For about fifteen years the performance marketing business ran on a reliable intake valve. Graduates arrived in June, learned the platforms by pulling reports nobody senior wanted to pull, built keyword lists, checked placements, wrote the first draft of the wrap deck, and three or four years later knew enough to run something. That valve is closing, and on Monday, September 28, 2026, Digiday reported that the people closing it are saying so out loud.

The setting was a town hall at Digiday Media's AI Marketing Strategies event on September 24, 2026, held under Chatham House Rules, which is why the quotes carry job descriptions rather than names. The most direct of them came from the owner of a small-to-mid-sized performance marketing agency: "We haven't hired college-level grads in some time, and I don't anticipate that we will be." The same executive did not dress it up as a philosophy. "It's not like I don't like college grads. I do... it's just a business" matter, he said, pointing at the tasks that used to justify the salary and now do not need a salary attached to them.

What makes the exchange worth reading twice is that the room did not agree with itself. One participant reframed the whole thing as a category error: "It seems like a learning and development challenge, not a technology challenge." Another put the obligation back on the people holding the budget. "The nature of work has to change, and I think as leaders, we have to be part of finding that right talent." A third suggested the training problem cuts in both directions, because the senior staff hold something the tools do not. "The older talent is going to have a lot of historical knowledge about like just how the sauce gets" made, that attendee said, while the juniors who are not being hired are the ones fluent in the systems now doing their old work.

The line that closed the session has been circulating in various forms for two years, and it has hardened into something closer to a job description than a warning: "AI isn't coming to replace your job. Someone who knows AI and has your experience is coming to replace your job." Experience plus fluency beats either alone. The difficulty, arithmetically, is that experience is produced by doing the work that fluency has absorbed.

Is this anxiety or measurement? Both, and the measurement side has been accumulating for some time. Digiday cited 2026 Pew Research showing that people aged 18 to 34 across many countries worry more about AI-driven job loss than older adults do, which reads less like a generational mood than an accurate reading of where the exposure sits. PPC Land has covered the harder version of that finding, a paper drawing on Stanford data that put the loss at 16% of employment in AI-exposed occupations for the youngest workers, a figure that concentrates almost entirely in the early-career band rather than spreading evenly across a workforce. The counter-case exists too, and PPC Land has covered the argument that AI may create more marketing jobs rather than fewer, which rests on new categories of work appearing faster than old ones disappear. Nothing in the September 24 town hall suggested those new categories are being filled by twenty-two-year-olds.

There is a quieter consequence underneath the hiring decision, and it concerns what a marketing practitioner actually knows. PPC Land examined it in a piece on briefs that arrive with no argument in them, the erosion that happens when the reasoning step is outsourced and only the output is reviewed. An agency that stops hiring juniors is not only saving a salary line. It is removing the years in which somebody learns why a recommendation is wrong, which is a different skill from noticing that it is. The town hall's advice on training staff was to "meet them where they're at," which is sensible and also assumes there is a them to meet.

The pattern is not confined to agencies. PPC Land reported in September that Amazon's human resources chief wrote 100,000 lines of code in a year after the company cut 30,000 jobs, an executive doing work that would previously have been requested from a team, which is the same substitution running from the top of the org chart rather than the bottom. Trade bodies have started issuing formal prompting guidance for marketing professionals, which treats fluency as a certifiable competence rather than a knack. Upskilling has become the standard answer. What almost nobody has answered is how somebody acquires the experience half of the formula once the tasks that generated it have been automated away.

A bill that counts the programmes nobody has counted

Four days before that town hall, a related problem arrived in the United States Senate with a price tag attached. During the week of September 21, 2026, Senators Lisa Blunt Rochester of Delaware and Ted Budd of North Carolina filed the National Talent Strategy Act of 2026, a bipartisan measure whose central observation is unglamorous and hard to dispute: the federal government runs more than 150 separate programmes spending over $250 billion a year on education, workforce development and childcare, and no single body holds a picture of what they collectively do.

The bill does not create a programme. It creates a process, which is both its weakness and the reason it has a chance of passing. Within 120 days of enactment, the legislation would stand up an American Talent Working Group of eighteen members: nine cabinet secretaries drawn from Labor, Agriculture, Commerce, Defense, Education, Energy, Health and Human Services, Homeland Security and Transportation, four agency heads, and representatives from five executive offices. Within a year of that, the Department of Labor would have to produce a Federal Strategic Talent Plan covering four years.

