Estimating incremental lift without user-level tracking

Park rangers who manage a mountain lake never tag individual drops of water. They track the lake's level day by day, measure how much each stream is pouring in, and keep an eye on rainfall and snowmelt. That's enough to tell which streams actually keep the lake full. The hard part comes in spring, when snowmelt sends every stream rushing at once and the lake rises fast. If a ranger opened a new channel from a nearby creek that same week, it'd be easy to give the new channel credit for water the snowmelt would've delivered anyway.

That's the position marketers are in now that user-level tracking has gotten so spotty. You can't follow every shopper anymore, but you can still see the whole lake: your total sales, your spend by channel, and the seasonal tides that move both. The question is whether your measurement can tell the difference between water your campaigns added and water that was coming regardless. That difference is incremental lift, and it decides whether your next budget increase grows the business or just pays for sales you would've gotten for free.

Key takeaways

• Incremental lift is the revenue your marketing added beyond what would've happened anyway, and you can estimate it from total sales and spend data without following individual shoppers.
• Privacy rules, Apple's App Tracking Transparency, and browser cookie blocking have left user-level tracking incomplete, and the pixel-based data that remains tends to over-credit channels that capture demand.
• Traditional marketing mix models (MMMs) also work from aggregated data, but their pie chart math struggles to separate ad impact from natural demand when brands spend more during busy seasons.
• Prescient's model treats marketing as a connected system in which campaigns build and convert an unseen pool of demand, which helps it credit the campaigns that created that demand.
• Because the model works with totals, it can capture halo effects across your website, Amazon, and retail stores without needing cookies or device IDs.
• In a benchmark with known right answers, our model's channel-level attribution error was about 70% lower than two widely used open-source MMMs, and its peak-season budget recommendation landed within about 1% of the ideal plan.
• Incrementality tests can add context, but they're accurate only for a specific place and time, so they work best as one input you can check against your MMM.

Why user-level tracking can't measure incremental lift anymore

For more than a decade, digital measurement ran on following people. Pixels, cookies, and device IDs let platforms and multi-touch attribution (MTA) tools stitch together who saw an ad, who clicked, and who bought. That chain has broken in a lot of places, and it was never built to measure incremental lift anyway.

Where the signal went

Several changes have chipped away at user-level data over the past few years. Each one removes a different link in the chain:

• App tracking permissions. Since iOS 14.5, Apple has required apps to ask permission before tracking people across other companies' apps and websites, and many users say no.
• Browser cookie blocking. Safari and Firefox block third-party cookies by default. Google ultimately decided to keep third-party cookies in Chrome, but that doesn't bring back the shoppers you can't see on other browsers and devices.
• Privacy laws. Regulations like the GDPR in Europe and the CCPA in California give people the right to opt out of tracking, or require consent before it happens, so even "trackable" audiences have gaps.
• Purchases you never could track. Shoppers who see an ad on their phone and buy at Target, Walmart, or on Amazon were never fully visible to your pixel to begin with.

The result is that user-level data now shows you a partial, uneven slice of your customers. And the slice you can still see isn't a random sample.

The last-click problem that was there all along

Even when tracking worked well, it mostly measured who was present at the moment of purchase. Last-click and many multi-touch models give the most credit to the final touchpoints before a sale, which tend to be branded search, retargeting, and other campaigns that capture demand that already exists.

Campaigns that create demand, like YouTube, connected TV (CTV), podcasts, and broad prospecting on social, usually do their work earlier and often without a click. Someone hears a podcast ad, remembers the brand a week later, and buys on Amazon. A pixel sees none of that. So user-level tracking has always tilted credit toward the bottom of the funnel, and losing signal only makes that tilt worse.

Why incremental lift needs a different question

Tracking tools answer the question, "Who touched an ad before they bought?" Incremental lift asks a harder one: "How much revenue did this marketing add compared to what would've happened without it?" Those are very different questions. A retargeting campaign can show up before thousands of purchases and still add very little, if most of those shoppers were already on their way to checkout.

You can't observe the "without it" version of your business directly, so it has to be estimated. The good news is that you don't need to see individuals to estimate it. You need to see the whole system clearly, which is where aggregated measurement comes in.

Why older aggregated models stumbled

MMMs have been around for decades, and they've always worked from totals instead of individuals: daily or weekly spend by channel, revenue, and outside factors like seasonality and promotions. That makes them naturally privacy-friendly. The problem is that many MMMs, including popular open-source ones, are built on math that has trouble separating ad impact from everything else.

The pie chart assumption

Most traditional MMMs treat revenue as a set of separate slices that add up to the total. In simplified form, it looks like this:

Revenue = base revenue + paid social contribution + search contribution + CTV contribution + ...

