A customer clicks an ad, returns through a search result, and buys after an email. The advertising platform and the email report can each associate that purchase with their own activity. The business still has one order.
Reconciling those reports against the order record settles how much revenue arrived. Attribution modeling then divides credit among the interactions the business can observe. A last-click rule gives the final eligible interaction the credit. A multi-touch model shares it across the recorded path.
Neither calculation, by itself, establishes how much revenue the business would have lost without the advertising. The customer might have bought anyway. Or an earlier ad might have introduced the product, with the final click doing little more than taking the customer back to it. The recorded path is compatible with both explanations.
That distinction changes a budget decision. A channel can receive credit for a sale without having created the demand behind it. Cutting a channel with little recorded credit can also remove something that helped the customer decide.
I start with revenue reconciliation because competing claims about the same order are already enough of a problem. Then I separate the next questions: which interactions preceded the purchase, what changed because a campaign ran, and what a different channel budget might produce. Those questions need different evidence. A more elaborate attribution model only helps when its answer matches the decision being made.
Last-click can reward the channel that collected existing demand
A customer types a company name into a search engine and clicks an ad. Last-click can credit that ad for the purchase, even though the customer already knew which company to look for. A referral, an earlier campaign or an existing relationship might explain why.
The search ad could still help. It might make the right product easier to find or change which seller gets the order. The attribution report can't settle that question from the final click alone. Reported return on ad spend, or ROAS, divides attributed revenue by ad spend. A high reported return leaves the causal question open.
Last-click gives earlier influences no credit
Consider a business buyer who hears a podcast mention, attends a webinar, consults a colleague, and eventually clicks a search ad. Last-click assigns the eligible conversion to that final interaction. Earlier influences receive no share under that rule, including influences the tracking system never recorded.
A budget based only on that view can favor the closing channel. The danger appears when the business cuts discovery activity because its direct conversion credit looks weak, then expects the same volume of interested buyers to keep arriving. The report hasn't demonstrated that those buyers will continue to appear.
Demand can rise while a campaign is running
A holiday email can coincide with an industry event that makes the product more attractive. Sales rise, customers click the email, and the email receives credit. Some of the increase may have happened without it.
Last-click is useful for describing the recorded closing interaction. That narrow job can matter when checking a checkout path or an email response. The trouble begins when the same number becomes evidence for how much business an extra advertising dollar will create. Sharing the credit across more touches gives a broader description, but the operator still has to justify the spending decision.
Multi-touch models change how the recorded credit is shared
Multi-touch attribution, usually shortened to MTA, allocates a conversion across several recorded interactions. The allocation can follow a fixed rule or a model fitted to conversion paths. Choosing among them means choosing which assumptions the reporting will use.
The common approaches answer different versions of the credit question:
| Approach | How it distributes credit | What the operator has assumed |
|---|---|---|
| Linear | Gives each eligible touch an equal share | Recorded touches deserve equal weight |
| Time-decay | Gives more weight to recent touches | Recency is a useful guide to contribution |
| Position-based | Gives specified shares to the first and last touches, then divides the rest | Discovery and closing deserve explicit priority |
| Data-driven | Estimates weights from patterns in the available data | Those patterns contain information useful for assigning credit |
A warehouse calculation can make these assumptions visible. Comparing a recency-weighted view with a position-based view shows which channels gain or lose credit when the rule changes. Large differences deserve investigation before a budget follows either result. The comparison has changed the arithmetic while holding the observed purchases fixed.
A modeled channel removal doesn't withhold advertising from real customers
A Markov approach represents transitions between recorded states in a buying path. Its removal calculation estimates how modeled conversion probability changes when a channel is removed. A Shapley approach allocates credit using a channel's modeled marginal contribution across combinations of channels. Both can expose patterns a fixed split would hide.
The calculation's sophistication doesn't remove missing interactions, selection effects or an incomplete customer identity. Ad viewers can differ from people who don't see it, including in their intention to buy. Removing a channel from a fitted model doesn't perform the experiment of removing it from the market.
Data volume matters, but there isn't a conversion-count threshold that guarantees a useful model for every business. Path diversity, tracking coverage and the stability of the underlying behavior matter too. Use a complex model when its assumptions and outputs can be examined well enough to improve the decision.
Platform controls and warehouse rules are different things
A rule that can be written in a warehouse may be unavailable in a platform's interface. Google Analytics removed first-click, linear, time-decay and position-based models in November 2023. Advice to switch GA4 to time-decay sends an operator looking for a setting that no longer exists.
Check the models a platform supports, and label custom reporting separately. A changed attribution setting alters the credit calculation. Evidence about what happens without the campaign requires a comparison outside that calculation.
Incrementality tests compare outcomes with and without a campaign
An incrementality test estimates the additional outcome caused by a specified advertising intervention. It needs a counterfactual: an estimate of what would have happened to the treated audience without that intervention. A comparison group provides evidence for that estimate.
Random assignment helps make treatment and control groups comparable before exposure. The analysis then compares outcomes on the same basis, such as conversion rate or revenue per eligible customer. Comparing raw revenue totals from differently sized groups would confuse audience size with advertising effect.
