Lookalike Audiences With AI: The Seed Is the Strategy

The lookalike audience most marketers learned - upload a customer list, pick a percentage, target the result - is being retired around them. Google stopped generating similar audiences in May 2023 and removed them from every campaign by that August. LinkedIn discontinued lookalikes on February 29, 2024, replacing them with predictive audiences. Meta still offers classic lookalikes but spends its energy steering advertisers toward Advantage+ audience, where your targeting choices become suggestions the AI may overrule.

I think the retirement is less dramatic than it looks, because the part of the lookalike that did the work was never the percentage slider. It was the seed. Every replacement product still starts from a list of your customers and finds people like them. The machinery changed owners, and the input is now the only lever you hold, which makes seed quality the whole strategy.

What actually changed, platform by platform

The pattern across platforms is the same: the similarity threshold went from a wall you set to a hint the system may ignore. Google’s newer lookalike segments for Demand Gen campaigns say it outright: “your seed list will act as a signal for Google’s models, and the system may reach highly qualified customers outside of similarity thresholds to help you reach your campaign goals.” Google has also dropped the minimum seed size requirement it once enforced. On Meta’s side, the Advantage+ audience page describes taking “your usual audience selections, such as a customer list, a gender, or an interest” and expanding beyond them.

Classic controls survive in two places worth knowing. Meta’s lookalike documentation still lets you set size from 1% to 20% of a country in 1% steps, requires only 100 people in the source audience while recommending 200 or more converters, and refreshes membership every three days while the audience is in an active ad set. TikTok’s lookalike audiences keep a Narrow, Balanced and Broad choice and want at least 1,000 users in the source. My read: the platforms kept manual lookalikes where advertisers still demand control, and everywhere else the lookalike dissolved into the bidding AI, the same absorption I trace in the PPC manager playbook.

The seed strategy

Since the seed is the surviving input, I build it like it matters, with three rules.

Seed with value, not volume. A seed of everyone who ever bought teaches the model what an average customer looks like, including the discount chasers and one-time refunders. My default seed is the top slice by lifetime value or repeat purchase, even when that shrinks the list toward Meta’s recommended 200 converters. Ten thousand mediocre examples produce a bigger, worse audience than eight hundred excellent ones, because the expansion machinery amplifies whatever pattern you hand it. Meta’s own guidance says to make the seed as large as possible, and I treat that advice the way I treat most platform advice: true for the platform’s delivery system, not automatically true for your margin.

Seed with recency. A customer file with 2021 buyers in it describes a market that no longer exists. I rebuild value seeds on a rolling window, twelve months for considered purchases and shorter for fast cycles, and I let the three-day refresh work with a live source (a synced CRM segment) rather than a one-time CSV that ages from the day it uploads.

Separate seeds by economics, not demographics. One seed of subscription customers and another of one-off buyers gives the model two honest patterns. One blended seed gives it a fiction. This is also the cheapest personalization decision left in paid social, since inside the campaign the AI now makes most of the others.

The companion rule sits outside the seed: exclusions. Feeding the platform your customer list for expansion while failing to exclude existing customers from prospecting is how AI targeting quietly re-buys people you already own, a measurement problem I covered in the Advantage+ setup guide, where the exclusion list does double duty as a targeting control and an honesty control.

Reading the platforms’ performance claims

Meta’s Advantage+ audience page reports median cost-per-acquisition improvements of 14.8% for upper-funnel, 9.7% for middle-funnel and 7.2% for lower-funnel outcomes, each “at a 99.9% confidence” per the page’s own footnoted studies. Those are Meta’s numbers about Meta’s product, and I read them the way I read Google’s Performance Max claims: plausibly true on average across millions of advertisers, and silent about whether your account is on the good side of the median. The honest test is the same as ever, a split of Advantage+ audience against your best manual lookalike with identical creative and budget, run past the learning period before judging, as laid out in the Advantage+ troubleshooting guide.

My results pattern, for what it is worth as one practitioner’s sample: AI expansion usually wins on cost per result when the seed is strong and the product is broad, and manual lookalikes still earn their keep when the audience is genuinely narrow, when compliance limits who may see the ad, or when I need the audience boundary fixed for a clean test.

When to still choose the manual lookalike

Three cases keep the classic tool in my accounts. Regulated or sensitive products, where “the system may reach outside your thresholds” is a legal problem rather than a delivery feature. Small markets, where a 1% lookalike is already most of the reachable buyers and expansion only adds noise. And experiments, because an audience the AI reshapes daily cannot serve as a stable test cell. Everything else has drifted to the expanded products, and fighting that drift costs more performance than it protects.

FAQ

What is the minimum seed size for a lookalike audience? Meta will build one from 100 people and recommends at least 200 converters for conversion-based seeds. TikTok wants 1,000 matched users. Google no longer enforces a minimum for lookalike segments. Practically, I care less about the platform minimum than about seed purity, and I would rather run 500 great customers than pad to 5,000 with marginal ones.

Are lookalike audiences dead? As a named product they are gone on Google and LinkedIn and de-emphasized on Meta. As a mechanism, seeding an algorithm with your best customers, they are more central than ever. Every AI audience product is a lookalike with the similarity dial taken away from you.

What percentage lookalike should I use on Meta? Where I still run classic lookalikes, 1% to 3% for performance and wider only for reach objectives. But the honest answer in most accounts is that Advantage+ audience with a strong customer-list suggestion has replaced the percentage question entirely.

Do lookalikes update automatically? On Meta, yes. Membership refreshes every three days while the audience is attached to an active ad set, but only against your source audience as it currently stands, which is why a synced source beats a static upload.

This guide is part of the paid media topic. The exclusion discipline comes from the Advantage+ setup guide, the claim-reading habit from the PPC manager playbook, and the wider practice starts at the complete guide to AI marketing.