Each slice is estimated as if it works on its own. Base revenue covers things like seasonality, long-term trends, and holidays, while each channel gets its own response curve.

When spend follows demand, the math can't tell them apart

Marketers don't spend randomly. They raise budgets right when demand is already climbing, around holidays, product drops, sales events, and seasonal surges. That's the snowmelt problem from our lake. When ad spend and natural demand rise together, the pie chart formula has no way to know which one caused the jump in sales.

In fact, there are endless ways to split that revenue between base revenue and media channels that all match your sales history equally well. Data scientists call this an identifiability problem, which is a technical way of saying the data alone can't settle the question. The forecast can look great while the channel credit underneath it is badly off, and we've written more about why a model can accurately predict revenue and still misattribute it.

Statistical penalties pick an answer, not the right answer

Open-source MMM frameworks know about this problem, and they try to handle it with regularization, which adds mathematical penalties to discourage extreme answers. Bayesian approaches do something similar with "priors," or starting assumptions.

But a penalty doesn't narrow all those equally good-looking splits down to the true one. It just gives the model a rule for choosing among them, like "prefer smaller numbers." That rule has nothing to do with how your marketing actually works. Adding more years of data doesn't fix it either, because if spend has always risen alongside demand, extra history doesn't give the model any new way to tell the two apart. It just makes the model more confident in whichever answer its penalties prefer.

How Prescient's model measures incremental lift from aggregated data

Our approach replaces individual tracking with something closer to how the park ranger thinks about the lake. Instead of slicing revenue into separate pieces, Prescient's model treats marketing as a system that changes over time, and it uses what we know about how marketing behaves to keep its answers realistic.

Modeling a pool of demand instead of individual shoppers

Underneath your sales numbers is something you can't see directly: a pool of demand and awareness for your brand. Your campaigns add to that pool, seasons and events raise or lower it, and shoppers draw from it when they buy. Revenue is the visible result.

Prescient's model estimates that unseen pool and how each campaign shifts it over time. Because base revenue and media effects come out of the same connected picture, the model doesn't have to arbitrarily decide how much of a holiday spike was "organic" and how much was "paid." And because it's working with the pool as a whole, it doesn't need to know who any individual shopper is.

Building real marketing behavior into the model

The model narrows down the possible answers using patterns that reflect how marketing really works, instead of arbitrary math rules. A few of the most important ones for measuring lift:

• Demand creation and demand capture. Some campaigns build interest, and others convert people who are already interested. Our model credits the campaign that created the demand, not just the one that happened to be there when someone bought.
• Halo effects across sales channels. Campaigns can drive revenue indirectly through another channel. A social ad someone sees without clicking can lead to a later Amazon purchase, and a CTV campaign can show up as more branded searches and in-store sales. Our model captures these from your sales totals, without cookies.
• Flexible response shapes. Many MMMs force every channel onto a preset curve that assumes returns start shrinking early. Our model lets each channel's response take the shape the data supports, so it doesn't call a channel saturated before it really is. (Our research on saturation assumptions explains why that matters.)
• Carryover that varies by channel. Ads keep working after they run, and some channels keep working longer than others. The model lets that lingering effect differ by channel instead of applying one fixed decay formula to everything.

None of these are hard-coded outcomes. They're guardrails that keep the model's estimates of lift within the range of what's realistic for a marketing system.

What the data needs to include

Since aggregated measurement looks at the whole lake, it works best when the data covers all the streams. That means more than ad spend and website revenue. The most useful inputs include:

• Spend by channel and campaign over time, so the model can see when and where marketing pressure changed.
• Revenue from every place you sell, including your website, Amazon, and retail partners, so halo effects have somewhere to show up.
• Promotions, pricing changes, and product launches, so discount-driven demand doesn't get mistaken for ad impact.
• Inventory and distribution changes, so a stockout or a launch in hundreds of new stores doesn't get pinned on whichever channel was spending at the time.

It also helps to have natural variation in your spending, like campaigns that start, pause, or scale on different schedules. The more your channels move independently of each other and of your busy seasons, the more clearly any MMM can separate their effects.

Where incrementality tests fit in

Geo-lift tests and holdout studies are a popular way to measure lift without user-level data. They turn ads on or off in certain regions and compare sales to regions where nothing changed. They can be a useful source of context, but it's important to understand what they can and can't tell you.

