Google published an Ads Decoded episode on September 2, 2026 in which John Chen, senior director of product management for ads measurement, described the lookback window that governs conversion counting as arbitrary, and set out the mechanism the company built to replace it.

In short

Google released a video interview explaining how its three measurement tools - attribution, incrementality testing and marketing mix modelling - are supposed to work together rather than compete. The company also detailed Qualified Future Conversions, a metric that watches for a follow-up action within seven days of a click and then counts sales for up to 180 days afterwards. If you run campaigns with long consideration cycles, this changes which purchases show up in your reports at all.

A 29-minute case against the single number

The episode, published on the Google Ads YouTube channel and running 29 minutes and 44 seconds, pairs Ginny Marvin, the company's ads product liaison, with Chen. It is available at https://www.youtube.com/watch?v=33cFdnC6k9A. At the time of writing it had drawn 1,988 views and 27 likes from a channel with 912,000 subscribers, and comments were disabled. A companion post on Google's Ads and Commerce Blog carried the same date and the headline "Build a measurement stack you can rely on to steer your campaigns."

The framing is unusually blunt for platform-published material. According to the blog post, evaluating a media strategy is difficult when real-time attribution, periodic incrementality tests and media mix models conflict. That is a concession that the three systems Google sells, gives away and documents do not produce a single agreed number, and that the divergence is normal rather than a fault to be repaired.

Chen returns repeatedly to an aviation metaphor. According to Chen, data is fuel and without it the aircraft cannot leave the ground, while attribution, incrementality and mix modelling are each a separate gauge. The conclusion he draws is that no single instrument lands the plane.

The distinct clocks each tool runs on

The episode separates the three approaches by cadence rather than by accuracy, which is the more useful distinction for anyone building a reporting calendar.

Attribution updates continuously. According to Chen, the number in the interface changes daily and supports real-time decisions, which makes it effective for tactical adjustment. Incrementality testing runs on a different clock entirely. Chen describes holdback studies as something an advertiser would run quarterly or annually, producing a periodic read on causality rather than a continuous one.

Mix models sit at the aggregate level. According to Chen, MMMs have existed for a long time, run every quarter or every year, and examine performance at an aggregate rather than an event level. He is explicit that the three sometimes agree and sometimes do not.

More data has not resolved the disagreement; it has multiplied the places disagreement can surface. According to Chen, measurement is getting easier because more tools and more data exist than ever, and harder because the data does not always sing the same song. He offers a concrete failure mode: real-time attribution can register something an incrementality test run three months earlier did not, while a mix model sees the big picture but cannot allocate at the keyword or creative level.

What Google means by data strength

Marvin raises the concept of data strength, which she attributes to an earlier Ads Decoded conversation with Eleanor Stribling, and which Google Analytics has spent the past year rebuilding around. Chen's answer is a checklist of transaction surfaces rather than a description of tooling.

Web transactions are sufficient for a business that only sells on the web. An app-based share of sales requires app data. A retail or partner presence requires offline data, and Chen names Macy's and Amazon as examples of places where a transaction can complete outside an advertiser's own systems. According to Chen, missing any part of that picture means missing part of the equation, because if the final transaction is invisible the effect cannot be measured, and what cannot be measured cannot be optimised against.

Above the purchase line, he lists add-to-cart events and site visits as upper-funnel actions that matter. On the media side, he draws a distinction that carries technical weight: click data is comparatively easy to assemble, view data is not, and cross-publisher view data has to be sourced separately from each platform before it can feed a single model. Google's own pipeline has documented boundaries here, including a seven-day gate on offline conversion uploads that keeps late-arriving records out of the data-driven model even when they appear in standard reporting.

Inside Qualified Future Conversions

The most specific product content in the episode concerns Qualified Future Conversions, introduced at Google Marketing Live 2026 alongside the Meridian integration into Analytics 360 on May 20, 2026. Chen's account of the reasoning is the clearest public description of the mechanism to date.

The starting complaint is about the lookback window itself. According to Chen, the practical difference between a conversion recorded on day 29 and one recorded on day 31 is negligible, and the only thing separating them is an artificial boundary determining which conversions get counted and optimised towards. "We think that's probably a little bit too arbitrary," he says.

The replacement is a two-stage test. Within seven days of clicking an advertisement, the user must take a further action that indicates the advertisement influenced them. Chen gives three examples: an attributed branded search, a visit to the advertiser's site, and an add-to-cart event. His illustration uses a Nike advertisement followed within a week by a search for Nike or for Jordan 1s on google.com, which he characterises as a strong signal that the advertisement played a substantial role.

