Creative diversity is a portfolio risk decision

Most teams decide how many creative concepts to run by counting how many their production pipeline can produce. That is backwards. The number of concepts live at any moment is a risk-appetite decision, the same kind of decision a finance team makes when it chooses how much cash to hold against a bad quarter, and it should be made by the people who own the growth target before anyone briefs a designer.

The thesis, stated plainly: concentration in creative buys you a better average month and a worse bad month, diversity buys the reverse, and the correct point on that trade depends on facts about your business that have nothing to do with creative. Teams that let capacity set the number are choosing a risk position by accident.

The mechanism: mean versus variance

When one concept is clearly winning, every dollar moved from a weaker concept onto the winner lowers blended CPI. Concentration therefore maximises efficiency in the present. It also maximises exposure to two events: the winner's fatigue cliff, and a shift in how a network's delivery model treats that concept. Both arrive without notice, and when either happens a concentrated account has nothing warmed up to take the spend.

Diversity is the insurance. Keeping several concepts live at meaningful spend means each one has passed learning phase, each has an audience model attached to it, and spend can move in days rather than weeks. The premium is paid every month in the form of a blended CPI that sits above what the single best concept could deliver.

UA Ledger's "Portfolio Theory for Creative: How Many Concepts to Run" covered the mechanics of that trade last March. What I want to add is the input the piece left implicit: who decides the acceptable variance, and on what basis.

The trade-off most coverage skips

Diversity carries a cost that does not show up as CPI, and it is the reason nominal diversity often turns out to be fake.

Every concept that is live on a network is paying an exploration tax. The learning phase, the minimum spend to exit it, and the ongoing share of impressions the model reserves for testing all scale with the number of active ads. Ten concepts at low spend each can cost more in exploration than three at scale, and the ten will often never leave learning at all.

Worse, automated bidding does its own concentrating. Feed a campaign ten concepts and the delivery model will typically push most impressions to one or two within days. Your reporting shows ten concepts live. Your delivered portfolio is closer to one. The insurance you think you are paying for is not in force, because the network has quietly cancelled the policy on the eight concepts it stopped serving.

Real diversity therefore requires structural separation: distinct campaigns or ad sets with their own budgets, so the model cannot collapse them. That costs more exploration, not less. Diversity is expensive in a way the concept count hides.

Three inputs that should set the number

Rather than starting from capacity, start from three facts and let them produce the number.

The first is tolerance for a bad month. If the studio runs on a monthly ROAS target with little buffer, one fatigue cliff can trigger a spend cut that costs more in lost momentum than a year of diversity premium. If the studio has runway and a patient board, concentration is affordable.

The second is the fatigue half-life of your genre. A concept in a puzzle account might hold for months; the same discipline in a hyper-casual account might have weeks. Short half-lives demand more live concepts because the cliff arrives more often.

The third is the share of spend running through black-box networks. The more of the budget sits in campaigns where you cannot see or control per-concept delivery, the less nominal diversity protects you, and the more you need separately budgeted structures to enforce it.

Sensor Tower's State of Mobile 2026, published last month, described the market's shift from new-user volume toward lifetime-value expansion, which makes a lost cohort harder to replace cheaply and so raises the price of a bad month.

A worked illustration

Take a mid-core account spending an illustrative 400,000 a month, or 4.8 million a year, on flat budget. Suppose the best single concept runs at a CPI of 3.00 and a portfolio of four separately budgeted concepts blends to 3.45. Concentration buys about 1.6 million installs a year; diversity buys about 1.39 million. The gap is roughly 209,000 installs, which at 3.00 apiece is a diversity premium of roughly 630,000 a year. That is the insurance bill.

Now price the risk it insures. Suppose the genre's concept half-life is about ten weeks, so the account expects roughly five fatigue events a year. In the concentrated case each event costs, say, three weeks at a CPI of 4.50 while a replacement is found and passes learning. Three weeks of spend is about 277,000. At 3.00 that would have bought around 92,000 installs; at 4.50 it buys around 62,000. The shortfall of about 31,000 installs is worth roughly 92,000 at the baseline CPI, so five events cost about 460,000 a year in excess.

In the diversified case each event costs a few days of reallocation at a much smaller penalty, call it 20,000, or 100,000 a year. Total cost of the diversified position: about 730,000. Total cost of the concentrated position: about 460,000. On these inputs, concentration wins on raw annual cost by roughly 270,000.

That is not the end of the argument, because the two figures arrive in different shapes. The diversity premium is paid evenly, about 52,000 a month, and a growth review can plan around it. The concentration penalty arrives in five lumps of about 92,000 each, during which the account's efficiency is 50 percent worse than plan for three weeks. In a business that reports ROAS monthly against a target with little slack, each lump is a missed month, and missed months trigger spend cuts, hiring freezes and board conversations whose cost never appears in the CPI column. The concentrated account is cheaper on average and more expensive on the months that decide whether the UA team keeps its budget.

Change one input and the answer flips. Shrink the half-life to five weeks, as a hyper-casual account might see, and ten events at 92,000 each cost 920,000, so diversity wins outright. Blend the portfolio at 3.20 rather than 3.45 and the premium drops to roughly 300,000, and diversity wins again. Stretch the half-life to twenty-five weeks and concentration wins by a wider margin still. Shrink the studio's buffer so a bad month forces layoffs and diversity wins regardless of the arithmetic. The numbers are invented; the point is that the decision is arithmetic on one consistent set of business inputs, not taste.

Staff to the number, not the other way

Once the number of separately budgeted live concepts is set, production capacity becomes a derived requirement. If the risk position says four concepts with a ten-week half-life, the pipeline must produce a validated new concept roughly every two and a half weeks, and the budget for that is part of the insurance premium, not a creative team request.

Put the decision in the growth review with the finance owner in the room. The creative team should be told how much variance the business can carry, and then asked what it takes to deliver it.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• Concepts vs Variants: Sizing a Live Creative Slate
• Creative concept portfolio sizing for a flat budget