# openai builds its own index; no channel means no yardstick > openai tests its own index against third-party serps and google opens three ad surfaces in one day; on the other side of the table, nobody can measure what they lost - edition: Monday, September 7, 2026 (2026-09-07) - notebook: product & markets - topics: media · market · brand · product - items: 8 from 6 sources - original: https://tonho.wtf/en/daily/2026-09-07-produto/ - portuguese edition: https://tonho.wtf/diario/2026-09-07-produto/ - authorship: written by an llm pipeline, reviewed and translated by antonio leandro (tonho.wtf) --- Two channel owners made a move on the same day, and in both cases the move is inward. OpenAI is testing its own index against third-party SERPs, and the name of the experiment leaked in Peec AI's research is almost a confession: `prefer-index-over-serp-v3`. Google, on Friday, opened three new inventory surfaces in 24 hours — Waze, YouTube and AI Mode. One internalizes distribution, the other monetizes every square meter it already controls. None of this is conceptually new: it's Netflix stopping mailing other people's DVDs. What holds the rest of the edition together is the underside of that table. Every item today is missing the same piece: the yardstick. GA4 files AI referrals as direct traffic and one publisher already sees 78% of its total drop into that bucket. Google's AI Mode now accepts exact and phrase match, but still has no segmentation in the report — meaning you may already be buying that traffic without knowing it. Brazilian beach tennis hosts more international tournaments than Italy and has no document that makes it legible to capital. And the watermark Anthropic announced, according to Fast Company, can only say that Claude was probably involved at some point. Friday's edition covered OpenAI selling the yardstick and Poppi betting on the channel nobody measures; today you can see the full pair. Whoever owns the channel profits from bad measurement. Whoever doesn't own it pays for it. ## product and growth **[Dopamine sites, the Korean shopping sites with no shopping](https://marginalrevolution.com/marginalrevolution/2026/09/south-korean-aspirational-markets-in-everything.html)** — Tyler Cowen links a story by Itika Sharma Punit about a format that reproduces the entire e-commerce ritual (search, review, cart, address, tracking) and ends in nothing: no money changes hands, no package arrives. It's a link post, I didn't read the original story, but the note alone is enough for the point: the shopping ritual and the purchase are two different products, and only one of them is the one that hooks you. Funnel optimizers treat the steps before checkout as cost to eliminate; here they are the whole product. ## brand and ip **[Shelf Life 126, on private label becoming an object of desire](https://thedieline.com/shelf-life-126-have-private-label-brands-become-more-desirable-than-their-counterparts/)** — the Dieline column opens with perception (grandpa's cabinet, the Charmin toilet paper the husband won't switch) and promises the thesis of a Pearlfisher report on the point where private label stops apologizing for being generic. Only the teaser is available, so what's left is the angle on record: if Kirkland and Trader Joe's became status, private label stopped being a discount and became positioning — and then whoever runs the shelf also runs the brand. ## media and attention **[ChatGPT is building its own search index](https://www.stackedmarketer.com/news/chatgpt-is-building-its-own-search-index-while-your-clicks-disappear/)** — beyond `prefer-index-over-serp-v3`, Peec AI's research mapped separate indexes for news, shopping, PDFs, YouTube and papers, plus a crawl experiment of up to a billion pages. chatgpt.com is the ninth largest site in the US, with 1.09 billion monthly visits and up 48% in a year, while Bing fell 50% and DuckDuckGo 16%. The detail that matters to anyone who depends on traffic: 95% of ChatGPT users still use Google, they just click less — forcing AI Mode drops external click-through by 18.8 points, news links by 12.5 and Reddit by 21.2. **[Google just opened three new places to spend budget](https://www.stackedmarketer.com/news/google-just-opened-three-new-places-to-spend-your-ad-budget/)** — Waze Ads reaches 42 countries, including the 27 in the EU, but only through Performance Max with store goals: it's a local intent play, for whoever has a physical store. Starting October 30, alcohol brands (and non-alcoholic drink brands, riding along) can run personalized campaigns on YouTube inventory, with the exception of Egypt, India, Indonesia and Poland. The pattern is the usual one — every Google surface ends up with ads — and what changes for media buyers is that the expansion now comes wrapped in an automated campaign, with no separate reporting. **["Did you use AI?" is the wrong question](https://www.fastcompany.com/91600126/did-you-use-ai-is-the-wrong-question-ask-what-did-you-use-ai-for)** — Faisal Hoque starts from a case with an explicit price: 14 publishers bid on a debut crime novel, the winner closed a $2 million deal, and when rumors of AI use came up it was the author's own agents who pulled the book. The text on the page didn't change a comma; what killed the deal was the suspicion about where it came from. The useful thesis for anyone running a team is to separate production (judge the result and demand accountability), evaluation (ask the person to defend the piece, don't run a purity test) and development — and in that third one there's a number: in a randomized study with devs learning a new Python library, those who had AI assistance did 17% worse on a later test of conceptual understanding, code reading and debugging. ## market and capital **[Beach tennis, a market with no owner](https://braziljournal.com/beach-tennis-um-mercado-sem-dono/)** — Brazil hosted 180 international tournaments in the sport in 2025, according to the Federação Paulista de Tênis based on the ITF calendar, and the state of São Paulo alone held more than the entire United States (49 to 45). Bernardo Stampa argues that what's missing isn't money or critical mass, it's the document: in American pickleball, the scene unlocked when the Sports & Fitness Industry Association started publishing an annual report on players and infrastructure, and in 2026 Apollo put in $225 million valuing the business at $750 million, according to CNBC. The report didn't create the sport, it made the growth legible. It goes for product categories too: as long as your market estimate swings across too wide a range, what you have is intuition, and organized capital doesn't buy intuition. ## reading **[The Rothschilds, revolutions and crisis: what 1873 teaches 2026](https://braziljournal.com/os-rothschild-revolucoes-e-crise-o-que-1873-ensina-a-2026/)** — a note by Artur Wichmann on *1873: The Rothschilds, the First Great Depression, and the Making of the Modern World*, by Liaquat Ahamed. The idea that sticks is the distinction between thesis and price: railroads, industrialization and trade integration were all real in 1869; what wasn't real was what people started paying for them — the Vienna Stock Exchange lost around 45% of its value in a single day, in May 1873, and more than $1 billion in American railroad bonds would end up in default. The Rothschilds' role in the book is that of the ones who knew how to sit out the euphoria. For anyone building product in the current cycle, the translation is direct: being right that the technology matters doesn't guarantee that the valuation, the CAC or the roadmap you backed on it are right. ## who wrote **[Arguments in Favor of AI Fair Use](https://kk.org/thetechnium/arguments-in-favor-of-ai-fair-use/)** — Kevin Kelly publishes his notes on why he considers LLM training fair use, and warns that he isn't building a closed case or listing the counterarguments. The central threads: training turns expression into latent space in a process he describes as irreversible, and the marginal contribution of any individual source is nearly incalculable because 99% of almost all material is redundant. I disagree with the redundancy part — it works well for the average corpus and badly for exactly what's rare, which is where the value sits for anyone producing original content — but it's the most organized inventory of the argument I've seen, and anyone with a brand and a catalog to defend gains more from reading this than from reading one more piece of outrage. ## stalled sources First Round Review (315 days), Calculated Risk (239), Gurwinder (253), Elad Gil (140), Y Combinator Blog (83), Anti-Mimetic (54), SVPG and Adjacent Possible (29). Elena's Growth Scoop, Intercom, Sherwood, Collab Fund and Kyla Scanlon close out 25 days.