The plan's required contents are where the bill touches the hiring question directly. Labor would have to evaluate every federal workforce programme for effectiveness, conduct a barrier analysis across eight priority areas, and build an interagency strategy addressing three things at once: cross-agency implementation, labour force participation, and workers displaced by automation or artificial intelligence. That last clause is the one that matters for anybody reading the September 24 town hall transcript. It is the first time the displacement the agency owners described has been written into a federal planning mandate as a named category rather than a hypothetical.

Accountability runs through familiar channels. Annual meetings would generate reports to the Senate Health, Education, Labor and Pensions committee and to the House Education and Workforce committee. Regional outreach engaging state and local leaders, industry, unions and tribal leaders would be mandatory rather than discretionary. The Government Accountability Office would evaluate implementation after four years, which places the first genuine scorecard somewhere in 2031 assuming the bill passes promptly, and later if it does not.

Several things are absent. The bill number, the exact filing date, the committee assignments and any cosponsors beyond the two senators were not disclosed in the materials available when PPC Land published on September 25, 2026. More substantively, the Bipartisan Policy Center Commission's recommendations on data systems and on modernising ONET, the occupational database that classifies what jobs consist of, were left out. That omission is awkward given the bill's own logic. A strategy for workers displaced by automation requires a taxonomy capable of describing which tasks within an occupation have been automated, and ONET in its current form describes occupations rather than tasks. The working group would be asked to plan against a map it has not been funded to redraw.

For the advertising industry specifically, the bill is a slow instrument pointed at a fast problem. The agency owner who said he does not anticipate hiring graduates was describing a decision already made, not one under consideration. A four-year federal plan delivered in late 2027 at the earliest will land several hiring cycles after the cohort that would have benefited from it entered the market. Still, it is the only instrument on the table that treats the question as structural rather than as something each employer settles privately with itself.

Seventy-three times a week, and nobody has licensed any of it

If the first two stories concern who does the work, the third concerns what the work is made of. On September 26, 2026, PPC Land reported on a Copyright Clearance Center and Outsell survey finding that senior executives generate roughly six times the potential unlicensed content sharing of individual contributors, and the underlying numbers describe an exposure that has been compounding quietly inside every large organisation.

The methodology first, because the figures depend on it. CCC and Outsell surveyed 570 full-time knowledge workers at companies with 1,000 or more employees between April 10 and May 8, 2026. Roughly 44% sat in the United States, with 28% across the German-speaking markets of Austria, Germany and Switzerland, and the remainder in the United Kingdom, the Netherlands, India and Japan. Margin of error was plus or minus 4.0% overall, widening to between 10% and 15% when the data is cut by vertical. Everything is self-reported, and the report says plainly that actual usage runs higher than what respondents were willing or able to recall.

The headline measure is what CCC calls velocity: the number of times per week an average employee shares published content through traditional channels in ways that may not be covered by a licence. In 2016 that number was 17.8. By 2025 it had reached 66, the year AI questions first entered the survey. In 2026 it stands at 73. Alongside it sit the AI-specific figures. Employees feed published content into AI tools 11 times a week. The resulting output reaches 96 people on average and is itself shared or distributed 12.5 times a week. And 51% of AI users report that content is arriving into those tools through automated feeds and application programming interfaces rather than through a person deciding to paste something.

The seniority gradient is the finding with teeth. Senior executives generate 111 potential unlicensed sharing instances weekly, six times the individual contributor rate. Their outputs reach 202 people rather than 96. Some 86% use AI daily. And 92% say they know their organisation's copyright policy, a figure that sits uncomfortably beside the 76% who report forwarding everything when time-pressed. Knowledge of the rule and adherence to it have come apart almost completely at the top of the building.

What is being fed in is exactly the material that carries licensing terms. Market research leads at 47%, websites at 46%, research papers at 41%, news at 40%. "Most organizations are still licensing for a world where content moves between people," said Lauren Tulloch of CCC. Grant Hunter of Outsell put the structural point more sharply: "Today, half of knowledge workers already have externally published content entering AI through automated feeds." A licence negotiated on the assumption that a human reads a document and forwards it to colleagues does not describe a pipeline that ingests the same document automatically and redistributes derivatives of it to 202 people.

CCC has been building commercial answers to this, including the addition of AI reuse rights to its annual copyright licence for US colleges, and the broader legal terrain was mapped by the US Copyright Office in its report on AI training. The gap the survey exposes is narrower and more practical than the litigation everyone watches. It is not about whether a model was trained on copyrighted books. It is about a marketing director pasting a subscription research report into a chat window on a Tuesday afternoon, receiving a summary, and sending that summary to two hundred colleagues, every week, at scale, under a licence written for photocopiers.

