4 Star × DataStudios · Artifact

Match-Key Analysis

Which conversions can we attribute, and why

Match-key reality & match rate — measured against live data (2026-06-22)

Author: Diego Sanz. The single most important question for the offline-conversions build is: of the deals that convert (Quote / Order), how many can we actually tie back to a Google ad click and upload? This was measured directly against the live Supabase data and the live 4 Star Google Ads account — not estimated.

How it was measured

Source: Supabase deals (lifecycle dates + rfq_source_id + value) joined to rfq_attribution_touches (the click data: gclid / msclkid / UTMs) by deal_id.
A conversion is uploadable only if it is a website RFQ (rfq_source_id = 1) and carries a gclid whose click is within Google's 90-day window before the conversion. Last-click attribution: the latest gclid click preceding the conversion.
Window: trailing 90 days, excluding the 2 most-recent (unsettled) days, per the founder's "never report on unsettled days" rule.
Reproduce: uv run python -m fourstar.pipeline --report.

Result (trailing 90 days, as of 2026-06-22)

Stage	Converted	Website	Uploadable now (gclid, in-window)
Quotes	6,981	4,297 (61%)	1,475 — 34% of web, 21% of all
Orders	1,165	468 (40%)	167 — 35% of web, 14% of all

Why the rest is not uploadable via gclid (this is the "which rows did not import" log, produced at extraction time):

Reason	Quotes	Orders
non-web RFQ (phone/email/fax/portal — no ad click exists)	2,684	697
web, but no gclid and no hashable PII	1,860	278
gclid click older than 90 days before the conversion	828	11
Microsoft/Bing click only (`msclkid` — handled outside Google Ads)	134	12

What this means

The gclid path works today. ~1,475 quote + ~167 order conversions per 90-day window can be uploaded deterministically right now. Real matched gclids + the real Quotes conversion action passed a live validate_only check against the 4 Star account (no errors), so the data, the account, and the event shape are all sound.
The biggest upside is Enhanced Conversions for Leads (ECL). ~1,860 web quotes (+278 orders) convert with no gclid but are still website leads. If hashed email/phone is available per deal, ECL recovers a large share of these (probabilistic, 63-day window). Open item: confirm a contacts source for email/phone in Supabase — not yet wired (the code supports the path; the data join is pending).
Long B2B sales cycles cost matches. 828 web quotes had a gclid but the click was

90 days before the quote — outside Google's hard limit. Nothing to do for those via gclid; ECL (63-day) won't help either. This is inherent to the business, not a bug.
Bing is real but small (134 quote / 12 order msclkid). Out of Google Ads scope; routes to the Microsoft/Bing leg if that scope is confirmed (still open with Mezar).

Open decisions surfaced by the data

Value per stage. rfq_value is frequently 0 at quote stage; order_total carries the order value. Confirm which $ value each conversion action should receive (bidding optimizes on it).
ECL PII source. Is per-deal email/phone available to join? It roughly doubles reachable conversions if so.
Bing scope in phase 1 (pending Mezar).

Status of the write path

Target = Data Manager API (offline + ECL moved off the Ads API on 2026-06-15). The DMA upload leg is built (fourstar.datamanager) with validateOnly dry-run and per-row failure parsing.
The live DMA dry-run is currently blocked by RAPT reauth on the datamanager OAuth token — the same Workspace session-control gate as the GA4 read. Unblocks when that policy is relaxed (or via a service-account key). The Ads-API adwords token is unaffected, which is how the real-account validation above was performed.