Founders Love Lookalike Audiences. Meta's AI Has Moved On.

The Lookalike Pitch

The logic is intuitive enough that it stuck: take your best customers, tell Meta to find people who look like them, and let the algorithm work. A 1% lookalike of your customer list felt like the closest thing to buying a guaranteed audience. Precise. Mathematical. Scientific-sounding. Founders loved it, and they still do.

Walk into any Meta ads conversation and you will hear the same playbook: build a lookalike from your purchasers, maybe layer in a second from your email list, run them against each other, pick a winner. The ad account audit checklist most agencies use was written for a version of Meta that no longer exists.

The machine learning model underpinning Meta's targeting has changed substantially over the past four years. The lookalike logic has not kept pace - not in how practitioners use it, anyway. Most founders are still running 2020 targeting strategy on a 2026 platform.

What Meta's AI Has Been Doing Since 2021

The conventional wisdom about lookalikes was formed during an era when Meta's interest-based and demographic targeting was the main lever, and the algorithm needed explicit audience signals to find buyers. You gave it a seed, it found matches, and the specificity of that seed was the competitive edge.

Three things changed that:

First, scale. Meta now processes purchase conversion signals from hundreds of millions of transactions annually across all the advertisers in its ecosystem. Your pixel tells Meta when someone buys from you, and Meta aggregates that signal across every advertiser who has a pixel - building a behavioral profile of what "someone about to buy something like this" looks like, independent of any individual advertiser's seed audience. By the time your pixel has a few hundred purchase events, Meta's model is training on your data plus behavioral patterns from millions of similar purchase paths.

Second, the privacy shift accelerated in-platform learning. iOS 14.5 degraded third-party signal quality everywhere, which pushed Meta to develop stronger first-party inference models. The platform got better at predicting purchase intent from on-platform behavioral signals - content consumption, engagement patterns, profile data - precisely because it could no longer rely as heavily on browser-level tracking. A counterintuitive outcome: iOS weakened the pixel but strengthened Meta's internal modeling.

Third, Advantage+ Audience. Meta's automated audience product was not just a UI update. It reflects how the platform now thinks about targeting: start wide, let delivery data constrain the audience dynamically, and trust the in-platform signals over manually specified parameters. Advantage+ Audience consistently tests against narrower manual audiences in Meta's internal benchmarks - and it does not need a seed.

The practical implication: a 1% lookalike of your customer list is Meta's externalized guess at who your buyers are, based on the customers you can identify. Meta's broad targeting is its internalized knowledge of who all buyers are, built on a vastly larger data set. In many categories, the internal model is now smarter than the external one you build for it.

This is exactly the argument behind letting your creative do the targeting work - the audience signal increasingly lives in who responds to your ad, not in the demographic box you drew around it.

Running the Test: LAL vs. Broad

The argument above is directional, not universal. Whether it holds for your specific account, product, and category is an empirical question - and the answer is a test, not an assumption.

Here is how to structure a clean comparison:

1. Use the same campaign, split at the ad set level. Two ad sets, identical budget, identical creative, identical bid strategy. One runs your best LAL (1-3% of a clean purchaser seed). The other runs broad - either Advantage+ Audience with no audience suggestions, or a manual ad set with no targeting restrictions other than age/location if relevant.
2. Match the seed quality before you run. If your customer list is messy - old addresses, one-time buyers, discounted purchasers - the LAL output will reflect that. If you are going to test a lookalike, seed it with your actual best customers: highest LTV, repeat purchasers, recent buyers. A clean signal going into the seed matters as much as seed size.
3. Run for 14 days minimum. The first week is learning phase noise. You need the second week of data to see stable delivery patterns. Calling a winner at day 4 is how you make expensive mistakes. The algorithm needs time - and the 50 weekly events threshold applies here too.
4. Evaluate CPA and CPM, not just ROAS. ROAS can be distorted by deal seekers and attribution overlap. CPA tells you what it actually cost to acquire a customer. CPM tells you how efficiently the algorithm is placing the ad - a lower CPM on the broad ad set usually signals that Meta is finding favorable inventory because it has strong purchase intent signals to work with.

