| Course | GB794 Advanced Topics in Marketing |
|---|---|
| Unit | Unit 4 |
| Paper type | Attribution analysis |
| Length | About 1,100 words, 6 pages double-spaced |
| Format | APA 7 student paper |
| School | Purdue University Global |
| Program | MBA |
| Updated | October 2026 |
Free sample paper for GB794 Unit 4
Sixteen Thousand Claims, Twelve Thousand Orders: A Unit 4 Attribution Analysis for an Online Grocer
[Student Name]
Purdue University Global
GB794: Advanced Topics in Marketing
Unit 4 Attribution Analysis
[Instructor Name]
[Date]
Thornbury Market, its campaigns and every figure are composite and illustrative; the methods studies cited are real.
Exhibit 1: Reconciling the Claims
Thornbury's two largest advertising platforms each report the online orders they believe their ads produced. For the last quarter, Meta's reports credited its ads with 6,800 orders and Google's with 9,200, a total of 16,000. The grocer's order database recorded 12,400 online orders in the same quarter, from all sources, including customers who came directly to the site and customers who received only the weekly email.
The surplus of 3,600 orders, the amount by which platform claims exceed every order the grocer actually received, is the first finding. It arises because each platform applies its own attribution window and counts any order that follows an ad within that window. Meta's default counts an order if the customer clicked an ad in the previous seven days or viewed one in the previous day. Google counts an order following a click within thirty days. A single order preceded by a Meta ad view on Monday and a Google search click on Wednesday is claimed by both. Neither platform sees the other's claim.
Exhibit 2: One Order, Four Accounts
Consider one composite order. A loyalty member watched an Instagram video ad for Thornbury's prepared meals on a Monday evening, without clicking. On Wednesday she searched for the grocer's name, clicked the paid search result at the top of the page, and that evening placed an order for 86 dollars.
Meta's account credits the order to the Instagram ad, because it was viewed within the attribution window.
Google's account credits the order to the search ad, because it was clicked within thirty days.
The site analytics tool's data-driven model divides credit between the two touchpoints and the direct visit that followed, assigning fractions by a statistical model of which paths tend to precede orders.
The marketing mix model, which estimates relationships between weekly spending by channel and weekly sales across the grocer's delivery zones, never sees this customer at all. It sees only that, in weeks when Instagram and search spending were higher, online sales tended to be higher, after accounting for seasonality, promotions and other factors.
None of these accounts asks the question that matters for spending decisions: would this customer have ordered if she had seen neither ad? She was a loyalty member who searched for the grocer by name. There is a good chance she would have.
Exhibit 3: What Each Method Can Answer
Platform attribution requires no extra data and is available immediately, but it answers only the question of which ads preceded an order. It cannot distinguish ads that caused orders from ads that customers saw on the way to orders they would have placed anyway, and each platform has an interest in claiming as much as possible.
Multi-touch attribution in the analytics tool combines touchpoints across platforms and avoids double counting within its own data. But it still distributes credit among touchpoints that preceded orders; it does not compare customers who saw ads with similar customers who did not.
Marketing mix modeling estimates the relationship between spending and sales from variation over time and across regions. It requires several years of weekly data and careful controls, and it can be confounded when spending rises in the same weeks that demand rises for other reasons, such as holidays. Its strength is that it does not depend on tracking individuals, which privacy changes have made less reliable, and it links marketing to sales outcomes in the way Hanssens and Pauwels (2016) argue firms need.
Experiments, such as geographic holdouts in which advertising is paused in randomly chosen regions, answer the causal question directly: what happens to sales when the ads are not there? They are costly, because they forgo advertising in some regions, and they require enough regions and time to detect an effect. Lewis and Rao (2015) showed that the sales swings of ordinary weeks dwarf the typical effect of a campaign, which is why experiments must be very large before they can measure advertising precisely.
What the Geographic Holdout Showed
Last spring, Thornbury paused Meta prospecting ads, those aimed at people who were not yet customers, in six of its twenty delivery zones for four weeks, chosen at random. Online orders in the paused zones fell, relative to the other fourteen, by an amount implying that Meta prospecting ads had produced roughly a third as many incremental orders as Meta's own reports had credited to them in those weeks.
Gordon et al. (2019) found a similar pattern on a larger scale. Comparing estimates from observational methods with the results of fifteen large randomized experiments at Facebook, they found that the observational methods frequently overstated the effects of advertising, sometimes by large margins, even when they used rich data on users.
The platforms reported what their ads touched; the holdout measured what the ads changed, and the second number was about a third of the first.
Why the Overstatement Is Not Fraud
It would be easy to read the surplus as dishonesty by the platforms, and that reading would be a mistake that leads to the wrong fix. Each platform's rule is disclosed, and each is a reasonable answer to a narrow question: did a customer who later ordered see or click our ad within a window? The rules overstate effects because many customers who see ads were going to order anyway, and because two platforms can each truthfully say their ad preceded the same order. Replacing one platform's reports with another's, or with a more elaborate attribution model, would not solve that problem. Only a comparison with customers or regions that did not see the ads can show what the ads changed.
Recommendations
Three changes follow. First, platform-reported conversions should no longer appear in the board report as revenue from advertising; they should be labeled as platform-attributed orders. Second, the marketing mix model should be recalibrated using the holdout result, so that its estimates for Meta prospecting are anchored to experimental evidence. Third, the grocer should run a holdout for branded search, its largest channel by claimed orders, because loyalty members who search for Thornbury by name are the customers most likely to have ordered anyway. A four-week pause in a random half of delivery zones would show how much of that channel's claimed revenue is incremental.
