diff --git a/README.md b/README.md
index 44b0bce..e11c80d 100644
--- a/README.md
+++ b/README.md
@@ -10,15 +10,35 @@ Coupa purchase-order email ingestion pipeline. SES receives Amazon PO emails, Cl
 4. The Lambda parses the email, sends it to Claude Haiku 4.5 for structured JSON extraction, and writes to DynamoDB.
    - `email_type: new_po` — conditional `PutItem` on `purchase-orders` (idempotent on `po_number`).
    - `email_type: cancellation` — `UpdateItem` marking the existing row `Cancelled`.
-5. LedgerFlow consumes the table via DynamoDB Streams → `po-sync`. This repo only writes.
+5. DynamoDB Streams (NEW_AND_OLD_IMAGES) on `purchase-orders` feeds two downstream consumers:
+   - **LedgerFlow** (`seahaven-slack-bot/po-sync`) — daily KB sync.
+   - **Verified-sites pipeline** (`po-ingest-site-extractor`) — real-time site address extraction (see below).
 
 A separate `po-web-ui` Lambda (Function URL, unauthenticated) renders a simple HTML dashboard scanning the table.
 
+### Verified-sites pipeline
+
+The `po-ingest-site-extractor` Lambda is triggered by the DynamoDB Stream on every PO INSERT/MODIFY. It:
+
+1. Extracts an Amazon facility site code from `ship_to.name` using a regex cascade (parentheses, `LLC - CODE`, `Station CODE`, `DS - CODE`) with a fallback to the first `line_items` description.
+2. Parses `ship_to.address` into structured fields (street, city, state, zip).
+3. Upserts to the `verified-sites` DynamoDB table — atomically increments `poCount` and appends the PO number to `sourcePOs`.
+
+POs with no extractable site code are logged and skipped.
+
+Backfill stats (initial run): 14,825 POs scanned → 9,900 with extractable site codes → 1,100 unique sites.
+
 ## Architecture
 
 - **IaC:** AWS CDK (Python), stack name `PoIngestStack`, region `us-east-1`.
-- **Lambdas:** `po-email-processor` (S3-triggered) and `po-web-ui` (Function URL). Python 3.12, 256 MB, 60s timeout.
-- **Storage:** S3 `po-ingest-emails-{AccountId}` with 90-day lifecycle expiry; DynamoDB `purchase-orders` (shared, not owned by this stack).
+- **Lambdas** (all Python 3.12, arm64, 60-day log retention):
+  - `po-email-processor` — S3-triggered, parses PO emails via Claude Haiku.
+  - `po-web-ui` — Function URL, HTML dashboard.
+  - `po-ingest-site-extractor` — DynamoDB Streams-triggered, extracts site addresses.
+- **Storage:**
+  - S3 `po-ingest-emails-{AccountId}` — 90-day lifecycle expiry.
+  - DynamoDB `purchase-orders` — owned by this stack, Streams enabled (NEW_AND_OLD_IMAGES).
+  - DynamoDB `verified-sites` — PK `siteCode`, GSI `by-state` on `state`.
 - **Secrets:** Anthropic API key in Secrets Manager at `po-ingest/anthropic-api-key`.
 - **SES:** adds the `PoEmailRule` to the existing `INBOUND_MAIL` receipt rule set (shared with `workorder-ingest`).
 
@@ -53,3 +73,13 @@ python scripts/reprocess.py --execute  # invokes po-email-processor for each
 ```
 
 Inserts are conditional on `po_number`, so re-processing existing POs is a no-op.
+
+## Backfilling verified sites
+
+The stream Lambda handles all future POs automatically. To backfill from historical PO data (one-time):
+
+```bash
+python scripts/backfill_sites.py
+```
+
+Uses the same extraction logic as the Lambda. Idempotent — safe to re-run.