Should Facebook push Groups harder in Feed?

Launch or notHard

Problem. A PM proposes ranking more Groups content into News Feed to grow Groups engagement. Should Facebook do it, and how would you decide?

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

"What's the actual goal — grow Groups participation, overall engagement, or retention? Groups is a means; which end are we optimizing?"
"How much more Groups content, and at whose expense — what gets pushed down in Feed to make room?"
"Is this reversible and rampable, or a big committed change?"

Lay out your approach

Frame it as a trade, not an add. Feed is zero-sum real estate: more Groups content means less of something else (friends' posts, publishers, Reels). The real question is whether a Groups impression is worth more than what it replaces.
Define success up front. Primary: total meaningful engagement and retention, not Groups engagement in isolation (that will obviously go up — it's not the point). Groups participation is a supporting metric.
Guardrails: friend-content engagement (don't starve the core), integrity (Groups can concentrate harmful content), and Feed satisfaction / "see less."
Decide with an experiment: cluster-randomized (Groups content spreads socially), read total engagement and retention with guardrails, over a days-in window for novelty.

Then check in: "I'd treat this as a real-estate trade judged on total engagement and retention, not on Groups engagement alone, with friend-content and integrity as guardrails. Want me to define the success metric and the test?"

Simulate the back-and-forth

Interviewer

Groups engagement will obviously rise if we show more Groups. Isn't that success?

You

No — that's mechanical, not value. If I show more of anything, engagement with that thing goes up. The question is whether total meaningful engagement and retention rise, because the Groups content is displacing other Feed content. If Groups impressions just cannibalize friend-post engagement one-for-one, we've done work for nothing. So the primary metric has to be total, with Groups as a supporting read.

Interviewer

Say total engagement is flat but retention is up. Ship?

You

That's actually a strong case to ship. Groups are known to drive return habits — belonging to an active community is a reason to come back. Flat near-term engagement with a real retention lift means we traded some interchangeable impressions for stickier ones. I'd want the retention effect to be real (not novelty) and the integrity guardrail clean, but retention is a more valuable outcome than raw engagement here.

Interviewer

How do you make sure the experiment isn't fooling you?

You

Two things. First, Groups content spreads socially — a friend's group activity can appear to others — so a naive user-level A/B leaks treatment into control; I'd use cluster randomization. Second, novelty: a new kind of Feed content spikes then settles, so I'd read the days-in curve and judge on the stabilized effect, not week one. And I'd watch integrity closely, since concentrating Groups content can raise exposure to problematic communities.

Land a clear conclusion

You

My recommendation: run it as a cluster-randomized test, judge on total meaningful engagement and retention with friend-content and integrity guardrails, and read it past the novelty window. Ship if total engagement holds or rises and retention improves with guardrails intact. Don't ship on the Groups-engagement number alone — that's guaranteed to go up and doesn't tell us whether we created value or just moved it.

Key summary — what really matters.

Feed is a zero-sum trade — judge Groups content against what it displaces.
Primary metric = total engagement + retention, not Groups engagement (that rises mechanically).
Retention is the real prize — Groups drive return habits.
Cluster-randomize (Groups content spreads socially) and read past novelty.
Integrity guardrail — concentrating Groups content can raise harmful exposure.

The cannibalization test. Let total engagement be interactions per user. Suppose in treatment we replace some friend/publisher impressions with Groups impressions:

Metric	Control	Treatment	Read
Groups interactions / user	2.0	3.0	+50% (mechanical)
Friend interactions / user	8.0	7.2	−10% (displaced)
Total interactions / user	10.0	10.2	+2%
D7 retention	62%	63.5%	+1.5pp

The Groups number screams success (+50%), but most of it was paid for by displaced friend interactions. The honest read is total +2% and retention +1.5pp — a modest but real win, driven by retention. If instead friend interactions had fallen to 6.5 (total −5%), it would be a loss dressed up by the Groups number.

Why cluster randomization. If a treated user's group activity becomes visible to control friends, control gets partially treated, shrinking the measured gap and biasing the effect toward zero. Randomizing clusters of connected users keeps treatment and control separated so the estimate is unbiased (see the A/B section's network-interference question).

Novelty check. Fit the daily treatment effect over the days-in window; if it decays from +6% (day 1) toward +2% (day 14) and stabilizes, report the stabilized +2%, not the launch spike.

Dimension	What interviewers look for	Below the bar ❌	Above the bar ✅
Problem framing	Whether you see Feed as a zero-sum trade, not an addition.	You treat "more Groups" as free upside.	You frame it as displacing other Feed content and ask what it replaces.
Methodology and math	Whether the success metric is total, not the thing you boosted.	You call rising Groups engagement success.	You make total engagement + retention primary, Groups supporting.
Business judgment	Whether you value retention and name integrity risk.	You optimize near-term engagement only.	You recognize retention as the real prize and Groups' integrity risk.
Deployment thinking	Whether you test it correctly.	Naive user-level A/B, read at day 1.	Cluster-randomized, guardrails, read past novelty.
Communication	Whether you give a conditional, defensible call.	"Yes, engagement goes up."	A clear ship rule tied to total engagement + retention + guardrails.

Interview tips

Strong-answer signal: refusing to accept the mechanically-rising Groups metric, making total engagement + retention primary, and testing with cluster randomization past novelty. Common trap: declaring success because the boosted surface's own engagement went up.