Marketing Analytics
Attribution models and windows, ad auctions, and incremental-value promotion.
16 questions
JuniorTheoryVery commonWhat is an attribution model in marketing analytics?
What is an attribution model in marketing analytics?
An attribution model is a rule for distributing credit for a conversion across the marketing touchpoints that preceded it. Common rules are last-click, first-click, linear, time-decay, position-based, and data-driven (Shapley or Markov).
Common mistakes
- ✗Confusing the model with the attribution window (the time span)
- ✗Naming only last-click and ignoring multi-touch models
- ✗Treating attribution as raw spend reporting, not credit assignment
Follow-up questions
- →How does last-click differ from a position-based model in credit?
- →Why might a data-driven model beat fixed-rule models?
JuniorTheoryVery commonROI, ROMI, ROAS, DRR — which is which, and which does the CFO mean by ROI?
ROI, ROMI, ROAS, DRR — which is which, and which does the CFO mean by ROI?
ROAS is revenue divided by ad spend, ignoring product margin. ROMI is margin-based return net of marketing cost, as a percentage. DRR is ad spend over revenue, the inverse of ROAS. A CFO saying ROI means profit-based return, so a strong ROAS can hide a negative ROI.
Common mistakes
- ✗Treating ROAS as a profit metric when it ignores margin
- ✗Confusing DRR with a channel count instead of a spend ratio
- ✗Assuming the CFO means gross ROAS by ROI
Follow-up questions
- →How can ROAS be 4:1 while the ROI is negative?
- →Why is DRR just the reciprocal of ROAS?
JuniorTheoryCommonWhat is an attribution window and how do you choose its length?
What is an attribution window and how do you choose its length?
An attribution window is the time span after an ad interaction during which a later conversion is still credited to that touchpoint. Choose it from the consideration cycle and the time-to-conversion distribution, balancing over- against under-crediting.
Common mistakes
- ✗Confusing the window with the attribution model itself
- ✗Using one fixed window regardless of purchase cycle length
- ✗Ignoring the time-to-conversion distribution when sizing it
Follow-up questions
- →Why would an expensive product need a longer window?
- →What goes wrong if the window is far too long?
MiddleTheoryCommonHow does the chosen attribution model affect business decisions, and which fits fintech?
How does the chosen attribution model affect business decisions, and which fits fintech?
The model shifts credit between channels, so it directly changes budget allocation, ROI assessment, and which channels you scale. Last-click overweights bottom-funnel; first-click overweights awareness. For fintech — long consideration, multi-touch, high LTV — prefer data-driven or position-based/time-decay over last-click.
Common mistakes
- ✗Claiming model choice does not move the budget
- ✗Defaulting fintech to last-click despite a long funnel
- ✗Confusing credit redistribution with changing the conversion count
Follow-up questions
- →Why does last-click starve upper-funnel channels in fintech?
- →What data do you need before trusting a data-driven model?
MiddleTheoryCommonLast-click, first-click, linear, data-driven — what does each one systematically over-credit?
Last-click, first-click, linear, data-driven — what does each one systematically over-credit?
Last-click over-credits the closing touch and starves awareness. First-click over-credits discovery and ignores closers. Linear over-credits filler touches and dilutes the pivotal one. Data-driven estimates each touch's marginal contribution but needs path volume and misses untracked touches.
Common mistakes
- ✗Swapping which end first-click vs last-click over-credits
- ✗Calling linear neutral when it dilutes the pivotal touch
- ✗Treating data-driven as unbiased regardless of tracking gaps
Follow-up questions
- →Why does last-click quietly defund upper-funnel channels?
- →What data does a data-driven model need to be trustworthy?
MiddleDesignCommonTwo channels both look profitable, but paid search pays back its CAC in about 4 months while paid social takes roughly 14. You are setting next quarter's budget split. Explain how channel-level LTV — not just CAC or payback speed — should drive the allocation: when a long payback is still worth funding, how the LTV:CAC ratio and a payback ceiling tied to your cash runway interact, and what number you actually shift budget on.
Two channels both look profitable, but paid search pays back its CAC in about 4 months while paid social takes roughly 14. You are setting next quarter's budget split. Explain how channel-level LTV — not just CAC or payback speed — should drive the allocation: when a long payback is still worth funding, how the LTV:CAC ratio and a payback ceiling tied to your cash runway interact, and what number you actually shift budget on.
Payback ranks how fast cash returns; LTV ranks how much a channel's customers are worth. A 14-month payback is fine if its LTV:CAC stays healthy and runway funds the wait; a 4-month payback with thin LTV can be weaker. Split on marginal LTV:CAC under a runway-driven payback ceiling.
Common mistakes
- ✗Ranking channels by payback speed while ignoring channel LTV
- ✗Treating a long payback as automatically unprofitable
- ✗Ignoring cash runway when choosing an affordable payback
Follow-up questions
- →When does a short payback still signal a weak channel?
- →How does your cash runway set the payback ceiling?
