These exceptions are necessary because a two-touch path has no middle to receive 20%.
The endpoint emphasis is a declared preference, not a discovered fact about buying behaviour. Use it to ask whether a conclusion depends on valuing discovery and completion more than the interactions between them.
Time decay
Time-decay attribution gives more weight to interactions closer to the outcome. Our version uses a seven-day half-life: an interaction seven days before conversion has half the raw weight of one at conversion, and an interaction fourteen days before has one-quarter.
The formula is raw weight = 2 ^ (−days before conversion / 7). Divide each raw weight by the total raw weight for that journey to obtain its credit share. Multiplying that share by the outcome value gives attributed value.
Seven days is a teaching parameter, not a recommendation for every business. A different half-life changes the allocation. Recency can be relevant to a decision, but an older interaction does not become less persuasive merely because a spreadsheet assigns it less weight.
Data-driven attribution
Data-driven attribution uses a platform’s model and available data instead of fixed shares chosen by the analyst. Google Ads describes comparing patterns among converting and non-converting paths, with a model specific to the advertiser. It also notes that data-driven and last-click results can sometimes be the same. See Google Ads’ explanation of data-driven attribution.
You cannot calculate Google’s proprietary allocations from four sample paths and a set of invented percentages. The companion therefore contains no “simulated Google DDA” column. If you later add platform results, keep them as a separately labelled export with their source, settings and extraction date.
Separate current Google options from legacy concepts
The following product information was checked against public official documentation on 8 October 2026. Model availability, reporting scope and account configuration are separate questions; verify all three before changing a live setting.
| Context | Documented choices | What to check |
|---|---|---|
| GA4 attribution reports | Data-driven; paid and organic last click; Google paid channels last click | Event scope, eligible channels and direct-traffic handling |
| Google Ads conversion actions | Data-driven and last click | The action’s creditable channels and use in reporting or bidding |
| This guide’s companion | Six explicitly defined rules | Educational arithmetic; no account connection or platform replication |
Google’s GA4 attribution overview lists the three reporting models and says direct visits are excluded from credit unless the path consists entirely of direct visits. That differs from our deliberately literal last-touch example, which credits the final direct visit.
Under GA4’s Google paid channels last-click option, credit goes to the last eligible Google Ads click; without a Google Ads click, it falls back to paid and organic last click. This is not interchangeable with “give credit to the last recorded visit”. Check the current definitions of the reporting choices when interpreting an export.
First-click, linear, time-decay and position-based attribution are no longer available in GA4, according to its key event attribution models report documentation. Google Ads also identifies those four models as retired and documents data-driven and last-click choices in its attribution model guidance. An older tutorial showing six selectable Google models is not a current implementation guide.
Google Ads currently says all conversion actions are eligible for data-driven attribution regardless of volume, while recommending at least 200 conversions and 2,000 supported ad interactions within 30 days for stronger modelling. Those are data-volume recommendations, not a universal business sample-size rule or a reason to invent extra conversion events.
Google Ads also documents that a conversion action’s attribution setting can affect bidding strategies using that conversion data. Treat a live model change as an operational change: record the prior setting, identify affected goals and bids, and define what you will review. Comparing exported reports can come first.
Compare the same fictional journeys under different rules
The following dataset contains four fictional completed orders worth $3,000 in total. Each order counts once. Values exclude tax and refunds solely to keep the illustration simple. Every listed interaction is eligible; the window is 30 days, the time zone is UTC, and no view-through interactions or modelled missing events are included.
| Journey | Recorded interactions, earliest to latest | Order value |
|---|---|---|
| J01 | Organic search at day −21 → paid search, non-brand at day −14 → email at day −7 → direct at day 0 | CAD 1,200 |
| J02 | Paid social at day −7 → paid search, brand at day 0 | CAD 600 |
| J03 | Direct at day 0 | CAD 300 |
| J04 | Email at day −14 → email at day −7 → paid search, brand at day 0 | CAD 900 |
The labels “brand” and “non-brand” are custom analysis categories in this fictional dataset. They are not claims about a platform’s automatic channel classifications. Direct is also a recorded category here, not a diagnosis of how the buyer originally discovered the business.
