A 2025 Lebesgue analysis measured an average ROAS of 113% for broad targeting versus 76% for lookalike audiences. The lookalikes also carried roughly 45% higher CPMs, per the same analysis. Lookalike audience targeting costs more and yields lower returns than handing the algorithm no audience at all. The model is the problem. It learns from seed lists of past converters, and privacy changes have spent five years shrinking and skewing those lists.

The deeper flaw predates the privacy squeeze. Lookalikes find people who resemble your existing buyers. Stop paying platforms to photocopy yesterday's customers.

Lookalike audiences don't find new customers; they photocopy old ones

Lookalike audience targeting cannot create new demand. The model is seeded with your past converters, so it can only find people who resemble buyers you already won. That is replication, not prospecting. For a CMO, the distinction decides where growth comes from.

A lookalike audience is a platform-built segment of users who statistically resemble your existing customers. Note what the model never asks: whether those people are likely to buy. It only asks whether they look like people who have already done.

Why the seed skews toward buyers you'd reach anyway

The bias starts in the seed. Seed lists come from performance campaigns that harvest a thin slice of buyers already shopping in the category. Ehrenberg-Bass research shows that around 95% of category buyers are out of market at any given time. A seed built from recent converters oversamples the active 5%. The lookalike then clones it.

Lebesgue's 113%-to-76% gap is that bias priced in. Broad targeting outperforms because it is not anchored to the active 5%. A lookalike pays extra to stay inside it.

At scale, the math compounds. Every dollar routed through a lookalike is optimized toward yesterday's buyer profile. Category growth lives with light buyers and first-time entrants, people who share almost nothing with your seed. Persona-based advertising starts from that wider pool instead: who should buy, not who already did.

Platforms profit either way. They charge for the modeling, report strong in-platform ROAS, and let the photocopy effect hide in your incrementality numbers.

Signal loss broke the seed: Why some lookalike match quality has quietly collapsed

Lookalike models need three inputs: observed conversions, matched identities, and deep behavioral histories. Privacy changes since 2021 cut all three at once. The seeds feeding the models are now smaller, stale, and skewed toward the shrinking pool of users who still permit tracking. The model runs fine, but the data under it broke.

The seed is thinner and more biased than it was in 2019

Apple made iOS tracking opt-in, and most users said no. Safari and Firefox had blocked third-party cookies years before Chrome ever moved, so much of the web was already dark. Every loss removes converters from the seed. What remains oversamples consenting users on trackable devices, a group that no longer mirrors your full buyer base.

Attribution broke alongside it. Only 18% of marketers feel very confident they can tie app installs to the right source, per Branch's 2025 State of App Growth survey. A seed built on misattributed conversions teaches the model the wrong customer.

Meta already voted with its product roadmap

Watch what the platform does, not what it sells. Meta now steers advertisers toward broad targeting and Advantage+, its automated audience product, and has removed key detailed targeting options in January 2026. Meta is conceding that seed-and-expand now performs worse than handing the algorithm everything and letting it hunt.

You cannot audit what "similar" means

The black-box problem is the gap between a model's confident output and your ability to inspect its inputs. A degraded seed throws no error. The audience still builds to full size, the dashboard looks normal, and nothing indicates that the input shrank or was skewed. You get confident-looking audiences with no visibility into what changed beneath the surface.

The same crumbling identity layer sits beneath connected TV. Privacy-safe CTV targeting was built without it. Signal loss is only half the indictment, though. Even a perfect seed measures the wrong thing.

Similarity is not propensity: The statistical flaw inside the model

Lookalike audience targeting scores resemblance, not intent. A shopper can match your best customer on age, income, and browsing habits and still have no reason to buy anything this year. The model measures correlation with past buyers. Purchase probability is a different number, and the platform never calculates it.

Propensity is the likelihood that a specific person will buy soon. Lookalike models skip it entirely. They rank people by similarity to a seed and stop.

Seed bias compounds with every refresh

Performance-channel converters are not average buyers. They skew toward deal-seekers and people who already knew the brand, because those are the users activation campaigns catch. The model then expands on the easiest demand rather than the largest.

Each refresh narrows the lens further. New converters arrive through lookalike campaigns, so the next seed resembles the last one even more. Statisticians call this selection bias. A photocopier calls it generation loss.

The pool you never reach is where the money sits. An Ehrenberg-Bass analysis of Dunnhumby data found that the bottom 80% of customers deliver almost half of a brand's sales over five years. Those light buyers share almost no traits with a converter seed.

Buying triggers live in situations, not profiles

Category entry points are the situations that pull a buyer into the market: a new baby, a lease renewal, a Q4 budget cycle. They are contextual and temporary. A demographic twin of your best customer with no trigger buys nothing. A low-similarity prospect with the right trigger buys this week.

