Advertising’s most famous measurement complaint is more complicated than the industry folklore suggests. The earliest known full attribution appeared in a 1919 speech in which Reverend Roy L. Smith credited department-store magnate John Wanamaker with the observation that about half of his advertising spend was wasted, but he could not tell which half. No direct written or spoken record from Wanamaker has been found, so the attribution remains uncertain.
More than a century later, the underlying problem survives. Marketing teams can open dashboards for paid social, connected TV, retail media, search and podcasts and find that every platform reports success. What those reports often cannot establish is how much of the reported activity would have happened without the advertising.
The central question is incrementality. Did an advertisement cause an action, or did it receive credit for an action the customer was already likely to take? Counting a conversion is relatively easy. Establishing the counterfactual is not.
Last-touch attribution credits the final recorded interaction before conversion and leaves earlier influences out of the calculation. Multi-touch attribution distributes credit across several interactions, but its output still depends on assumptions about which touchpoints matter and how much weight each should receive. Neither approach automatically separates correlation from causation.
This is the first honest thing to say about digital ROI: the dashboards are not necessarily lying. They may simply be answering a narrower question than the one the business thinks it asked.
Retail media makes that limitation particularly visible. A platform can sell the advertising, control the transaction data and report the campaign’s performance. An independent marketing mix model may reach a more conservative conclusion because it uses different data, assumptions and time horizons. The disagreement is not proof that one side fabricated its number, but it is evidence that neither number should be treated as self-explanatory.
Nielsen put the scale of retail media’s growth into commercial terms in its January 2025 analysis of retail media attribution. It reported that worldwide retail media advertising spending was projected to grow by nearly $100 billion between 2020 and 2025, including a predicted 21.8% increase during 2025. Nielsen’s 2024 Annual Marketing Report also found that 68% of global marketers considered retail media more important to their strategies than one year earlier.
The same analysis explains the walled-garden problem. Amazon, Home Depot, Instacart, Kroger, Target and Walmart operate retail media networks in the United States, while Carrefour, Flipkart, Mercado Libre, Lazada, Tesco and Zalando are among the international players. These systems can give advertisers valuable first-party transaction data, but their proprietary structures make consistent comparisons across networks difficult.
Nielsen also acknowledges its own position as a measurement provider. Its case for third-party measurement is that an independent party has less incentive to inflate or downplay campaign performance than the platform selling the inventory. That does not make any measurement provider infallible, but it does make the source of each estimate relevant.
When platform reports and independent models diverge, the disagreement should become part of the analysis. It can reveal different attribution windows, definitions of conversion, audience assumptions or methods for estimating incremental demand. Hiding that divergence behind one blended ROI figure creates certainty that the evidence does not support.
Brand-building adds another difficulty because its effects may accumulate outside the reporting window. A September 2025 Search Engine Journal guide notes that brand-building results are not always immediately visible and recommends measuring awareness, consideration, conversion and loyalty rather than relying on one short-term performance metric.
Search itself now demonstrates why that broader view matters. In February 2026, Ahrefs published an updated analysis of 300,000 keywords. It found that the presence of a Google AI Overview correlated with a 58% lower average click-through rate for the top-ranking page, based on December 2025 data. Its previous study, using March 2025 data, had estimated a 34.5% reduction.
The distinction matters. Ahrefs reported a correlation for pages occupying the top search position within its sample. It did not establish that every page, query or industry experiences an identical decline, and the result should not be presented as a universal causal estimate.
A separate Ahrefs study of 75,000 brands found that branded web mentions, branded anchor text and branded search volume were the three factors most strongly correlated with brand visibility in AI Overviews. Ahrefs explicitly cautioned that correlation does not equal causation, but the result still strengthens the commercial case for tracking brand presence beyond conventional paid conversions.
Brand strength is therefore becoming part of the infrastructure on which SEO, paid search and AI-search visibility operate. Its returns may appear in direct traffic, branded searches, unaided recall and later conversions rather than in the dashboard attached to the original campaign.
This is where finance and marketing often talk past each other. Finance wants a defensible return attached to a specific investment. Brand teams are managing an asset whose value may become visible only across several channels and reporting periods. Both concerns are legitimate, but neither can be resolved by forcing every outcome into a single short-term attribution model.
A more defensible measurement system uses a portfolio of methods. Last-touch reporting can support immediate tactical decisions. Multi-touch models can illuminate recorded customer journeys. Marketing mix models can estimate aggregate channel effects, randomized lift tests can help calibrate causal assumptions, and brand studies can examine awareness and recall.
Each method has limitations. The practical goal is not to find a flawless model but to compare methods with different weaknesses. When several independent approaches point in the same direction, confidence should rise. When they disagree, the disagreement should remain visible until the assumptions behind it are understood.
Privacy changes make that work harder, although the familiar description of an inevitable cookieless future is now outdated. Google ended its plan to phase out third-party cookies in Chrome in April 2025. Chrome users retain cookie choices rather than facing a universal cutoff.
Other restrictions still limit user-level tracking. GDPR governs personal-data processing in Europe, CCPA provides privacy rights in California, and Apple’s App Tracking Transparency rules require permission before an app can track users across other companies’ apps and websites or access the device’s advertising identifier. The result is not the disappearance of measurement, but a more fragmented view that increasingly depends on consent, aggregated data and modeled estimates.
The same discipline applies to social reporting. Business.com’s guide to data-driven social strategy warns against tracking every possible metric without a clear objective. Engagement data can help refine content and targeting, but it becomes useful only when the selected indicators correspond to the campaign’s actual goal.
The pattern underneath all of this is straightforward. Every recorded impression can be counted, but not every impression can be assigned a reliable economic value. Retail media, connected TV and AI-mediated search widen that gap because each arrives with its own data boundaries, attribution framework and commercial incentives.
The honest position for a marketing leader in 2026 is that ROI is rarely one number produced by one tool. It is an estimate built from independent methods, calibrated through experiments where possible, checked against long-term brand signals and treated cautiously whenever one source claims more certainty than its methodology can provide.
The old advertising complaint survives because the uncertainty never disappeared. The industry built better counters, faster dashboards and more detailed models. It still has to decide which numbers describe genuine influence and which merely describe activity that was easiest to observe.