One order, six allocations
For J01, the model changes the allocation while the order remains $1,200. The table shows attributed value rather than conversion-count credit; divide any amount by $1,200 to obtain its share of the single conversion.
| Teaching model | Organic search | Non-brand paid search | Direct | |
|---|---|---|---|---|
| First touch | 1,200 | 0 | 0 | 0 |
| Last touch, including direct | 0 | 0 | 0 | 1,200 |
| Last non-direct | 0 | 0 | 1,200 | 0 |
| Linear by touchpoint | 300 | 300 | 300 | 300 |
| Position-based 40/20/40 | 480 | 120 | 120 | 480 |
| Time decay, seven-day half-life | 80 | 160 | 320 | 640 |
The time-decay calculation uses raw weights of 1/8, 1/4, 1/2 and 1. Their total is 15/8, so the normalized shares are 1/15, 2/15, 4/15 and 8/15. Multiplying by $1,200 produces $80, $160, $320 and $640. No additional revenue is created by changing those weights.
Notice the interpretation shift. First touch directs attention to organic discovery. Last non-direct highlights email. Literal last touch points to a direct return. Linear keeps all four interactions visible. These are alternative descriptions of the same recorded order, not six independent pieces of evidence that a channel caused a sale.
The same issue appears in the four-order total
| Channel | Last non-direct | Linear by touchpoint | Linear minus last non-direct |
|---|---|---|---|
| Organic search | 0 | 300 | +300 |
| Paid search, non-brand | 0 | 300 | +300 |
| 1,200 | 900 | −300 | |
| Direct | 300 | 600 | +300 |
| Paid social | 0 | 300 | +300 |
| Paid search, brand | 1,500 | 600 | −900 |
| Total | 3,000 | 3,000 | 0 |
In J04, linear attribution assigns $300 to each of three interactions. Because two interactions are email, email receives $600 for that journey. Combined with J01’s $300 email allocation, that gives $900. Aggregating the two email contacts into one before applying the model would produce a different result.
Across all four orders, the brand-search allocation falls from $1,500 to $600, a 60% decrease in attributed value. The business has not lost 60% of its sales. Its four orders and $3,000 total are unchanged. When a channel moves from zero credit to $300, report the dollar difference; a percentage increase from a zero base is undefined.
The companion includes touchpoint results and channel totals for all six models. Credit fractions sum to one per journey, and each model preserves four conversions and $3,000 before rounding. Displayed decimals can create tiny apparent differences, so sum full-precision values before formatting currency.
Use model disagreement to choose the next investigation
A model comparison is most useful when it changes the question you ask next. Do not select whichever column makes a preferred channel look strongest. Choose the main reporting view in advance, then use an alternative to identify conclusions that are sensitive to the rule.
When the question concerns discovery
In the fictional portfolio, organic search receives $1,200 under first touch but zero under last non-direct. That gap suggests inspecting the early recorded interactions: what pages were visited, whether those interactions belong to the same eligible outcomes, and whether the segment contains new or returning customers.
It does not justify the sentence “organic search generated $1,200 that would otherwise disappear”. A first-touch rule guarantees credit to the earliest eligible interaction whether that interaction changed the purchase decision or merely recorded an existing intention.
A practical next step is to examine a defined set of relevant landing pages and downstream outcomes. Keep the selection method visible so the team does not choose only attractive paths. Include records with short paths, unknown sources and unsuccessful enquiries where the available evidence permits meaningful comparison.
When the question concerns completion
Brand search receives the largest last-non-direct allocation in the fictional portfolio. That can help explain where recorded journeys end. It does not establish that brand advertising created the underlying demand, nor that removing it would leave the same outcomes.
Before changing spend, examine the actual decision: maintaining coverage, changing a bid target or testing additional investment. The relevant evidence might include conversion definitions, eligible channel coverage, commercial outcomes and a feasible experiment. The model supplies a reporting perspective; it does not supply the missing comparison with an unexposed group.
When data are sparse or unstable
For a small team with few completed outcomes, elaborate attribution can produce detailed-looking numbers without resolving the decision. Prefer a clearly defined baseline, a limited sensitivity comparison and a visible uncertainty note. Track the actual count of outcomes alongside attributed fractions.
If one large fictional contract determines a channel’s ranking, show the result with and without that contract as a sensitivity analysis. Label the exclusion and retain the full result. This reveals concentration; it is not permission to remove inconvenient data from the official report.
For SaaS, distinguish signup, activation, paid conversion and renewal before interpreting channel credit. The SaaS marketing playbook provides the broader growth-motion context. An attribution allocation for signups should not silently become a claim about retained recurring revenue.
