Direct answer: do not begin by asking an ad platform to “find new customers only.” First write an Acquisition Eligibility Contract across CRM, commerce, identity, and media. It should classify each case as eligible new, existing, lapsed with a separate treatment, or measurable unknown; retain effective dates and confidence; and specify whether the action is exclusion, retention routing, acquisition eligibility, or no activation. AI can reconcile records and prioritize exceptions. Deterministic rules should govern permission, purchase state, offer scope, destination list, expiration, and authority to change campaigns.
The waste is less obvious than a dashboard suggests. A repeat buyer can click a prospecting campaign after using another email, changing a phone number, clearing identifiers, purchasing in a store, or aging out of a list. At the same time, excluding “all customers” can block a buyer who purchased a different product, a B2B contact without an active contract, or a lapsed customer who needs a distinct win-back journey. The objective is not the largest list. It is a defensible commercial policy applied to incomplete identities.
Use an Acquisition Eligibility Contract
- Unit — person, household, or account; brand, product, market, and channel covered by the decision.
- Identity — permitted customer ID and keys, source, match confidence, merge rule, and an explicit unknown state.
- Qualifying event — paid order, active subscription, contract, store purchase, recognized marketplace sale, or another approved fact.
- State — eligible new, existing, lapsed, partner or employee, fraud or test, and unknown.
- Recency — latest event, commercial window, calculation time, and classification expiration.
- Permission — purpose, applicable basis, markets, allowed fields, retention, and destination processor.
- Treatment — exclude, allow, route to retention, observe in shadow mode, or send to human review.
- Destination — account, platform, campaign, list, and naming rule; never an ownerless loose file.
- Latency — time from purchase to CRM update, state calculation, upload, processing, and application.
- Evidence — sent, accepted, matched, active, applied, and outcome, each with a timestamp and reconciliation.
The contract turns “existing customer” into a testable decision. A person can be existing for brand-level acquisition but eligible for a new business line; an enterprise can have a contract with one subsidiary and remain a prospect for another. Marketing owns the objective, Product and Sales define scope, CRM maintains state, Privacy limits use, and Media applies treatment. If the boundary cannot fit into a rule, document it and test examples rather than hiding it inside a longer prompt.
Treat an exclusion list as a control, not universal truth
Google describes Customer Match as a way to connect first-party records with identifiable platform users. Match rate indicates how much of an uploaded list could be linked to accounts and helps diagnose formatting and data viability; it is not a performance measure. Google also notes that unmatched records are normal and processing can take up to 48 hours. “Accepted file” therefore does not mean “every customer protected before the next impression.”
Google's lifecycle goals distinguish acquisition, retention, and re-engagement. Acquisition can use customer lists, tags, and autodetection. Yet its troubleshooting documentation states that even New Customer Only ads may sometimes reach returning customers when status was uncertain at serving time because of technical or privacy limitations. Reporting can also contain unknown or estimated states. Use the platform goal as one execution layer; retain the company's economic definition and unclassified share in your own data.
Meta's customer-list terms also require the advertiser to have the necessary rights, permissions, and lawful basis to disclose and use the data. Hashing helps protect matching and transfer; it does not create permission or repair an incompatible purpose. Across platforms, verify accounts, campaigns, geography, special categories, minimum sizes, membership duration, refresh, opt-outs, and export roles. Never send free-text notes, sensitive attributes, or fields that are not needed for an authorized match.
Model four states instead of forcing a binary answer
Eligible new means there is no qualifying event in scope and the minimum evidence supports the decision. Existing means a valid purchase or relationship exists and policy removes the person from prospecting. Lapsed means history exists but recency allows a separate win-back treatment, budget, creative, and metric. Unknown means identity, coverage, or latency cannot support a classification.
Do not automatically convert unknown to “new” to improve a column. Show its spend, conversion, and composition. If unknown grows after a checkout, tag, CRM, or consent change, pause expansion and investigate. For ecommerce, Shopify's operational distinction—first order versus prior order history—is a useful starting point, but the business still must decide how cancellations, returns, tests, multiple stores, marketplaces, subscriptions, and B2B purchases affect state.
