Did the TV campaign actually drive incremental trips?

Measure impactHard

Problem. Uber ran a big TV brand campaign. Awareness rose and lots of trips are "attributed" to it. Leadership asks whether it actually drove incremental trips. How do you measure it?

Before you reveal: say your answer out loud, as if you were in the real interview — get your reasoning across clearly first. There is no single correct answer: reading what the interviewer is really after and defending your own thinking is what makes an answer strong.

Clarify the question first

"Attributed how — a model, or brand-lift survey? Both over-credit."
"Can we vary the campaign by geo (some markets on, some off), or was it national?"
"Outcome — incremental gross bookings, net of what would happen anyway?"

Lay out your approach

Attribution and brand-lift over-credit; use a causal design. TV is upper-funnel and can't be user-randomized, so I'd measure incrementality with a geo experiment (matched markets with/without the campaign) read by diff-in-differences, or synthetic control (a treated market vs a weighted synthetic control of representative markets) if the campaign was city-level.
Measure gross bookings, not awareness — the incremental trips the campaign caused, net of baseline.
Power on markets (few, noisy) — matched pairs, long windows, adstock (effects carry over after the flight).
Read it as a marketplace effect (net trips, not requests) and feed the result back to calibrate the MMM.

Then check in: "Attribution and brand-lift over-credit TV, so I'd measure incrementality with a geo experiment (matched markets) or synthetic control on gross bookings, powered on markets and accounting for adstock. Want me to walk the geo design?"

Simulate the back-and-forth

Interviewer

Awareness rose and trips are attributed — isn't that impact?

You

No. Awareness rising doesn't prove trips, and attributed trips include riders who'd have booked anyway — TV is broadcast to everyone, so any trip in the following weeks gets "associated" with it. Both massively over-credit. The only honest measure is the incremental trips versus a counterfactual where the campaign didn't run. Since I can't randomize which individuals see TV, I'd randomize (or quasi-randomize) at the market level: run the campaign in some matched markets and hold it out of others, and compare gross bookings.

Interviewer

How do you design and power the geo test?

You

Match markets on pre-period bookings, size, and trend, assign some to the campaign and some to holdout, and read the difference-in-differences in gross bookings over the flight and after (TV has adstock — the effect builds and decays, so I'd measure a window past the flight, not just during). For a single big market, I'd use synthetic control: build a weighted combination of untreated markets that tracks the treated market's pre-period bookings, and read the post-period gap. Power is on the number of markets (few, noisy), so I'd use matched pairs, longer windows, and state the minimum detectable lift honestly — with few markets I can only detect a sizeable effect.

Interviewer

What do you conclude and do with it?

You

The incremental gross bookings the TV campaign caused (treated minus control), and the ROI (incremental bookings ÷ spend), with the causal assumptions stated. I'd read it net of the marketplace (did demand become trips or surge in constrained markets), and — importantly — feed the incrementality result back to calibrate the MMM, so the model's TV coefficient matches the experimental truth and we stop relying on inflated attribution. If the incremental ROI clears the bar, TV paid off; if the geo test shows little incremental lift despite the awareness bump, it didn't. The senior signal is refusing attribution/brand-lift and measuring incremental gross bookings with a geo/synthetic-control design.

Land a clear conclusion

You

So attribution and brand-lift over-credit TV. I'd measure incremental gross bookings with a geo experiment (matched markets, diff-in-differences) or synthetic control, accounting for adstock, powered on markets with an honest MDE, read net of the supply constraint, and feed the result back to calibrate the MMM. TV paid off only if the incremental ROI clears the bar — awareness and attribution alone don't prove it.

Key summary — what really matters.

Attribution + brand-lift over-credit TV — use a causal design.
Geo experiment (matched markets, diff-in-differences) or synthetic control — TV can't be user-randomized.
Measure incremental gross bookings, account for adstock (carryover past the flight).
Power on markets (few, noisy); state the MDE honestly.
Read net of the marketplace; calibrate the MMM to the result.

Geo diff-in-differences. Gross bookings, campaign markets vs matched holdout:

	Pre	Campaign + adstock window	Change
Campaign markets	100	114	+14
Matched holdout	100	107	+7
DiD (incremental)			+7%

Bookings rose +14% in campaign markets, but +7% was the common trend (holdout), so the TV-attributable incremental lift is +7% — divide by spend for incremental ROI. Attribution might have claimed the full +14% (or more, crediting anyone who booked after seeing TV); the geo design isolates the causal +7% and calibrates the MMM.

Dimension	What interviewers look for	Below the bar ❌	Above the bar ✅
Problem framing	Whether you distrust attribution/brand-lift.	"Awareness up, worked."	Incremental gross bookings vs counterfactual.
Methodology and math	Whether you use a causal design.	Attribution model.	Geo DiD / synthetic control; adstock.
Business judgment	Whether you read net of marketplace.	Requests/demand.	Net trips; supply constraint.
Deployment thinking	Whether you calibrate the MMM.	One-off number.	Feed incrementality into MMM; power honestly.
Communication	Whether you convey incrementality.	"+14%."	"Incremental +7%; ROI vs spend."

Interview tips

Strong-answer signal: measuring incremental gross bookings with a geo experiment/synthetic control (adstock-aware), and calibrating the MMM to the result. Common trap: accepting attribution or brand-lift as the TV campaign's impact.