The Carve-Out Cohort Turnaround
To fix a load-bearing product, secede from the metrics system with a small randomized cohort.
- Difficulty
- Expert
- Time to result
- ~months to results
- Steps
- 7
- Confidence
- 95%
Cohen's method for what he calls 'minus-one-to-one' products: turnarounds, as distinct from zero-to-one or scaling. The core insight is that turnarounds are hardest internally, not externally, because entrenched flows, metrics, and dependent teams fight any change. Rather than fighting escalations, you carve out a randomized cohort of members, build the target experience only for them with full liberty, and let evidence of behavior change do the organizational persuasion.
Origin
Developed by Tomer Cohen while leading the LinkedIn feed transformation (a role that did not exist — there was no unified feed team and no feed PM until he asked for it). He first tried changing the whole experience and failed, hitting 'walls after walls after escalations' for months before inventing the carve-out.
Core principles
- 01Minus-one-to-one (turnaround) is its own discipline and is mostly an internal-politics problem, not a product problem.
- 02In a company where everyone depends on a surface for their traffic, every mass experiment on that surface scares the whole system and consumes the leader in escalations.
- 03A randomized carve-out lets you build without hurting company numbers at scale — a country of members living in a different product.
- 04Redefine the surface's purpose first; the purpose is what you build backwards from.
- 05Evidence of a growing pie ends the arguments about how the pie gets sliced.
- 06Behavior change takes months, not a week. Give the cohort real time.
How to run it
- 1
Declare the new purpose of the surface
Before touching the product, state what it is emphatically NOT. For the LinkedIn feed: not a springboard for other products, not a traffic jump-start, not an upsell feed. What it IS: people who matter talking about things I care about professionally — knowledge exchange that gets the right views to the right experts.
Pro tip The negative definition does the real work; it revokes every other team's implicit claim on the surface.
- 2
Own it — ask for the role nobody wants
Assemble a dedicated team with a single owner whose success is the surface's success. Cohen asked for the feed role when nobody cared about it and no unified feed team existed, then reassembled the team around it.
Pro tip Put a PM and team on it with a goal where somebody's neck is on the line for this one problem, and nothing else.
- 3
Unify the engine team into the product team
Identify the actual engine of the product (for the feed it was AI, then a centralized org sitting outside any product team, building to a different objective) and bring it into one unified SWAT team aimed at the new peak.
Watch out The engine team pulling in a different direction is not bad intent — it is what they were told to do. Fix the objective, not the people.
- 4
Carve out a randomized member cohort
Take a fixed randomized slice of members (LinkedIn: two million) and declare them yours. Build the target experience for them with full liberty. Because of the scale of the base, the cohort's swings do not damage the company's overall numbers, so no other team's goals are threatened.
Pro tip The cohort must be randomized so results generalize when you roll out.
Watch out This is only available if you can shield the cohort from other teams' metric accountability. Without that shield you are back in escalations.
- 5
Define bespoke downstream metrics for the cohort
Do not measure the surface with top-level sessions or page views — on a landing surface those count bystanders and bypassers, and any shift can move them. Set unique metrics on downstream, high-value engagement, and on both sides of the marketplace (creation and consumption health), designed so they can later be extended company-wide.
Pro tip Pick metrics that easily extend out to the whole product once you have proven the case.
Watch out Top-level engagement counts on a default landing surface are near-meaningless and will let a bad build look fine.
- 6
Run negative tests for the sake of learning
Deliberately run experiments you expect to lose — e.g. playing a purely promotional feed forward organically over time, or ad-load tests — purely to establish causal evidence about what degrades the experience.
- 7
Let the evidence roll it out
After months of dramatic behavior change in the cohort, present it as proof that the pie grows rather than gets re-sliced, then extend the experience and its metrics to the whole member base.
Pro tip The internal argument shifts from 'who loses traffic' to 'how do we bring this out' the moment a randomized cohort is visibly outperforming.
Watch out Expect this to take months. It was not overnight and not over a week.
In the wild
LinkedIn's feed began as an activity feed (who changed jobs, who connected to whom) and drifted promotional. Cohen tried to change the whole experience and spent his time in escalations as every dependent team's traffic numbers moved. He carved out two million members, gave them a feed built purely for professional knowledge exchange, and ran it for months.
→ Dramatic behavior change in the randomized cohort gave hard evidence the experience grew the pie. It was extended to all members and became the turnaround Lenny describes as 'so underappreciated' — a feed he now checks 10+ times a day and which drives most of his newsletter traffic.
Before the carve-out, Cohen spent a few months attempting the change on the overall experience. Because every team relied on the feed for traffic, numbers shifting up and down scared the whole system and teams freaked out about missing goals.
→ He was hitting walls and escalations and building nothing. This failure is what produced the carve-out method — and Cohen notes he did not yet have the CPO authority to simply mandate a short-term metrics hit.
Common mistakes
Running the turnaround at full scale first
On a surface other teams depend on, mass experiments trigger goal panic across the org. The leader ends up spending their time in escalations instead of building a great product.
Assuming a turnaround is mostly a market-perception problem
The market perception is real but secondary. Most of the difficulty is internal: entrenched flows, processes, and metrics. You have to change the inner workings of the system to make it work.
Measuring the carve-out with top-level engagement
On a default landing surface, counting engagers counts bystanders. Only downstream, high-value engagement metrics tell you whether the new experience actually works.
Is it for you?
Best for
A product leader inheriting a strategically important but declining surface inside a large company, where other teams depend on it and every experiment triggers escalations.
Not ideal for
Early-stage products with small user bases (a carve-out leaves too little signal) or companies whose CEO has already granted a mandate to accept two years of degraded numbers.
From the transcript
“one of the main things we've done was to really set the new purpose for it”
“this is not a a springboard for other products”
“what I did was I carved out two million members and I said those are my members”
“we have this cohort that is doing extremely well which was a randomized cohort”
“spending my time in escalations instead of actually building a great product”
“we've done some negative tests to prove some stuff out test for the sake of learning”
“because the feed is the first thing you land on I can't just count how many folks engage with the feed because then I'm counting…”
From the episode
How LinkedIn became interesting: The inside story
Tomer Cohen (CPO at LinkedIn)