How to Optimize Google Ads — Data-Driven

My articles premiere on my WeChat account, then sync to CSDN, Zhihu and this blog. How do you optimize Google ads? Every beginner asks. However you run Google ads, what lands in the end is data. So I strongly recommend optimizing from data — not from "brand" or "user mindshare" hand-waving, which, bluntly, wastes money and time.

1. Define the analysis goal

Whether something broke or the business needs mining, analysis starts with a clear goal — one satisfying the SMART principle; the S, M, A and R are mandatory.

SMART goals for Google ads

If the goal isn't specific

Discuss with your manager and peers whether it can be sharpened. Example: "use ads to build a powerful brand effect." What is a powerful brand effect? What's the metric? Not specific.

If the goal isn't measurable

Assuming the goal is right, convert unquantifiable notions into metrics — typically through statistical decomposition, labeling, dimensional splits. Example: "cultivate high-potential talent." What is potential? Learning ability, imitation, synthesis, logical expression — each dimension can get its own tests and scoring.

If the goal isn't achievable

Change the frame: expand resources outward, or sort out inward what's actually doable. Example: iOS 14 cut device tracking IDs. Can we get answers from AppsFlyer externally, or fix our internal attribution? Complaining to Apple is an option too — with near-zero success probability.

If the goal isn't relevant

Blindly fixing goals around things you don't understand diverges from reality — coordinate closely with the departments, people and business lines the goal touches. Example: ads in Brazil suddenly went wild — is it something in the Brazil business? Go find out together. Overall, the goal of analysis is to use data logic to find a specific, measurable fact that is within our power to change — and our subsequent effort should influence that fact's course.

2. Start from the business process

Process idea 1: big to small

Market → whole business → channel → project → individual.

Process idea 2: MECE, pyramid principle

Work both horizontal and vertical dimensions — no overlaps, no gaps, every node connected.

Process idea 3: seize the key nodes and their blast radius

No-overlap-no-gap implies many nodes, but not every node deserves attention. Each datum influences a limited stretch of the chain; identify the key node per stage and know the range and boundary of its influence. Try the sandbox below — how would you analyze it?

Analysis sandbox

Background: my Brazil Google ads ROI spiked abnormally, ~+180%. Analysis flow:

Layer	What to check	Status
Market	Whole-market volatility?	Normal
Site	Brazil channel volatility?	Normal
AF attribution	AppsFlyer attribution anomalies?	Normal
SEM channel	Big swings in Brazil?	Abnormal
SEM channel	Whole channel swinging?	Abnormal
SEM sub-channels	Only Google abnormal?	All of SEM abnormal
Brazil ops	Brazil coordination abnormal?	Normal
My ads	Brazil campaigns abnormal?	Normal

Reasonable inference: market, macro, ops and ads all normal → likely AppsFlyer attribution feeding into internal data went wrong; escalate to engineering.

3. The three axes of data analysis

The industry-standard trio:

1. Compare

Compare the changed data within the same dimensions (time, country, delivery type, cycle): what exactly differs, by how much, beyond experience? Example: "temperatures fluctuate a lot lately." Compare morning-evening deltas against the historical deltas and you'll know.

2. Segment

A. With a confirmed anomaly, walk the business composition piece by piece

Example: ROI keeps falling. Check campaigns are serving, check country-level ROI trends, check actual ROAS gaps. If all clean, inspect landing pages and keywords.

B. Decompose the anomalous metric's structure

Example: a campaign's spend goes weird. Spend decomposes into time, creative, audience, keywords — split the anomalous period's structure to find which components broke, and you've found the root drivers.

3. Trace to the source

With the anomalous segment identified, trace the factors behind it. Case: October 2020, several SDS campaigns spent abnormally. Compare, segment: the split showed UK keyword spend exploding at 1 a.m. Trace: keyword analysis found 80% of spend on "iphone12", "iPhone" — utterly irrelevant keywords. Action: excluded the runaway keywords across all global campaigns, reported the anomaly to Google, investigated why those keywords triggered the ads.

4. Validate the data

After analysis delivers conclusions, two situations commonly arise:

1. Misjudgment

The actions your conclusion implies don't fix the real problem. Example: analysis finds Google's ROAS bids clearly correlate with ad ROI; in practice, raising ROAS bids doesn't necessarily lift ROI — it can lower it. This is correlation vs causation. Correlation is easy to observe; causation needs more dimensions and data to establish. It's a hard problem — which is why every analysis needs validation to test whether you misjudged.

2. Attribution is messy

As above, causality is hard precisely because outcomes have multiple causes with different ranges of influence. Analyses typically surface several high-impact factors. So: never seal the verdict immediately. Validate repeatedly through subsequent operations — first to catch misjudgment, second to measure each factor's range and weight.

5. Document everything

Keep records — documents and spreadsheets — of the data, the analysis and the validation, noting each stage's impact. It speeds up localization when things break, eases handovers, and professionalizes control of each business stage. That's the data-driven approach to Google ads optimization — really just a starting point, since these methods are the standard toolkit of research and analysis generally. In short: treat every campaign as an experiment, not as a launch that must return big. Not everyone can accept that. Great media buyers are forged with money — they know where it hurts and which doors are locked because countless ROI-1 and ROI-0.x campaigns taught them. That price tag is exactly what makes many bosses flinch.