Make data gaps visible before adding sophistication
A calculation can be internally correct and still describe an incomplete picture. List the missing evidence beside the model, including whether the gap affects the outcome total, the path, the channel label or the business interpretation. Those are different problems with different repairs.
| Gap | Possible distortion | Useful response |
|---|---|---|
| An outcome has no linked path | Observed journeys represent only part of the business total | Keep an explicit unmatched bucket and investigate the join |
| Campaign labels are missing or inconsistent | Credit lands in an unknown or incorrect category | Retain raw labels and document the mapping repair |
| Duplicate interactions or outcomes | Extra touches gain shares, or one sale is counted twice | Apply documented deduplication before allocation |
| Offline influence is not recorded | Digital interactions appear to explain the whole decision | State the missing influence without inventing a touchpoint |
| Identity is fragmented | One person’s activity appears as separate paths | Document available identity coverage and permitted joins |
| Recent outcomes are still developing | Early comparisons mix mature and incomplete cohorts | Use a defined maturity rule and retain extraction dates |
Do not fill an unknown source with “direct” merely to make a table balance. In the companion, a record with no eligible interactions stays unallocated. That differs from a record explicitly containing a direct interaction. Preserving this distinction makes a reconciliation possible.
Keep that teaching convention separate from platform labels. GA4’s attribution-report definitions say Direct can also appear when no path data are available, such as for data import. Preserve the platform’s original value and explain its meaning rather than assuming every Direct row proves a recorded direct visit.
Modelled platform results also need a clear label. Google explains that GA4 can estimate key events that cannot be directly observed and combine modelled and observed information in reporting. Its documentation on modelled key events also says channel attribution data can update for up to 12 days after a conversion is recorded. A changing recent total is therefore not automatically evidence of a broken calculation.
The distinction between observed and modelled does not mean one is always useful and the other useless. It changes the kind of statement you can make. An aggregated estimate is not a recovered record of a particular person’s exact journey.
Work only with information the business is entitled to use, and minimise identifying details in analysis exports. The worksheet needs evidence references and definitions, not customer names or email addresses. More collection is not the automatic solution to every measurement gap.
Reconcile the totals before interpreting the channel split
Reconciliation asks whether you understand the relationship between two reports. It does not require every system to produce the same number. Begin with the system that defines the chosen business outcome, then bridge to the measurement report using explicit, non-overlapping adjustments.
A practical reconciliation sequence
- Match the outcome. Compare purchases with purchases, or qualified opportunities with the same qualification stage.
- Match the population. Align geography, products, customer type, included channels and excluded records.
- Match the time basis. Record the time zone, outcome date versus interaction date, extraction date and maturity cut-off.
- Match counting and value. Check duplicates, refunds, currency conversion and whether values are actual or estimated.
- Identify coverage differences. Record unmatched outcomes, unavailable dimensions and modelled components.
- Compare the allocation. Only now examine model-driven credit shifts within the comparable population.
Time basis deserves special attention. GA4’s model-comparison documentation distinguishes event time from ad interaction time. Google Ads likewise documents reporting-time and coverage differences between attribution reports and standard campaign reports. Consult the GA4 model-comparison report controls and Google Ads attribution reporting boundaries instead of assuming identical date selectors create identical populations.
Also distinguish the lookback rule from the displayed reporting period. A monthly outcome report can allocate credit to interactions from the prior month. Google’s key event lookback-window guidance explains that the window determines how far back interactions remain eligible and that changes apply going forward. Record the actual setting rather than assuming an old export used today’s window.
A fictional bridge that leaves uncertainty intact
Suppose a fictional order ledger contains 120 completed orders. The comparison excludes eight fully refunded orders and seven orders outside the selected market. Those groups are defined to be mutually exclusive. That leaves 105 eligible net orders. Ninety-seven have an accepted path match and eight remain unmatched: 120 − 8 − 7 = 105, then 105 = 97 + 8.
The path-match coverage is 97 ÷ 105, or approximately 92.38%. This is a coverage measure for that fictional reconciliation. It is not model accuracy, consent coverage or the percentage of sales caused by marketing.
If the attribution export contains 99 matched purchases while the bridge expects 97, record an unexplained difference of two. Investigate event duplication, identity joins, date boundaries and inclusion rules. Do not subtract two from a channel to force agreement. An unresolved difference is a finding, not an embarrassment to hide.
