Why We Stopped Using Lookalike Audiences, and What Replaced Them.

We removed lookalike audiences from most of our Meta prospecting campaigns in late 2024. Not because lookalikes stopped appearing in Ads Manager. They still do. We removed them because Meta's algorithm had effectively made them irrelevant, and in many cases, actively counterproductive. The platform was already overriding our audience definitions based on what it read in the creative. Restricting to a 2% lookalike was telling an algorithm that knows billions of behavioral signals to ignore most of them.

Since making the switch, open audience campaigns, broad targeting with Advantage+ enabled and no interest restrictions, have outperformed their lookalike equivalents consistently. The shift was not a theory we ran toward. The data pushed us there.

This piece explains what changed, why it changed, and what the replacement framework looks like in practice across the accounts we manage.

How Lookalike Audiences Used to Work

For about a decade, lookalike audiences were the backbone of Meta prospecting. You built a high-quality seed: a purchase list, a value-based custom audience, or an email list of high-LTV customers. Meta analyzed the shared characteristics of that group and served your ads to users who matched statistically similar behavioral and demographic patterns.

The logic was sound. If your best 1,000 customers shared certain signals, finding people who matched those signals should surface your next 1,000 customers. For years, it worked well enough that DTC brands ran two or three lookalike audiences, a retargeting stack, and careful exclusions to prevent overlap. That was the standard playbook.

Performance generally held up through 2019 and 2020. Advertisers who refined their seeds and understood the match-rate tradeoffs, tighter 1% lookalikes for quality, broader 5-10% for volume, had a meaningful edge over brands running generic interest targeting. Then several things changed at once, and lookalike performance started degrading in ways that were easy to attribute to the wrong cause.

Three Things That Broke the Model

iOS 14 and the ATT prompt. Apple's App Tracking Transparency framework, introduced in early 2021, required explicit opt-in for cross-app tracking. Opt-in rates landed around 25% across most consumer categories. The off-platform behavioral signal Meta could observe to build and refresh lookalike models dropped significantly. Seeds became less representative. Statistical models became noisier. A purchase list that was highly predictive in 2020 generated a fuzzier match in 2022 because Meta had less data to work with when mapping those customers to similar users.

Third-party signal erosion. Beyond iOS, browser-level privacy changes, including Safari Intelligent Tracking Prevention, Firefox tracking protection, and Chrome's third-party cookie deprecation, further reduced the off-platform signals feeding into Meta's matching algorithms. The inputs into lookalike model construction became progressively degraded over a multi-year period.

The algorithm itself changed. This is the one most people underweight. In mid-2025, Meta finished rolling out Project Andromeda across most campaign objectives and placements. Andromeda did not just change the delivery optimization. It changed the fundamental logic of how Meta decides who sees an ad.

Under the pre-Andromeda system, the delivery model started with your audience definition, found users within that pool, and evaluated which ad to show them. Audience first, creative second.

Under Andromeda, the sequence reversed. The algorithm starts with your creative, reads it using computer vision and semantic analysis, infers what kind of user would respond to that content, and then finds those users across the full Meta graph. Your audience parameters are now a secondary input. Restricting to a 1% lookalike means capping the algorithm's exploration based on seed-similarity scores computed before Andromeda had the richer creative signal to work from.

We have written about what Andromeda changed across account structure more broadly at Meta Algorithm Changes 2026: What Andromeda Did and How to Adapt Your Ads. The short version: it is a meaningfully different system, and most tactics built for the old system actively fight it.

What the Data Shows

Independent analysis measuring 3,014 advertisers across the Andromeda rollout period found that broad targeting delivered 49% higher ROAS compared to lookalike targeting (Confect.io, 2026). That is not a marginal difference. A 49% ROAS gap between two structuring choices, with matched budgets and matched creative, is the kind of result that ends the theoretical debate and forces a practical one.

Controlled split tests have repeatedly shown Advantage+ Shopping beating identical manual campaigns running the same creative on the same budget. We see the same pattern across our own accounts, which is why broad targeting is now the default rather than the experiment.

Meta's own internal data, cited by Top Growth Marketing's test analysis, points to a 17% average cost-per-conversion improvement for established stores when Advantage+ Shopping replaces manual campaign structures, for accounts generating at least 50 weekly purchases.

