The ready-made answer — from Google, from ChatGPT, from the voice assistant — takes away people who only wanted the answer. Today that loss shows up with opposite signs in two places. At the Atlantic, visits from Google are falling and subscribers from Google are rising, because the people who disappeared were the ones who weren’t going to subscribe. At Wikipedia, human traffic falls about 8% a year, and that is the whole problem: the people who came only for the answer were also the ones who spotted the error in the article and fixed it. The difference isn’t in the size of the drop. It’s in where the conversion happens. When the thing you need from the visitor still happens with fewer visits, the drop works as a filter. When it only happens with them on your page, the drop eats the product.
The other side of the same shift is the customer who builds the interface in-house. Ploomes lost customers to CRMs built with vibe coding and saw some of them come back when the system broke and there was no one to fix it. Teresa Torres answers people who have already built their own pipeline with Claude with a tally: 4 evals and 16 experiment variants for a single complaint. Pagani, of Ploomes, says what stays standing is what sits behind the screen, and the Stratechery teaser says Salesforce has already dropped the UI as a moat. My read: when the screen belongs to someone else (the agent, Google, the customer), what you sell is whatever takes work to maintain behind it. And the Commoncog piece puts the brakes on the panic: in the five cases it looks at, nobody died of stubbornness. Redoing the whole roadmap, as Ploomes did, is necessary and guarantees nothing.
product and growth
Ploomes lost customers to vibe coding, and some came back — Ploomes customers built their own CRM with vibe coding and canceled. Some came back because, when the system breaks, there’s no one on the other end. CEO Matheus Pagani doesn’t treat this as relief: he says the CRM of funnel, records, tasks and history is over and that the value moves to data and context. On the management side, the company swapped four-year planning for three-month cycles and redid almost 100% of the roadmap mid-year. For anyone selling SaaS, the churn criterion that sticks is time: if setup takes too long, the customer runs the numbers and decides it’s worth more to build in-house.
Teresa Torres: 4 evals and 16 variants for one complaint — A Vistaly beta customer got an AI-generated opportunity solution tree with a flattened branch, full of opportunities with no hierarchy, and wanted a “clean up” button. Teresa Torres chose to find the cause instead, and that took three weeks, 4 new evals and 16 experiment variants. The excerpt I read shows the start of the method: a code-assertion eval that measures the shape of the tree (children per node, depth) and an LLM-as-a-judge that looks for missed groupings, calibrated against hand-made labels. That judge had 100% recall and a specificity the text itself calls terrible. The dossier’s text stops there, so I don’t know which variants fixed it. The product argument, though, is already made: what separates the tool from the homemade pipeline is the calibration, not the prompt.
brand and ip
Wikipedia versus the answer engine — By the Ahrefs data cited in the interview, ChatGPT cites Wikipedia more than almost any other source, behind only Reddit. Meanwhile, the site’s human traffic falls about 8% a year, and bot traffic climbs to the point of straining the servers. Zach McCune, who heads global brand at the Wikimedia Foundation, describes Wikipedia as more visible than ever and less known, with drops in awareness and favorability in brand health tracking. The risk isn’t revenue, it’s editors: corrections happen on the site, and fewer people arriving means fewer people correcting. The response comes on two fronts. One is the 25th-anniversary campaign by the agency Kin, built from screenshots of about 700 pages to put editors in front of the page. The other is a real-time enterprise API, so that LLM companies pay for usage and the bill doesn’t fall only on the 8 million donors, who give $11 on average.