The connection to the hiring story is not decorative. The junior analyst who used to read the research report and write the summary was also, incidentally, the person who knew where the report came from and what the subscription permitted. Remove that role and the provenance step goes with it.

A judge sends five scraping claims back to state court, and takes fair use off the table

The fourth story is what happens when the provenance question reaches a courtroom. On September 26, 2026, PPC Land published a detailed account of Judge Trina L. Thompson granting Reddit's motion to remand its case against Anthropic to state court, a procedural ruling with a substantive consequence: the dispute will now be decided under contract law, where fair use is not available as a defence.

The route here was long. Reddit's user agreement took effect on October 15, 2020, with Sections 3 and 7 governing access. In May 2024 Reddit sent Anthropic a cease-and-desist letter. In July 2024 an Anthropic spokesperson said Reddit had been on the company's crawler blocklist since mid-May; Reddit subsequently alleged more than 100,000 accesses after that statement. Reddit filed suit in San Francisco County Superior Court on June 4, 2025 as case CGC-25-625892. Anthropic removed it to federal court on July 3, 2025, where it became 3:25-cv-05643-TLT in the Northern District of California. Judge Susan Illston stayed proceedings on July 7 pending mediation, which ended without settlement on August 1. Reddit moved to remand on August 29. Judge Illston recused herself on October 1, 2025, Judge Thompson took the case, and Anthropic disclosed its Alphabet, Google and Amazon affiliations. Supplemental questions went out in late January 2026, a tentative order issued on March 20, oral argument ran 54 minutes in Courtroom 9 on March 24, and the twelve-page order was signed on March 28 and entered on March 30, terminating the federal case. The hearing transcript was filed on April 2 with access restricted until July 1, 2026, which is why the reasoning has surfaced publicly only now.

The legal test is the two-prong preemption analysis under the Copyright Act. Prong one asks whether the material falls within copyright's subject matter, and the court found that Reddit's repository of discussions does, citing Best Carpet Values v. Google from 2024 and Craigslist v. 3Taps from 2013. Prong two asks whether the state claims assert rights equivalent to those copyright protects. Here Judge Thompson found an extra element in each of Reddit's five claims, which is why all five go home to state court.

Breach of contract protects conditional access and restricted purposes rather than copying as such. Unjust enrichment rests on bypassing technical safeguards, violating contractual access restrictions and misrepresenting compliance, which the court called qualitatively different from copying. Trespass to chattels turns on alleged server impairment from crawling volume. Tortious interference concerns Reddit's privacy covenants owed to its users. Unfair competition under California Business and Professions Code section 17200 involves intentional misrepresentation and affirmative deception. The precedents the court leaned on are the scraping line rather than the copying line: Craigslist v. Autoposterpro from 2009, Yu v. ByteDance from 2023, Altera v. Clear Logic from 2005, and MDY Industries v. Blizzard from 2010, the last of which held that anti-bot provisions create contractual covenants that live outside copyright.

Two findings extend beyond this case. The first concerns robots.txt. Section 7 of Reddit's agreement prohibits accessing data "by any means (automated or otherwise)" outside what is permitted, and states that "scraping the Services without Reddit's prior written consent is prohibited." The order notes that this language converts robots.txt from a voluntary protocol into a contractual condition, quoting the agreement's grant of permission "to crawl the Services in accordance with the parameters set forth in our robots.txt file." A file that has been advisory since 1994 becomes, when incorporated by reference this way, a term of a contract.

The second concerns browsewrap. Reddit's agreement is linked rather than click-through, which normally weakens enforceability. Judge Thompson found Anthropic bound anyway, reasoning from actual knowledge: the alleged 100,000-plus accesses after a public denial, Anthropic's own user agreement containing similar restrictions, its public statements about honouring robots.txt, and its continued scraping after the May 2024 letter. Anthropic, she wrote, "engaged in conduct sufficient to infer actual knowledge of the terms of the User Agreement and to form an implied-in-fact contract with Reddit."

Counsel's positions are on the record. Ragesh Tangri of Morrison and Foerster, for Anthropic, conceded at argument that Section 7 "draws the line" at scraping, accepting that bots were permitted while scraping was not. Richard Worcester of Quinn Emanuel, for Reddit, offered a bookstore analogy: a visitor may browse during business hours, "but what they absolutely cannot do is break the window at nighttime and go in." Kelsey Falkenberg of Reddit's team supplied the prong-two argument when the judge signalled which issue concerned her.

