Marketing Analytics for Google Ads: How to Know If Your Changes Actually Worked

When it comes to marketing analytics, there’s a gap nobody in paid media talks about enough: most Google Ads “optimizations” are never actually measured.

A keyword gets paused. CPA looks better the following week. Someone writes it up as a win. But was it the keyword? The competitor who went dark for three days? Normal weekly variance? A longer conversion window that hadn’t closed yet?

This is the central problem with Google Ads change impact measurement, and it’s costing companies real money. Not because they’re making bad changes, but because they have no way to tell the difference between a good change and a lucky week.

Why Marketing Analytics Often Gets Change Impact Wrong

You make a change. You watch the numbers. You draw a conclusion. Then you repeat that pattern for months, building a mental model of “what works” that’s only loosely connected to what’s actually happening.

We’ve seen this across dozens of accounts. A client will be convinced that pausing broad match keywords was what turned their CPA around, when the data shows it started improving three days before the change went live. Another will undo a campaign structure consolidation because “performance dropped,” not realizing the drop was entirely explained by a seasonal dip in search volume that hit every account in that category at the same time.

The instinct to act is right. The marketing analytics framework, usually, isn’t there.

What’s missing is a pre/post analysis with actual statistical grounding. The same basic approach used in clinical research and A/B testing, applied to the messy, noisy world of Google Ads.

Why “Before and After” Isn’t Enough in Digital Marketing Analytics

Comparing last week to this week sounds reasonable. It isn’t.

Google Ads performance has natural variance. An account spending 5000 € per month might see CPA swing 15–20% week to week with zero changes made. If you’re evaluating a change against that backdrop without accounting for that baseline noise, you’ll find patterns everywhere. Most of them false.

The right question isn’t “did CPA go up or down after this change?” It’s: “is the observed difference large enough, and based on enough data, that we can rule out random chance?”

What a confidence interval actually tells you

A confidence interval answers exactly that question. If the interval around your observed CPA change includes zero, you cannot conclude the change had any real effect. If it comfortably excludes zero, say at a 90% or 95% confidence level, you have actual evidence.

This isn’t academic. It’s the difference between scaling something that works and doubling down on something that didn’t.

Not every metric moves at the same speed

Clicks respond to changes in hours. Conversions can lag by days or weeks depending on the sales cycle. An ad copy test evaluated after 48 hours is almost meaningless. The same test evaluated after three weeks, with enough conversion volume, is something you can act on.

Any serious marketing analytics framework has to account for sample size, baseline variance, and conversion lag. Separately, for each metric. Most dashboards don’t.

How We Built a Marketing Analytics Tool for Change Impact

The problem we kept running into at Resaco was simple: Google’s own API shows the change history for 30 days. If you want to understand whether a campaign restructure from eight weeks ago actually moved the needle, that data has to be checked manually in the Google Ads UI.

So we fixed that first. Every change event (ad group pauses, bid strategy switches, keyword additions, budget adjustments) gets pulled from the API and written to BigQuery in real time. Permanently. That gives us a full audit trail for every account we manage, going back as far as we need.

From there, the tool works in two directions.

Connecting changes to performance automatically

For each change event, the tool pulls the relevant performance window before and after, matched to the right entity level. A campaign budget change is measured at campaign level. An ad group pause is measured at ad group level. A keyword match type edit is measured at keyword level.

The output is a ranked view of changes with observed performance deltas, confidence levels, and sample sizes. Not “CPA improved 12%.” More like: “CPA improved 12%, confidence 88%, based on 340 conversions across a 21-day window.”

That’s a number you can make a decision with.

Flagging when the data can’t be trusted

The harder problem is confounders: other changes that happened in the same window that make it impossible to isolate the effect of any single action. The tool flags these automatically. If a bid strategy change, a new negative keyword list, and a budget increase all went live within the same week, no honest analysis can separate their individual contributions. The tool says so clearly, rather than producing a tidy number that looks clean but isn’t.

This is the part most marketing analytics tools skip. We don’t.

The model underneath

The engine behind the analysis is a state space model called Unobserved Components. The idea is straightforward even if the name isn’t: it decomposes your pre-change performance into a trend, a weekly seasonal pattern, and noise. Once those components are fitted to the pre-period, the model projects forward and produces a counterfactual.

What would clicks, conversions, or CPA have looked like if nothing had changed? The post-period actuals are then compared against that projected baseline, not against last week’s raw numbers. The confidence interval you see is the uncertainty band around that counterfactual. A narrow band with the actual line sitting clearly outside it is a strong signal. A wide band where actual and counterfactual overlap is the model telling you it can’t tell. That’s a useful answer too.

Marketing Performance Analytics in Practice

The most immediate impact is on account reviews. Before recommending anything new, we look back at the last 60–90 days and verify what the previous changes actually did. Not what they looked like they did. What the data, with appropriate confidence levels, actually supports.

This changes the conversation with clients. Instead of a list of optimizations completed, we can show: here’s what we changed, here’s the measured effect, here’s how confident we are. When something worked, we explain why and what we’ll replicate. When something didn’t, we say that too.

Clients notice the difference. It’s the gap between a vendor who “manages your Google Ads” and one who can explain with numbers exactly what’s driving results in your account.

The Bigger Picture: Incrementality in Marketing Analytics

This kind of thinking isn’t new to marketing science. The debate around attribution (what actually caused a conversion) has been running for years. Google’s own research on incrementality and the broader shift toward marketing mix modelling are responses to the same fundamental problem: last-click attribution lies, and most “performance improvements” are never really proven.

What’s changed in 2026 is the context. Smart Bidding and Performance Max have removed a lot of the tactical levers that used to keep analysts busy. You can’t manually bid every keyword anymore. What you can do, and what increasingly separates good agencies from average ones, is measure the strategic decisions more rigorously.

Which match types are actually contributing incrementally? Which campaign structures hold up when you stress-test them statistically? Which “optimizations” are just noise?

Change impact analysis doesn’t answer all of those questions. But it’s a practical, account-level starting point, and it’s available right now, without a data science team or a six-figure modelling project.

The accounts that will perform best in the next few years aren’t necessarily the ones with the smartest bidding strategies. They’re the ones that get better at learning from what they’ve already done.

Want to build that kind of measurement into your account? Get in touch: resaco.fi/ota-yhteytta/