Lookalike audience generation

A usable lookalike for bank acquisition starts with a seed you can describe in plain language: who is in it, why they belong, and who was stripped out. Embedding-based clustering can surface similar profiles from the existing book. It does not invent lifetime value, and it does not license you to ship a black-box "high value" list to a media platform.

Quality means a segment built from customers a performance marketer or CRM lead can describe aloud: product held, tenure band, channel of origin, payment behavior that is not fee-driven, and an open relationship. If the cluster exists only as a vector neighborhood, it is not ready for paid media.

A lookalike expands reach toward people who resemble a named seed. It does not replace propensity-to-open scoring at the application gate, and it does not set a credit limit.

Existing customers, marketing opt-outs, and people who failed or were declined at KYC do not belong in the seed or in the export. If they leak through, you pay to reach people you already serve, people who asked you to stop, or people you already refused.

Define the seed with named fields from the portfolio, not with a reconstructed lifetime-value number.

Include product and relationship attributes you actually sell: primary current account, active saver, first-time mortgage customer, or a named combination. Add tenure and activity bands a reviewer can explain: months since open, recent card-present or logged-in activity, presence of a standing order. Add funding and usage that is not punitive: inbound salary or equivalent regular credit, debit-card spend in a normal range, and no persistent unauthorized overdraft. Add channel and geography you are allowed to advertise in, using original acquisition channel only when it is still a fair analog for the campaign footprint.

Strip vulnerability and exclusion flags at seed time. Any marker that the customer is in financial difficulty, under a care protocol, or otherwise out of scope for acquisition creative removes the record.

Do not compute a synthetic lifetime value and then cluster on it. Fee-heavy current accounts often look profitable because unauthorized overdraft charges and returned-item fees dominate the contribution line. A seed built on that profit profile teaches the lookalike to find people who will overdraw, not people who will hold a healthy primary relationship. Name the attributes you want: regular inbound credit, low unauthorized-overdraft incidence, tenure past the early months. Drop fee-derived profit from the feature set.

CRM platforms such as Salesforce typically hold the relationship and consent objects needed to assemble those attributes. Credit-reference and identity vendors such as Experian typically sit on the KYC, affordability, and bureau side. Treat them as classes of system: CRM for who the customer is to you and whether they can be marketed; bureau and KYC for whether they were ever an applicant you declined. You still define the join, the suppressions, and the human-readable description.

One working picture, not a measured case: a performance team wants more primary current-account holders who resemble the "best" existing book. The first cluster, driven by contribution margin, concentrates on accounts with repeated unauthorized overdraft fees. The revised seed is an open primary current account, tenure of at least twelve months, regular inbound credit consistent with a salary or benefit pattern, no persistent unauthorized overdraft in the observation window, and no vulnerability flag. That seed is smaller and less "profitable" on a fee view. It is the one you can describe to conduct review and to the media buyer.

Work on cross-sell propensity at account opening can inform which products you mention in creative after a lookalike click. Do not fold that propensity into the seed as a hidden value score.

Run suppressions on the seed and on the final audience file before anything leaves the bank.

Match existing accounts on identifiers you already use: party ID, hashed email, hashed phone, and device or login IDs where they are lawfully held. Include joint parties and recently closed relationships if policy treats them as current customers for marketing. A lookalike that retargets people who already hold the product is not acquisition.

Honor marketing preference, channel-level consent, and suppression lists for mail, email, SMS, and paid social where your policy extends. If someone opted out of advertising but remains a customer, they stay out of the seed and out of the hashed export. Do not refresh them in because clustering ran on a warehouse snapshot older than the preference center.

Anyone who applied and was declined, withdrawn for identity failure, or left in a failed KYC state is not a seed member and is not a lookalike target for the same product family. Exporting them asks a media platform to find more people like applicants you already refused. Keep declined-applicant files in the suppression set, not in the positive seed.

If your conduct framework excludes customers in arrears programs, garnishment, or documented vulnerability from acquisition lookalikes, apply that exclusion at seed time.

If the campaign is actually reactivation of former customers, use churn-to-win-back targeting and its own consent rules. Do not mix closed-book win-back into an acquire lookalike seed.

After suppressions, the remaining seed should still meet the media platform's minimum size and still be describable. If suppressions gut the cluster, the cluster was the wrong seed, not a reason to loosen KYC or opt-out rules.

A human reviewer (the performance or CRM lead, plus conduct or compliance as your governance requires) reads a short seed card before export. The card lists named inclusion attributes and the observation window, explicit exclusions (fees-as-profit, vulnerability, declined KYC, existing accounts), match keys that will be hashed, destination platforms, and the product the ads will promote. It states that this file is media targeting only.

If the reviewer cannot paraphrase the seed without referring to a mystery score, the export waits. Embedding clusters are a discovery aid. The approved seed is the named attribute definition, not the raw nearest-neighbor list.

Export only hashed identifiers the platform contract allows. Keep raw national IDs and full account numbers out of the file. Record the snapshot date so CRM can reproduce who was in the seed if a complaint arrives. Upload the hashed seed as the source audience, then ask the platform to expand from that source. Do not substitute a pixel-based high-value event that you cannot describe.

When the platform returns an expanded lookalike, do not treat that expansion as evidence of creditworthiness. People who resemble salary-funded current-account holders are still unknown applicants. Limit setting, especially for thin files, belongs with thin-file credit limit calibration, not with the media audience.

Three failure modes look like performance wins until someone inspects the seed.

Seeding on profit that is actually overdraft fees overweights penalty income unless you strip it. The lookalike then finds people whose behavior resembles fee-paying overdrafters. Fix that in the seed definition, not in bid adjustments.

Exporting current customers happens when a party ID or email-hash match is missing, or the extract is stale. Suppress on the way out, and sample the matched audience against the live customer table after the first upload.

A media audience is not an origination policy. Do not auto-approve, raise limits, or skip KYC because someone arrived through a lookalike. Do not use declined-applicant similarity as a positive seed. The acquire stage ends at qualified traffic and a complete application.

Keep the job on a schedule that re-reads consent and KYC suppressions, not only embeddings. Re-describe the seed when the product mix or fee policy changes. If you cannot name who you are cloning, you are not ready to buy the traffic.