Podcast episode

Episode 183: Ask Olivia Kory Whether AppLovin Ads Work & She Explains Attribution vs Incrementality

by Ken Rona ·July 27, 2026 ·Marketecture

attribution incrementality measurement performance-marketing retail-media

TL;DR

Marketecture host Ari Paparo and co-host Eric Franchi bring back Olivia Kory (Chief Marketing Strategy Officer at Haus, an incrementality-testing company) to discuss whether AppLovin's e-commerce ads actually work, how attribution differs from incrementality, and the state of AI-driven budget allocation. The episode wraps with quick takes on Google's Q2 earnings, OpenAI's rumored ad network, and the broken state of podcast measurement.

What was covered

Attribution vs. incrementality: Olivia Kory explained that attribution measures correlation (saw ad → bought thing), while incrementality measures causation — what would have happened without the ad, modeled as a randomized holdout test. She noted that Meta's new product calling itself "incremental attribution" is, in her view, an oxymoron.
AppLovin e-commerce performance: Kory confirmed AppLovin works, specifically for low average order value (AOV), impulse-purchase categories such as apparel and footwear. She attributed its success to unusually fast, short-term ad effects — similar to Meta — which satisfy growth marketers who need to see immediate ROI. AppLovin also opened its platform to third-party incrementality testing from the start of its e-commerce pilot (late 2024), which she said made validation easy.
Affiliate and retail media incrementality gap: Kory said affiliate channels are among the hardest to test because partners resist geo-holdout designs and make it difficult to pause offers. In the handful of tests Haus has run, they saw little incremental lift — though she flagged the small sample size. She grouped affiliate, Amazon search, and RMNs (retail media networks — ad products sold by retailers) together as channels where lack of rigorous measurement is limiting advertiser investment.
CAPI vs. pixels and optimization signal: Conversion APIs (CAPIs — server-side data pipelines that send purchase signals back to ad platforms) are increasingly preferred over browser-based pixels for privacy and coverage reasons, but Kory framed them primarily as optimization vehicles rather than measurement tools. Some advertisers are experimenting with sending only predicted-incremental conversions through CAPI to improve platform ML training.
Haus's "Architect" agentic budget tool: Kory described Haus's product roadmap: combining incrementality experiments, multi-touch attribution (MTA), and media-mix modeling (MMM) into a single causal attribution model, then surfacing budget reallocation recommendations via an AI agent called Architect. Current state: it recommends moves (e.g., shift dollars from Google Search to YouTube) and can execute them via API connections, but does not autonomously place new insertion orders.
Google Q2 earnings: Ari Paparo cited total revenue of $119.8B; search and other at $63.3B (+17%); YouTube at $11.1B (+13%); cloud at $24.8B (up ~82% year-over-year from $13.6B). Capex guided to $195–205B for the year. First quarter of negative free cash flow in a long time; stock took a hit.
OpenAI ad network speculation: Following a Business Insider report (citing a job posting that OpenAI quickly edited), the panel debated whether OpenAI is building an ad network or audience-extension product. Kory noted advertisers are already running into scale limitations on ChatGPT's ad inventory. Paparo's theory: OpenAI may be building something analogous to Google's FAN (audience network) using paid-user behavioral data.
Podcast measurement: Paparo argued the shift of podcasts to video has worsened measurement — Spotify, Apple, and YouTube all use different video technologies, none robustly supports video ad insertion, and geo-holdout testing is nearly impossible for host-read buys. Kory confirmed Haus's best workaround is a time-based pre/post forecast comparison.

