Funnel, the Stockholm-based marketing data company, today pointed marketers to a 96-page guide arguing that most of the confusion around marketing measurement comes from asking a method to answer a question it was never built for. The Definitive guide to marketing measurement, promoted in an email sent today as part of a follow-up series to Funnel's Playing defense survey, sorts the problem into three questions - what happened, did it work, and what should happen next - and walks through 18 approaches, from dashboards and attribution models to controlled experiments, marketing mix models and AI agents.

In Short

Funnel has put out a long guide that explains what each way of measuring marketing can and cannot tell a business, and it sent marketers to it by email today. It matters because many budget decisions still lean on attribution reports, which the guide says hand out credit for sales without showing whether the ads caused them. Its central case is that the method has to fit the question, and that a healthy average return can hide extra spending that loses money - though several of its statistics come from vendor blogs, and some of its platform details are out of date.

One budget question, three different answers

The email, signed by Christopher Van Mossevelde, Funnel's content director, opens with a line familiar from any budget review: "Should we increase the budget?" According to the email, that "sounds like one question" but can mean three separate inquiries. Reporting shows spend, activity and recorded results. Incrementality testing estimates whether the marketing made a difference by working out what would have happened without it. Marketing mix modeling and scenario planning explore what might follow from spending more.

"These approaches serve different purposes," the email states. It directs readers to the guide's second chapter, "The questions that drive measurement," as the place to start.

The guide builds on the same split. It groups measurement questions into three categories: diagnostic ("what happened?"), causal ("did it work?") and predictive ("what should we do next?"). "The foundational premise of this guide is that different questions require different measurement approaches, and to treat them as interchangeable is the root cause of confusion," according to its executive summary. It opens and closes with the same phrase: "Measurement is a discipline, not a dashboard."

The email describes itself as the third of five in a series following up on Playing defense, the survey Funnel published with Ravn Research on October 29. That report, covered by PPC Land on October 31, found that 36% of in-house marketers had kept funding channels they believed were not working, and that only 9% combined marketing mix modeling, incrementality testing and attribution. The guide reads as the methods manual behind that diagnosis.

What the document is

The guide is undated. The most recent events it describes took place in January 2026, and its budget figures come from Gartner's 2025 CMO Spend Survey, published in May 2025. It runs to 18 chapters and a closing section, followed by 100 numbered references.

There is no original research in it. Almost every number comes from another organisation: Gartner, McKinsey, BCG, the IPA, Nielsen, Google, academic economists, and a long list of software vendors and agencies. Funnel's own products are not named anywhere in the text. The company appears only in the reference list, as the source behind three citations.

Each chapter closes with a short statement of when the method in question should be used. That structure makes the guide easy to scan, and it is the most concrete expression of its main argument: no single method covers every decision.

Why the numbers stop adding up

The opening chapter sets out the problem as a gap between marketing reports and the profit and loss account. "Marketing dashboards report high conversion rates, yet the finance team sees no corresponding growth in net revenue," it states.

The supporting figures are drawn from several surveys. According to the CMO Survey, as reported by MarTech, 65.7% of chief marketing officers name data integration as their biggest martech management challenge. The guide cites Funnel's own blog for two more: 55% of US marketers believe poorly integrated data has caused revenue loss, and 34% of CMOs say they do not trust their data at all.

The trust problem extends to how data is used. According to Gartner, 33% of marketers admit cherry-picking data to support choices they had already made, 24% rely on gut feeling and 24% reject data-driven recommendations altogether. The guide notes that these figures are self-reported, so "the actual figures may be worse." Its reference list dates that Gartner survey to 2024; the Gartner release it links to is dated September 15, 2022.

Two numbers from the IPA Databank frame the financial stakes. Return on investment has risen by roughly 4% since COVID, the guide states, while net profit is down 11%. "In other words, ROI looks better on paper while the actual financial contribution of marketing has declined."

Budgets have not grown to meet expectations. According to Gartner's 2025 survey, marketing budgets flatlined at 7.7% of company revenue, 59% of CMOs reported insufficient budget to execute their strategy, and 61% said marketing was viewed as a profit center, up from 54% a year earlier.

