fix(attribution): Google/GA4 model availability + accuracy polish (post-#473)

Summary

Accuracy review of the attribution skill shipped in #473. The core advice is solid; one important mistake would make an AI recommend Google Ads / GA4 settings that no longer exist, plus a few smaller wording and handoff fixes. Draft PR: https://github.com/coreyhaines31/marketingskills/pull/508

Findings

High

1. Google / GA4 model availability — The skill says “use position-based when data-driven attribution doesn’t have enough volume,” but Google Ads and GA4 removed position-based (and first-click, linear, and time-decay) as choices in 2023. Those tools only offer data-driven and last-click now. Position-based is still a useful idea you can apply on your own data — it just isn’t a Google dropdown option.

Medium

1. “Six models” vs seven rows — The table lists last non-direct plus data-driven; last non-direct should be described as a last-touch variant, not a separate school of thought.
2. Data-driven volume guidance — The old Google gate (~3,000 interactions / ~300 conversions) is fine as history. Today Google recommends about 200 conversions and 2,000 ad interactions in 30 days for good quality (still no hard block).

Low

1. Calendly — Unlike SavvyCal, Calendly doesn’t take free-form metadata. Pass an anonymous id through a standard UTM field (or salesforce_uuid).
2. Ghost ads — Clarify that the auction still runs; your ad is withheld and the “would have served” impression is logged.
3. influencer-marketing — Branded-search / direct blind spots should point to attribution, not only ai-seo.

What already checks out

Don’t sum platform numbers; use CRM/store as the conversion count; platform ROAS isn’t incremental ROAS; direct / branded / dark social / AI blind spots; time-decay math; PSA vs ghost ads; PostHog identify/merge guards; analytics skill boundary.

Proposed fix scope

Accuracy fixes above, plus a short note on how to adjust click-based attribution reports when you have holdout or survey evidence (channel-level override). Explicitly not expanding into new MMM-copilot or ads-automation skill territory.