When Main Street Sports Group announced in April that it would wind down — the corporate successor to Diamond Sports, the operator of the FanDuel Sports Network portfolio, the last big load-bearing beam of the regional sports network era — twenty franchises lost their local broadcast home in a single stroke. The bids they have been fielding since are not rescue offers. They are haircuts. DAZN has been courting the orphaned teams with rights-fee cuts of as much as 40 percent. Per SBJ’s reporting on the scramble, Fubo has been offering minimum guarantees in the range of $8–15 million per team; Gray, Scripps, and Nexstar have been coming in below $10 million. And then there is the line in Yahoo Sports’ write-up of that reporting that I cannot pretend to read neutrally: “Even the free streaming service Victory+ is making a play for these rights.”
I should stop here and say plainly: I am the Chief Strategy Officer of Victory+. Everything about the company in this essay comes from third-party reporting — SBJ and the trade press — not from inside the building, and nothing here should be read as the company announcing anything.
The coverage of this repricing has treated it, almost uniformly, as a story about rights fees. Who pays $12 million, who pays $8 million, who demands a 40 percent cut, which league centralizes and which one doesn’t. I think that framing misses the actual contest. When every bid on the table is a haircut, the rights line stops being where the war is won. The war moves to the cost side — specifically, to the question of who can produce seventy or eighty professional-grade game telecasts a year for the least money. And the cost of producing a game telecast is, right now, being rewritten by a wave of AI production tooling that almost nobody covering these deals is pricing in.
The cheapest truck wins. That is the thesis. Let me earn it.
The shape of the vacuum matters, because it determines how long the window stays open.
Main Street’s wind-down, telegraphed by SBJ in January and executed at the end of the NBA and NHL seasons in April, is the terminal event of a collapse I documented at length in my Q2 research report on local sports rights. The one-paragraph version: the RSN model ran on cable-bundle subscriber fees; the bundle collapsed; a $9.6 billion leveraged acquisition became the largest concentrated repricing event in the history of American sports media. MLB got out early — the league’s clubs terminated en masse in February, and MLB Local Media now produces local telecasts for 14 of the 30 clubs, with the Angels buying their RSN outright and selling Angels.TV at $74.99 a season, the Braves launching BravesVision, and the commissioner’s office openly targeting all 30 clubs’ local rights in one national in-market package by the end of the 2028 season, per CNBC.
Baseball, in other words, answered the vacuum by absorbing the production function into the league. The NBA and NHL teams stranded this April have no such apparatus waiting. What they have is a menu of discounted bridge offers — and a league office, in the NBA’s case, actively assembling a centralized in-market streaming package targeted at the 2027-28 season, and urging its stranded teams toward one-year deals in the meantime. YouTube TV, per Awful Announcing and SBJ, is interested in the local package only if twenty or more teams come bundled together. The message under the message is that the current scramble is an interregnum. Whatever deals get signed this summer are placeholders until the map gets redrawn.
Some of the placeholders are revealing. Scripps signed the Pistons in May — its first NBA team, reported in the $8–10 million range, with a DTC app promised and details to follow. Gray extended the Suns and Mercury for four years over the air — a long-term re-up of what had been the first modern NBA deal to leave an RSN for broadcast television. DAZN, having walked away from rescuing Main Street itself, announced in April that it would merge ViewLift into its own operation — absorbing, in one roughly $100 million transaction, the streaming plumbing behind NESN, Monumental, Altitude, Space City, and the Chicago Sports Network.
A one-year deal at a 40 percent discount is not a partnership. It is a holding pattern. Everyone in this market — the broadcast groups, the streamers, the league offices — is buying position ahead of the real repricing in 2027-28. Which is exactly why the cost side matters so much right now. The operator who can run profitably during the holding pattern, at haircut rights fees, on one-year terms, with no guarantee of renewal, is the operator who gets to be standing there holding relationships and infrastructure when the big package finally gets assembled.
To see why production cost is the live variable, you have to look at what the RSN model actually paid for.
The regional sports network was funded by a per-subscriber fee — $5 to $8 a month, every month, from every cable household in the team’s territory, whether that household watched a single game or not. I walked through the mechanics of this in the Q2 report; the short version is that the marginal fan, the household that paid the fee without watching, was the entire pricing structure. And that structure funded a production standard built to national-broadcast expectations: the truck outside the arena, the camera positions, the full crew, the announce team, the replay and graphics operators, the pre-game and post-game shows wrapped around 70-plus telecasts a year. I am not going to pretend there is a clean public number for what an RSN-grade telecast costs per game — operators guard those numbers, and the setups vary enough that a single benchmark would mislead. What is public, and structural, is this: the production line was priced for a world in which millions of non-viewers paid for it.
That world is gone. The cost line it funded is still here.
This is the inheritance problem at the center of every deal being negotiated this summer. A team leaving the RSN system does not just lose a rights fee; it inherits, one way or another, a production obligation that was scaled to someone else’s business model. Either the team pays for production itself, or the league absorbs it (MLB’s answer), or the new rights partner takes it on inside an $8–15 million all-in number that has to cover rights and trucks and people. Every one of those paths runs through the same question — what does it actually cost to put a professional game on a screen in 2026? — and the honest answer is that the number is falling fast, unevenly, and mostly out of view of the people writing about the deals.
I wrote in the spring about how the production-cost collapse rewired the economics of minor league broadcasting — how the mid-tier broadcast that used to require an RSN budget can now be produced at a fraction of historical cost, and how the decision-makers of 2020 cut teams whose economics had already flipped without anyone noticing. What has changed since I wrote that essay is the arrival of a specific new layer on the cost curve. The AI layer.
Here is where the production stack actually stands, tier by tier, as of this summer.
At the youth and lower tiers, fully automated production is not an experiment; it is the default. Pixellot’s automated camera systems, distributed through partners like NBC’s SportsEngine, were on track to capture 1.5 million games in 2025, per SBJ’s reporting on the company’s latest raise — a company-supplied count, but the order of magnitude is the point. Spiideo launched AI Highlights in May, built with the AWS Generative AI Innovation Center. FloSports has deployed what it calls an “AI Producer” on FloWrestling — a virtual camera system with no human operator at all. The pattern across all three is the same and worth stating precisely: fully automated AI production exists today only where there was never a crew to replace. Nobody at the youth tier lost a job to Pixellot, because the alternative to the robot camera was no camera.