Once that seven-day qualifying signal fires, the counting horizon opens. According to Chen, the system then looks at how many conversions occurred within the next thirty, ninety or one hundred and eighty days, up to six months. The design intent is stated plainly: a clear signal inside one week licenses the counting of conversions up to 180 days out.

The consequence for reporting is structural rather than cosmetic. A metric that credits a click with a purchase made nearly six months later will produce different channel rankings than one that stops at 30 days, and the difference will be largest exactly where consideration cycles are longest. Chen notes that shorter cycles also benefit, because signals previously invisible become countable. Marvin extends the logic to lead generation, where qualified leads along the customer journey would enter the same data set.

Meridian's descent from the data science team

The second product thread concerns access. Mix modelling has historically required an engineering and data science team simply to stand a model up before anyone could examine the output, and Chen acknowledges that legacy directly.

He describes three successive attempts to lower that barrier. Meridian was released as open source so that a smaller company with a handful of engineers or data scientists could run it. Meridian Studio, which arrived earlier this year on Google Cloud Platform, lets the module run on BigQuery data the advertiser already holds. Meridian inside Google Analytics is described as an attempt to make the process turnkey: log in, follow a few steps, and, assuming the data connections already exist, run the model. Marvin confirms this refers to Google Analytics 360, the enterprise tier.

Chen's stated objective is that time shifts from building and validating the model to reading its output. Marvin adds that the Scenario Planner, introduced earlier in the year, was built for marketers rather than data teams, and describes that work stream as bringing Meridian to marketers instead of requiring marketers to route through data teams.

She also relays an anecdote from an agency chief executive whose technical staff built a client model, identified undervalued channels, redistributed budget and recorded the client's best month. Neither the agency nor the client is named, and no figures accompany the claim.

Chen adds a note on maintenance: reach and frequency innovations and Google query volume features have already been folded into Meridian, with the platform versions intended to spare teams the work of importing each update into their own model. Those mathematical foundations were extended in September 2025 with non-media variables and channel-level contribution priors, and again in May 2026 with Meridian GeoX and Meridian Studio.

A bonus episode released alongside the main conversation features Patrick Gilbert and Nechama Teigman of AdVenture Media discussing their implementation of Meridian. According to the blog post, the companion episode covers how AI coding tools remove technical barriers, why data variance matters more than budget size, and how to plan a strategic test.

When the gauges disagree

Marvin puts the tug-of-war question directly: when an incrementality test and platform attribution conflict, which becomes the north star.

Chen declines the premise. According to Chen, a north star implies a single direction, and using either tool alone for that purpose would be difficult. His case against relying on incrementality alone is economic rather than methodological. Causal studies answer the right question, he says, telling an advertiser what the advertisement delivered that would not have arrived otherwise. But they are bounded in three ways: they take effort and time to set up without contamination, they reflect only the period in which they ran, and they cost money. According to Chen, every held-back advertisement forfeits sales that might otherwise have occurred, which affects the bottom line and is why most customers cannot run holdouts continuously.

He raises seasonality separately, asking how relevant a study run in spring would be for Black Friday.

The case against attribution alone is the mirror image. Attribution runs continuously and streams data suited to real-time analysis, but according to Chen it cannot say which of the counted conversions were incremental, or whether the conversion would have occurred without the advertisement. His conclusion is that following either as a north star leads in materially different directions, and that the practical answer is to use both alongside judgement and experience to triangulate.

The divergence Chen describes has a name in the econometrics literature. Endogeneity is the condition in which advertising spend correlates with the unmeasured drivers of demand, which is why an observational dashboard, a multi-touch model and a mix model can each be internally consistent and still disagree about the same campaign. Independent work has quantified the gap. Research by a Zalando scientist published in August 2026 found that standard mix modelling practice overstated paid search returns by roughly 2.5 times against experimentally established ground truth.

The three-step move off last click

Asked how to measure channels that generate no click, Chen identifies the problem as one of the industry's largest. Last click is easy, he says, and what it misses is upper-funnel media that primes a customer to convert. He names YouTube, TikTok and CTV as formats that a last-click model discards.

The sequence he sets out has three parts. First, move the attribution model off last click to a data-driven model that builds a full path across events and includes views as well as clicks, which requires sourcing view data from Facebook, TikTok and YouTube separately and consolidating it. Second, run incrementality studies to calibrate that attribution model, testing which media, formats, creatives and keywords proved most effective. Third, build a triangulation layer with a mix model running more frequently than quarterly.

