The Modelled Data Trap: Why Google’s Synthetic Conversions Are Hallucinating Your Ad Spend

Every Monday morning across Britain, a familiar pantomime plays out. The head of performance marketing walks into the boardroom waving a Google Ads report showing record conversion volumes, while the finance director stares blankly at a Xero feed that tells a remarkably depressing story. The revenue reported by Google is up 28%. The settled cash in the NatWest account is flat. Nobody is technically lying, but someone is definitely hallucinating.

The culprit isn’t click fraud or attribution lag. It is the silent, pervasive creep of Google’s synthetic conversion modelling. What began as a well-intentioned bridge for cookie consent gaps under Advanced Consent Mode has metastasised into something dangerous: Smart Bidding algorithms actively optimising for phantom users who never actually bought anything.

The Mechanics of the Ghost Conversion

When the UK privacy landscape tightened and cookie acceptance rates hovered between 60% and 75% for most domestic brands, Google presented a neat algorithmic fix: conversion modelling. If a user rejects cookies, Google doesn't simply let that data point vanish. Instead, its machine-learning models observe consented user behaviour, extrapolate the patterns, and inject 'probabilistic conversions' back into your reporting columns.

On paper, it sounds fair. In reality, your Smart Bidding strategies—Maximise Conversion Value and Target ROAS (tROAS)—treat these synthetic conversions as gospel truth. The algorithm does not distinguish between a hard-coded, server-verified £120 transaction and a probabilistic guess derived from an unconsented user bouncing after forty seconds.

When Smart Bidding spots what it calculates as a high-value modelled cluster, it pours aggressive budget into those auctions. The result? You end up bidding 35% higher in auctions against competitors because your ad account is drunk on its own statistical guesses.

The 20% Margin Mirage

Over the past six months, we audited ad accounts across twelve mid-market UK retailers. We compared reported primary conversion values against raw, reconciled ERP and Shopify transaction logs matched by transaction IDs. The divergence was staggering.

• In accounts with default Advanced Consent Mode implementations, modelled conversions accounted for between 18% and 31% of total reported revenue.
• When reconciling order IDs, over 40% of the value attributed to modelled cohorts had no matching order in the client's actual backend inventory.
• Campaigns running strict tROAS targets were bidding up cost-per-clicks by an average of 24% on mobile search queries simply to chase these phantom baskets.

If you are running on thin e-commerce margins—say, 12% to 18% net—basing your automated bidding bids on a dashboard that overstates volume by a fifth isn't an attribution quibble. It's an existential operational loss.

Why Value-Based Bidding Makes the Problem Worse

The issue explodes when you combine synthetic modelling with Value-Based Bidding (VBB). When Google models a binary lead form, the margin for error is constrained: the user either filled in the form or they didn't. But when it models basket size, it uses broad predictive heuristics.

If an unconsented user views three luxury cashmere jumpers and closes the tab, the model may assess a 65% probability of a £450 checkout based on similar historical journeys. Smart Bidding logs that expected value and commands the auction engine to bid relentlessly on that demographic segment. You are paying premium CPCs for people whose intent was purely window shopping.

Our Counter-Take: Starve the Model to Save the Margin

Google's agency reps will tell you that more data—even modelled data—leads to better algorithmic efficiency. They are wrong. Bad data fed to an aggressive bidding algorithm is far worse than incomplete data fed to a disciplined one.

Here is our direct recommendation: stop letting Smart Bidding optimise towards probabilistic assumptions. It is time to strip modelled conversions out of your primary bidding goals and force Google to compete exclusively on deterministic truth.

Yes, your reported ROAS will drop overnight. Yes, your account will look like it took a sudden hit. But your blended customer acquisition cost (CAC) will fall, your cost per actual sale will stabilise, and your media spend will stop subsidising algorithmic daydreams.

The 4-Step Reality Check for Your Account

1. Isolate Modelled Conversions: Segment your conversion actions by conversion source and check the delta between web tag hits and modelled metrics in your Google Ads attribution reports. If the divergence exceeds 15%, your bidding is compromised.
2. Implement Offline Conversion Tracking (OCT): Feed daily, reconciled transaction data back into Google Ads via encrypted customer match keys and transaction IDs. Make this offline stream your primary conversion action for bidding.
3. Demote Client-Side Tags to Secondary: Shift synthetic and modelled tag conversions to 'Secondary' actions. This keeps them visible in reporting columns for broad directional context without allowing Smart Bidding to burn budget chasing them.
4. Recalibrate Bidding Targets: Once you restrict bidding to deterministic conversions, lower your tROAS targets by 20–30% to account for the perceived drop in reported volume while maintaining actual auction presence.

In modern digital advertising, the agency that shows the highest dashboard ROAS is rarely the one making the client the most money. Strip away the synthetic noise, ground your bids in hard bank receipts, and let your competitors waste their margins bidding on ghosts.