The outcome also has to match the decision. A study of ad recall or brand awareness can inform an awareness objective. A revenue allocation needs evidence about purchases or revenue. Calling both studies "lift" doesn't make their outcomes interchangeable.
Geography can define the comparison
A geo experiment changes advertising in some regions while retaining a comparison elsewhere. Google's geo-experiment research describes randomly assigning non-overlapping regions to treatment and control. Historical demand helps assess whether the regions make a useful comparison.
The design still needs scrutiny. A local promotion, stock shortage or different seasonal pattern can affect one region more than another. Customers can cross regional boundaries, and advertising can reach people outside the intended area. A matched-region study without random assignment needs a stronger argument that other differences aren't explaining the result.
Fertilizing a plant in sunlight and comparing it with an unfertilized plant in a dark corner mixes two changes. The measured difference can't isolate the fertilizer. An advertising comparison needs equivalent care about what differs between groups.
Audience holdouts require care about who would have seen the ad
Platform-run experiments can assign eligible users to different conditions. A control group receives no campaign ad, or a different ad, depending on the design. The operator needs to know how assignment, delivery and outcome measurement work before interpreting the reported difference.
Ghost-ad methods address a particular delivery problem. The original methodology identifies control-group counterparts of exposed consumers: people who would have received the advertiser's ad had they been in treatment. A ghost ad is a measurement device for that comparison. It doesn't mean every control user sees a placeholder.
A public-service-announcement test uses another creative as a control. That introduces its own question: does the alternative ad provide a suitable baseline, and does delivery select comparable people? A PSA test and a ghost-ad design aren't interchangeable labels for any platform lift study. Check the actual experiment offered and its eligibility conditions.
An inconclusive result can leave the spending question open
Test size and duration depend on conversion frequency, variation in outcomes, the effect worth detecting and the purchase cycle. A small holdout can produce an estimate too uncertain to support the proposed decision. Failure to detect lift then leaves open both a commercially useful effect and an effect too small to justify the spend.
Specify the outcome, assignment, analysis window and decision rule before looking at the result. Inspect uncertainty alongside the estimate. Apply the finding to the intervention tested. A result from one audience, spend level and period needs further justification before it becomes a rule for the whole channel.
Mix modeling uses aggregate history to inform channel budgets
A campaign experiment answers a bounded question. A business allocating money across channels also needs an estimate of how revenue might respond to a different spending mix. Marketing mix modeling, or MMM, examines aggregate outcomes and channel activity over time to help answer that question.
An MMM relates revenue or another business outcome to media activity, with controls for influences such as trend and seasonality. It can also model delayed advertising effects and saturation. Saturation means an additional unit of spend produces less additional response as spending rises. A channel's historical average return can therefore be a poor guide to the return on its next dollar.
History needs enough variation to distinguish the channels
A model has difficulty separating channels whose spending always rises and falls together. A promotion that coincides with higher advertising spend creates another attribution problem unless the analysis accounts for it. Good predictive fit alone doesn't establish that the channel-effect estimates are right.
Useful history needs variation in spend, relevant seasonal coverage and consistent outcome definitions. A fixed minimum number of months can't guarantee those conditions. A new channel with little history gives the model little evidence about its response curve, and a changed market can make older relationships less useful.
Experimental results can help challenge or calibrate modeled channel effects where their scope matches. Budget scenarios should carry uncertainty and stay within a range the available evidence can support. A fitted curve beyond the spending levels the business has observed is a stronger extrapolation.
Open-source packages still need an analyst's judgment
Meta's Robyn uses ridge regression and other techniques to estimate media response. The repository provides R and Python options and identifies the Python version as beta.
LightweightMMM took a Bayesian approach. Its repository now says it's unsupported and directs users to Meridian. Choose a supported package for a new build.
Open-source software can make the modeling available without buying a specialty platform. It still leaves the business responsible for data preparation, assumptions, diagnostics and interpretation. The output earns a place in a budget discussion when someone can explain why the model assigned the effects it did and what evidence would make that allocation change.
The measurement methods need a shared account of business outcomes
I wouldn't begin by buying an attribution platform. I'd begin by making sure the business can follow its recorded conversions through to the outcome it cares about. A lead-generation business needs to distinguish a form submission from a qualified lead and a completed sale. An ecommerce business needs an order record that can account for refunds and cancellations.
Reconciliation gives the reports a common basis. Match events to the appropriate order or customer identifier where the business has permission to do so. Check time zones, conversion windows and revenue definitions. A platform can count an assisted conversion inside its own window while another platform counts the same purchase. Adding their attributed revenue can exceed the business's sales.
A normalized warehouse can retain those separate claims without counting the order twice. It also makes the rules inspectable: what counted as a conversion, which interactions were eligible, and how the model distributed the credit.
