# openai sells the yardstick; poppi bets on the channel nobody measures

> openai's global ad stack, pepsico's $1.7bn media account and a paper measuring 82.1% consumption above plan: the whole day is about who controls attention's yardstick

- edition: Saturday, September 5, 2026 (2026-09-05)
- notebook: product & markets
- topics: media · advertising · product · market
- items: 9 from 8 sources
- original: https://tonho.wtf/en/daily/2026-09-05-produto/
- portuguese edition: https://tonho.wtf/diario/2026-09-05-produto/
- authorship: written by an llm pipeline, reviewed and translated by antonio leandro (tonho.wtf)

---

A job market paper measured how much attention escapes the viewer's own plan: in microdata from an American short drama platform, paying users consume 82.1% more than they intended. The number is good, but the instrument is better. What revealed the intent was the non-linear top-up menu — the very price structure the platform built to extract more is what makes it possible to audit how much it extracted. The design that produces the excess is the same one that makes it visible.

That puts the rest of the day in order. OpenAI didn't launch a format: it launched a pixel, a Conversions API, GAID and reporting per carousel card, which is to say the yardstick without which an advertiser doesn't move money from a test budget to a media plan. PepsiCo pulled $1.7 billion in media out of Omnicom after twenty years and left creative, sports marketing and PR there — what left is the part that gets measured, what stayed is the part that doesn't. The Live Nation segment table that Web Smith takes apart says the same thing in margin: the show pays little, the ticket pays well, the sponsorship pays a lot. And at the opposite end are the two channels with the best apparent return today, precisely the ones nobody has instrumented yet: sorority rush and the artist who skipped the festival and became one. Yesterday I wrote that the rented channel is getting expensive; today's answer is that the cheap channel is the one that doesn't have a yardstick yet — and it's worth it for as long as that lasts.

## launches

**[OpenAI expands ChatGPT Ads and builds out its ad stack](https://searchengineland.com/openai-expands-chatgpt-ads-globally-and-builds-out-its-ad-stack-487125)** — It's going to countries in Europe, India, the Middle East and North Africa, and it comes with the plumbing: custom audiences above 5 million members with mixed identifiers, GAID as a supported identifier, a pixel accepting hashed phone, name, region and zip, impressions and clicks per carousel card in Ads Manager (and the explicit note that those impressions aren't billable). The next announced step is an optimization model that accounts for view-through conversion, charged per impression. On Aug 31 the subject here was $1bn in annualized revenue; the news now is a different one and harder to copy — attribution. No media director moves budget to a channel that doesn't close conversion, and that's what's being built, not the format.

## product and growth

**[Short Videos, Big Self-Control Problems](https://marginalrevolution.com/marginalrevolution/2026/09/short-videos-big-self-control-problems.html?utm_source=rss&utm_medium=rss&utm_campaign=short-videos-big-self-control-problems)** — The paper is by Renjie Bao, a Princeton job market candidate according to Cowen, and the mechanism is the point: short units renew the temptation over and over, and the temptation lasts longer than the unit. The structural estimate gives an average horizon of 11.2 minutes — short next to a whole soap opera, long next to a one-minute episode, and that gap is where the 82.1% excess lives. The counterfactuals point to a bigger decision unit, a cap by default and breaks. Two things for anyone building: the consumption unit is a design lever as much as the content is, and non-linear pricing works as an instrument for inferring intent. If your product has a top-up menu, a plan or a bundle, you're already collecting the data on how much the user planned to spend — and you can compare it with what they spent.

## brand and ip

**[Five years of RushTok and brand ROI only grows](https://www.marketingbrew.com/stories/sorority-rush-brand-sponsorships-poppi-lucky-charms-strategy?utm_source=&utm_medium=syndication&utm_campaign=feed)** — Poppi will send more than 1 million cans and 20,000 pieces of merch to sororities this season, and that accounts for 30% to 40% of the brand's back-to-school budget, according to its VP of culture. The origin detail matters more than the number: the college team was put together three years ago because the requests were coming from inside, inbound from the chapters themselves. Lucky Charms, meanwhile, got in for a reason the brand didn't create — the clover was already showing up organically in content from Kappa Delta chapters — and activated with a balloon and a bracelet station, no creator contract, betting that the decoration is the content. The ROI the sources describe is relational and long-term, which is to say: not attributable. That's the price of getting into a channel before the yardstick arrives, and also the reason it's still cheap.

## media and attention

**[Publicis wins PepsiCo's global media account](https://adnews.com.br/post/publicis-groupe-vence-conta-global-de-midia-da-pepsi-co-de-us-1-7-bilhao-e-deixa-concorrencia-da-coca-cola)** — $1.7 billion, the end of more than twenty years with Omnicom in markets like the US and the UK, and consolidation into a single operating model, "One PepsiCo", across more than 200 markets. Omnicom keeps creative, sports marketing and PR. The instructive part is the structural cost of the holding model: because of account conflict, Publicis dropped out of Coca-Cola's global media, data and technology review, estimated at around $4 billion, where it was up against WPP — and it already handled Coke's media in North America. Winning 1.7 meant giving up the shot at 4. Anyone selling services into concentrated categories lives with this arithmetic: every new client is also an exclusion, and the pipeline has a ceiling that doesn't depend on quality.