But the tooling does not stay at the tier where it was proven. Earlier this year AWS launched automated reframing of live horizontal feeds into vertical video — and the early adopters are not youth soccer platforms. They are Fox Sports Digital and NBCUniversal. This is, to my eye, the clearest pure production-cost-compression product in the market right now: a task that used to require human editors cutting vertical versions of live content, done by software, at the tier of the biggest broadcasters in the country. It is not a demo. It is deployed, on premium feeds, by companies whose entire business is premium feeds.
The premium tier itself tells a parallel story, and an instructive one. The World Cup that just ended was produced — all 104 matches, roughly 9,000 hours of content — through about 60 servers running as a private cloud in Dallas, with signal processing done in software on commodity hardware, per SVG’s excellent reporting from the IBC. Almost none of that is AI, and the trade press’s habit of writing “AI” when it means “software” has muddied this story badly. Fox called the tournament the biggest production in its 32-year history, and the AI it did announce sat at the derivative edge — the AWS vertical reframing of its digital feeds — not in the telecast itself. The humans stayed in every creative chair. What dissolved was the hardware underneath them: the dedicated truck, the purpose-built signal chain, the physical infrastructure that made every additional telecast expensive. The truck became software. That is the precondition for everything else, because once production is software-defined, AI features slot into a pipeline instead of requiring a forklift.
And at the top of the automation curve sits ESPN’s SC For You — the AI-personalized SportsCenter, built on WSC Sports’ engine, processing something like 16,000 game feeds a year with 25 tagged studio shows and a staff of 162 writers and editors feeding the pipeline, per SBJ’s reporting last November. I cite it not because a local operator needs to replicate it, but because it demonstrates the shape of the mature system: not zero humans, but a small number of humans supervising an automated content line whose per-unit output cost would have been unthinkable five years ago.
So the stack, top to bottom: software-defined infrastructure proven at the World Cup tier, AI reframing deployed at the Fox/NBCU tier, AI highlights and automated cameras proven at the youth tier and climbing. The direction of travel is unambiguous. Automation enters where there was never a crew, gets hardened at volume, and moves up-market one tier at a time — and the tier it is arriving at next is exactly the tier the RSN collapse just vacated.
Now put the two halves together, because the arithmetic is the essay.
A local rights deal at a 40 percent haircut is a catastrophe if your cost structure is the old one. The legacy operator — the broadcast group running RSN-grade trucks, the team production department staffed to 2019 expectations — cannot make $8 million a year work against a cost line built for a $5-per-subscriber world. This is why the incumbent bids all cluster at the same uncomfortable altitude: below $10 million, hedged, one year at a time. The bidders are not being cheap. They are being honest about what the old production model lets them afford.
But the same deal is a different proposition entirely for an operator whose production stack was born after the collapse. If your game telecast runs on software-defined infrastructure instead of a truck, if your highlights are cut by a model instead of a night shift, if your vertical and social derivatives are generated by an AWS service instead of an edit bay, then your cost per telecast is not the incumbent’s cost minus ten percent. It is a structurally different number. And at a structurally different number, a haircut rights fee stops being a haircut. It becomes a margin.
When the rights fee takes a 40 percent haircut, the war is not won on the rights line. It is won on the production line.
I want to be careful about what I am and am not claiming. I am not going to invent a per-game cost comparison; no operator has published one and the configurations vary too much for a single number to be honest. The claim is structural: the spread between old-stack and new-stack production cost is now wide enough to decide who can survive the terms this market is actually offering. Two bidders can look at the same $8 million-a-year, one-year, no-guarantee deal, and for one of them it is value-destroying and for the other it is a business.
There is a larger irony sitting on top of this, and it is my favorite fact in the whole landscape. Institutional capital has spent the spring pricing sports as AI insurance — Axios reported in May that sports deals are becoming “AI protection,” JPMorgan’s Mary Callahan Erdoes called sports “the antithesis of AI” in June, and PitchBook reported within the same five weeks, citing Arctos, that private equity is betting on live events as the thing AI cannot replace. The thesis is right as far as it goes: the game itself is the one media asset a model cannot synthesize. But nobody making that argument has priced its corollary. AI cannot replace the game — what it can replace is a large share of the cost of televising the game. Production-cost compression works entirely in favor of whoever holds the rights and entirely against whoever’s margin was the production. The same technology wave that makes the rights more valuable makes the legacy apparatus around the rights less necessary. Sports is AI insurance for the asset owner and AI exposure for the truck.
There is a second half to this story, and it is the part I find genuinely strange.
I spent part of this summer looking for a single shipped AI product feature at the local and team level of American sports streaming. Not a press release, not a partnership announcement, not an “exploring opportunities” quote — a feature a fan can touch. The search covered the broadcast groups building local portfolios, the TEGNA-style hybrids, MLB’s league-operated local channels, the ViewLift-built platforms that DAZN just absorbed, and the team-branded DTC apps that launched into the vacuum. I found nothing. No AI personalization, no automated highlights surfaced to fans, no catch-up features, no conversational anything. I want to hold this finding at its correct weight — it is an absence of evidence from targeted searching, not proof of a negative — but the absence is broad, and it is consistent.
Meanwhile, at the national tier, the product shelf is filling up fast. Amazon ships Rapid Recap and Key Moments on NBA games. ESPN ships Catch Up to Live and a personalized SportsCenter. Peacock is beta-testing AI-driven vertical cropping of live NBA broadcasts. The national platforms are lapping the field on product while the entire local war is being fought on a single axis: price. Every bid in SBJ’s reporting — the 40 percent haircut, the $8–15 million, the sub-$10 million broadcast offers — is a price position. Not one operator in the local market is currently differentiated by what their product does.
I read that gap as the actual opportunity hiding inside the post-RSN scramble. Fighting on price is what markets do when no one has anything else to fight with. The first local platform that ships national-grade product — highlights that find the fan, catch-up for the viewer who tunes in down two goals in the second period, a local telecast that behaves like it knows who is watching — gets to compete on something other than the size of its haircut. In a market where every offer looks like every other offer minus a million dollars, being the one platform a fan actively prefers is worth more than any line item in the deal memo.
