Golden Data

Issue #17 · September 2, 2026

+53% Wasn't Enough

Paul BrownWrites Golden Data from Ogden, Utah5 min read

On September 3, The Trade Desk told its employees the company would be about 15% smaller the next morning. The 8-K followed on September 4: $39 million to $51 million in severance charges, booked in the third quarter. The stock lost 4% on the day, on top of a year-to-date decline of roughly 64%. It is the last move in an ad-tech earnings season that fit inside 48 hours in early August and produced a result stranger than the layoff: the company with the best growth number got one of the worst stock reactions.

Criteo went first, before the open on August 5. Revenue down 11%, full-year outlook cut, stock down 24% by the close. AppLovin reported after that same close: revenue up 53%, powered by its Axon ad engine, and $20 million short of the estimate. The stock lost 19% the next session. The Trade Desk closed the window on August 6 after the bell: revenue up 3%, and a third-quarter guide of at least $650 million, which works out to a decline of about 12% year over year. The stock fell 28% on August 7.

Magnite reported in the same window and went the other way. Revenue up 11%, a beat, a raised full-year outlook. The stock went from $20.67 to about $24.32 on August 6, up roughly 18%.

Three growth numbers, great to terrible, and three stocks down 19% to 28%. One 11% and a stock up 18%. The market is pricing a number. It just isn’t the one in the headline. It’s the next one, and it has stopped giving the AI story any credit for the gap between the two.

What’s actually moving in the market

The Trade Desk, August 31 to September 4. Kokai Zuma, the latest release of the AI platform that is supposed to be the turnaround, was unveiled August 31. The cuts came four days later. On the August call, CEO Jeff Green had already said revenue growth was “below our expectations and below the standard we hold ourselves to.” Cutting staff after a growth miss is the ordinary move. Cutting staff four days after launching the product that is meant to fix the growth miss is the one that tells you how the third quarter is going.

AppLovin, August 5 and 6. Revenue of $1.92 billion against a $1.94 billion estimate, and a third-quarter guide about half a percent light. CEO Adam Foroughi attributed the miss to the timing of improvements to the company’s ad models, which were lighter than usual in the quarter, with a stronger step-up arriving after quarter end. Read that again. The company with the best AI ad engine in the group told investors its growth depends on when the model happens to get better. The stock gave up 19% the next day.

Criteo, August 5. An 11% revenue decline, and a full-year outlook cut to a 10% to 12% decline in contribution ex-TAC. The OpenAI partnership now covers more than 2,000 brands on ChatGPT, and management said it will not produce meaningful revenue before 2027. A partnership announcement did not move the number that mattered, and the company said so itself.

Magnite, August 6 and September 4. Up about 18% on a beat and a raised outlook. Then down about 1% on September 4, pulled along by The Trade Desk’s news despite having no news of its own. Sentiment moved a company that had just reported the only clean quarter in the group.

What I’d build in the revenue org this week

The lesson for anyone running a revenue function that sits on top of an AI optimization layer, in ad tech or anywhere else, isn’t which vendor to bet on. It’s that a vendor’s own before-and-after number for what its AI improved is not evidence. AppLovin’s CEO just explained why: the lift arrives in uneven steps, so a single before-and-after is mostly measuring which step you landed on.

I’d build a holdout harness before trusting the next lift claim, whether it comes from an ad platform, a CRM’s AI features, or an internal build.

Data shape: campaign- or account-level spend and yield, already landing in the warehouse from the ad server’s or CRM’s reporting API, tagged by whether that unit is currently exposed to the AI feature in question.

System shape: hold out a fixed 10% of accounts on the legacy path. No exceptions, no choosing which 10%. Run it for a window long enough to clear normal week-to-week noise.

Eval shape: compare yield per dollar between treatment and holdout. Don’t act on the difference until it clears a significance threshold set before the test started, not after the vendor’s number looked good.

Latency budget: weekly. Not real time.

Cost shape: a warehouse job and a dashboard. The only optional spend is one bounded model call to summarize the week’s delta in plain language, which costs closer to nothing than to anything.

Instrumentation: a standing chart of treatment-versus-holdout yield, refreshed weekly, with the significance check computed rather than eyeballed.

Two weeks in, success looks like having your own number instead of the vendor’s slide before the renewal conversation. Credible failure looks like a holdout too small to reach significance in the window available. That is also a finding. It is a better one than the slide.

Sources

Send me an email and we will talk. If something here landed close to what you're working on, the door is open. No calendar funnel, no pitch deck — I read every note that comes in.

Doing the work rather than deciding what to build? Crafting is the column for that chair.

Email meRead more Signal

← All Signal issues · Drafted with Claude · Edited by Paul Brown