The stated logic of that third step is that a mix model tends to value upper-funnel media more accurately while attribution tends to value lower-funnel activity more accurately, and that the two together cover the full funnel.

Cost data, connectors and the ROAS calculation

On the mechanics of getting cross-channel data into Google Analytics, Chen distinguishes three import types with different destinations inside the stack. Event-level connectors carry clicks and views from individual publishers, and he names Snapchat, Facebook and Pinterest as examples. That data feeds the attribution model. Aggregate-level imports carry longer historical series, and his example is two years of Facebook spend and impressions alongside Pinterest and CTV figures. That data feeds the mix model. Media cost forms a third category, which can also arrive in aggregate.

According to Chen, the three combined produce strong attribution results, strong mix model results, and, with the cost layer present, the ability to calculate return on ad spend and to run planning towards the best return at a given spend level.

Explaining a lower number to the client

The stakeholder problem is the one most practitioners face first. Marvin frames it precisely: how does a practitioner held to a return or cost-per-acquisition target explain to a client accustomed to one-to-one attribution that a holistic model showing lower immediate returns is more accurate.

Chen concedes the counterintuitive shape of the argument and reaches for a metaphor he credits to an unnamed practitioner. According to Chen, an advertiser that only harvests will eventually run out of things to harvest, which makes planting for the next cycle necessary. He accepts that a Black Friday return will look better than June spend, and offers three approaches.

An attribution model that credits the upper funnel adequately produces a return figure that can be defended as it stands. A causal study yields an incremental return figure that similarly accounts for upper-funnel contribution. The third is a deliberate manual adjustment: Chen describes an advertiser who, seeing a June return three times what Black Friday would deliver, reduces the target by a third on the understanding that the June spend pays off in November.

On the holiday period itself, Chen separates two requirements. Real-time data strength in the three to seven days before Black Friday determines whether optimisation systems can detect the differentiation happening in that window. The nine months preceding it determine which media planted the demand, a judgement he traces to incrementality tests and mix models indicating that an investment pays off across six to nine months.

Why this matters for the marketing community

The measurement stack Google described on September 2 is not new in its components. What has changed is that the company is now publishing a hierarchy in which no component is authoritative, at a moment when its automated bidding systems consume the output of all three.

That combination creates a practical tension. Attribution output feeds Smart Bidding directly, which means the model an advertiser selects is not only a reporting choice. Google has spent 2026 adjusting the surrounding parameters: conversion lookback windows became editable to any integer between 1 and 90 days on August 11, and attribution settings became configurable per conversion action in January. Qualified Future Conversions extends the same trajectory by decoupling the counting horizon from a fixed window altogether.

Scepticism about the underlying methods is well documented. IAB research published on February 7, 2026 found that up to 75 percent of buy-side decision-makers rated attribution, incrementality testing and mix modelling as underperforming. Meta reached a comparable conclusion from a different direction, placing randomised experiments above modelled attribution in the hierarchy set out in its suite of truth framework.

The governance question sits underneath all of it and the episode does not address it. Meridian is built by the company that also sells a large share of the media the model evaluates, a structural tension PPC Land examined through Meta's Robyn in April 2026. Chen's triangulation argument reads one way as an answer to that critique: three imperfect instruments constrain each other more effectively than one confident number. It reads another way as an observation that all three instruments now come from the same vendor.

Timeline

Summary

Who: John Chen, senior director of product management for ads measurement at Google, interviewed by Ginny Marvin, the company's ads product liaison, on the Ads Decoded series. A companion episode features Patrick Gilbert and Nechama Teigman of AdVenture Media.

What: A 29-minute video and accompanying blog post setting out how attribution, incrementality testing and marketing mix modelling are intended to operate together, with detailed explanations of Qualified Future Conversions, the seven-day qualifying signal that triggers a counting horizon of up to 180 days, the three routes into Meridian, and the data imports required to support each layer.

When: Published September 2, 2026. Qualified Future Conversions and the Meridian integration into Google Analytics 360 were announced on May 20, 2026 at Google Marketing Live.

Where: The Google Ads YouTube channel and the Google Ads and Commerce Blog, with the products spanning Google Ads, Google Analytics 360, Google Cloud Platform and the open-source Meridian repository.

Why: Automated bidding consumes the output of the measurement stack, which makes the choice of attribution model an operational decision rather than a reporting preference. Google is arguing that no single tool should be treated as authoritative, at the same time as it supplies all three tools in the recommended combination.