Attribution: credit, not cause

The guide's sharpest argument concerns attribution, which it calls "the default measurement framework for digital marketing for more than a decade." Its objection is not that attribution is wrong, but that it answers a narrower question than most organisations think. "The distinction between credit allocation and causal measurement sits at the very heart of why marketing dashboards can look strong while growth stagnates," according to the executive summary.

Usage of simple models remains high. Citing ProperExpression and OWOX BI, the guide states that 41% of marketers use last-touch attribution and 44% rely on first-touch. It walks through the rules-based alternatives - linear, time-decay and position-based, which assigns 40% of credit to the first touch, 40% to the last and spreads the remaining 20% across the middle - and concludes that none of them "can be validated against ground truth because they are assigning credit, not measuring impact."

Multi-touch attribution was supposed to fix that. The options have since narrowed. Google announced in April 2023 that it would remove the first-click, linear, time-decay and position-based models from Google Ads and Google Analytics 4, noting that fewer than 3% of Google Ads web conversions used them, and Merchant Center followed with the same four models in November 2023. That left last-click and data-driven attribution.

Google's data-driven model gets more nuanced treatment than the rest. It "uses a counterfactual approach, comparing the probability of conversion with and without each touchpoint to assign credit," the guide acknowledges, before listing its constraints: a 90-day lookback window that truncates longer journeys, credit assignments marketers can see without the reasoning behind them, and no view of offline activity, engagement inside walled gardens such as Meta and TikTok, or CRM pipeline interactions. "A more sophisticated algorithm distributing credit across an incomplete picture of the customer journey doesn't close the gap. It just allocates within it more precisely."

The guide's illustration is a familiar one in performance marketing. "A retargeting ad might receive last-touch credit for a sale, but if the customer was already heading to checkout, the ad intercepted the conversion rather than caused it."

It also points at the platforms' response to weaker tracking. Google and Meta have invested in modeled attribution, using machine learning to estimate conversions that tracking no longer sees. That fills gaps, the guide states, "but it also means marketers are increasingly relying on the platforms' own models to evaluate the platforms' own media." Its recommendation is unambiguous: attribution suits optimisation within digital channels, but should not be used as evidence of causal impact or as the primary metric reported to senior executives.

Out-of-date platform details

Some of the context around attribution has aged. The guide says Google "has delayed full cookie deprecation in favor of user choice prompts." Google in fact reversed course on cookie deprecation in April 2025 and said it would keep third-party cookies in Chrome. In October 2025, it went further and said it would retire most Privacy Sandbox technologies, including the Topics and Protected Audience APIs, after adoption failed. The guide's chapter on targeting nonetheless presents those two APIs as "the platform-level response" to declining third-party tracking.

Experiments and their failure rate

Where attribution assigns credit, experiments test cause. The guide quotes economist Hal Varian's description of randomised trials as "the gold standard for causal inference," because a randomly assigned control group serves as an unbiased estimate of what would have happened anyway.

The chapter is candid about how often tests disappoint. Citing Stefan Thomke of Harvard Business School, it states that only 10% to 20% of Google's and Bing's online experiments generate positive results, and that for every successful online experiment "nearly ten do not." Booking.com is the cultural example: 75% of its core employees take part in experimentation. The obstacle, the guide argues, is rarely technical. Running a trial means deliberately withholding marketing from part of an audience, and Thomke's line about the HiPPO - the highest-paid person's opinion - is cited as the thing that stalls it.

For A/B testing, the numbers are sobering. Most teams use a 95% confidence threshold, yet only about 20% of experiments reach it, according to the guide. Rules of thumb such as 5,000 visitors per variation and at least 100 conversions are described as "helpful guardrails rather than hard requirements." Adoption is wide: 77% of organisations run A/B tests on their websites, 60% on landing pages, and 93% of US companies on email. According to the guide, a test usually needs between two and six weeks: anything shorter risks catching a blip, anything longer invites seasonal noise.

