Ads, Revenue, And Marketplace Tradeoffs

What's being tested

Interviewers are probing whether you can reason about a two-sided marketplace where revenue, advertiser value, and user experience are jointly optimized rather than maximized independently. Meta cares because small ranking, pricing, or ad-load changes can move billions in revenue while also affecting long-term retention, advertiser trust, and content ecosystem health. The key skill is not reciting ads metrics; it is choosing the right objective, identifying countervailing metrics, and explaining how you would make a launch decision under causal, marketplace, and product constraints. Strong answers show you can separate short-term monetization wins from sustainable marketplace improvements.

Core knowledge

Meta’s ads system is a marketplace among users, advertisers, and organic content. A typical ranking objective combines expected advertiser value and platform value:
$\text{score} = \text{bid} \times p(\text{action}) \times \text{quality adjustment}$
where the action may be click, conversion, install, lead, or purchase.
Revenue is not the same as value. Common revenue metrics include impressions, CPM, CPC, CPA, total revenue, ARPU, and revenue per session. Advertiser value metrics include ROAS, cost per incremental conversion, conversion rate, retention of advertisers, budget utilization, and advertiser churn.
User-side costs are often indirect and lagged. Increasing ad load can raise short-term revenue but hurt sessions, time spent, feed depth, posting, sharing, or long-term retention. A strong answer distinguishes immediate engagement effects from cumulative fatigue and ecosystem quality effects.
Expected revenue per impression is often approximated as $eCPM = 1000 \times pCTR \times CPC$ for click-optimized ads, or $eCPM = 1000 \times pCVR \times CPA$ for conversion-optimized ads. In practice, ranking may include calibration, relevance, predicted negative feedback, and policy constraints.
Auction mechanics matter. Generalized second-price auctions and Vickrey-Clarke-Groves-style intuition separate allocation from pricing, but real ad systems use reserve prices, pacing, budget constraints, quality penalties, and predicted action rates. A change can affect both who wins and what they pay.
Budget pacing is central to marketplace interpretation. If high-quality advertisers are budget-constrained, better ranking may not increase total revenue; it may spend the same budgets more efficiently. If demand is unconstrained, better prediction can increase revenue by allocating impressions to higher-value advertisers.
Marketplace experiments can violate SUTVA because users, advertisers, and auctions interfere with one another. If one advertiser is treated, control advertisers may face different prices or win rates. For auction changes, cluster randomization by user, advertiser, geo, or auction segment may be necessary.
Incrementality beats attribution when evaluating advertiser value. Last-click or view-through attribution can over-credit ads that would have converted anyway. Better methods include randomized conversion lift tests, geo experiments, ghost ads, PSA tests, and holdout-based incremental ROAS:
$iROAS = \frac{\text{incremental revenue for advertiser}}{\text{ad spend}}$
Guardrail metrics should cover all sides of the marketplace. For users: retention, sessions, time spent, negative feedback, hide/report rates, survey quality. For advertisers: CPA, ROAS, conversion volume, budget depletion, churn. For Meta: revenue, long-term value, policy violations, ad quality.
Ads changes often have heterogeneous effects. A change may help large performance advertisers but hurt small businesses, or improve revenue in high-demand geos while worsening user experience in low-demand segments. Always slice by country, placement, device, advertiser size, objective, and new versus mature users.
Short-term A/B tests can understate long-term effects. Ad fatigue, advertiser learning, budget reallocation, and user retention may take weeks or months. For major monetization changes, propose ramped experiments, long-running holdouts, ecosystem monitoring, and post-launch backtesting.
Optimization should use constrained decision-making, not a single metric. For example: maximize revenue subject to no statistically or practically significant degradation in retention, ad quality, or advertiser ROI. This is often more defensible than saying “launch if revenue is up.”

Worked example

For “How would you evaluate increasing the number of ads in News Feed?”, a strong candidate would first clarify whether the change increases ad load globally, changes the ranking threshold, or adds an extra ad slot at a specific feed position. They would state the core tension: the experiment may increase impressions and short-term revenue, but could reduce user engagement, ad effectiveness, or long-term retention. The answer should be organized around four pillars: define success metrics, design the experiment, evaluate marketplace effects, and make a launch decision with guardrails.

For metrics, they might propose revenue per user, ad impressions per session, CPM, and total revenue as primary business metrics, while tracking user guardrails such as sessions, feed depth, hides, reports, survey quality, and 7/28-day retention. On the advertiser side, they should monitor CTR, CVR, CPA, ROAS, conversion volume, and budget utilization because more supply may reduce prices but also lower marginal ad quality. For experimental design, a user-level A/B test is a natural starting point, but the candidate should explicitly mention interference: adding ad supply for treated users may change auction competition and advertiser pacing across both treatment and control.

A key tradeoff to flag is marginal revenue versus marginal user cost. The first added ad impression may monetize well, but later impressions often have lower relevance and higher fatigue, so average revenue can look healthy while marginal quality deteriorates. A strong close would say: “I would launch only if incremental revenue is positive after accounting for engagement and retention guardrails, advertiser ROI does not materially degrade, and the effect is robust across key segments. If I had more time, I would run a longer holdout to estimate retention and advertiser adaptation effects.”

A second angle

For “How would you decide whether an ads ranking model improvement should launch?”, the framing shifts from ad quantity to allocation quality. The candidate should focus less on ad-load fatigue and more on whether the model better predicts valuable actions and improves auction efficiency. Offline gains such as AUC, log loss, calibration, or normalized cross-entropy are useful diagnostics, but they are not launch criteria because auction feedback, advertiser bidding, and user behavior can change online. The launch decision should compare revenue, advertiser outcomes, and user quality metrics in an A/B test, while checking calibration by segment and whether the model over-optimizes clickbait-like ads. The same marketplace logic applies: a model that raises CTR but lowers conversion quality or increases negative feedback may not be a true improvement.

Common pitfalls

Analytical mistake: optimizing only for revenue. A tempting answer is, “If revenue per user goes up and the result is statistically significant, launch.” That misses long-term user and advertiser costs; a better answer frames revenue as the objective only under constraints on retention, ad quality, CPA/ROAS, and negative feedback.

Communication mistake: listing metrics without a decision rule. Candidates often name CPM, CTR, CVR, retention, and ROAS but never explain how those metrics determine action. Interviewers want to hear a concrete launch framework, such as “ship if revenue increases by X with no practically meaningful decline in 7-day retention or advertiser ROI, and ramp gradually while monitoring heterogeneous effects.”

Depth mistake: ignoring marketplace interference. A naive A/B test assumes treatment and control are independent, but ad auctions share advertisers, budgets, and pacing systems. For ranking, pricing, or supply changes, mention possible spillovers and propose mitigations such as geo experiments, advertiser-level clustering, auction-level diagnostics, or long-term holdouts.

Connections

Interviewers may pivot from here into experimentation under interference, causal inference for ad incrementality, ranking-model evaluation, or auction design. They may also ask about long-term metric design, heterogeneous treatment effects, budget pacing, or how to diagnose a revenue increase that coincides with worse user engagement.