The cheapest truck wins the current round. The smartest software wins the next one. The rare operator positioned to hold both ends of that sentence at once is the one to watch.
My Q2 research report ended on a deliberately open question. The RSN era is over — that part is settled — and four replacement models are competing for the freed rights: league-direct production, free over-the-air broadcast, team-owned DTC, and free ad-supported streaming. The question the report posed was which replacement model captures which markets, on what terms, and at what valuation when local rights next reprice.
This essay is my answer, or at least the first half of one. The model that captures the market is not a distribution model at all. Distribution is the visible layer — OTA versus DTC versus FAST versus league app — and the report is right that most franchises will end up running hybrid stacks of two or three. Underneath every one of those stacks sits the same invisible layer: a production line, priced either for the world that ended in April or for the one that started. The operators built on the old line will keep bidding what the old line lets them bid, and they will keep losing ground a million dollars at a time. The operators built on the new one can pay the same haircut prices, clear margin at them, survive the one-year holding pattern, and arrive at the 2027-28 repricing with working infrastructure, fan relationships, and a P&L that works — while everyone else arrives with a truck they can no longer afford to park.
Twenty teams became free agents this spring. Nobody bidding for them is going to win on the rights line, because there is nothing left to win there — every number is a haircut and everyone knows it. The deals will be announced on the rights line. They will be won on the production line.
The cheapest truck wins. Then the truck stops being a truck at all.
]]>Run the inventory. Amazon’s NBA coverage — year one of an eleven-year deal — ships Rapid Recap, a roughly two-minute AI-generated catch-up that plays a late tuner everything they missed and then hands them into the live feed, alongside Key Moments, real-time AI highlight curation. ESPN’s app ships Catch Up to Live. Peacock spent this spring beta-testing AI that crops a live NBA broadcast into vertical video in real time. Different companies, different leagues, different stacks — the same product. Every deployed AI feature in a major sports app compresses time.
Now run the other inventory, the one from a decade of conference keynotes: the recommendation layer. The discovery engine that would learn your teams, your players, your tolerance for a blowout, and serve you the right game at the right moment. It mostly never arrived. The closest thing at scale is ESPN’s SC For You, an AI-personalized SportsCenter that launched in beta in August 2025 — and even that personalizes on the teams you already told it you follow, not on anything it discovered about you.
I don’t read that as a failure of the technology. I read it as the technology telling us something true about the product.
Catching up is a bounded problem with a clear reward. The system knows which game you opened, knows what you missed, knows where now is. Solve it and the payoff is immediate and legible: you are back in the room while it still counts. Recommendation on live sports, by contrast, fights the calendar. The game is on when it is on. A fan does not need an algorithm to discover that their team plays tonight — the schedule made that decision months ago, and the fan organized their evening around it. The product of live sports is synchrony, being in the moment with everyone else who cares, which is the whole reason it holds attention that nothing else can. An engine that suggests other games is solving a problem the fan does not have. An engine that gets them back to now is solving the only one they do.
The industry is starting to say this out loud. Adweek came back from Cannes last month with the temperature check in its headline: sports fans want better experiences, not better algorithms. Personalization earns its keep when it simplifies — when it removes the distance between opening the app and being inside the game — not when it curates.
A decade of product roadmaps promised Netflix for sports. Fans kept asking for a faster road back to live. The features that shipped are the ones that listened.
]]>That was Carson Kelly’s record on challenged ball-strike calls as the Cubs’ catcher when ESPN checked in this May — twenty-one calls overturned, four lost. Baseball Savant scored the work at 2.3 runs above expectation. Not runs from hitting. Not runs from framing, blocking, or throwing out runners. Runs from knowing the strike zone better than the umpire standing behind him, and proving it, in public, twenty-one times out of twenty-five.
A year ago that skill did not exist as a measurable thing. There was no column for it, no run value, no leaderboard. Now there is a public dashboard.
Half a season into Major League Baseball’s robot-umpire era, this is what the machine has produced: a new market, a new skill, a new anxiety — and, the part I keep coming back to, a changed game on the thousands of pitches where the machine did nothing at all.
The ABS challenge system went live on Opening Day this season, and its design is worth pausing on, because the design is most of the story. MLB did not hand ball-strike calls to the machine. The umpire still calls every pitch. What the league added is an appeal court. Each team gets two challenges per game and keeps any challenge it wins. Only the pitcher, the catcher, or the batter can invoke it — no dugout signals, no manager theatrics, no video coordinator whispering into a headset. The player taps the top of his cap within roughly two seconds of the call, the Hawk-Eye tracking layer that already powers Statcast renders the verdict over a T-Mobile 5G link, an animation goes up on the board, and the game moves on. The whole exchange takes seconds.
Notice what the design refuses to do. It refuses to make the machine the default. It refuses to give the decision to anyone who was not on the field feeling the pitch. It refuses to allow unlimited appeals, which means every challenge is a bet placed with a scarce resource. This is a league that has spent the decade making deliberate product decisions — the pitch clock, two games every Friday night sold to Apple as a habit rather than a broadcast — and the challenge system carries the same fingerprints. The machine was installed the way you would install a guest you were not sure about. Powerful, but on probation, and only allowed to speak when spoken to.
Half a season in, the guest has rearranged the furniture anyway.
Start with the totals, because they undercut the loudest prediction about this system before we get to the interesting part.
By mid-April, when CBS Sports ran the numbers, players had issued 1,050 challenges — 4.05 per game, covering roughly 1.4% of all pitches — and won 54% of them.
Sit with that success rate for a moment. These were not random appeals. Challenges are scarce, losing one costs you the resource, and the only people allowed to burn one are the three players with the best view in the building. These were the calls that professional hitters and catchers felt surest the umpire had missed. And the umpires were still right almost half the time.
The pre-season fear — that the machine would expose umpires as frauds and hollow out the human element — inverted on contact with data. The half-season record reads closer to a vindication: on the most contested one-and-a-half percent of pitches, the humans split the difference with a tracking system accurate to fractions of an inch. Half a season of this data has, if anything, increased my appreciation for the umpires. The machine did not dethrone the umpire. It gave the umpire a grade, and the grade was better than the folklore.