Notable claims & predictions

Olivia Kory: "AppLovin ads work fast — the effects are short-term and immediate. That is not the case with most ad platforms. I don't even recommend a CTV test under eight weeks." (Implication: AppLovin's ML model is optimized for speed in a way most DSPs are not.)
Olivia Kory: Advertisers with AOV under $50 spend three times more on TikTok than advertisers with AOV above $50 — Haus data point suggesting impulse-purchase categories dominate TikTok's ad mix.
Olivia Kory on affiliate/RMN incrementality: "I wonder if [affiliate, RMNs, Amazon search] are limiting their advertising TAM by not offering incrementality. Advertisers are holding back dollars because they're skeptical and we haven't had a way to get evidence."
Ari Paparo on Google's trajectory: Plugging Q2 cloud growth numbers into ChatGPT, Paparo was told Google's cloud revenue could surpass its advertising revenue in the "early 2030s" — raising the prospect of Google being thought of primarily as a cloud company.
Ari Paparo on OpenAI's ad strategy: "OpenAI may be building something like FAN (a publisher audience-extension network) using paid-user behavioral data — that could be pretty hot." He also noted the inventory constraint: GPT queries are so long and context-rich that there isn't enough ad inventory to absorb advertiser demand even today.
Ari Paparo on podcast measurement: "The podcast world is getting significantly worse [for measurement] because the movement to video has broken all of it." He argued streaming audio (e.g., Spotify programmatic) is meaningfully more measurable than podcast, and buyers should not conflate the two.

Fact check

Paparo: Google cloud revenue up ~82%, from $13.6B (Q2 prior year) to $24.8B (Q2 this year). The figures directionally match Alphabet's reported Q2 2025 Google Cloud segment results. The ~82% growth rate, however, is higher than consensus and deserves scrutiny — actual year-over-year growth was substantial but closer to the 28–30% range in recent quarters. The $24.8B figure Paparo cites as a Q2 2025 number is worth verifying; the show's figures may conflate annual or trailing figures with a single quarter. Verdict: the specific revenue figures and the ~82% growth rate are unverified from the transcript alone; listeners should confirm against Alphabet's official filing before repeating.
Paparo: "Google's total ad revenue is about a billion dollars a day" ($81.6B ads + $11.1B YouTube ≈ $90B / ~90 days in a quarter ≈ ~$1B/day). The arithmetic is roughly correct as a cocktail-party heuristic. Verdict: true as an approximation.
Kory: "Meta's new ad product for optimization is called incremental attribution." This appears to refer to Meta's "Incremental Attribution" optimization feature. Meta has indeed rolled out incrementality-oriented tools, but calling the product precisely "incremental attribution" and characterizing it as a contradiction may slightly misname or oversimplify the product. Verdict: unverified exact product name; the framing that it blurs correlation and causation is a contested opinion, not a verifiable fact.
Paparo (via Claude): "There are 1,960 FAST channels globally." This figure came from a Claude query, not a primary source, and was presented as such. FAST (Free Ad-Supported Streaming TV) channel counts vary widely by definition and data source. Verdict: unverified; self-disclosed as AI-generated estimate.
Incentive flag — Olivia Kory on affiliate/RMN incrementality: Kory's company sells incrementality testing. Her framing that affiliate, RMNs, and Amazon search are suppressing advertiser spend by not offering incrementality is plausible but also directly serves Haus's commercial interest. The claim that "advertisers are holding back dollars" from these channels due to measurement gaps is asserted without data. Listeners should weight accordingly.

Full analysis

Olivia Kory, Haus's Chief Marketing Strategy Officer, went on Marketecture and told Ari Paparo and Eric Franchi something most measurement vendors won't say out loud: AppLovin's e-commerce ads actually work. The bigger tell in the conversation is which channels she says don't stand up to a holdout test, and why that gap is quietly capping how much money advertisers will pour into them.

Reversibility: Type 2 for any single operator. Measurement is a dial you can turn. But the industry-wide shift toward causal testing is Type 1, and it's already moving.

What's actually being decided: Whether the channels that refuse rigorous incrementality testing (affiliate, retail media networks, Amazon search) can keep growing on faith, or whether the buy side finally makes proof of causation a condition of the budget.

Timeline: No hard forcing function. This plays out over the next few budget cycles as CFOs get louder about ROI.

The Market Analyst. AppLovin gets validated here, and it matters that the validation is independent. Kory confirmed AppLovin opened its e-commerce pilot to third-party incrementality testing from day one, late 2024. That's a company that wants to be measured because it knows it passes. Compare that to the channels Kory grouped as measurement-shy: affiliate, RMNs, Amazon search. For an informed outsider: the ad channels willing to be graded are pulling ahead of the ones that hide the test. AppLovin's edge is speed. Kory said its effects are short-term and immediate, unlike a CTV test she won't even run under eight weeks. Fast, provable ROI is exactly what wins budget in a tight year.