On the creative side, the impact of diversifying assets is sharper than most people expect. One advertiser documented a cost per result drop from $86 to $13.87 within 24 hours by expanding the creative library in a campaign that had been running the same few ads for several weeks (Social Media Examiner, 2026). No audience changes. No budget changes. Creative only.

One note on the last row: Advantage+ Shopping requires conversion data to optimize efficiently. For accounts generating fewer than 20-30 purchases per month, the algorithm takes significantly longer to stabilize, and a hybrid approach, testing broad alongside a 1% lookalike, can make sense while purchase volume builds. The comparison is increasingly one-sided as spend and signal volume increase.

The Framework We Run Instead

Here is the actual structure we use for Meta prospecting across the accounts we manage:

One Advantage+ Shopping Campaign as the primary vehicle. One campaign, one ad set. The platform handles budget allocation and audience discovery. We apply no interests, no age restrictions, no gender restrictions. Geo is the only inclusion parameter, country-level, occasionally state-level for brands with regional distribution or compliance requirements. You are giving the algorithm room to work.

An existing customer budget cap. Meta's research confirms that including a cap of at least 10% for existing customer spend improves both ROAS and cost per result. We typically set the split at approximately 70% new customers and 30% existing. This keeps the algorithm prospecting aggressively while preventing it from burning budget re-converting people already on your list.

Exclusions only, no inclusions. We exclude existing customers from prospecting and exclude prospecting audiences from retargeting. We add nothing to targeting. We only remove things. This is the opposite of how most teams were trained to think about audience management. The old model was about finding the right people. The current model is about getting out of the algorithm's way.

A separate manual retargeting campaign. We keep retargeting outside ASC as its own campaign. It targets website visitors, add-to-cart abandoners, and high-engagement video viewers with creative calibrated for higher intent. Running retargeting separately from ASC gives us cleaner performance data on each and prevents ASC's own retargeting activity from blurring the picture. The separate manual campaign catches the highest-intent pool with specific message sequencing.

Why Creative Is Now the Targeting Signal

If Andromeda decides who sees your ad based on what it reads in the creative, then creative has become your targeting parameter. The algorithm infers the right audience from the content of the ad, and your audience dropdown is a secondary input at best.

An ad that speaks directly to the specific experience of a first-time dog owner will reach first-time dog owners. An ad that references the frustration of spending on supplements that never get finished will reach health-conscious buyers who have that exact relationship with supplements. You do not need to define that audience in Ads Manager. The creative does it.

The first implication is volume. Under the old model, three to five solid ads rotated into a defined audience was workable. Under Andromeda, three ads give the algorithm three creative signals. That is a narrow reading of your product's potential buyers. We aim for fifteen or more active ads in any meaningful prospecting campaign, covering different hooks, formats, angles, and pain points. More creative surface area means the algorithm can find more distinct buyer segments.

Static images account for approximately 60-70% of conversions on Meta across account types (Social Media Examiner, 2026). The assumption that short video has displaced static for all prospecting is not accurate. Format diversity matters more than format dominance. Catalog ads showed 23% higher ROAS and 37% better cost-per-purchase compared to static ads in Andromeda-era accounts (Confect.io, 2026), likely because they align well with the algorithm's entity-matching approach.

The second implication is that unpolished, founder-led, and user-generated creative has outperformed studio production in most categories we manage. Partly because polished production reads as an ad immediately and gets scrolled past. Also because raw, authentic creative carries more semantic signal for Andromeda to parse. A founder explaining the product in an uncut 60-second video tells the algorithm a great deal about who this product is for and what kind of buyer it is trying to reach.

Our creative briefing process now starts with the angle: who is this ad for, what specific frustration or desire does it address, and what is the concrete claim being made. Format and production quality are downstream decisions. The angle is the targeting signal.

The Signal Foundation That Makes It Work

Open audience and Advantage+ Shopping only outperform when the algorithm has clean, high-frequency conversion signal to optimize toward. Andromeda's creative-first model depends on finding buyers efficiently. If your pixel fires inconsistently, your Event Match Quality score is below 6, or Conversions API is not running alongside the browser pixel, the algorithm is working with incomplete data.

This is where we most commonly see ASC underperform despite correct structural setup. The campaign is right. The creative library is broad. But the signal is degraded, so the algorithm cannot find buyers efficiently. CPM climbs. Frequency goes up on the wrong users. ROAS disappoints. The team concludes open audiences do not work for their brand. The actual problem is a broken signal layer.