MiddleDesignCommonYou tripled the budget in your best channel and its CAC doubled. That is channel saturation with diminishing returns. Design how you find the efficient frontier — how you map the channel's marginal-CAC-versus-spend curve, where you stop pouring in money, and how you reallocate the freed budget across channels to hold blended efficiency.
You tripled the budget in your best channel and its CAC doubled. That is channel saturation with diminishing returns. Design how you find the efficient frontier — how you map the channel's marginal-CAC-versus-spend curve, where you stop pouring in money, and how you reallocate the freed budget across channels to hold blended efficiency.
Sweep spend in steps and measure marginal CAC — the next customer's cost, not the average — since saturation shows first at the margin. The response curve bends up as the channel saturates. Stop where marginal CAC exceeds its LTV or another channel's; reallocate until it is equal.
Common mistakes
- ✗Steering by average CAC when saturation shows at the margin
- ✗Equalizing average CAC instead of marginal CAC across channels
- ✗Denying diminishing returns exist for a winning channel
Follow-up questions
- →Why does marginal CAC rise before average CAC as spend grows?
- →How do you keep the response-curve sweep from confounding seasonality?
MiddleTheoryCommonChannel CAC vs blended CAC — how can the blended number hide the one channel that is killing you?
Channel CAC vs blended CAC — how can the blended number hide the one channel that is killing you?
Blended CAC spreads total acquisition spend over all new customers, mixing paid and organic; channel CAC is spend over one channel's customers. The blend hides a paid channel bleeding well above its LTV behind a cheap organic flood, so it looks healthy. Judge each channel on its own CAC.
Common mistakes
- ✗Steering budget by blended CAC and missing a losing channel
- ✗Assuming organic zero-cost makes the blend equal channel CAC
- ✗Comparing CAC across channels without their differing LTV
Follow-up questions
- →When is a high channel CAC still acceptable to keep funding?
- →How does an organic surge flatter your blended CAC?
MiddleDesignCommonYou run national TV for a retail brand and the CMO wants proof of its true incremental effect, not an attributed number. You can turn TV on or off by region and have two years of weekly regional sales. Design a geo holdout — how you pick treatment and control regions, how long you run it, and how you read out incremental sales and incremental ROAS.
You run national TV for a retail brand and the CMO wants proof of its true incremental effect, not an attributed number. You can turn TV on or off by region and have two years of weekly regional sales. Design a geo holdout — how you pick treatment and control regions, how long you run it, and how you read out incremental sales and incremental ROAS.
Split comparable regions into treatment (TV on) and control (held out), matched on baseline sales. Run a pre-period plus a treatment window longer than the purchase lag. Read out with difference-in-differences: incremental sales is treatment minus control net of the pre-gap, incremental ROAS that over TV spend.
Common mistakes
- ✗Using best vs worst regions instead of matched comparable ones
- ✗Running too short to cover the purchase lag
- ✗Reading out attributed conversions instead of the diff-in-diff lift
Follow-up questions
- →How does region count drive the power of the test?
- →What does spillover between adjacent regions do to the readout?
MiddleTheoryCommonAttribution credits brand search with 30% of revenue; a holdout says 3% — which do you act on?
Attribution credits brand search with 30% of revenue; a holdout says 3% — which do you act on?
Act on the 3%. Attribution credits the last tracked touch, but brand-search users mostly typed your name and would have bought anyway — capture, not creation. The holdout measures causal lift by switching the channel off for a random group. Fund the incremental value, not the attributed one.
Common mistakes
- ✗Trusting attributed credit as if it were causal
- ✗Missing that brand search mostly captures existing demand
- ✗Averaging attribution and incrementality as if interchangeable
Follow-up questions
- →How would you design the brand-search holdout cleanly?
- →When could the attributed and incremental numbers actually agree?
MiddleDebuggingOccasionaliOS ATT and cookie loss halved paid-social attributed revenue overnight — a real collapse or a measurement collapse?
iOS ATT and cookie loss halved paid-social attributed revenue overnight — a real collapse or a measurement collapse?
Measurement collapse, not real. Company revenue, orders, and spend hold flat, and the ~$1M that left paid-social reappears in direct/unattributed. ATT and cookie loss broke iOS tracking, so real conversions now land as 'direct.' The tell: attributed revenue moved to unattributed while totals held.
Open full question →Common mistakes
- ✗Reading an attribution drop as a real sales drop
- ✗Missing that lost revenue moved into unattributed/direct
- ✗Cutting budget before a holdout confirms incrementality
Follow-up questions
- →Why does ATT push conversions into the 'direct' bucket?
- →How would a geo holdout settle the question here?
SeniorTheoryOccasionalA brand-lift study shows +8 points of awareness — what can it mean for revenue, and why do marketing and analytics disagree?
A brand-lift study shows +8 points of awareness — what can it mean for revenue, and why do marketing and analytics disagree?
Awareness is a leading, upper-funnel proxy: +8 points may lift revenue only through lagged, fuzzy links with no fixed conversion rate. Marketing values the leading signal; analytics wants the causal revenue number. Bridge them by tying brand lift to a holdout or MMM as a future-sales hypothesis.
