The $35M test that proved Meta wasn’t working for Uber

This is the marketing-science leg of Brandformance, co-hosted by Pranav Piyush and Sundar Swaminathan. Every two weeks they break down the concepts behind modern marketing measurement — incrementality, marketing mix modeling, brand versus performance — in plain, practical terms, and bust a few myths along the way.

Pranav Piyush is the co-founder and CEO of Paramark, where he’s building modern marketing measurement — incrementality testing and marketing mix modeling — for growth teams. A long-time marketing leader, he hosts Brandformance.

Sundar Swaminathan is a marketing data science and experimentation advisor to consumer-tech scaleups and the creator of the ExperiMENTAL newsletter and podcast. He previously built and led brand data science at Uber — measuring over a billion dollars of brand spend — and earlier worked on the debt desk at the US Treasury.

The gist

• Measurement doesn’t fundamentally change between brand and performance — it’s the same structure on different time horizons. What changes is how much certainty you need at your stage.

• The hard part of incrementality usually isn’t the math — it’s the organizational story and the willingness to act on it.

• Saturated brands advertise to sustain, not just to grow. Brand equity decays, so staying top-of-mind is the job.

• Daily dashboard-checking is mostly an “illusion of action.” Pick a weekly proxy you can actually steer on instead.

• AI makes great marketers superpowered and is a trap for juniors who skip the reps. Keep judgment, taste and ideation human.

The $35M test: when Meta wasn’t incremental

Sundar’s signature story is from his time leading rider-acquisition performance marketing at Uber in the US and Canada. Investigating weeks of swinging CAC, he found that for that specific channel and audience, Meta was virtually non-incremental — and the team handed roughly $35 million of budget back to the business.

The surprising part wasn’t the finding; it was the lack of friction. A ten-slide deck — no fancy regressions, just normalized CAC and spend moving in lockstep with seasonality — was enough to win approval to run an incrementality test. He credits Uber’s culture of experimentation, plus context he only learned later: the company had recently lost a large sum to ad fraud, so leadership was already primed to question what actually worked. His takeaway, repeated often: he faces more scrutiny over $100,000 decisions than $100 million ones, and that gap between insight and action is the real bottleneck.

Why saturated brands keep spending

When a brand is fully penetrated — everyone already knows it — what is media even doing? Sundar points to a natural experiment from leading brand data science at Uber: an Uber Eats campaign in the UK lifted awareness, and when the media was switched off, awareness fell back to pre-campaign levels within weeks. The lesson is that a lot of brand spend isn’t about growth; it’s about sustaining.

For a saturated brand, the job is staying top-of-mind so you’re the default choice — the app a rider opens first, the Coke that’s tied to pizza. Pranav frames the other side: brand equity isn’t a savings account that compounds on its own. It decays, so you have to keep depositing. Both agree the measurement question shifts with market share — early on, chase transactions and share; once you’re saturated, measure whether you’re sustaining the number-one spot in people’s minds.

Measurement doesn’t change — your time horizon does

The myth Sundar pushes back on hardest: that brand and performance require fundamentally different measurement. In his view it’s the same structure — same CPMs, same click-through dynamics — just observed on different time windows. A “transactional” ad with a CTA can still be a great brand ad; the real difference is often just whether you optimize for conversions or reach.

What actually changes is the level of certainty you need, which depends on stage and spend. A single-channel startup hunting product-market fit should be throwing things at the wall and looking for big, obvious pre/post swings — if you’re not seeing 20–30% movement, you’re not being bold enough. As you scale, you layer on more rigorous tools — diff-in-diff, geo tests, MMM — not because the toolkit is different, but because you can afford and need more confidence.

The illusion of action

Despite a decade of advances in incrementality, geo testing and MMM, marketers still open the dashboard every Monday to see what happened in the last 48 hours — while knowing that judging yesterday by today’s numbers is close to pointless. Why? Partly because the rest of the business trades weekly, so marketing is dragged onto the same cadence. And partly, Sundar argues, because humans are bad with silence; not looking at a dashboard feels like not doing your job.

His fix isn’t to go dark for six months — it’s to find the right proxy. Don’t stare at revenue daily; watch a leading metric that tells you whether you’re steering in the right direction, and make small weekly adjustments without mistaking activity for certainty.

A human-in-the-loop bet on AI

Asked about AI and agents, Sundar makes a deliberately contrarian bet: he keeps his newsletter and posts AI-free, and invests in the human part. He uses AI to speed up things he already understands — writing a near-perfect SQL query, drafting a summary, acting as a search engine — but won’t outsource ideation, because letting the model think for you reverts your work to the average.

The pattern he sees: AI makes great marketers superpowered and bad ones more replaceable. Analysts still need business context to know when an output “doesn’t feel right,” and juniors build that muscle by doing the work. Pranav adds his own framing — treat AI like a calculator (of course use it, but you’re still the operator), and let real life inspire the work rather than regurgitating what’s already on the internet. Both land on the same place: human-in-the-loop wins, because people don’t actually want to hand over the decisions that matter.

Quote snacks

• “I faced more scrutiny over a $100,000 decision than I do over a $100 million one.” — Sundar

• “Brand equity decays over time. It’s not a savings account that keeps compounding on its own.” — Pranav

• “Measurement between these various types of marketing doesn’t actually change. It’s the same structure, just different time horizons.” — Sundar

• “We as humans are not good with silence — and not looking at dashboards every day is that version of silence.” — Sundar

• “AI makes really great marketers superpowered. But if it becomes your thinking center, it’s an average brain by definition.” — Sundar

• “I’m betting big on the humans.” — Pranav

Why it matters

This is the brand-versus-performance debate viewed through the measurement lens. Sundar’s argument — that the two aren’t different measurement problems, just different time horizons and certainty thresholds — cuts through a lot of the religious war between brand and performance camps. It reframes the question from “which is better” to “what do you need to be confident of, at your stage, on what timeline.”

It also grounds the AI conversation. The hosts are bullish on AI as an accelerant but skeptical of handing over judgment — the same discipline that separates a real incrementality finding from a dashboard twitch. The throughline: respect the time horizons, keep a human in the loop, and measure what actually moves the business.

Practical next steps

Match your measurement to your stage. Early on, look for big pre/post swings; as you scale, layer on diff-in-diff, geo tests and MMM — same toolkit, higher certainty.

Pick a weekly proxy. Instead of staring at revenue daily, choose a leading metric you can actually steer on, and resist mistaking activity for certainty.

Pressure-test your biggest channels. Saturation and habit hide waste — run an incrementality test on the spend you’re most sure about.

Measure sustain, not just growth. For saturated brands, track whether you’re holding top-of-mind, because brand equity decays without investment.

Use AI as an accelerant, not a brain. Lean on it for SQL, drafts and lookups; keep ideation, taste and business judgment human.