**[The Business of Fest](https://2pml.com/2026/09/05/fest/)** — Web Smith takes the two-night show at Soldier Field (more than three hours, 60 songs, 15 guests, 70,000 people, sold out) and treats what Billboard called "threw their own festival" as an economic fact: no sponsor on the bill, no Live Nation or AEG, with one-off stadium dates operated by IKON Presents instead of a routed tour. The contrast comes from the segment table in Live Nation's record quarter: concerts at roughly 5% adjusted operating margin, ticketing near 39%, sponsorship and advertising around 67% — the show is traffic, the brand deal is the product. And the top-line records hide the middle: per Pollstar, the average gross per show in North America fell 7.8% to $1.14 million and the average ticket fell 2.4% to $122, with stadium tours going from eighteen to eleven. What stuck with me most is Coachella's payment plan, which went from 18% of GA buyers in 2009 to about 60% in 2025: with a signup fee of roughly $41 across about 100,000 people, that's more than $4 million before anyone pitches a tent. That isn't checkout convenience, it's a working capital instrument. (The text I managed to read is long and cuts off mid-section; the full read is at the link.)

## market and capital

**[They paid $470k for a "boring" business](https://www.fastcompany.com/91596754/boring-business-everything-went-wrong-ai)** — Andrea Palacio and their husband bought a landscaping company in South Florida chasing passive income. Days after closing the seller vanished, 10 of the 12 employees quit and revenue fell from $70k to $40k a month, with personal assets pledged as collateral on the SBA loan. A year of 80-hour weeks brought the revenue back, and a $750k offer (a 60% premium) was turned down for lack of another income source; six months out of the country later, the company account had $2.50, which they attribute to the manager siphoning funds. In June the operation did $75k a month with seven employees. The rule left standing is the day's deal structure lesson: don't buy if the seller won't finance part of it, because without a seller note the seller has no incentive at all in the transition. The AI automation the article celebrates came later — and what was missing at the start was more banal, a KPI dashboard and someone looking at it.

## reading

**[Disposable Workers, by Paul Osterman](https://www.fastcompany.com/91595370/america-disposable-workforce-growing-hidden-costs)** — The MIT Sloan professor emeritus sums up their own new book in five points, and what holds the thesis up is a national survey of their own with more than 6,000 workers: 35% of the American workforce would be in disposable arrangements, counting contractors, freelancers and what they call marginal workers, people on payroll but off any career track. The figure that's most useful to anyone building a team is the disaggregated satisfaction: freelancers report higher satisfaction than standard employees, contractors and marginals report much less, and that same group declares far less willingness to put in extra effort. There's a measured externality too — hospitals that outsource cleaning show up with higher infection rates, because a contractor isn't plugged into the in-house communication networks. And the argument about AI is finer than the usual one: it isn't the elimination of jobs that pushes disposability, it's the uncertainty about how many people and which skills will be needed in two years.

## who wrote

**[Herman Bessler, of Templo, in propmark](https://propmark.com.br/mercado/agencia-fatura-hora-e-fee-a-ia-comprime-hora-avalia-herman-bessler-do-templo/)** — The thesis sentence is the best formulation of the problem I've read this month: "Agencies bill hours and fees, AI compresses hours, and without redesigning what you charge for and how you charge it the productivity gain doesn't turn into margin for the agency, it turns into a discount argument in the client's hands at the next renewal". Whoever charges for the input loses when the input shrinks. The numbers from Templo's study with more than 400 leaders explain why the problem hasn't blown up yet: about 84% use AI through chat and less than 3% in tools integrated into full workflows; 82.6% of companies increased usage in the last year, but only 31.5% declare high maturity, and process automation scored 35 out of 100. The expensive bottleneck, Bessler says, isn't in the creative work, it's in the assembly line around it — RFP reading, briefing, approval. And the structural warning holds for any product team: cutting the junior base solves cost in 2026 and creates a capacity problem in 2031.

**[TBM 438, by John Cutler](https://cutlefish.substack.com/p/tbm-438-the-myth-of-workplace-consensus)** — The thesis is measurement error, not culture: in a company you observe three stacked layers — norms carrying the weight of power, people's adaptation to them, and the people themselves — and you see almost only the second, attributing it to the third. Adaptation isn't endorsement, and the loop closes because your own adaptation becomes one more data point for everyone else to conclude that everybody agrees. It's the same vice as the rest of the day, turned inward: confusing what's visible with what's prevalent. If you research users rigorously and your read on the team comes from hallway observation, it's the read on the team that's wrong.

## stalled sources

Fourteen sources from this desk without publishing, several already terminal: First Round Review at 313 days, Gurwinder at 251, Calculated Risk at 237, Matthew Ball at 201, Elad Gil at 138. Of the group still breathing, SVPG and Adjacent Possible are at 27 days, Elena's Growth Scoop and Kyla Scanlon at 23.