Broadcast and outdoor media cannot be split person by person, so the guide turns to geography. Meta's open-source GeoLift builds a synthetic control from a weighted combination of untreated markets and needs no user-level tracking. The most striking result in this section comes from a case study published by Measured, a measurement vendor: a retailer's known-audience holdout test found its catalog programme delivered 14% incremental lift in revenue, against a 40% contribution claimed by vendor attribution reporting.

Incrementality and the walled-garden problem

Incrementality is described as the question at the centre of experimentation: how many conversions would not have happened without the marketing. According to the guide, over half of performance marketers now use incrementality testing and 36% plan to increase their investment. Those figures come from INCRMNTAL, a company that sells incrementality measurement. A typical test holds back 10% to 20% of the target audience.

The chapter introduces iROAS, which counts only revenue from conversions that would not otherwise have occurred. The threshold that matters is 1.0: above it, a campaign generates more incremental revenue than it costs. The guide cites iROAS results "ranging from 253% to 1,609% across advertisers"; its source is a Moloco blog post about retail media advertising. A Think with Google example puts a beauty brand's incremental ROAS on Performance Max at £6 for every £1 invested.

Platform lift studies have become cheaper to run. The guide states that Google lowered the minimum spend for incrementality experiments from $100,000 to $5,000 and refined the methodology to deliver conclusive results 50% more often. Its reference list dates that change to 2024. Google announced the cut from thresholds approaching $100,000, along with up to 50% more conclusive results, on November 11, 2025, after PPC Land had reported a $5,000 minimum for Google Ads incrementality tests in May 2025.

The guide is blunt about the limits. Holding back ads forgoes revenue during the test. Smaller advertisers may never reach statistical significance. A single test captures a window, not a trend. And there is a structural conflict: "When Google or Meta runs a lift study measuring the effectiveness of campaigns on their platform, they are both the measurement provider and the media vendor. They're essentially grading their own homework."

Marketing mix modeling's second life

Marketing mix modeling dates from the 1960s, and its revival is a privacy story. Because it works on aggregate spend and outcomes rather than individual journeys, it survives the loss of cookies and device identifiers. The guide collects adoption figures: 49% of marketers worldwide use it, according to EMARKETER; 56% of US ad buyers were putting more focus on it heading into 2025; The Drum reported that 76% of UK advertisers use it; and a Feld-M survey found 61.4% of marketers want better or faster models.

The technical core is adstock, the carryover of advertising effects after spend stops. Brand campaigns tend to decay more slowly than performance campaigns, which is why, the guide argues, mix models are better suited than attribution to capturing upper-funnel effects. The models usually work with "a few hundred observations," which makes them sensitive to how many variables are included.

The guide contrasts the slow, opaque annual projects of the past with what it calls a "living system." It cites best practices the IAB published in December 2025 recommending weekly data refreshes and monthly model retraining, and credits open-source tools such as Meta's Robyn and Google's Meridian with moving modeling in-house. Google opened Meridian to all marketers in January 2025; PPC Land has since examined who benefits when a platform builds the model used to judge its own media, the same conflict the guide identifies in lift studies but does not apply to these tools.

On limitations, the guide is direct. Mix models are "correlation-based not definitively causal." If search spend always rises in the busiest month, the model may credit search for natural demand. Channels that scale together are hard to separate. And the method lacks granularity: it is "virtually useless for deciding which specific ad set to pause this afternoon." According to BCG, 80% of marketers said robust scenario-planning capabilities would be the main driver of wider adoption.

That caution matches recent research. In August, a Zalando researcher found that a standard mix model overstated paid search ROAS by 2.5 times on simulated data, reporting 10.61x against a true 4.20x.

The next dollar

The guide's chapter on saturation and marginal returns may be its most practical for search and social buyers. Every channel, it argues, has a point where the next dollar earns less than the last. The first thousand dollars might generate 100 conversions, the next 80 and the one after that 50, because the most responsive people convert first and frequency rises among the rest.