Retain source brand, as-of time, and policy version. Recalculate with idempotent events: one order cannot create two customers, a canceled order cannot remain a first purchase, and a profile merge must be reversible. The identity workflow in https://makinai.co/insights/en/ai-customer-identity-resolution-crm-duplicates explains why a false merge can suppress the wrong person while a missed match can keep a known customer in acquisition.
Build a reconciled pipeline, not occasional uploads
Start from systems of record: orders, subscriptions, contracts, POS, and CRM. Normalize only authorized identifiers; compute state by brand, product, and market; produce versioned lists; transmit through supported integrations; ingest diagnostics; and reconcile the active destination population with the eligible source population. A dashboard should show calculated, sent, accepted, matched, active, applied, and removed counts, plus p50/p95 delay and destination failures.
Do not accept the CRM as truth without tests. Shared profiles, corporate accounts, indirect sales, guest checkout, and returns change meaning. Do not make the ad platform the customer master either. It knows what it can observe or estimate under its rules. The company must retain the reason, supporting event, and rule that authorized treatment.
AI belongs in the unresolved set: propose likely record pairs, cluster rejection causes, detect match-rate drift, summarize exceptions, and prioritize review. It should not invent identity, declare someone new without evidence, remove a suppression, or export data. Every suggestion needs source, confidence, triggered rule, and human disposition. Sensitive cases, high-value accounts, conflicts, and policy changes stay with accountable people.
Run a six-week pilot before giving bidding the signal
- Week 1 — select one brand, market, acquisition objective, and two or three campaigns; baseline platform-labeled new, returning, and unknown customers.
- Week 2 — approve the unit, qualifying event, recency, permissions, four states, owners, SLA, and stop criteria.
- Week 3 — connect orders and CRM, resolve deterministic keys, produce versioned lists, and test replay, exclusion, removal, and rollback.
- Week 4 — run in shadow mode; compare internal eligibility with platform diagnostics and conversions without changing budget.
- Week 5 — apply policy to a controlled share of media with a comparator, spend limits, and a separate lapsed-customer route.
- Week 6 — wait for the defined maturity window, reconcile contribution, reach, and errors, then fix, expand, simplify, or stop.
Accept the project on coverage and decision quality, not gross list size. Measure coverage of known purchasers, reviewed-sample precision, false suppressions, existing customers still exposed, purchase-to-suppression delay, failures, opt-outs, complaints, and exception hours. For the business, compare observable new customers, cost per eligible new customer, margin, and payback using https://makinai.co/insights/en/ai-customer-acquisition-margin-payback-value. Preserve a test: fewer existing-customer impressions do not by themselves prove incremental savings, just as high ROAS does not prove demand creation—the distinction developed in https://makinai.co/insights/en/high-roas-brand-demand-ai-incrementality.
Cost rises with brands and markets, order systems, stores, marketplaces, weak keys, product-specific policies, volume, refresh frequency, destinations, review, privacy, and experimental design. Daily batch can be enough for a long buying cycle; high-frequency commerce may need faster events. Lower latency is valuable only when it reduces mistaken exposure by more than it adds infrastructure and operating cost. Fund maintenance because fields, APIs, permissions, and commercial definitions change.
Separate acquisition from retention without abandoning the customer
Removing someone from prospecting does not mean silencing every relationship. Route eligible customers to service, expansion, loyalty, or reactivation with their own objectives. In https://makinai.co/insights/en/ai-customer-reactivation-crm-without-discounts, reactivation is a contribution-margin decision rather than acquisition under another label. Prevent the same person from receiving a “first purchase” offer and a loyalty treatment with conflicting economics.
Marketing owns the objective and new-customer definition; CRM and Data own identity, state, and latency; Media owns application and diagnostics; Analytics owns the comparator; Finance owns margin and payback; Privacy owns purpose and transfer; the agency owns configuration and monitoring. If your account cannot explain why a buyer was classified as new, bring one campaign, the current definition, and an anonymized workflow sample to MAKINAI at https://makinai.co/services/en/digital-marketing-media-performance-growth-agency. We can scope the contract and smallest pilot before AI expands acquisition.