Double-Counting Detection.

// 01How double-counting happens

Trace a sale's journey:

1. User sees Meta Reels ad Monday, doesn't click but remembers the brand.
2. Wednesday searches brand on Google, clicks ad, lands on site, doesn't buy.
3. Friday sees product on TikTok, clicks, buys.

Who sold it? Three platforms, three answers:

• Google Ads: "Wednesday click → Friday sale, 7-day window" → 1 conversion mine.
• Meta: "Monday view → Friday sale, 7-day-click + 1-day-view" → 1 conversion mine.
• TikTok: "Friday click → sale" → 1 conversion mine.

Reported total: 3 conversions. Reality: 1. Double-counting = 200% inflation.

// 02How big is the typical gap?

Brand setup	Typical double-count
2-platform e-com (Google + Meta)	15-25%
3-platform e-com (+ TikTok)	25-40%
4+ platforms (display, email...)	30-50%
Brand-heavy account (big branded share)	40-60%

So a brand spending $100K/mo, reporting "5x ROAS," may actually run 3.5-4x. Decisions on the fiction = real loss risk.

// 03Detection method

Step 1: Establish ground truth

• Shopify / e-com backend report.
• Stripe / payment processor.
• GA4 (one conversion per browser; cookie loss creates small under-count).

Step 2: Sum platform reports

Same period, sum Google Ads + Meta + TikTok + ... reported conversions.

Step 3: Compare

Example: Shopify 850 sales; Google Ads 320 + Meta 410 + TikTok 280 = 1,010 platform total. Double-count = (1,010 − 850) / 850 = 19%.

// 04Separating organic / non-attributed sales

Second factor that pollutes the math: organic sales. If someone visits directly without seeing an ad, they shouldn't go to any platform — but if they're inside attribution windows, they get claimed.

Solution: holdout / incrementality test

1. Turn off ads in a specific region for 2-4 weeks.
2. Track sales there.
3. Difference (ads on vs off) = real incremental contribution.
4. Platform reports may overstate by 30-60%.

Detail: Geo-Lift Testing Guide.

// 05Build unified attribution

Stop trusting platform reports — build your own unified attribution. Three approaches:

Method 1: Last-non-direct click (simple)

Take GA4 / web analytics last-click as truth. Limitation: ignores TOFU contribution; can't credit a video viewer who later purchases.

Method 2: Data-driven attribution

Use GA4 or Google Ads data-driven attribution. ML-based, distributes credit proportionally across touchpoints. Good single-platform; insufficient cross-platform.

Method 3: MMM (Marketing Mix Modeling)

Combines all channels + organic + external factors (season, price, competition). Extracts real incremental contribution. Detail: MMM Guide. Most robust, but needs data team + 3-6 months.

// 06The server-side fix

Server-side GTM reduces double-counting two ways:

• Single-source fan-out: server sends sale event to Google, Meta, TikTok with the same ID. Each platform says "already counted" — deduplication kicks in.
• Conversion ID unification: assign every sale a unique order_id; pass it to every platform. Platform-level dedup activates.

Implementation: Server-Side Tracking Migration Guide.

// 07What to do after correction

The real ROAS will likely come in 20-30% below platform reports. Don't panic — you know the truth now, not less.

• Revise target ROAS: the old "4x target" was platform-reported 5x. Recompute properly: margin-based ROAS target.
• Recalibrate tROAS bidders: drop platform target numbers to match reality.
• Revisit budget allocation: which platform actually contributes more? MMM output guides this.
• Question brand search incrementality: brand campaigns are often 30-50% subsidized — ROAS looks great, real impact lower.

// 08Common pitfalls

• "Last-click is enough": last-click zeroes TOFU, gives 100% credit to brand. Solves double-count, tells the wrong story.
• Shrinking attribution window isn't the fix: 7-day → 1-day reduces double-counting but increases missed conversions. Real problem is cross-platform deduplication.
• "GA4 right, platforms wrong": GA4 has its own cookie loss. Good for comparison, insufficient for truth.
• Treating it as a product-team problem: the marketing team that decides budget owns the validation; product helps with server-side, but analysis is marketing's job.