Just ran the latest Gemini attribution report for our e-commerce platform. According to this masterpiece of data science, our engineering blog is now responsible for 200% of last quarter's revenue.
Yes, you read that right. Our three readers (hi mom) somehow generated double the money we actually made. Meanwhile, our actual paid search campaigns are showing negative attribution. Brilliant.
* Last touch? Nonsense.
* Data-driven? More like data-hallucinating.
* Spent two days checking our BigQuery pipeline. It's not us.
So, is this just our instance, or is everyone's attribution model now running on magic and wishful thinking? Fighting the urge to just chuck it and go back to basic UTM tags.
fight me
Welcome to the beautiful world of fractional attribution. It's not you, it's the model. They're all glorified guesswork. When you split one conversion across five touchpoints, someone's math eventually adds up to more than 100%. I've seen it a thousand times.
Your problem likely isn't BigQuery, it's your model's rules. Negative attribution on paid search is the real red flag, that usually means you've got conversion data being applied retroactively or there's a messed-up weighting rule that's allowing credit to be stolen from one channel and given to another. Check your lookback windows and any custom credit rules you've set, that's where the ghosts live.
Going back to UTMs isn't chucking it, it's sanity. At least you know what you're counting, even if it's incomplete.
Exactly. I've found the root cause in these models is often the intersection of *custom rules* and *global conversion counting*. For instance, if you have a rule that says "direct traffic gets 10% credit if any other channel is present," and then you also have a standard linear fractional model, you can easily get double-counting because the rules engine and the base model operate in different layers, each applying its own slice of credit without knowing the other has already acted.
Your point about retroactive conversion data is also key. If a user completes a purchase, then later opens a blog post from a re-engagement email, and the lookback window is overly generous, that blog post can sometimes get assigned a share of a conversion that already happened. The system sees a touchpoint within the window and applies the fractional logic, even though the causation is backwards.
- Mike
Spot on about the layering issue. I've seen similar chaos when someone sets a custom rule in Google Analytics *and* applies a Data-Driven model in the same view. Each layer thinks it's the only one doing the math.
Your retroactive example is a classic. We fixed a version of that by setting a "conversion lock" timestamp in our pipeline. Once the purchase event is logged, any future touchpoints within the lookback window are ignored for that session. It's a bit brute-force, but it stopped our support docs from "generating" sales.
Dashboards or it didn't happen.
Check your timezone settings in the attribution tool vs your data warehouse. We had >100% attribution because the model was using UTC and our conversions were stamped in PST. The date overlap caused double counting across reporting periods.
Negative attribution points to rule conflicts, like others said.
null