Document what each adjustment means and retain its evidence reference. If one adjustment can overlap another, calculate the union of excluded records or apply a mutually exclusive priority rule. Subtracting the same order twice creates a neat-looking but incorrect bridge.
Never sum the credited conversions from several ad platforms and call the result unique business sales. Different systems may credit the same outcome under their own rules. Deduplicate against the chosen business counting unit where the available identifiers support it; otherwise describe the overlap as unresolved.
Use incrementality evidence for causal budget questions
Attribution asks how to distribute credit within a measurement system. Incrementality asks how outcomes differ because an activity took place. These questions can support each other, but one is not a substitute for the other.
A fictional retargeting campaign can receive substantial credit from people who were already close to buying. Another fictional campaign can influence a buyer early and receive little closing credit. Neither pattern proves whether the campaign created additional sales. The missing evidence is what comparable people would have done without the activity.
Google’s Conversion Lift documentation describes a controlled comparison between treatment and control groups to measure incremental outcomes. It also states that Conversion Lift is not available for every Google Ads account. Access and feasibility must be checked separately; a guide cannot promise that every business can run the same study.
If the decision warrants an experiment, define the eligible population, outcome, treatment, comparison group, duration and analysis plan before examining results. Consider whether volume, spillover between groups or simultaneous changes could make the comparison difficult to interpret. A qualified analyst should help design material investment tests.
Where a credible experiment is not feasible, make a narrower claim. You can report observed outcomes, attribution sensitivity and commercial constraints while stating that incrementality is unresolved. A before-and-after improvement can be informative, but seasonality, pricing, competitor activity or sales capacity could also have changed.
A model disagreement can identify a test candidate. In the fictional portfolio, the large brand-search credit shift suggests a question about demand capture and additional value. It does not answer that question. The next action should be proportionate to the evidence and the cost of a mistaken decision.
Use the model-selection worksheet and journey files
Download the attribution model-selection worksheets and fictional journey calculations (ZIP, 13 files)
The companion pack turns the discussion into a repeatable review. It contains blank and worked model-selection worksheets, fictional input journeys, calculated credit by touchpoint, channel totals, a reconciliation template and calculation code. The worked entries are examples to replace, not recommended settings for every business.
Start with the blank selection worksheet. Complete the business decision, outcome definition, population, available inputs and missing evidence. Then name the primary reporting model and the comparison model. Explain what each one emphasises, what it omits and which decision it can support.
Use the evidence-reference column for the report name, saved export or internal specification that supports an entry. Use the decision-or-limitation column to explain the consequence. “Source labels incomplete” is a finding; “defer the channel budget ranking until labels are reconciled” is its decision consequence. Assign an owner and a review date so the limitation does not disappear in the next presentation.
Next, inspect J01 in the fictional journey file and compare it with the article’s allocation table. Inspect J04 to see how repeated channels accumulate credit. Change one assumption at a time when exploring alternatives: the model, window, direct-traffic rule or path definition. Changing several at once makes the source of a credit shift unclear.
The CSV files open in spreadsheet software without an account connection. Their calculated result files are static values; they do not recalculate when you edit an input CSV. The included instructions explain how to use the supplied JavaScript calculator with the JSON input file, or reproduce the formulas in a spreadsheet. Keep original examples separate from your own working data.
Arithmetic checks were run on the synthetic calculation code, including total conservation, repeated channels, single-touch and two-touch paths, direct-only paths and missing paths. These checks validate the stated teaching rules. They do not validate a live GA4 property, Google Ads account, CRM integration or spreadsheet application.
Present a decision with its limits attached
A useful attribution report can be brief. State the business outcome, comparable population and reporting contract. Show the main allocation, the sensitivity comparison and the reconciliation status. Finish with the supported action and the evidence still needed.
A fictional decision note could read: “Across four synthetic orders worth CAD 3,000, brand paid search receives CAD 1,500 under last non-direct and CAD 600 under linear attribution. Total value is unchanged. The allocation is sensitive to the rule, so this comparison supports further investigation of brand search’s role; it does not establish incremental revenue or justify a budget cut by itself.”
For real reporting, add the evidence owner and review date. If a model change affects an operational platform, keep its effective date separate from the date of the management decision. If the outcome definition changes, show that break in the series rather than presenting a seamless trend.
The aim is a measurement process that makes decisions easier to explain: one chosen outcome, visible assumptions, conserved totals, known gaps and a next step that the evidence can support.
To discuss the assumptions behind your measurement plan, contact Canada Create.
