Your ROAS Is Lying to You Now
Key Takeaway: Google's AI-driven auctions have made ROAS an unreliable primary metric. It looks good on dashboards while potentially directing spend toward demand that already existed, not demand you created.
There is a metric on almost every PPC dashboard in every marketing team right now that is producing a number. That number looks like signal. It is mostly noise.
ROAS, return on ad spend, was built for a world where you could trace a specific input (a bid on a keyword) to a specific output (a conversion). Google's AI-driven campaign systems, particularly Performance Max and AI Max, do not work that way. They make thousands of micro-decisions about targeting, creative selection, placement, and bidding in real time, based on signals you cannot see. The causal chain between your inputs and your results has been fundamentally broken.
This does not mean your campaigns are not working. It means you can no longer tell whether they are working or why.
What AI-Controlled Auctions Actually Changed
The shift to AI-driven auctions did not just change bidding. It changed the unit of measurement.
In the old model, you bid on keywords, controlled placements, and could run A/B tests on specific variables. The feedback loop was slow but legible. You changed one thing, waited for data, drew a conclusion. The process was manual and imprecise, but at least the cause-and-effect relationship was traceable.
With Performance Max and similar systems, Google's models are optimizing toward a conversion objective using signals across Search, Shopping, YouTube, Display, Gmail, and Maps simultaneously. The keyword as a controllable variable barely exists. The model decides where your ad appears, to whom, and at what price, based on patterns across your account history, your creative assets, and real-time demand signals.
A recent analysis from Search Engine Journal documented a pattern that should concern every PPC manager: AI campaigns are very efficient at capturing existing demand. They find people who were already likely to buy and show them your ad at the right moment. The ROAS looks exceptional. But the conversions might have happened anyway.
That is not attribution confusion. That is a strategic problem.
The Metrics That Actually Measure What You Think You Are Measuring
The fix is not to abandon AI-driven campaigns. It is to build a measurement framework that accounts for what AI can and cannot tell you.
Three shifts are necessary.
First, extend your attribution windows. User journeys with AI-assisted research are longer than they used to be. People ask ChatGPT about a product category, read several sources, come back to search, see your ad on day 60, and convert. A 30-day attribution window misses most of that path. Moving to 60-to-90-day windows gives a more accurate picture of how your campaigns are actually contributing to revenue.
Second, replace ROAS with contribution margin at the product level. A campaign with a 4.2x ROAS on a low-margin product is worse than a campaign with a 2.8x ROAS on a high-margin product. ROAS is agnostic to the commercial reality of your product mix. Profitability metrics are not.
Third, run incrementality tests. Geo holdout testing, where you suppress ads in a control geography and compare conversion rates to your treatment geography, is the most reliable way to determine whether your campaigns are creating new demand or capturing existing demand. Google has made this testing more accessible and less expensive than it used to be. Use it.
The Blended CAC Nobody Wants to Calculate
There is one more number that most marketing teams are avoiding because it is uncomfortable.
As AI Overviews, zero-click search results, and AI-generated answers reduce the share of searches that result in organic website visits, the real cost of acquiring a customer is rising for most businesses. Paid search picks up some of what organic used to deliver. Email or social picks up more. But the total spend required to acquire a customer, across all channels combined, is higher than a channel-specific ROAS number will ever show you.
A recent breakdown from Search Engine Journal framed it as blended customer acquisition cost: the total marketing spend divided by net new customers, regardless of which channel touched them. It is an uncomfortable number to calculate because it implicates your entire marketing investment. It is also the only number that reflects operational reality.
I wrote about the related shift in organic visibility in a previous edition on AI Overviews reducing organic traffic. The PPC measurement problem and the organic visibility problem are two faces of the same structural change: the AI intermediary layer between users and content is absorbing value that previously accrued to paid and organic search separately.
Marketing teams that do not calculate blended CAC are optimizing for dashboard aesthetics, not commercial outcomes.
FAQ
Why is ROAS no longer a reliable primary PPC metric?
Because AI-driven campaign systems like Performance Max make targeting and bidding decisions across multiple channels simultaneously, and they are optimized to find conversions efficiently, including conversions from people who were already likely to buy. ROAS measures conversions, not whether those conversions were incremental, meaning new demand you actually created.
What attribution window should companies use for AI-era PPC campaigns?
Most marketers should extend from the traditional 30-day window to 60 to 90 days. AI-assisted research creates longer, more complex decision paths. A user may interact with AI tools, organic content, and paid ads across several weeks before converting, and a short attribution window simply misses most of that journey.
What is blended customer acquisition cost and why should marketers track it?
Blended CAC is your total marketing spend divided by net new customers, across all channels. It matters because organic traffic is declining as AI search features absorb more clicks, and paid channels are compensating for that loss. Channel-specific ROAS does not capture this dynamic. Blended CAC does.