The fix is to repair signal first, then let the campaign run. What strong signal looks like in practice:

Pixel and CAPI running simultaneously, with server-side CAPI events matching pixel events. No duplication, no gaps. Check Event Match Quality in Events Manager, not in the campaign dashboard.

EMQ score above 7 on purchase events. Below 6 and the algorithm is working with partial identity matches. Above 8 and you have strong signal. Most accounts have room to improve EMQ by adding hashed email and phone to the CAPI event payload.

At least 50 weekly purchase events in the account before committing fully to Advantage+ Shopping optimization. Below that threshold, the algorithm takes longer to stabilize and CPAs will spike early before settling.

Consistent event naming across the funnel. Add-to-cart, initiate checkout, and purchase events need consistent naming across platforms and placements. Multi-product stores often have naming inconsistencies that break the funnel signal.

When all four are in place, the open audience and ASC comparison to lookalike campaigns almost always favors open. When one is broken, the comparison is compromised and the wrong structure may appear to be the better one.

How to Transition Away from Lookalikes

If your account is currently structured around multiple lookalike ad sets at different match percentages, the transition does not have to be abrupt.

Run a direct comparison first. Set up an ASC with a matched budget against your existing lookalike prospecting campaigns. Run the same creative across both. Give it three weeks minimum before reading results. The data usually speaks clearly enough that the decision is not political.

Consolidate progressively. If the ASC is winning, shift budget incrementally rather than switching all at once. Most accounts see performance improve as budget concentrates in the winning structure and the algorithm gets more signal to work with.

Rebuild the signal layer in parallel. Before or during the structural consolidation, audit EMQ scores, verify CAPI is running correctly, and check that purchase events fire with complete customer data: email, phone, city, zip, name. This takes a few days to implement and significantly improves ASC optimization quality.

Keep exclusions in place throughout. The transition does not remove exclusions. Existing customers stay excluded from prospecting. The retargeting campaign stays separate. These are not audience inclusions, and they serve a different function.

One thing we do not do: we do not pause existing campaigns during a test. We run both structures simultaneously with budget splits and let the comparison produce a number. The test decides. Leave the opinion out of it.

What This Frees Up

Removing lookalike management frees up a meaningful portion of account time. Under the old model, a significant chunk of account management went into audience work: building new seeds as customer lists aged, refreshing lookalikes, testing match percentages, managing overlap exclusions between multiple ad sets, reviewing audience size changes as signal degraded.

That work is largely gone. Which sounds like less work. What it actually is: a shift in where the work goes.

The complexity moves to creative. Building new angles. Reviewing ad performance by hook, format, and concept. Identifying when a winning creative starts to show fatigue signals and briefing a replacement before the decline shows up in CPA. Auditing EMQ in Events Manager. Verifying CAPI consistency. These are the active levers on the accounts we manage now.

Account structure also simplifies in ways that compound. Fewer ad sets means fewer overlap issues, cleaner budget data, and easier attribution. One ASC doing the heavy lifting makes it easier to see what is actually working. This connects to a broader point about how paid efficiency should be measured. Platform ROAS does not capture whether the click is profitable after accounting for cost of goods, operational overhead, and customer LTV. We went deep on this in A 4x ROAS Can Still Lose You Money, and the full payback framework is in The Number That Tells You If Paid Actually Pays Back.

Frequently Asked Questions

Lookalike audiences made sense when Meta's delivery model was audience-first. That model has changed. Andromeda reads your creative and finds your buyers from the full behavioral graph, drawing on more signal than any seed audience you build in Ads Manager.

Restricting the algorithm with a lookalike cap means paying for a system that could find your best customers and then actively limiting its reach. The replacement is structurally simpler: one Advantage+ Shopping Campaign, broad audience, diverse creative library, clean conversion signal. The complexity moves from audience configuration into creative strategy, which is where creative decisions have more impact on CPA than audience configurations ever did.

If your Meta account is still built around multiple lookalike ad sets at different match rates, the good news is that simplifying the structure is a low-risk move. You are removing a constraint on an algorithm that has better information than your seed audience does. Run the test. Let the comparison decide. In most accounts with meaningful purchase volume, the data has been pointing the same direction for the past year.