The distribution of who wins challenges is where the texture starts. In spring training, the defense won 60% of its challenges while batters won only 45%. That gap makes intuitive sense once you say it out loud — the catcher watches the pitch into his glove from directly behind it, while the hitter is processing a ball moving at close to a hundred miles an hour with an at-bat’s worth of adrenaline in his bloodstream. Conviction and accuracy are different skills. The batter feels certain. The catcher knows.
If the story ended there — a well-designed appeals system, a modest error-correction rate, a shrug from the fans — this would be a footnote in the history of sports officiating.
It does not end there.
In late April, the Associated Press published a piece I have been thinking about ever since. Players around the league, it reported, believe the challenge system is shrinking the effective strike zone. Umpires now get instant feedback on every challenged call, and — in the words of Diamondbacks catcher James McCann — the zone is “tighter in general,” the calls “much more uniform.” Hitters, sensing the shift, have adjusted — laying off the borderline pitches they used to defend against. And walks spiked: 9.8 percent of plate appearances through the season’s first month, a rate the league had not seen since 1950. The AP was careful with the causality — early-season walk rates always run high, and nobody can prove ABS is the reason — but even adjusted for the calendar, the jump from last season was, in its word, massive.
Hold the two numbers side by side. The machine directly touches about 1.4% of pitches. The behavioral shadow it casts touches all of them.
The machine's biggest effect on baseball has not come from the calls it overturned. It has come from the calls it never made.
The mechanism is not mysterious; it is incentives. For a century, an umpire’s borderline strike call was an assertion. It could be argued with, but it could not be falsified — not in the moment, not in front of the crowd, not on the scoreboard. The challenge system changed the epistemic status of every close call. Any borderline strike is now, potentially, two seconds and a cap-tap away from being rendered as an animation that shows the ball missing the zone by half an inch, in front of thirty thousand people and a broadcast audience. Umpires are professionals with reputations, and professionals with reputations respond to falsifiability the way anyone does. They retreat toward the calls that cannot embarrass them.
This is the observer effect, and it is not new. Physicists have known for a century that measuring a system disturbs it. Management theorists have known nearly as long — the famous studies at Western Electric’s Hawthorne plant found that workers’ output changed when researchers watched them, regardless of what the researchers actually changed. Being observed is itself an intervention. Baseball has now run the cleanest version of this experiment that professional sports has ever produced: install a perfect observer, let it act on 1.4% of events, and watch the other 98.6% move.
I wrote earlier this year about the measurement problem in creator economics — the way measurement frameworks quietly reshape the businesses they claim to merely describe. That process usually takes years, because dashboards work on humans slowly, through budget meetings and performance reviews. Baseball compressed it into weeks. The umpires did not need a quarterly review to internalize the new incentive. They felt it the first time a cap-tap turned one of their strikes into a ball on the videoboard.
And here is the detail that makes it art: nobody decided this. MLB did redraw the zone’s formal geometry for the machine — a precise slab, seventeen inches wide, pegged to a percentage of each batter’s height — but the shrink the players describe is not in the rulebook. The humans shrank the effective zone, in anticipation of the machine, one flinched borderline call at a time. The walk spike is not an output of the technology. It is an output of the psychology the technology created.
Every new measurement creates a market in whatever it measures, and this one formed with startling speed.
Carson Kelly’s 21-4 is the headline case — an 84% success rate against a league average of 54%, worth 2.3 runs above expectation by Baseball Savant’s accounting. Two-point-three runs is not a rounding error; it is real value, produced by a skill nobody could price a year ago. And because successful challenges are retained, the skill compounds: a catcher who challenges well does not just win calls, he preserves his team’s optionality deep into games, which is worth something extra in every close ninth inning. The system’s design accidentally invented a portfolio-management job and handed it to the catcher.
The market noticed immediately. By early July the Associated Press was writing up Hunter Goodman’s challenge record the way wire services used to cover hitting streaks. When challenge skill gets the feature-story treatment inside one season, it has already crossed from curiosity to attribute — the kind of thing that shows up in scouting reports and, eventually, in arbitration hearings. Because the dashboard is public, every front office can see which of its players spend challenges well and which ones torch them on wishcasting. Somewhere in every analytics department, there is now a spreadsheet ranking players by challenge discipline, and I would bet the spread between the best and worst organizations on that spreadsheet is wider than anyone expected.
Every market has a short side, and the short side of this one is pitch framing. For a decade, framing was the sabermetric community’s favorite hidden skill — catchers who could receive a borderline pitch so smoothly that the umpire called it a strike were quietly worth wins, and teams paid for the craft accordingly. The challenge system attacks framing from both ends at once. The stolen strike can now be challenged and un-stolen. And the shrinking zone means there are fewer borderline calls being awarded in the first place. The early ESPN and FanGraphs comparisons point the same direction: the elite framers are among the system’s first losers. A skill that took the analytics movement a decade to price is being repriced in a single season.
Though not everyone agrees on the direction. Will Smith — the Dodgers catcher, a man whose living sits directly on this fault line — offered the counterpoint in the Los Angeles Times in May: under a challenge format, framing becomes “more important, in a way.” I read his logic like this: the machine only speaks when invoked, which means the human umpire still calls more than 98% of pitches, and a catcher who can still win those borderline calls forces the other team to spend scarce challenges disputing them. Framing used to steal strikes. Now it can also tax the opponent’s challenge budget. Whether that nets out positive is exactly the kind of question the next half-season of data will answer, but the fact that the league’s best-paid catchers are openly theorizing about it tells you how fast the ground is moving under them.
The umpires, meanwhile, are living the other side of the observer effect. Former umpires have described to reporters the particular anxiety of working under the system — every close call now carrying the possibility of instant, animated, public contradiction. It is worth saying that the umpires’ union agreed to all of this inside a routine collective bargaining agreement. There was no labor war over the robot. The fight everyone predicted simply did not happen; the anxiety got absorbed as a working condition, the way surveillance usually does.
There is one more number in this dataset, and it is the one the enthusiasts skip.
When MLB detailed the system’s accuracy — figures given to The Athletic and reported widely in April — the league said the tracking is accurate to within 0.39 inches at a 95% confidence level, and within 0.48 inches at 99%.