Use each method at the level its evidence supports
An operator can combine frequent reporting with less frequent experiments and budget planning. Each method has a different job:
| Decision | Evidence to use | Limit to retain |
|---|---|---|
| Investigate recorded channel and campaign performance | Reconciled outcomes and path attribution | Credit depends on the observation window and allocation rule |
| Estimate the effect of a specified campaign change | A suitably designed incrementality test | The estimate applies to the tested conditions |
| Explore a different channel budget | Aggregate mix modeling, checked against other evidence | Response curves depend on history and model assumptions |
A quarterly experiment can be a planning habit, but conversion volume and the decision at stake should determine whether the proposed test is informative. A platform's conversion-lift study applies to its supported intervention. It can't automatically test whichever unrelated channel happens to have the largest budget.
A test can also challenge the attribution report without supplying a universal replacement coefficient. A result about one retargeting audience doesn't establish how much of every paid-social conversion was incremental. Keep the adjustment tied to the audience, period and spend conditions studied.
Better conversion capture improves observation
Browser-side tracking can miss events. Server-side ingestion and platform matching can recover some measurement, subject to the system's design and the data the business is permitted to use. Google's enhanced conversions supplement measurement with hashed first-party conversion data.
Check consent, matching and deduplication before treating a larger event count as an improvement. Sending the same purchase through two routes can create a new counting error. Better capture also leaves the causal question intact: a recorded conversion can still have happened without the ad.
The case studies document changes without isolating their revenue effects
The measurement discipline has to apply to case studies too. A business can grow after a change without that growth isolating the change's contribution. My audit, solar and metal-art records show why examining the data can change a decision. They don't document a controlled incrementality estimate for each engagement.
Forensic audits found several reasons to change the spend
The audit work started with account exports, segmented by campaign objective, audience, device and funnel stage. I reconciled performance back to revenue and examined structural issues, misattribution traps and creative gaps. The output was a prioritized action plan.
The case record reports that, in many accounts, 10–30% of spend could be reallocated or cut without hurting revenue. That finding included several kinds of waste. It would overstate the record to attribute the whole opportunity to last-click, or to promise another business the same percentage. Check the account structure and revenue connection before deciding that the latest creative is the problem.
The solar portfolio made comparison across accounts possible
The solar consulting engagement involved an agency with dozens of separate customers. Each account produced its own performance data, while the agency needed to see which clients were doing well and which needed attention. I helped assemble the data infrastructure and built a dashboard that brought the accounts together.
The case record describes tracking offers, creatives, audiences and funnel structures across contexts, then using the findings in playbooks. A shared view can reveal differences an isolated account report misses. A pattern across accounts is still an observational finding. The work described here doesn't establish a controlled estimate of portfolio-level lift.
The metal-art work changed what we tested and whom we addressed
The metal-art account needed a more useful view of its buyers. I exported Facebook advertising data to CSV and worked in Google Sheets, using macros and calculations to organize it. I grouped creative and headline performance into cohorts and examined demographic and device patterns.
Women aged 45 and above were a strong group in that analysis. I also saw stronger reported performance among iPhone users. We tailored advertising to the groups and separated campaigns by age and gender. The messaging emphasized family and personalized gifts rather than copying a competitor's personality-led approach.
The case record reports growth from roughly $10K to $150K in monthly revenue over roughly 90 days. Audience targeting, creative and campaign structure changed together. The revenue result doesn't identify how much each change caused, or how much additional revenue came from advertising.
Examine the cohorts in the actual account, use the differences to decide what to test, and keep the reported business result separate from the causal effect the evidence can establish.
Start with a budget decision and specify what would change it
A measurement project needs a decision the business can act on. Pick a channel or campaign whose budget is under consideration, and write down what the available evidence would need to show for spending to increase, decrease or stay unchanged. Revenue reconciliation and experiment design then have a job to do.
The steps depend on one another:
- Define the business outcome. Specify whether the decision concerns orders, net revenue, qualified leads or completed sales. Use that outcome consistently in the comparison.
- Reconcile the reported conversions. Match platform claims against business records where permitted, remove duplicate events, and document differences in windows and definitions. Keep unknowns visible.
- Identify the unanswered question. A dispute over conversion credit may need a reporting correction. A question about the effect of withholding spend needs a counterfactual. A channel-budget scenario may need aggregate modeling as well as experimental evidence.
- Choose a feasible comparison. Check assignment, contamination, outcome volume and the effect worth detecting before committing to a test. A study unable to distinguish the alternatives relevant to the decision can't resolve that decision.
- Record the result with its uncertainty, then act within its scope. An inconclusive result can justify gathering more evidence or limiting exposure. A useful effect estimate can justify a bounded budget change with follow-up measurement.
A software purchase comes after those requirements are clear. Buy a platform when its data connections, operating workflow or team coordination solve a defined constraint. Building a custom model carries maintenance and analyst costs too. Compare both options against the work the business needs performed. A more expensive attribution calculation still needs an appropriate basis for causal claims.
The first useful output can be modest: a reconciled revenue view, a named measurement gap, and a decision about the next comparison. That's enough to stop two channel reports from being treated as two independent sales, while preserving the question neither report could answer.
Return to the customer who bought after the email. Count the order once. Use the recorded interactions to understand the path. Before deciding how much more to spend, examine evidence of how the outcome changes when the advertising changes.