Single-platform lookalike models cannot see triggers. They see static profiles, frozen at the moment of someone else's purchase. Contextual targeting reads the buyer's current moment instead of another buyer's history.

The price compounds the flaw. That roughly 45% CPM premium from Lebesgue's analysis buys resemblance that carries no intent signal at all. You pay more to fish a smaller pond.

Persona targeting inverts the model: Start with who should buy, not who already did

Persona-based targeting starts with defining the buyer who should purchase next, then finds those people across the open internet. A persona is an audience built from independent signals: category behavior, content consumption, life stage, and location patterns. No conversion log required. The seed problem disappears.

There is no seed at all, and that single inversion changes what the model can see. Lookalikes expand one platform's record of who has already bought. A persona, by contrast, describes the conditions of a future purchase rather than a resemblance to buyers a platform already logged. Forrester argues the 95-5 rule is "more like the 85-15 variable", with in-market buyers in many B2B tech markets closer to 15 to 30%. Persona targeting reaches those buyers while they form preferences, not after a competitor's pixel logs the sale.

What changes when you define the buyer first

The source diversity is the point. Behavioral, contextual, psychographic, and location data fail independently, so no single privacy change can blind the model. Projections from Statista put global contextual ad spend on pace to more than double by 2027. The market is already funding the input personas run on.

Activation breaks the walled garden, too. A lookalike lives and dies inside the platform that built it. A persona is portable: the same definition reaches a buyer on CTV at 8 PM., streaming audio during a commute, and native placements at lunch. One programmatic advertising platform that carries a single persona across channels also produces cleaner reach and frequency data. That makes CTV measurement far easier to defend to a CFO.

The transition playbook: Moving budget from resemblance to relevance

Treat the shift as a staged reallocation, not a hard cutover. Audit how much prospecting spend runs through seed-based audiences, test persona targeting against the incumbent with a holdout, then move new-demand budget in stages. Most brands can complete the move within a few quarters without disrupting the in-quarter pipeline.

Step 1: Audit your lookalike dependence

Pull every prospecting campaign and tag its audience source. Two numbers matter: the share of new-customer budget running through lookalike audiences, and the age of the oldest seed. A seed built before 2021 predates ATT. It describes a buyer pool that no longer exists, which is why the audit belongs next to your acquisition costs.

Put them side by side. Genesys Growth has documented a 60% increase in CAC over five years across a wide sample of businesses. If costs climbed while your audiences stayed "optimized," the audit usually explains why.

Step 2: Run the head-to-head

Pit one persona-targeted open internet campaign against your incumbent lookalike prospecting. Hold out a matched control group. Score both on net-new customer acquisition, not last-click ROAS. Lookalikes win last-click contests by harvesting buyers who would have converted anyway. A proper incrementality test strips out the subsidy and shows which budget created demand.

Run it for eight to twelve weeks. One purchase cycle is the minimum honest window.

Step 3: Reallocate in stages

Keep lookalikes where resemblance still earns its CPM: cart abandoners, lapsed-customer win-backs, and other jobs adjacent to retargeting. Move the new-demand budget to persona-based reach in 20-30% increments per quarter, keeping the head-to-head scoreboard constant.

Expect the gap to widen. Roughly 69% of advertisers believe Chrome's third-party cookie changes will hit harder than GDPR and CCPA, per the IAB's State of Data report. Every consent prompt that ships shrinks the seed a little further, and lookalike audience targeting decays right along with the signal underneath it. Persona definitions hold. They never needed that signal in the first place.

How Agility builds personas without a seed list

Agility's precision brand advertising runs on this inversion. Persona targeting starts with who should buy. Our personas draw on 1,000+ third-party data sources: behavioral, demographic, purchase intent, and geo. The resulting audience definitions layer more signals than standard platform tooling can combine natively, the kind of build most teams would have to assemble by hand, automated here. No conversion log seeds the model, so the photocopy effect never starts.

The other three pillars carry the persona to market. Creative excellence tests six levers per ad, from value proposition to art and imagery. Media buying places the persona across CTV, streaming audio, display, and native, the open internet, where 61% of online time happens. Measurement science runs the matched holdout tests.

The holdout numbers settle the debate over resemblance. A national outdoor retailer ran a PSA holdout against Agility's persona-targeted campaigns and measured a 2.4x incrementality lift among brand-exposed households. CAC dropped 52%. Lookalike dashboards report harvested demand. A placebo holdout counts only the buyers a campaign created.

Run that head-to-head for one purchase cycle. Score net-new customers, not last-click ROAS, and let the holdout decide where your prospecting budget belongs.

See what precision brand advertising looks like for your brand at agilityads.com/test-precision-advertising.