Locally accurate, globally inaccurate

An incrementality test measures one channel, in specific regions, during a specific window. At Prescient, we describe tests as locally accurate but globally inaccurate. They can give you a reasonable read on that one moment, but they can't tell you how the same channel performs in other markets, other seasons, or alongside every other campaign you're running. Regional differences, local events, and shoppers moving between test and control areas can also skew results.

Tests also don't fix the structural problem described earlier. If an MMM's basic setup can't separate ad impact from natural demand, plugging a test result into it just forces that result through the same flawed math.

How Prescient uses test results

We don't treat incrementality tests as the final word, but we don't ignore them either. When a brand has run a test, Prescient shows how accurate the model is with and without that result included, so the brand can see whether the test makes the overall picture better or worse. The brand makes the final call. If a team trusts a test even though it lowers the accuracy score, it can be incorporated as a prior.

How our model performed against known right answers

Testing attribution with real brand data is tough, since nobody knows the true answer. So our research team built a simulated marketing environment in which virtual shoppers responded to spend, seasons, and events, and every channel's true contribution was known in advance. The simulation was built to include the hardest case for aggregated measurement: spend that ramps up right alongside peak demand.

We ran Prescient's model and two widely used open-source MMMs across 10 separate datasets. Because not every MMM measures at the campaign level, the comparison was run at the channel level. Here's what we found:

• About 70% lower attribution error. Our model's average error was about 31,000 in the simulation's revenue units, compared to about 99,000 and 110,000 for the two open-source models.
• Consistent credit across the funnel. On a scale where 1.0 means perfect agreement with the true results, our model held per-channel scores between 0.90 and 0.97 for both demand creation and demand capture channels. The open-source models' upper-funnel scores often fell below 0.6.
• Less holiday inflation. The open-source models often overstated holiday effects by 20% to 30%, while our model's base revenue estimates stayed close to the true values.
• Better peak-season budget calls. During a simulated period modeled on Black Friday and Cyber Monday (BFCM), one open-source model recommended overspending by about 81% and the other recommended underspending by about 11.5%. Our model's recommendation landed within about 1% of the ideal budget.

The takeaway: aggregated data can support accurate estimates of lift, as long as the model's structure matches how marketing really works.

How the three approaches compare

The table below compares user-level attribution, traditional open-source MMMs, and Prescient's model on the issues covered above.

Where Prescient comes in

Prescient's marketing mix model was built for omnichannel brands that need to measure lift across digital, retail, and offline channels without relying on user-level tracking. It works from aggregated spend and sales data, measures halo effects across your website, Amazon, and retail partners (including retail halo effects), and credits the campaigns that created demand instead of the ones that simply showed up at checkout. Daily model updates and campaign-level attribution mean you can see what's working while you can still act on it.

That clarity carries into planning. Our Media Forecaster feature uses what our MMM has learned about your seasonality to show what happens when you scale a channel or shift budget between channels, so your next move is based on the lift your marketing actually adds. If you want to see which streams are really filling your lake, book a demo.

FAQs

What is incremental lift in marketing?

Incremental lift is the extra revenue, sales, or conversions your marketing generated beyond what would've happened without it. If you earned $1.2 million during a campaign and estimate you would've earned $1 million without it, the incremental lift is $200,000. Because you can't directly observe the "without it" version of your business, lift always has to be estimated, either through a model, a test, or both.

Can you measure incremental lift without cookies or pixels?

Yes. Marketing mix models estimate lift from aggregated data, like daily spend by channel and total sales across every place you sell, so they don't need cookies, pixels, or device IDs. The catch is that the model's structure matters a lot. A model that can't separate ad impact from seasonal demand will still produce numbers, but they may give credit to the wrong channels.

Is MMM better than multi-touch attribution for measuring lift?

For measuring lift across your whole business, an MMM has some clear advantages. It isn't affected by tracking gaps, it can capture purchases on Amazon and in retail stores, and it doesn't automatically favor the last touchpoint before a sale. Multi-touch attribution can still be useful for in-platform optimization, but it measures who touched an ad, not how much revenue that ad added.

Are geo-lift tests the most accurate way to measure lift?

Geo-lift tests can give a reasonable read on one channel in specific regions during a specific window, but they're locally accurate and globally inaccurate. Results can be skewed by regional differences, local events, and shoppers moving between test and control areas, and they don't tell you how a channel performs in other seasons or markets. They're most useful as one input you can check against your MMM, rather than as the final answer.

How much data do you need to measure lift with an MMM?

There's no single magic number, but variety in your data matters as much as volume. An MMM learns the most when channels start, pause, and scale on different schedules, and when your data includes sales from every place you sell plus promotions and distribution changes. Years of data in which every channel moves up and down together can be less useful than a shorter history with more independent movement.