Upwave today opened a closed beta for Predictive Brand Lift, a measurement approach that estimates campaign impact from modelling rather than from consumer surveys, following more than a year of development with selected customers.

In Short

Brand lift has always been measured by asking people questions: one group sees the ad, another does not, and the difference in their answers becomes the score. Upwave now offers a second route that skips the questions entirely and predicts the same outcome from patterns it has recorded across thousands of previous studies. That matters because small campaigns rarely gather enough survey responses to produce a usable result, so a modelled estimate is the only way many of them get measured at all.

What Upwave put into closed beta

The San Francisco company said today that Predictive Brand Lift entered a closed beta program after more than a year of development, testing, and refinement with selected customers. According to Upwave, the product is a surveyless approach to brand lift measurement, and it is currently available to a restricted group of customers and partners rather than to the general market.

The technical claim rests on three inputs. According to Upwave, Predictive Brand Lift applies the company's proprietary Bayesian modelling, artificial intelligence, and years of Observed Brand Lift measurement gathered across more than 3,000 brands, and it produces a lift estimate without fielding consumer surveys. The same figure appears in the company's own description of itself, which states that Upwave has measured brand outcomes across more than 3,000 brands.

Bayesian methods have become the default statistical grammar of advertising measurement over the past two years, and the reason is practical rather than fashionable. A Bayesian model carries prior beliefs into the estimation and updates them against new evidence, which allows it to return an answer where a purely frequentist approach would return nothing usable for lack of sample. Google's open-source marketing mix model, Meridian, rests on the same architecture, and its documentation frames priors as the mechanism by which teams import domain knowledge from experiments into a model rather than relying on correlations alone. Upwave's release does not describe what priors its model uses, how they were set, or how a prediction is validated against an observed result on the same campaign.

Two methodologies now sit alongside each other in the product line. Observed Brand Lift combines the survey methodology with analytics to produce direct measurement of campaign impact, according to the company, while Predictive Brand Lift covers cases where surveys are not run. Upwave frames the pair as complementary, giving advertisers the option to measure a campaign with whichever approach fits its objectives, audience, and scale.

Chris Kelly, chief executive of Upwave, tied the beta to a coverage argument rather than an accuracy argument. "We're opening the door to a new generation of Brand Lift measurement," Kelly said in the announcement. "Historically, various constraints meant advertisers could only measure a fraction of their campaigns. By offering both Observed Brand Lift and Predictive Brand Lift, we can extend Brand Lift to campaigns that previously fell below measurement thresholds, generating more insights and a more complete understanding of brand performance."

Kelly also addressed the development timeline. "Predictive Brand Lift has been in development behind the scenes for more than a year as we've validated the science and refined the methodology," he added. "As we expand access through closed beta, we're excited to work with more customers as we continue defining the future of Brand Outcomes measurement."

Upwave said it will continue working with customers over the coming months to validate performance across additional campaign types, industries, and media environments while refining the product ahead of broader availability. No date for general availability was given, and the release names no beta participants.

The threshold problem the beta is aimed at

The constraint Kelly refers to is arithmetic. A conventional brand lift study needs an exposed group and a randomised control group, a survey fielded to both, and enough completed responses in each to make the difference between them statistically defensible. Vendors therefore set impression floors, and campaigns below those floors go unmeasured.

Those floors have been falling, and the movement has been visible across the market for the past year. LoopMe set a one million impression threshold when it made its brand lift product available inside The Trade Desk Measurement Marketplace on 2 September 2026, against competitor minimums of three to four million impressions in many markets. Utiq priced its Deterministic BrandLift product in Germany from campaign budgets of 50,000 euros when it went live on 8 October 2025.

Lowering a threshold and removing it are different propositions. A one million impression floor still requires a survey to be fielded and answered; it simply requires fewer completes than a four million impression floor. A modelled estimate has no floor of that kind, because nothing is being asked of anyone. What it has instead is a dependency on the representativeness of the training data. A campaign type, industry, or media environment absent from the several thousand brands in Upwave's history is a campaign the model has no direct evidence for, which is precisely the gap the beta period is described as addressing.

Survey infrastructure under pressure

The economics behind surveyless measurement are not specific to one vendor. Response rates on panel-based research have been falling for years, panels cost money to maintain, and the wait for results has drawn consistent criticism from buyers.

That criticism is documented. Basis described the conventional model bluntly when it embedded Cint measurement into its platform on 16 June 2026, noting that traditional brand lift solutions sit outside the campaign execution workflow, require separate manual setup, and deliver results days or weeks after a campaign has ended. An advertiser receiving results ten days after a month-long campaign closed cannot act on them until the next cycle. Cint went further the following day, merging brand lift and sales lift into a single in-flight dashboard with transaction data supplying the second signal.

Speed has been the industry's first answer to the problem. illumin and Cint cut study launch time by roughly 90 percentwith an integration dated 14 May 2026, moving results from weeks to minutes. Coverage has been the second. DISQO's Outcomes Lift, named Vevo's preferred brand outcomes partner in May 2026, connects impression data to observed digital behaviours through a matched identity layer instead of asking panel members what they remember.