New Balance sues Decathlon over the Kiprun logo — The suit was filed on Tuesday the 15th in federal court in Massachusetts. According to New Balance, Kiprun’s angular “K”, mirrored on eight models sold in the US, reads as its “N”. The request includes stopping use of the symbol, recalling and destroying the shoes and paying damages. New Balance approached Decathlon in January, and the answer was that the design is a K. The lesson I take for anyone creating a brand: the logo has to be tested the way it will appear on the product, mirrored and from the side, not just in the file. New Balance also has a history of going to court over the N ($1.5m in a Chinese court in 2017, New Bunren in 2019, Michael Kors in 2021). In my read, that consistency is part of the brand equity the company says it has invested hundreds of millions of dollars in.
media and attention
The Atlantic gets more subscribers from Google on fewer visits — Nicholas Thompson, the magazine’s CEO, told the Channels podcast that visits from Google are falling while subscribers from Google are rising: Google is sending fewer people who weren’t going to subscribe. For a subscription business (1.6 million paid subscribers, profitable since 2024), it’s proof that visits and revenue no longer move together. With AI, two contracts show the design. The OpenAI one has clauses on how much the model can summarize, whether it can quote an author and whether it can imitate their style. The one with Parallel Web Systems, Parag Agrawal’s company, runs on micropayments and pays more for content that’s hard to replace or used in high-value agent work, instead of a flat fee per access. The Parallel deal came about when Thompson noticed that most of the traffic leaving his own computer was already coming from his agents. The traffic ChatGPT sends back to the Atlantic, meanwhile, was, in the CEO’s words, surprisingly small.
Design Anthology and print as an advantage — The magazine was born in 2014, in Hong Kong, to cover Asia-Pacific design as a region. Today it has print, digital subscriptions, events, branded content and a YouTube channel with more than 500 thousand subscribers. The stated decision is not to chase pageviews: protect scarcity and sell ads on long-term relationships, not audience size. The reader sits between B2B (architects, designers) and the wealthy consumer luxury advertisers want, and YouTube serves as the front door to the rest. The part explaining why print is the center comes after the dossier’s cutoff: the title promises that argument, and I didn’t read it.
market and capital
Kinea: the discount isn’t in stocks, it’s in rates — The report takes apart “the stock market is cheap” with a mechanism. The Brazilian market trades at about 8 times projected 12-month earnings, versus 10 times in the past, but the risk premium over government bonds is around 4%, below the historical average of 5%. In other words: the low multiple comes from a high risk-free rate, not from a generous premium for taking risk. The other half of the explanation is earnings. In dollars, listed companies’ earnings are 36% below 2010. Over ten years, nominal earnings of domestic companies grew 3.5% a year, below average IPCA inflation of 5%, while those of commodity producers grew 16.5% a year. The report ends with a portfolio suggestion, which stays out of this edition. What’s useful for anyone building here is the mechanism: a compressed local multiple is the price of the cost of capital and of earnings that don’t compound, not a bargain.
reading
Commoncog: how to handle disruptive change — Cedric Chin discusses five cases commissioned for the series: Swatch in the quartz crisis, Amazon in the dot-com bust, Ford in the global financial crisis, Blockbuster versus Netflix and the death of BlackBerry. The first point takes apart the explanation everyone has ready, that the disrupted die of stubbornness. In all five cases, nobody died that way. Omega and even Rolex launched quartz watches, Blockbuster nearly killed Netflix with its online operation, and Mike Lazaridis, RIM’s co-CEO, took apart an iPhone a few days after launch. What changes for anyone building product is the diagnosis: “not falling behind” isn’t a plan, because the companies that died didn’t stand still either. The piece announces three points, the dossier excerpt only reaches the first, and the reactions to each case are members-only.
who wrote
Ben Thompson: Salesforce, agents as UI and the race to headless — Today’s analysis is subscribers-only, and what I read was the teaser: Salesforce is abandoning the UI as a moat, and Ben Thompson thinks the decision is very smart because the UI is disappearing for everyone. The title talks about agents as the interface and a race to headless, but I didn’t read the argument itself. I’ll note the coincidence: on the same day, Matheus Pagani, of Ploomes, was saying in Brazil that what stays standing in CRM is what sits behind the screen.
stalled sources
First Round Review (324 days), Gurwinder (262), Calculated Risk (248), Elad Gil (149), Y Combinator Blog (92), Anti-Mimetic (63), Adjacent Possible (38), Kyla Scanlon (34), Sherwood News and Collab Fund (30 each), Granted (19).