When Lookalikes Still Outperform

The case for lookalikes is not dead. It is just more specific than founders typically apply it.

The common thread: lookalikes add most value when Meta's native model lacks enough in-category purchase signal to self-direct. You are filling an information gap. When Meta already has abundant signal for your category, you are duplicating something it can do itself - and often introducing unnecessary constraints that limit delivery efficiency.

When Broad Takes It

The flip side is just as specific. Broad targeting - whether Advantage+ Audience or a genuinely unrestricted ad set - tends to outperform LALs when several factors align:

High purchase data category. Fashion, beauty, supplements, consumer electronics, home goods, DTC food and beverage. Meta has processed billions of purchase signals in these categories. Its internal model of "who buys this" is richer than anything you can build from your own customer list.

Budgets above $200 per day. Broad targeting is more efficient at scale because it can course-correct quickly based on delivery signals. Below that threshold, a LAL gives the algorithm a useful head start before it has enough conversion data to self-optimize.

Accounts with 1,000+ purchase events. Once your pixel has that volume of signal, Meta's model has a strong internal map of your buyer. The LAL seed becomes redundant because the algorithm already knows who converts for you - it learned it from your own event history.

When you are running Advantage+ Shopping. ASC does not accept targeting inputs at all. If you are running Advantage+ Shopping campaigns, the LAL question is moot - the algorithm decides. And in competitive consumer categories, ASC's broad mandate frequently outperforms the tighter audience you would have specified manually.

The 2026 Prospecting Stack

The right answer is not "use broad" or "use LALs" across the board. It is building a prospecting architecture where each tool is doing the job it is actually suited for.

For most DTC brands spending $5,000 or more per month on Meta, the stack looks like this:

Layer 1: Broad prospecting as the volume driver. An Advantage+ Audience ad set or an unrestricted manual ad set running your best evergreen creative. This is your reach engine. No seed, no lookalike restrictions. Let Meta find buyers from its own signal. This is where most of your budget should sit once the account has enough conversion history. The bid strategy here matters - Lowest Cost on broad prospecting lets the algorithm work; Cost Cap on broad with a thin event history leads to underdelivery.

Layer 2: LTV lookalike as a precision overlay. A separate ad set seeded from your top-tier customers - not all purchasers, but your highest-LTV cohort. This is not replacing broad; it is running in parallel to find a specific buyer profile that broad might de-prioritize in favor of volume. Allocate 20-30% of prospecting budget here. If this ad set consistently beats Layer 1 on CPA over 30 days, shift budget toward it. If not, trim it.

Layer 3: Retargeting as the converter, not the prospector. Custom audiences of visitors, add-to-cart, and past purchasers. This is where LALs do not belong at all - you are targeting people who already know you, not trying to find new ones who look like them. The retargeting layer needs different creative, different messaging, and different bidding logic than prospecting. Run them separately and do not let them bleed into each other.

Lookalike audiences are a tool for narrowing the algorithm's search. Once the algorithm knows your buyers better than any seed you can provide, the tool becomes a constraint rather than a guide.

The mistake most founders make is not running LALs - it is only running LALs, or defaulting to LALs without testing the alternative. The 1% lookalike feels like it should work because the logic is tidy. But performance marketing does not reward tidy logic. It rewards the thing that generates the lowest CPA at the highest volume, and in 2026, for most accounts in consumer categories, that is broad targeting with strong creative doing the qualification work.

If your creative is built to speak directly to the person you want to convert - not a generic message for a broad demographic - then the audience targeting becomes less load-bearing. The right hook earns a lower CPM by pulling the right people into engagement. That is a more durable edge than a lookalike percentage you set once and forget.