The guide separates four ratios. ROAS is revenue divided by ad spend. Profit on ad spend (POAS) uses profit, accounting for margins and variable costs. The marketing efficiency ratio (MER) divides total revenue by total ad spend. Marginal ROAS divides the extra revenue from the next increment of spend by that increment.

The worked example is illustrative, not drawn from campaign data. A channel spends $1,000 and generates $3,000, a ROAS of 3.0. Spend rises by $500 in expectation of $1,500 more. Revenue grows by only $200. "Marginal ROAS on that incremental $500 is 0.4, or $0.40 per dollar. The extra spend is unprofitable even though the overall return still looks strong." By PPC Land's calculation, the blended ROAS in that example falls from 3.0 to about 2.13 - lower, but still comfortably above 1.

That raises the question the guide wants budget holders to ask. If one channel sits at a marginal ROAS of 0.5 and another at 2.5, money is being left on the table, according to the guide, and the remedy is to shift spend until marginal returns are roughly equal, using response curves from a mix model. Waitrose's experience, as described to The Drum, is cited as the shift in thinking: there is no single source of truth, "but a suite of truths."

Forecasts, finance and the brand problem

Forecasting is presented as the point where measurement "starts earning its keep with finance teams." The guide's examples are hypothetical but specific: a forecast that $2 million of brand TV spend "will generate between $5M and $7M in incremental revenue over 12 months, based on these assumptions and subject to these constraints" is described as far more useful than a single-point ROAS. According to BCG, only a third of respondents track the long-term impact of investment on brand equity.

The gap between marketing and finance is quantified through research by BCG and Google reported by The Drum: only 11% of advertisers use shared KPIs between the two functions, and only 18% actively bring effectiveness insights into cross-team decisions. The IPA Bellwether report found that only 53% of marketers formally track financial return. The IPA also warns that ROI is "a potentially dangerous way to measure effectiveness" because cutting the budget is an easy way to raise it in the short term.

The guide lists customer acquisition cost, customer lifetime value, payback period and contribution margin as the translation layer finance teams expect. It states that a CAC-to-CLV ratio "below 1:3 often signals" overpaying for customers, a phrasing that leaves the direction of the ratio ambiguous.

Brand measurement gets its own chapter, anchored in Les Binet and Peter Field's analysis of the IPA Effectiveness Databank. Their central finding is an optimal split of roughly 60:40 between brand building and sales activation. Long-term brand investment delivers approximately twice the profit of activation alone, and emotional campaigns are almost twice as likely to drive top-box profit growth. Brand effects can take six to twelve months to show, which the guide calls an accountability problem for organisations on quarterly cycles. Measuring only short-term ROAS, it warns, undervalues brand activity and leads to overinvestment in activation, slowly depleting the brand equity that makes activation work.

Targeting, uplift and the questions numbers can't answer

On targeting, Nielsen Digital Ad Ratings data cited in the guide puts the average on-target rate for digital ads in North America at 65%, meaning more than a third of impressions miss the intended audience. Nielsen's One Ads meta-analysis found the share of campaigns using closed targeting doubled from 15% to 30% between 2022 and 2024, campaign frequency ranged from 1.01 to 48.06, and only 18% of digital campaigns used more than one platform or publisher.

The section on AI-driven targeting relies on weaker sources. The claim that Meta's Advantage+ campaigns are "the primary growth engine for 65% of US advertisers" is referenced to the Insider Intelligence homepage rather than a specific report. Figures on a 32% drop in cost per acquisition from consolidated Meta campaigns and 113% higher ROAS for broad targeting over lookalikes come from Madgicx, which sells Meta ads software. The statement that AI Max early adopters saw "up to 27% higher conversion volume with no increase in spend" is cited to an agency blog. Google's own figures, published with the AI Max launch in May 2025, were 14% more conversions or conversion value at a similar CPA or ROAS, rising to 27% for campaigns mostly using exact and phrase match keywords - and independent tests later that summer found more modest results.

