4 Star × DataStudios · Artifact

Match-Rate POC

gclid + ECL, measured on live data

POC: Enhanced Conversions for Leads (ECL) — measured match-rate uplift (2026-06-22)

A proof of concept built against live data: add a hashed-PII fallback (Enhanced Conversions for Leads) for the web conversions that have no gclid, using the email/phone now confirmed in contacts. Re-measured against the real Supabase data.

Result (trailing 90 days, live data)

Stage	gclid-only (deterministic)	+ ECL-eligible	Combined uploadable	Uplift vs gclid-only
Quotes	1,475 (34% of web)	+1,249	2,724 — 63% of web	+85%
Orders	167 (35% of web)	+254	421 — 89% of web	+152%

After ECL, the web conversions we can't reach at all collapse to 77 quotes / 10 orders (unmatched_no_key). The remaining misses are now mostly structural: out-of-window (long sales cycles) and non-web RFQs (which have no ad click by definition).

Read this honestly (deterministic vs probabilistic)

gclid matches are deterministic (~100%): a matched gclid conversion will register.
ECL-eligible is submittable, not guaranteed: Google matches hashed email/phone to a click probabilistically — typically ~30–70% depending on first-party data quality. So the incremental matched conversions ≈ ECL-eligible × that match rate, not the full +1,249 / +254. Even at the low end (~30%), ECL roughly adds ~375 quote and ~75 order matched conversions per 90 days on top of gclid — a material, conservative gain.
The table counts what we can submit; the realized lift is measured after go-live from Ads diagnostics + reconciliation.

How eligibility is decided (`fourstar.source.eligible_pii`)

A contact's email/phone is used only if the contact is consented and valid: - not deleted, not do_not_contact, suppressed_at is null; - email used only when present and email_is_valid; phone used when present.

Window: ECL uses the 63-day offline window (tighter than gclid's 90), with rfq_date as the click proxy (the no-gclid web visit). Out-of-window rows are flagged, not sent.

What changed in the code

source.py: deals → contacts join (batched), consent/validity filter, populates email/phone + rfq_date on no-gclid web rows only.
matching.py: ECL branch uses rfq_date as the click proxy for the 63-day window.
pipeline.py: report now shows gclid / +ECL / combined + the uplift line.
25 unit tests (incl. consent-filter + ECL-window cases), ruff clean.

Reproduce: uv run python -m fourstar.pipeline --report --days 90

Next to make ECL production-ready

Confirm the Quotes/Orders actions have Enhanced Conversions for Leads enabled and the account has accepted customer-data terms (else CUSTOMER_NOT_ACCEPTED_CUSTOMER_DATA_TERMS).
Validate ECL events against the live account (extend scripts/validate_account.py to the hashed-PII path), then send via the Data Manager API userData.userIdentifiers leg.
Measure realized match rate post-go-live and feed it into reconciliation.