Upwave's beta represents a third answer: keep the survey methodology as one product, and sell a modelled estimate as the other. The difference is not cosmetic. Observed lift and predicted lift answer the same question with different evidentiary standing, and a buyer comparing two numbers in the same dashboard has to know which is which.

Where the number comes from matters

Broad dissatisfaction with measurement gives modelled approaches an opening. IAB research published in February 2026 found up to 75 percent of buy-side decision-makers rating current measurement methods as underperforming, a figure that has been cited repeatedly in vendor positioning through the year.

Dissatisfaction with an existing method is not evidence that a replacement is more accurate. Recent academic work on marketing mix modelling, covered in the same reporting, found synthetic-data conditions under which a mix model overstated paid search return by a factor of about 2.5, a reminder that a model calibrated on historical patterns inherits whatever correlations sat in that history. The same logic applies to a lift model trained on prior studies: it can only reproduce relationships it has already seen.

This is why the distinction between survey measurement, incrementality testing and modelled estimation is doing real work rather than serving as vocabulary. Brand lift measures survey-reported changes in recall, awareness or purchase intent between an exposed group and a control. An incrementality test measures behaviour instead, usually through a holdout study that withholds advertising from a randomised group. A predictive model measures neither directly. It estimates what a study would have found had one been run.

Upwave has not published the validation evidence behind the product. Kelly's statement that the science was validated over more than a year is a claim about internal process, not a disclosed result, and no accuracy figures, correlation coefficients, or backtest comparisons against observed studies appear in the announcement. Vendor-supplied performance claims of this kind have been a recurring feature of measurement product releases, and the absence of published methodology is the standard gap.

Upwave's position going into the beta

The company describes itself as the only business entirely focused on measuring and optimising brand lift driven by advertising, operating an artificial-intelligence-driven software platform with real-time measurement across CTV, digital, retail media, social, streaming audio, linear, and addressable channels. It is based in San Francisco and New York and backed by Silicon Valley venture investors, according to the company.

That channel list explains part of the commercial logic. Streaming is where measurement demand has grown fastest and where survey delivery is hardest, because a television screen is a poor surface for a questionnaire. Upwave's benchmark data already circulates widely in connected television: research published with IRIS.TV in April 2025 recorded campaigns using IRIS-enabled data at twice the awareness lift, three times the ad recall and five times the favourability of standard CTV benchmarks, figures that have since been repeated in announcements from Viant, IPG Mediabrands, and Xumo. The company's tactic norms function as a reference point other vendors cite, which is a position with obvious value and equally obvious dependence on the underlying data remaining survey-grounded.

Upwave's integrations reach beyond streaming. Wondery built the company into its podcast measurement suite in May 2024, covering both baked-in and dynamically inserted audio advertising without additional tracking pixels. According to the company's own materials, Predictive Brand Lift precedes two other 2026 product moves: AI Campaign Insights Sessions, powered by an agent the company calls Bayes, dated 12 August 2026, and a deepened partnership with MNTN dated 18 June 2026.

What this changes for buyers

Measurement supply is consolidating around fewer, broader vendor relationships. The Video Advertising Bureau's refreshed measurement directory catalogued 20 vendors in July 2026, and demand-side platforms have been absorbing measurement into the buying interface rather than leaving it as a separate procurement decision. Google exposed 24 conversion lift metrics through its advertising API in August 2026, though both the lift resources and the associated insights service require account allowlisting.

Against that backdrop, a surveyless product creates a reporting question that will outlast the beta. If observed and predicted lift results appear in the same interface, campaign reports, agency scorecards, and upfront negotiations will carry numbers produced by materially different methods. Labelling conventions, not statistics, will decide whether the distinction survives the journey from dashboard to slide.

The other open question is pricing. Survey fielding is a variable cost that scales with completes; inference against an existing model is not. Nothing in today's announcement addresses how Predictive Brand Lift will be priced relative to Observed Brand Lift, and the closed beta gives no indication of when that will be settled.

Upwave has not said how long the beta will run, how many customers are participating, or which campaign types remain unvalidated. Those are the details that would let a buyer judge the product rather than the positioning.

Timeline

Summary

Who: Upwave, the San Francisco and New York based brand outcomes measurement company led by chief executive Chris Kelly, together with the selected customers and partners admitted to the beta.

What: A closed beta of Predictive Brand Lift, a surveyless brand lift product built on proprietary Bayesian modelling, artificial intelligence, and Observed Brand Lift data accumulated across more than 3,000 brands. It runs alongside the survey-based Observed Brand Lift rather than replacing it.

When: Announced today, 9 September 2026, after more than a year of development and testing with selected customers. Validation across additional campaign types, industries, and media environments continues over the coming months, with no general availability date given.

Where: Inside the Upwave platform, which covers connected television, digital, retail media, social, streaming audio, linear, and addressable inventory. Access is limited to selected customers and partners.

Why: Survey-based studies require impression volumes many campaigns never reach, leaving a large share of advertising unmeasured. A modelled estimate removes that floor, and the beta is positioned to test whether the estimates hold across campaign types the model has not been trained on.