The guide's more durable point is about what changes when algorithms choose the audience. The question shifts from "did we reach the right people?" to "did the algorithm find more valuable people than we would have targeted manually?"

Uplift modeling, the next chapter, tries to answer that at the level of individuals. It sorts customers into four groups: Sure Things, who buy regardless; Lost Causes, who never will; Persuadables, who buy only when marketed to; and Sleeping Dogs, for whom marketing reduces the chance of a purchase. The catch is fundamental. As statistician Donald Rubin set out in 1974, nobody can observe the same customer both seeing and not seeing a campaign, so individual uplift can never be validated directly.

A chapter on qualitative research covers focus groups of four to twelve people, in-depth interviews of 30 to 90 minutes, and creative pre-testing systems such as System1's Spike and Star ratings. Its argument is that quantitative methods show what happened and, with effort, whether marketing caused it, but rarely why.

Maturity, triangulation and the case for agents

The guide's later chapters move from methods to organisation. It uses LiveRamp's four-stage model - Nascent, Building, Developing and Leading - and then cites BCG for the finding that 76% of advertisers across Europe, the Middle East and Africa remain in the first two stages, using the stage names from the LiveRamp model. The more striking BCG figure is that, of the organisations that invested in experiments, "68% failed to use the results to change a single decision." The guide's verdict: "Organizations are buying the engine but forgetting to build the steering wheel."

BCG's research also found that technical accelerators, such as data infrastructure and modeling capabilities, have about twice the impact of organisational ones on business outcomes. The guide adds a caution attributed to BCG's Joanna Stringer in The Drum: if the finance team does not believe the numbers, insights "might as well be dust."

Triangulation is presented as the industry's standard answer. Google's framework assigns attribution to the micro level, incrementality to the meso level and mix modeling to the macro level; BCG describes a "four-legged stool." The guide insists that running several models is not enough. When a mix model says a channel underperforms and a lift test says it drives significant sales, "Simply averaging the two answers is intellectually lazy. Discarding one is wasteful." It describes three reconciliation mechanisms: informed starting assumptions, also known as Bayesian priors, which let experiment results seed a mix model; multi-criteria evaluation, which flags a model that assigns 80% of contribution to a channel receiving 5% of spend, however good its statistical fit; and governance processes that decide which signal wins for a given decision.

The chapter on econometrics frames every method as an attempt to estimate the same unobservable counterfactual. Varian's example is cities with a strong interest in hockey, which show both high advertising for hockey equipment and high sales of it. A simple regression would credit the ads, when the cause is the interest. The guide lists randomised trials, difference-in-differences, regression discontinuity, instrumental variables and synthetic controls as the toolkit.

The final chapter turns to agentic AI. It cites a Precedence Research forecast that the agentic AI market will grow from $7.55 billion in 2025 to $199.05 billion by 2034, a compound annual growth rate of 43.84%, and Gartner data from 2024 showing 65% of CMOs expect AI to dramatically change their role within two years. According to BCG, only 5% of marketing leaders who use generative AI solely as a tool report significant gains on business outcomes.

Two industry milestones are cited, both to the same WebProNews article. The guide says PubMatic launched AgenticOS in December 2025; PubMatic announced it on January 5, 2026. It also cites the IAB Tech Lab's January 2026 roadmap extending OpenRTB, AdCOM and VAST for agent-based systems.

The guide is cautious about autonomy. A system that recalibrates models automatically is valuable, it argues, but "A system that automatically reallocates millions in spend based on those recalibrations, without human review of the assumptions, is risky." It raises the accountability question directly - "who is responsible? The data team that built the system, the marketing leader who approved its use or the vendor who sold it?" - and warns that organisations building on "a single vendor's agentic platform may find it difficult and expensive to switch."

How the evidence holds up

The guide's structure is sound and its caveats are frequent. Its sourcing is uneven.

More than a dozen of its 100 references point to blogs published by software vendors and agencies, including several companies that sell measurement products, such as Lifesight, Measured, INCRMNTAL, Sellforte and Prescient AI. Some figures are cited without the original study: the 41% and 44% attribution usage numbers come via Lifesight, and the Advantage+ figure via a homepage.