Read that as an engineer and it is a triumph: sub-half-inch precision on a ball moving a hundred miles an hour. Read it as an epistemologist and it is a confession. The machine has an error bar. Roughly one call in twenty could be off by more than four-tenths of an inch — and some of the calls the machine overturns are calls the umpire missed by less than that.
Which means that on the closest pitches, the challenge system is not necessarily replacing a wrong answer with a right one. It is replacing one uncertain measurement with another uncertain measurement that happens to carry more authority. The umpire’s zone came with visible fallibility — a face, a stance, a history of blown calls you could yell about. The machine’s zone comes with an error distribution nobody can see and an animation rendered in the crisp, confident graphics of certainty. The animation never wobbles. The confidence interval never makes the broadcast.
I do not raise this to argue the machine is bad at its job. It is astonishingly good at its job, and materially better than any alternative on the closest calls. I raise it because the difference between accurate and authoritative is exactly the gap every league adopting these systems is about to fall into. We did not install truth behind home plate this season. We installed a measurement — a very good one — and agreed, collectively, to treat its output as truth because arguing with an error bar is unsatisfying. The strike zone did not become objective in 2026. It became settled, which is a different thing, and the fact that the sport is happier with settled than it ever was with human is one of the more revealing findings of the whole experiment.
The diffusion has already started. The SEC ran ABS-style challenges at its 2026 baseball tournament in May — which means there are now college players building challenge discipline as a skill before they ever sign a professional contract, the way previous generations built framing. The attribute is propagating downward faster than the infrastructure is.
And it will not stop at baseball. Adam Silver said in May, on the record, that NBA out-of-bounds calls “will be done by an AI-automated system with cameras lined around the court” — taking, in his words, the whole category of objective calls out of the referees’ hands — and the league has had Sony’s Hawk-Eye tracking installed league-wide since 2023 to do it. Gary Bettman wants AI to resolve “where exactly is the puck.” Cathy Engelbert says the WNBA is looking at the same AI playbook as its “big brother,” the NBA. Every commissioner in American sports is now, in effect, promising to hire the observer.
Here is what I would tell each of them, from baseball’s half season of data.
The system will work. The calls will be more accurate, the appeals will be fast, the fans will absorb it within a month, and the labor war you are budgeting for will probably never arrive. All of that is the easy part, and none of it is the point.
The point is that you are not installing a tool. You are installing a witness. And the consistent lesson of every watched workplace since the Hawthorne plant — now confirmed at the scale of a major professional sport — is that the watched do not stay the same. Your officials will shade their judgment toward whatever the machine cannot contradict. Your athletes will discover, within weeks, which behaviors the machine rewards, and a new skill market will form around gaming the appeal before your competition committee has scheduled its first review meeting. The second-order effects will be larger than the first-order ones, they will arrive faster, and they will not be the effects you designed.
Baseball’s machine was given the narrowest possible mandate — speak only when spoken to, rule on one pitch at a time, touch 1.4% of the game. It still moved the walk rate of the entire sport, repriced a decade-old catching skill, created a new one from nothing, and changed how every umpire in the league experiences a two-strike count.
Every measurement system changes the thing it measures. The leagues lining up behind Silver’s promise are not adopting a technology; they are agreeing to be observed, and observation is never neutral. The machine behind home plate had made its first thousand calls by the middle of April. Its real work is everything the humans did because it was watching.
]]>I want to sit with that, because the tournament we just watched — all 104 matches of it, in my case — was covered for 39 days as the “AI World Cup,” and the description is wrong in a way that matters.
Here is what actually happened, per Sports Video Group’s on-site reporting from the International Broadcast Centre in June. FIFA and its host broadcaster HBS ran feeds from sixteen venues across three countries into the Dallas IBC, with a second, non-live hub in London. Every stage of signal processing ran as software on commodity hardware over SMPTE ST 2110 — roughly 9,000 hours of content across the tournament, every match in UHD HDR. The public cloud touched only the social and digital clips. Everything that mattered lived on those sixty servers.
That is a genuine engineering revolution. The production truck — the great rolling cathedral of broadcast television — became software. It is the biggest architectural change in live sports production in decades, and it happened at the biggest tournament in the sport’s history.
And there was a human in every creative chair. Fox, which called this the biggest production it has ever done, kept its AI where the viewers scroll — personalization inside the Fox One app — not in how the broadcast got made. The directors were people. The replay operators were people. The camera operators were people, standing behind cameras, at sixteen stadiums.
So where did the “AI World Cup” come from? The sponsors. Lenovo’s Football AI Pro — the same AI match analyst handed to all 48 competing teams — arrived by press release. Salesforce’s tournament AI arrived as sponsorship announcements, with no shipped product anyone independently verified. The trade press took the releases and wrote “AI” where the accurate word was “software.”
Where AI genuinely runs a sports production, it is at the other end of the pyramid — the youth and amateur tiers where there was never a crew. Pixellot automates camera work for youth games at enormous scale — more than a million streamed games a year, by its own count. Spiideo launched AI highlights in May. FloSports put a camera with no operator on live wrestling. AI is automating production where there was nobody to replace — which is exactly where automation makes sense, and exactly not where the headlines put it.
The two words have very different budgets attached. “Software” is a cost line: engineers, ST 2110 infrastructure, a facility in Dallas, a plan measured in years. “AI” is a story: a sponsorship deliverable, a press release, an adjective for the earnings call. The most consequential transformation in live production this decade just shipped at this World Cup, and the industry covering it filed it under the wrong word.
]]>The next day, FIFA affirmed the call. VAR and the semi-automated system had functioned exactly as designed; the touch was real; the goal was correctly disallowed. Croatian fans, and a healthy share of neutral ones, reached the opposite verdict: robbed.
Here is the uncomfortable part. Both sides are almost certainly right. The machine detected something real, and the fans were shown nothing they could check. Automated officiating’s problem in 2026 is not accuracy — by any honest reading of the evidence, the machines are better at this than we are. The problem is legitimacy, and the institutions running the machines are failing at the one thing that has ever bought legitimacy: showing the governed the evidence.
