marketingengineer.

↳ marketingengineer / field note

Smart Bidding is only as smart as its inputs

Stop tweaking tROAS targets and start auditing what the algorithm eats. A field note on value rules, conversion quality and the inputs nobody checks.

A pattern I keep seeing in account audits: weeks of energy spent nudging tROAS targets up and down, zero energy spent on what the algorithm is actually eating.

Smart Bidding is a function. Garbage in, confidently-optimised garbage out.

The three inputs that matter

Conversion completeness. If 15% of conversions never reach Google Ads (see the server-side tagging teardown), the algorithm learns from a biased sample. Fix the pipe before touching any target.

Conversion value honesty. Most shops send revenue. Revenue is a lie the moment margins differ per category, return rates differ per product, or new-versus-returning matters to you. We switched one account from revenue to a margin-weighted value — same campaigns, same targets — and watched the algorithm quietly shift budget toward the products that actually make money. ROAS looked worse. Profit was up 14%.

Conversion latency. If your sales cycle is three weeks, last week’s spend has reported almost none of its conversions yet. Judging — or worse, letting an automated rule judge — on a seven-day window punishes campaigns precisely when they’re working.

The one-hour audit

  1. Pull backend conversions and Ads-reported conversions for the same 30 days. The gap is your data leak.
  2. Take your top 20 products. Multiply each conversion value by actual margin. If the ranking changes, your value signal is misleading the algorithm.
  3. Plot days-to-conversion. Set your evaluation window past the 90th percentile.

No scripts, no tooling, one spreadsheet. I’ve yet to run this on an account and find nothing.

Targets are a steering wheel. Inputs are the engine. Most accounts I open have someone wrestling the wheel of a car that’s running on fumes.