Funnel's three self-citations deserve attention. One supports the claim that "At Cannes Lions 2024, industry experts estimated that just 6% of advertising spend drives real results." The reference for that sentence is Funnel's own blog post on marketing mix modeling, not a Cannes session or a named expert. The guide does qualify the claim, adding that the other 94% is not necessarily wasted.

Several references share a URL for different claims. The same Google support page is cited for both the $5,000 threshold and a separate Google and BCG finding that 80% of US senior marketing analytics professionals saw a high revenue impact from incrementality insights. The same WebProNews article backs both the PubMatic and IAB Tech Lab claims, and the same EMARKETER article supports both mix-model adoption rates and a point about AI transparency. Dates in the reference list are also unreliable: beyond the Gartner survey, the Google threshold change is labelled 2024 when Google announced it in 2025.

None of this undermines the guide's core argument, which is well established and does not depend on any single statistic. It does mean that individual numbers should be traced to their original sources before they are repeated in a board presentation.

Why this matters for marketers

The guide arrives as the measurement debate has become crowded. On October 20, the IAB released a draft Cross-Channel Measurement Decision Guide for public comment, designed for exactly the situation Funnel describes: mix models, incrementality tests and attribution disagreeing, with the business decision as the starting point. The two documents arrive at similar answers from different directions.

Funnel's own data suggests few teams are there yet. Its previous survey edition in December 2025 found 86% of in-house marketers could not determine each channel's impact. A Winterberry Group survey sponsored by Lifesight found only 9% of US marketing executives said they excel at unified measurement. CIMM and the 4As found senior marketers struggling to trust their own cross-platform measurement.

The guide's warnings about attribution also echo evidence PPC Land has covered. An essay drawing on 663 Facebook experiments argued that standard ad measurement is often off by a factor of three. Silverback Strategies' chief executive ran a public apology campaign over inflated ROAS figures in November 2025. Platforms are responding with their own tools, such as TikTok's Attribution Portfolio, which raises the guide's own question about who is marking whose homework.

For paid search and social buyers, the most direct implication is the marginal-returns argument. Most reported ROAS figures are averages. The guide's point is that an average can stay comfortably profitable while the latest increment of budget loses money, and that the difference shows up only when spend is varied and the response measured. Funnel's data on US search spend rising 122% to $100 million a month across its customers is a reminder of how much budget now rides on that distinction.

It is also worth noting who is making the case. Funnel sells marketing data software, and a guide arguing for integrated, continuously refreshed measurement systems serves a company in that business. The guide names no product, and much of its advice - run experiments, distrust platform self-grading, report ranges rather than single numbers - would apply whichever vendor a team uses.

Timeline

Summary

Who: Funnel, a Stockholm-based marketing data company, which published the guide and promoted it in an email from Christopher Van Mossevelde, its content director. The guide draws on research from Gartner, BCG, McKinsey, the IPA, Nielsen, Google, academic economists and a range of software vendors.

What: The Definitive guide to marketing measurement, a 96-page, 18-chapter document that sorts measurement into diagnostic, causal and predictive questions and explains what each method - from dashboards and attribution to experiments, incrementality, marketing mix modeling, uplift modeling, triangulation and agentic AI - can and cannot show. It contains no original data and does not name Funnel's products.

When: Funnel promoted the guide today, November 9, 2026, in the third of five follow-up emails to its Playing defense survey of October 29. The guide itself is undated; its most recent cited events are from January 2026.

Where: The guide is published online by Funnel and is not tied to a single market, although many of its statistics describe the US, the UK and Europe, the Middle East and Africa.

Why: Funnel argues that measurement confusion comes from treating methods as interchangeable, that attribution assigns credit without proving cause, and that average returns can hide unprofitable marginal spend. Its sourcing leans on vendor blogs, includes self-citations, and describes Google's Privacy Sandbox APIs and cookie plans in ways that Google's 2025 decisions have overtaken.