I have watched every match of this tournament, the way I have watched every match of every World Cup since 1990. I want the calls to be correct. I carry no nostalgia for the linesman’s guess. And when the Croatia decision landed — on the broadcast, in the group chats, across every second screen — what I watched was not a technology failing. It was an institution declining to make its case.
The system that made the call is genuinely remarkable. FIFA’s advanced semi-automated offside runs on skeletal tracking of every player on the pitch, a 500Hz inertial sensor inside the adidas Trionda match ball, and per-player 3D avatars built from full-body scans of every squad in the tournament. When the system detects a clear offside, the alert goes straight to the on-field officials’ earpieces. It does not even pass through the VAR relay anymore. The design goal is obvious: kill the four-minute wait, deliver certainty at the speed of play.
It has not quite gotten there, even for the people wearing the earpiece. The Guardian interviewed World Cup assistant referee Micheal Barwegen in June, and he described the experience from inside the loop: “The computer has to think, and it’s super fast, but [on the field] it feels like forever.” The humans, it turns out, still out-call the machine on speed. The machine wins on certainty. Everyone on the pitch is now waiting for the same verdict the rest of us are.
And how certain is it, exactly? Here is a telling detail. What FIFA has disclosed is a threshold: the upgraded system now sends its audio alert when a player is more than ten centimeters offside, down from fifty at the last World Cup. What it has not published is an error bar — how confident the system is in any given ten-centimeter judgment, how often it might be wrong, and by how much. The precision of the most consequential officiating instrument in world football is, as of this writing, an undisclosed number.
Which brings us back to the Matanović touch. It is below the threshold of human perception. The replay cannot show it because the cameras did not resolve it; the only witness is a sensor inside the ball, sampling five hundred times a second, legible exclusively to the people who operate it. The call may be perfect. It is also, to every human being who does not work for FIFA, unfalsifiable.
That is a new condition in sports.
FIFA does release evidence — sometimes. The tournament has produced offside animations for plenty of decisions, rendered avatars frozen at the moment of the pass, the kind of exhibit that settles arguments — even for calls nobody disputed. But on June 13, when an offside check stood between Switzerland and a penalty against Qatar, the avatar animation never came. FIFA blamed a “brief technical outage,” and the evidence it finally produced, four and a half hours after the call, was a pair of old-fashioned hand-drawn VAR lines the BBC judged unconvincing. Gary Neville was calling the withholding “like a dictatorship” on air before the statement even landed. The avatar exhibit for that call has never surfaced. Animations for the routine cases, a four-hour vacuum for the contested one.
It scarcely matters that the cause was a malfunction rather than a decision; the audience only sees the pattern. Evidence that appears instantly for the routine calls and hours late for the contested ones teaches the audience exactly what the disclosures are for. A court that publishes its reasoning only when the ruling is easy is not somewhat legitimate. It is performing legitimacy, and everyone watching can tell.
By early July the discourse had found its register. Forbes asked “Has Technology Ruined The World Cup?” the day after the Croatia decision; the BBC’s The Inquiry had devoted an episode in late June to “Is technology ruining sport?” The framing is wrong in both cases — the technology is working, probably better than any officiating instrument in the history of the game. But when mainstream outlets reach for the word ruined about a system with a near-flawless accuracy record, that tells you something important about where the deficit actually is. It is not in the sensors.
It is worth saying that the machine giveth too. On June 19, the United States had a Freeman goal awarded against Australia after the system’s automatic review overturned an on-field offside flag. Nobody demanded the sensor data. Nobody ever does when the call goes their way — which is precisely why the disclosure policy cannot be improvised match by match. Legitimacy is never tested by the calls in your favor. It is built, in advance, for the day the machine rules against you on evidence you cannot see.
Now look at baseball, which is running the same experiment with the opposite disclosure posture.
Since Opening Day, MLB has been live with the ABS challenge system: Hawk-Eye tracking underneath, two challenges per team, retained if successful. Only the pitcher, catcher, or batter can challenge, with a tap of the cap inside about two seconds — no dugout committees, no replay huddles. By mid-April, per CBS Sports, teams had thrown 1,050 challenges at the machine, at a rate of about 4 per game and covering roughly 1.4 percent of pitches. The success rate: 54 percent.
Two design choices separate MLB’s system from FIFA’s, and neither is about accuracy. First, MLB published its machine’s fallibility. The league’s stated precision — figures given to The Athletic — is 95 percent confidence within 0.39 inches, 99 percent within 0.48. Read that again: the machine has an error bar wider than some of the margins it overturns. MLB printed that fact anyway. Second, the whole dataset is public. Every challenge of the season sits on a live dashboard at Baseball Savant, auditable by anyone with a browser, and the animation of every challenged pitch plays on the broadcast and the ballpark videoboard within seconds of the cap-tap. The evidence is not released on appeal. It is the product.
Has radical disclosure destroyed faith in the system? The opposite, as far as anyone can measure. One survey of two thousand Americans in May found 21 percent saying ABS will make them watch more baseball, and 54 percent saying it moves the needle not at all — automation as furniture, not crisis. Challenge skill has become a scouted attribute: Cubs catcher Carson Kelly went 21-4 on challenges through mid-May, worth 2.3 runs above expectation per Baseball Savant. Fans argue about whether their catcher challenges well, the way they argue about whether he frames well. The machine got absorbed into the sport’s argument culture instead of ending it.
The most interesting finding in the half-season data is stranger still: the machine changed the game without making a call. Players told the AP in April that the effective zone has gotten tighter — umpires now get instant feedback on every challenged call, and the borderline strike is disappearing; hitters have adjusted, and walks through the season’s first month ran at 9.8 percent of plate appearances, a rate the league had not seen since 1950 — a jump the AP called massive even adjusted for April’s usual walk inflation. The machine’s behavioral shadow is bigger than its interventions. That is a real cost, and worth its own essay. But note that we know about it — players discussed it openly, reporters quantified it, the league’s own public data made the analysis possible. In an open system, even the second-order problems get litigated in daylight.
Baseball shows its work. Football files the work in a drawer and issues a statement.
If you suspect the answer is simply to keep humans in the chair, tennis ran that experiment this season, and it is worth watching closely, because tennis ran both arms of the trial at once.