The commercial stakes sit in the margins of the record. Google reportedly pays Reddit around $60 million a year for data licensing, an arrangement that has drawn antitrust counterclaims of its own. Reddit booked $36 million of AI data licensing revenue in a single quarter. Anthropic settled piracy claims in Bartz for $1.5 billion in September 2025. Reddit has also sued data scrapers and Perplexity, and faces SerpApi's push to kill a separate DMCA suit.

The order is not a clean document. PPC Land catalogued the irregularities: the caption describes the motion as the defendant's when Reddit filed it, the conclusion refers to doubts about "Reddit's right of removal" when Anthropic removed, the trespass section attributes an opposition to the wrong party, burden statements are reversed in places, the signature date falls on a Saturday, a December 2025 docket entry carries the date 12/17/2028, and lead counsel appears variously as Richard Worcester, Corey Worcester and Richard C. Worchester. None of that disturbs the holding. It does suggest a twelve-page order produced under pressure.

For anybody licensing content in either direction, the practical shift is the removal of fair use from the argument. Terms of service, drafted years before generative models existed, are now the operative instrument.

What publishers built while the traffic left

The last story is the one that has been running longest, and on Monday, September 28, 2026, Digiday published its recap of the Digiday Publishing Summit held September 14 to 16 in Miami. The useful part is not the traffic numbers, which have been documented to exhaustion. It is the inventory of what has actually been built.

The traffic numbers set the floor. Kristin Roberts, president of USA Today Media, said Google referral traffic to the company's sites has fallen from more than 70% to roughly 40%. Steve Horowitz, president of the technology and shopping division at Ziff Davis, put his exposure precisely: 35% of revenue tied to web traffic, 17.5% to search specifically. Those are disclosures of dependency rather than complaints about it, and they arrive alongside PPC Land's reporting that news publishers have lost half their Google search traffic in two years and that smaller publishers have shed as much as 60%.

What has been constructed in response splits into three categories. The first is product sold to other companies. Mike Peralta, chief revenue officer at Future Publishing, described Future Optic, a generative engine optimisation product now serving more than thirty clients with renewals attached. That is a publisher selling visibility services into the same AI surfaces that reduced its own traffic, and it lands in a market PPC Land has tracked closely, from Adobe's enterprise AI visibility tool to Similarweb's dual tracking platform, against a baseline where only 16% of brands track AI visibility at all. It is also a market with attrition: Lorelight's founder shut down his AI visibility tracking tool earlier this year. Thirty clients with renewals is a real business, and a small one.

The second category is infrastructure. Francesca Barber, executive vice president of product, audience and AI at Politico, described implementing the Sanity content management system, the sort of replatforming that reads as housekeeping until the requirement becomes publishing into formats nobody specified three years ago. At The Wall Street Journal, Taneth Evans, head of digital, announced a new multimedia role covering video and audio, with Samantha Henig hired to fill it and Leital Molad serving as senior director of shows. Semafor's head of video, Adam Banicki, now reports to Ben Smith and Rachel Oppenheim. These are org-chart facts, and org charts are where a strategy becomes payroll.

The third category is the licensing question, which remains unsettled in both directions. Adam Greenberg, vice president of strategic partnerships at The New York Times, called licensing "underdeveloped," which is a striking word from a company with a signed deal and active litigation. Sally Shin, executive vice president of growth and partnerships at CNBC, said her company has signed no LLM deals at all. Ziff Davis, whose divisional revenue exposure Horowitz quantified, is separately suing OpenAI. Compensation terms and attribution mechanics remain inconsistent across every deal that exists.

Esther Cohen, director of audience and subscriptions at The Verge, offered the framing that has quietly become the default among publishers who moved early: treat Google traffic as "extra." It is a demotion of a channel that was, for twenty years, the business model. And Liz Goff, co-founder of Puck, described a compensation model giving journalists equity rather than only salary, which converts a cost line into partial ownership and is the clearest statement at the summit that the unit of production has changed.

Threaded through all of it is the same question the agency owners were arguing about in the town hall four days later, and the same question CCC measured at 73 times a week, and the same question Judge Thompson answered in twelve pages. Somebody produced the material. Somebody else is using it. The terms on which that happens were written for a set of arrangements that no longer describe how anything moves.

Also noted

By the numbers

$250bn - Annual federal spend across more than 150 programmes that one Senate bill would place under a single four-year plan. Source

111 - Weekly instances of potentially unlicensed content sharing per senior executive, against 96 people reached by the average AI output. Source

100,000+ - Accesses Reddit alleges Anthropic made after the company publicly said it had stopped. Source

101% - Year-over-year growth in telemedicine applications to LegitScript, the certifier gating health ads in eleven markets. Source

32% - Share of their own CRM records that B2B sellers reckon are inaccurate, incomplete or out of date. Source