Common mistakes
- ✗Converting awareness points straight into a revenue percentage
- ✗Dismissing awareness as wholly unrelated to revenue
- ✗Bridging the gap by picking a flattering attribution model
Follow-up questions
- →How would you test whether the awareness lift becomes real sales?
- →Why is there no fixed awareness-to-revenue conversion rate?
SeniorDesignOccasionalYour CMO wants one measurement stack for a large budget across TV, paid social, and search. Explain when you reach for a media-mix model (MMM), when for geo-lift experiments, and when for multi-touch attribution — what each answers, their blind spots, and what MMM specifically needs (history, spend variation, correct controls) to be trustworthy rather than a curve-fit.
Your CMO wants one measurement stack for a large budget across TV, paid social, and search. Explain when you reach for a media-mix model (MMM), when for geo-lift experiments, and when for multi-touch attribution — what each answers, their blind spots, and what MMM specifically needs (history, spend variation, correct controls) to be trustworthy rather than a curve-fit.
MMM regresses sales on spend plus controls: it covers offline, privacy-safe channels but needs years of history and real spend variation, or curve-fits. Geo-lift gives causal lift for togglable channels. Attribution is granular, yet correlational. Use MMM for the mix, geo to validate, attribution to tune.
Common mistakes
- ✗Treating multi-touch attribution as if it were causal
- ✗Trusting an MMM fit with no spend variation or controls
- ✗Assuming geo-lift can measure an always-on national channel
Follow-up questions
- →How do you triangulate when MMM and geo disagree on a channel?
- →Why does an MMM with no spend variation just curve-fit?
SeniorDesignRareA new channel has run for eight months in a business with only three regions. That rules out a clean geo experiment (too few regions for power) and a media-mix model (too little history). Leadership still wants evidence the channel drives sales incrementally, not just captures existing demand. Describe what you actually do: which lightweight causal tools you reach for, how you stage them, and how you communicate the confidence level of the answer.
A new channel has run for eight months in a business with only three regions. That rules out a clean geo experiment (too few regions for power) and a media-mix model (too little history). Leadership still wants evidence the channel drives sales incrementally, not just captures existing demand. Describe what you actually do: which lightweight causal tools you reach for, how you stage them, and how you communicate the confidence level of the answer.
Use platform-native lift and holdout tests: conversion-lift or ghost-ad experiments, a randomized user holdout, or on/off pulsing read as a time series. Triangulate several weak signals, not one. Pre-register the effect and report a confidence range as directional evidence, not MMM-grade precision.
Common mistakes
- ✗Forcing an MMM on far too little history
- ✗Treating attribution as proof of incrementality
- ✗Running a geo test with too few regions for power
Follow-up questions
- →How would you pre-register the expected lift here?
- →Why does triangulating weak signals beat one point estimate?
SeniorDesignRareDesign an algorithm to decide which of 6 ad-promoted products to additionally boost with internal banners. Six federal ad campaigns run simultaneously for a month for six products across many channels; CTR, CR, and the money flow generated after the target action all matter. You can additionally promote products with internal banners. The goal is to maximize the overall business effect. Which business effect do you optimize, how do you build the algorithm that decides which product to boost in the moment, and what does the scoring formula and feedback loop look like?
Design an algorithm to decide which of 6 ad-promoted products to additionally boost with internal banners. Six federal ad campaigns run simultaneously for a month for six products across many channels; CTR, CR, and the money flow generated after the target action all matter. You can additionally promote products with internal banners. The goal is to maximize the overall business effect. Which business effect do you optimize, how do you build the algorithm that decides which product to boost in the moment, and what does the scoring formula and feedback loop look like?
Optimize incremental margin or revenue (the money flow), not vanity CTR. Rank products by expected incremental value per banner slot: traffic × CTR × CR × margin per conversion. Allocate slots live with a multi-armed-bandit policy that keeps some exploration, feeding observed conversions back to update each estimate.
Common mistakes
- ✗Optimizing CTR instead of incremental margin or revenue
- ✗Static equal split with no live reallocation
- ✗Pure exploitation with no exploration of other products
Follow-up questions
- →Why is incremental value, not total value, the right target here?
- →How does the bandit balance exploring a cold product?
SeniorTheoryRareHow does the truthful ad-auction mechanism VCG work, and how is it tuned?
How does the truthful ad-auction mechanism VCG work, and how is it tuned?
VCG (Vickrey-Clarke-Groves) charges each winner the externality it imposes — the loss in welfare it causes other bidders by taking a slot. This makes truthful bidding a dominant strategy: a bidder maximizes payoff by bidding its true value.
Common mistakes
- ✗Confusing VCG pricing with first-price (pay-your-bid)
- ✗Missing that truthful bidding is the dominant strategy
- ✗Forgetting quality/relevance weighting and reserve prices
Follow-up questions
- →Why does charging the externality make truth-telling optimal?
- →What does a reserve price protect the auction from?