The French Open is the last Grand Slam still using human line judges. Amélie Mauresmo defended the choice on the grounds that “the machine is not 100 percent reliable.” She is right — no machine is, as MLB’s published error bar cheerfully concedes. And then her tournament produced the worst line-call scandal of the tennis season, a missed call in Ruud–Fonseca on May 31 that had Jim Courier saying on air that electronic line calling “makes far fewer mistakes than humans.” The clay-court distrust runs the other way too: Elena Rybakina, after a no-mark call went against her in Madrid this spring, said she “won’t trust it at all.” And the worst automation incident at Wimbledon remains last July’s Pavlyuchenkova–Kartal game, in which the electronic system missed a call because a human had switched it off — the club apologized. The failure modes have migrated from judgment to operations, but they have not disappeared.
So: the humans miss calls, the machines get switched off, the players trust neither, and the tournament that kept the humans got punished for it hardest. Neither arm of the experiment produced trust, because trust was never a function of who makes the call. It is a function of whether the governed can check the call. Tennis’s electronic calls at least leave a mark on a screen. The French Open’s human calls leave a mark on the clay — which is exactly why players stand and point at it. Everyone in that sport understands, instinctively, that the argument is settled by shared evidence. The only unforgivable state is the call with no mark at all.
For half a century, the story of officiating evidence was a story of democratization.
Instant replay’s deal with the audience was radical when it was new and invisible by the time I was watching cricket and football as a kid in India: everyone in the bar saw what the referee saw. Slow motion, then more angles, then super-slow-motion, then every phone in the stadium. The entire argument culture of modern sports — talk radio, the pub table, the group chat, first-take television — is built on shared access to the same frames. The referee had authority, but the evidence was common property. Even the VAR era, for all its agonies, kept the deal intact. The freeze-frame with the lines drawn on it was mockable, contestable, endlessly relitigated — but it was visible. You could hate the lines and still see them.
Cricket, the sport I grew up arguing about, understood this better than anyone and turned it into theater. When a review goes upstairs, the evidence plays on the giant screen for the entire ground — the ball-tracking arc, the audio spike on the edge, each stage of the verdict revealed to sixty thousand people at once, who gasp and groan at every frame like a jury that has been handed the exhibits. The decision review is one of the most-watched moments in the modern game precisely because the crowd is inside it. Nobody in that stadium is asked to trust a sensor they cannot see. The evidence is the show. That is what a tracking system looks like when the institution running it decides that the audience is a party to the proceedings rather than a security risk.
Sub-perceptual tracking quietly reverses the whole fifty-year arc. The decisive fact in Croatia–Portugal is not in any video. There is no frame to freeze. The evidence has moved from the screen — where everyone could see it — into a data stream only a priesthood can read. Officiating evidence, democratized for two generations, is being re-aristocratized in a single tournament cycle. You are no longer shown the proof. You are asked to trust that it exists.
Verifiability, not accuracy, is what legitimacy runs on.
The tempting frame here is man versus machine, and I have written before about the danger of binaries — this one obscures more than it reveals. Because the men signed up. NFL officials ratified a seven-year labor deal by a 116-4 vote this spring, months after a season in which the league ran virtual first-down measurement league-wide. MLB’s umpires accepted ABS inside a routine collective bargaining agreement. Barwegen, the World Cup assistant referee, says the system makes his job easier in some ways. The officiating labor war that a decade of think pieces predicted simply never happened. The real division in automated officiating is not man versus machine. It is open versus closed.
And closed is about to scale. Adam Silver said in May that NBA out-of-bounds calls “will be done by an AI-automated system with cameras lined around the court.” Every major league is walking toward the same architecture — tracking layers that adjudicate below the threshold of what a camera can show a fan. Whether the governed get to see the evidence is being settled right now, league by league, mostly by default, one press release at a time.
This essay publishes on a Friday. The World Cup final is two days later, on Sunday, at MetLife Stadium. Somewhere in that match — a tight offside, a disallowed goal, a penalty awarded off a touch nobody in the stadium perceived — there is a decent chance the machine will speak. It will very probably be right. The question FIFA has two days to think about is whether anyone will believe it.
The fix does not require a task force. It is a policy, and it fits in a sentence: publish the evidence, every time. Every semi-automated decision releases its full case at the moment of the call — the animation, the sensor reading, the margin, the confidence interval — as a matter of routine rather than discretion. Not only when the exhibit is flattering. Not withheld when the margin is embarrassing. Not on appeal. The institutional instinct says that showing uncertainty invites doubt. Baseball just spent half a season demonstrating the opposite: MLB published an error bar wider than its overturned margins, put every decision on a public dashboard, and fans metabolized it into the ordinary texture of the sport within weeks. The disclosure absorbed the doubt. The drawer manufactures it.
Sports adopted these systems because the calls matter — because a World Cup turns on a fingertip, and getting the fingertip right is worth building a 500-samples-a-second sensor into the ball. But the call was never the product. The shared conviction that the call was fair is the product, and that conviction has only ever been built one way, in courtrooms and elections and penalty boxes alike: show the governed the evidence.
On Sunday, the machine will almost certainly be right. Whether anybody believes it is a decision FIFA still gets to make in advance.
Make the right call. Show us.
]]>It is styled in the same broadsheet system as the rest of this journal — Newsreader at 19 pixels, vermilion link color, the same short centered rule between sections.
But I had to get Claude to figure out a workaround.
Buttondown’s free plan blocks two things. The first is RSS-to-email automation — the feature that watches a feed and sends each new post as an email when it appears. That is the entire reason most writers use Buttondown. The second is custom CSS — the feature that lets the email actually look like the website it came from. Both blocks surfaced as crisp 400-series errors during setup, complete with neat error codes and a link to upgrade.
These are not arbitrary blocks. They are precisely the two features that produce the most willingness to pay. A free-tier limit is a pricing experiment — vendors gate the features customers want most, because that is the line where the credit card comes out. I do not begrudge them the experiment.
I just had time, and Claude Max subscription.
The publish workflow on this site already runs on a cron. Every Friday morning at 14:30 UTC, it promotes the oldest draft from _drafts/ into _posts/, rewrites the date in the front matter, builds the site, and deploys to GitHub Pages. Adding a final step to that workflow that POSTs the same post to Buttondown’s API — bypassing the dashboard’s RSS-to-email feature entirely — was about thirty lines of YAML and one hundred and fifty lines of Python.