References
Gordon, B. R., Zettelmeyer, F., Bhargava, N., & Chapsky, D. (2019). A comparison of approaches to advertising measurement: Evidence from big field experiments at Facebook. Marketing Science, 38(2), 193-225. https://doi.org/10.1287/mksc.2018.1135
Hanssens, D. M., & Pauwels, K. H. (2016). Demonstrating the value of marketing. Journal of Marketing, 80(6), 173-190. https://doi.org/10.1509/jm.15.0417
Lewis, R. A., & Rao, J. M. (2015). The unfavorable economics of measuring the returns to advertising. The Quarterly Journal of Economics, 130(4), 1941-1973. https://doi.org/10.1093/qje/qjv023
What the GB794 Unit 4 assignment asks for
Unit 4 of GB794 asks you to analyze how your organization's marketing results are attributed and what each method can answer. Reconcile the platforms' claims against your own order records and explain any surplus. Follow one representative order through every system that credits it, stating what each account says. Compare the main method families on data required, question answered, main weakness and cost. Report any experimental evidence your organization has, describing the design and what it implied. Bring in published research that compares observational estimates with experiments. Explain why overstatement occurs without assuming bad faith, since the fix depends on the cause. Close with specific recommendations for reporting, for calibrating models and for the next experiment to run. About eight pages with three exhibits is a common scope for this attribution assignment.
How this GB794 Unit 4 example is built
Exhibit 1 sets 6,800 Meta claims and 9,200 Google claims against 12,400 recorded orders and traces the 3,600 surplus to each platform's click and view windows. Exhibit 2 follows one 86 dollar order through Meta, Google, the analytics tool's data-driven model and the mix model, and asks whether the loyal shopper would have ordered anyway. Exhibit 3 compares platform attribution, multi-touch attribution, mix modeling and experiments, citing Hanssens and Pauwels and Lewis and Rao. The geographic holdout result follows with Gordon and colleagues' evidence, and the highlighted sentence contrasts what the ads touched with what they changed. A section explains why the overstatement is not fraud, and three recommendations close the paper.
Where the points sit in the GB794 Unit 4 rubric
The GB794 attribution analysis is graded on whether it explains rather than complains. Instructors give credit for a reconciliation that quantifies the gap between claims and orders and traces its source. Following one order through every system is weighted because it shows concretely that each method answers a different question. A comparison of method families should be fair to each, naming strengths as well as weaknesses. Experimental evidence earns significant marks when its design is described and its implication stated. Published comparisons of observational and experimental estimates add authority. Explaining the cause of overstatement accurately shows judgment, since the remedy depends on it. Recommendations should be specific enough to adopt next quarter. Sources should be cited accurately in APA 7.
Common GB794 Unit 4 mistakes, and how to avoid them
Many GB794 attribution papers declare one method best without asking which question each answers. Others accuse platforms of inflating results without explaining the window rules that produce the surplus. Reconciliations against actual orders are often missing, though they are the simplest evidence of double counting. Students frequently describe experiments without their design or results. Recommendations tend to call for better attribution models, which cannot solve a causal problem. Reconcile claims with orders, follow one order through each system, compare methods by the question they answer, report experimental evidence with its design, explain why overstatement happens and recommend specific changes. Bring your platform reports and order totals, and we can work through the reconciliation with you.
Get a GB794 Unit 4 example written to your instructions
Pull the platform reports and order totals you are able to share, or describe the channels if the data are restricted, and add the Unit 4 prompt with its rubric. Reconciled against real orders, the analysis states each method's reach accurately and reports any experiment's uncertainty. It usually takes 24-48h, and a first sample carries no fee. The paper above is an original model document written by our desk, not a submitted student paper and not an official Purdue University Global document.
More GB794 papers and MBA samples
- GB794 Unit 1: Discussion Board Post
- GB794 Unit 2: Consumer Study Critique
- GB794 Unit 3: Experiment Design Memo
- GB794 Unit 5: Measurement Validity Review
- GB794 Unit 6: Seminar Reflection
- GB794 Unit 7: External Validity Memo
- GB794 Unit 8: Field Study Proposal
- GB794 Unit 9: Evidence Summary
- GB794 Unit 10: Marketing Research Paper
- GB793 Unit 4: Empirical Article Appraisal
- GB792 Unit 4: Asset Pricing Memo
- GB791 Unit 4: Measurement Comparison
- GB740 Unit 4: Organizational Assessment Plan
GB794 Unit 4 questions, answered
Is marketing mix modeling more reliable than multi-touch attribution?
Neither is reliable for every question. Mix models see all channels, offline ones included, but lean on historical variation in spend that may be tangled with seasonality and promotions. Attribution sees individual paths but not what would have happened without the ads. Experiments answer the counterfactual directly for the tested channel and period, which is why current practice uses them to calibrate the other two.
Why do platform-reported conversions exceed actual orders?
Each platform counts a conversion whenever its own rules connect it to an ad, using click and view windows it sets itself. When a shopper touches two platforms before buying, both can legitimately claim the same order. Summing platform reports therefore overstates the total, and reconciling them against the retailer's own order records makes the overlap visible.
Can the analysis rely on a platform's own lift study?
It can, provided the paper describes the design. Platform conversion lift studies randomize users into exposed and holdout groups, which gives them a genuine counterfactual, but the platform runs the test and reports the result. A doctoral paper notes that dependence, compares the estimate with independent evidence such as a geo holdout, and reports whatever uncertainty the study provides.
Where can I find a free GB794 Unit 4 sample paper?
A complete GB794 Unit 4 attribution analysis is on this page, reconciling platform-claimed orders with actual orders and comparing attribution, mix modeling and a geographic holdout.
What is marketing mix modeling?
It estimates how sales relate to spending by channel using weekly data over time and across regions, controlling for factors such as seasonality and promotions, without tracking individual customers.