high-LTV repeat buyers from California
who lapsed in last 60 days

{
  "suggested_name": "CA Lapsed VIP · 60d",
  "estimated_user_count": 1247,
  "event_filter": {
    "all": [
      { "event": "purchase", "count_gte": 3,
        "window_days": 365 },
      { "geo.region": "US-CA" },
      { "ltv_usd_gte": 850 },
      { "no_event": "purchase",
        "window_days": 60 }
    ]
  },
  "platforms": ["meta","google","tiktok","klaviyo"]
}

The shape of the gap

Every growth team has the same loop: marketing asks for a cohort, analyst writes Postgres / dbt / Looker SQL, ops re-pulls hashed identifiers, someone uploads to Meta, someone else mirrors into Klaviyo. Four tools, three hand-offs, two days. The audience definition lives in a Slack message; the implementation lives in four places.

The cost of the gap

Win-back campaigns lag by 2-3 weeks. Audience definitions drift between platforms (Klaviyo's 'lapsed' isn't Meta's 'lapsed'). Identifier hashing gets done wrong on at least one platform. And every cohort change is another SQL ticket — so teams stop iterating.

Why retention matters

Most cohort definitions are time-windowed: 'lapsed 60d', 'high-LTV 365d', 'YoY repeat buyers'. If your warm event store has 30-day retention (Lite) or 90-day (Starter), 'lapsed 180d' literally cannot be expressed. Scale's 365-day retention is the floor for a working audience system. Pro's 2-year retention is the floor for YoY churn-prediction baselines.

// workers/api/src/ai/audience-compiler.ts
const AudienceFilter = z.object({
  suggested_name: z.string().max(64),
  estimated_user_count: z.number().int(),
  event_filter: z.object({
    all: z.array(EventClause).min(1),
  }),
  platforms: z.array(
    z.enum(["meta","google","tiktok","klaviyo"])
  ).min(1),
});

await anthropic.messages.create({
  model: "claude-opus-4",
  tools: [{
    name: "compile_audience",
    input_schema: zodToJsonSchema(AudienceFilter),
  }],
  tool_choice: { type: "tool", name: "compile_audience" },
  messages: [{ role: "user", content: nlDescription }],
});

Meta Custom Audiences

Sends batches keyed on email_sha256, phone_sha256, fbp, external_id. SHA-256 lowercased per Meta hashing spec. CUSTOM source, 30-day retention attribute.

Google Customer Match

OfflineUserDataJob with hashedEmail, hashedPhoneNumber, hashedFirstName + hashedLastName + countryCode + postalCode. Crm-based user list, upload + close in one job.

TikTok Audience

EMAIL_SHA256, PHONE_SHA256, IDFA_SHA256, GAID_SHA256. file_paths upload via TT MMP, then create custom audience referencing the file id. EU residency respected.

Klaviyo Profiles

Profile attributes only (Klaviyo dedupes on raw email). External_id mirrored into custom property. Adds the audience as a Klaviyo segment via segment-create on first sync.

Win-back lapsed customers (180d)

Define: purchased ≥1 in last 365d AND no purchase in last 180d. Sync to Meta + Klaviyo. Needs 365-day retention to look back far enough — Scale tier covers it. On Lite (30d) or Starter (90d) the cohort definition can't even compile because the lookback exceeds your retention window.

High-LTV VIP exclusion list

Define: ltv_usd ≥ 1000 AND purchases ≥ 5 within 365d. Sync as exclusion to Meta and Google ads to stop wasting acquisition budget on existing repeat buyers. Estimated count surfaces in dashboard before sync — no platform call until you confirm.

Cart-abandon high-intent (7d, ≥$200)

Define: add_to_cart in last 7d AND no purchase since AND cart_value_usd ≥ 200 AND viewed_product_count ≥ 4. Sync to Meta CAPI as audience and Klaviyo as flow trigger. Refreshes every 6h via the cron worker.