The Skeptic. Kory sells incrementality testing. Her thesis that affiliate, RMNs, and Amazon search are "limiting their advertising TAM" by not offering holdouts is, conveniently, an argument for buying more Haus. The fact-check flags it: no data behind "advertisers are holding back dollars." And her affiliate finding of little incremental lift came from a handful of tests she herself called a small sample. For a non-specialist: the person telling you these channels are unproven makes her living proving things. That doesn't make her wrong. It means her central assumption, that skepticism is suppressing spend rather than the channels just being cheap and easy, is untested.

The Operator. The practical takeaway is CAPI, and it's not what most think. Conversion APIs, the server-side pipes that send purchase data back to ad platforms, are winning over browser pixels. But Kory frames them as optimization fuel, not measurement. Feed the platform your conversions and its machine-learning model gets better at finding buyers. That's the trap. If you send your purchase signal into AppLovin or Meta to improve their targeting, then use their reported conversions to judge them, you're grading the platform with data you handed it. The clever move Kory mentioned: some advertisers send only predicted-incremental conversions through CAPI. Optimize on causation, not correlation. Most teams aren't set up to do that Tuesday morning.

The Customer / End User. The advertiser here is the growth marketer who needs to show a number this quarter. That's why AppLovin and Meta win: fast, visible ROI. It's also why RMNs get a pass. Retail media is easy to buy, sits next to the transaction, and reports a tidy ROAS. Nobody wants to run a geo-holdout that might reveal half of it was going to happen anyway. Kory's honest point cuts both ways. Advertisers aren't demanding incrementality from RMNs because the current numbers look good enough and testing is a hassle. The skepticism she describes is real but passive.

The CFO. Attribution measures who bought after seeing the ad. Incrementality measures who bought because of it. The gap between those two is where ad budgets go to die. Kory's affiliate finding, little incremental lift, is the nightmare case: paying a commission on sales that would have happened anyway. If that holds at scale, affiliate and a chunk of RMN spend is a rebate on existing demand dressed as new growth. The reason nobody acts on this is that the proof is expensive and the channels resist the test. A CFO who forces one clean holdout on the biggest "unmeasured" line item will learn more than a year of ROAS dashboards.

The tensions.

Skeptic versus Market Analyst: is the measurement gap actually suppressing spend, or is that a vendor's sales pitch? Kory asserts advertisers hold back dollars from unmeasured channels. But RMN spend keeps climbing anyway. If skepticism really capped budgets, retail media wouldn't be the fastest-growing line in the deck.

Operator versus Customer: CAPI is sold as measurement progress, but it mostly makes the platforms better at optimizing themselves. The advertiser thinks they're getting cleaner data. They're mostly feeding the machine that grades its own homework.

What this hinges on. One belief: that advertisers will eventually demand causal proof from the channels that currently avoid it. The council leans skeptical that this happens fast. AppLovin proves the upside of volunteering for the test, but RMNs and affiliate have every incentive to keep the lights dim, and buyers have shown they'll spend anyway. Before acting: run one geo-holdout on your largest unmeasured channel. Not a vendor's model. Your own test. If the lift is there, keep spending. If it's not, you just found your budget for next year.

Prediction: No major retail media network (Walmart Connect, Amazon Ads, Target Roundel, Kroger) will open its inventory to standardized third-party geo-holdout incrementality testing on AppLovin's terms before the 2026 holiday budget commitments lock in Q3 2026.

Confidence: Medium. RMNs have every incentive to avoid a test they might fail.

Why: AppLovin volunteered for third-party testing because its fast, short-term effects pass, and Kory confirmed that made validation easy. Retail media networks are in the opposite position: their reported ROAS sits right next to the transaction, so a clean holdout risks showing a big share of those sales would have happened anyway, exactly what Kory found with affiliate. A channel growing fast on flattering attribution numbers has no reason to invite a test that could shrink its budget, and buyers keep spending without demanding one. The opposite outcome, an RMN opening its books to standardized causal testing, only happens under pressure that doesn't yet exist.

Revisit by 2026-10-15: We're right if no top-four RMN has publicly committed to standardized third-party geo-holdout testing by then. We're wrong if any of Walmart Connect, Amazon Ads, Target Roundel, or Kroger announces one.

Listen to the episode →