The Python script does the boring, necessary things. It reads the front matter for the title. It strips the Jekyll chrome that does not render in email — the marginalia asides, pull-quotes, custom deposition blocks. It rewrites relative links to absolute ones, because /2024/09/09/foo.html is meaningless in a Gmail tab. It converts kramdown footnotes into a numbered Notes section, because Buttondown’s Markdown renderer does not handle them natively.
It then renders to HTML and inlines the broadsheet styles per element. Inline styles are the workaround for the second wall — the same rules documented in bin/newsletter.css (the file the script would PATCH onto the newsletter object if the free plan would accept it), copied into a Python dict because the API returned css__not_allowed.
Inline styles are more reliable anyway. Outlook strips <style> blocks and ignores @import. Inline is the only form every email client honors without negotiation.
The whole pipeline cost an afternoon. The plan I am paying for is $0. The next plan up at Buttondown is $108 a year.
The math is not interesting on its own. What is interesting is that the same shape of math holds for a lot of SaaS product walls. The features that get gated are usually the features that someone with a couple of hours and access to an LLM can replicate against the public API. Vendors price for the median customer, not the curious one, and certainly not for the AI-enabled customer.
Whether to do that or to pay for the gate is a personal call. I happen to find the workaround slightly more enjoyable than the upgrade — the workaround leaves a small, well-documented Python file in the repo that does exactly what I need, and nothing else. The upgrade would have left a recurring line item.
If this dispatch arrived in your inbox, that script is what got it there.
]]>Three seasons in, it is worth saying plainly. YouTube TV won NFL Sunday Ticket — carrying it since the 2023 season on a seven-year deal reported at roughly $2 billion per year — because it is the only bidder that understood the package was not really about the games. It was about the couch.
DirecTV had the games from 1994 through the 2022 season — almost three decades. The bid that lost was not a bad bid. On paper, the economics were close. On paper, the distribution footprint was close enough. On paper, you could have argued either way.
Paper is not where football gets watched.
Football gets watched on a TV, by a person holding a remote, at one in the afternoon on a Sunday, with people walking in and out of the room. That is the product. Whoever wins that moment wins the package.
I have written before about what the remote control actually does now. It is not a channel changer. It is a launcher. The buttons on it — the ones literally printed on the hardware — are sold to the streaming apps that paid to be there. The home screen is the new dial.
A sports package that lives inside the TV’s native interface — one that opens when you turn the TV on, one that a guest can find without asking for your login, one that surfaces the 1pm window without you having to remember which of four apps you subscribed to — wins before it has shown a single play.
That is the variable that gets systematically undercounted in bidding decks. You can model churn. You can model ARPU. You cannot model the hesitation of a sixty-five-year-old who stopped watching because the login flow felt like a dental form.
The direct-to-consumer sports apps — the ones each league keeps trying to float — are not losing on content. The content is often better than what cable carried. They are losing on the ten seconds between sitting down and seeing a score.
Ten seconds is an eternity in a living room. The DTC app that asks for re-authentication on a new TV, the one that does not remember which team you follow on which device, the one that buries the live game three taps deep under a “Browse” menu — those frictions are not small. They are the whole competition.
YouTube TV inherits the native TV grammar. Guide. Channels. Pause. Rewind. Record. The experience of watching Sunday Ticket through it is the experience of watching football on TV — which is the only experience the package was ever really selling.
Every league negotiating a next-cycle package is running the same math. Reach. Price. Exclusivity windows. Addressable ads.
The variable that is not on the spreadsheet, and that decided this one, is whether the winning bidder has built something a person can turn on without thinking.
DirecTV had that once. The satellite dish was, in its era, the most seamless sports product in America — turn on the TV, the game is there. YouTube TV rebuilt that feeling for a house that no longer has a dish. The rest of the industry spent the same years building apps that required a fan to remember a password.
One of those strategies was going to win Sunday Ticket. The surprise is not that it did. The surprise is that anyone expected otherwise.
]]>Most of the writing on this number reads as grief. The publishing industry has lost something it used to have, the loss is real, and the response has been to point out that the loss is real, repeatedly, in increasingly precise percentages.
I want to write the structural version instead.
Zero-click is not a thing that is happening to content. It is a re-pricing of content categories. Some categories were always going to be answered in a sentence — what time the Super Bowl starts, who won the 2014 World Cup, the boiling point of water, the capital of Bolivia. Those queries used to land on a publisher’s page because Google had not yet figured out how to deliver the answer without sending you there. Google has now figured it out. The query no longer needs the page.
The category was never the page. The category was always the answer.
The mistake, going back to the early 2010s SEO arbitrage era, was treating those queries as content. They were not content; they were lookups. A lookup with a clean answer is a lookup that the platform will eventually internalize. The fact that publishers built entire businesses on aggregating lookups was a side effect of Google not yet being good enough to skip the publisher.
The categories that survive zero-click are the ones whose value cannot be compressed into a sentence. Investigative reporting. Long-form analysis. Reviews that synthesize hundreds of data points into a judgment. Personal essays whose argument depends on the writer’s specific authority. Newsletters that subscribers buy because they want one person’s voice in their inbox three times a week. Anything where the judgment is the product, not the lookup.
This is not a comforting thought for most of the publishing industry, because most of the publishing industry was never selling judgment. It was selling lookups dressed up as articles. The dressing-up was the business model. Zero-click is removing the dressing.
The cohort that gets to keep operating in the zero-click era is the cohort whose product was never a commodity in the first place. It is also a smaller cohort. The Verge does a different kind of journalism than the SEO factories whose traffic is down 85% or more. The outlets that have held up better against AI Overviews are the ones whose readers come direct, not through search — the ones that built a relationship rather than a funnel.
There is a real business in zero-click, and it does not look like the old business. It looks like a smaller audience that pays more, comes back more, and trusts the byline. It looks like a newsletter, or a paid Substack, or a private community, or a podcast with a sponsorship instead of programmatic ads. None of that scales to AdSense numbers. All of it survives.
The grief stage of zero-click is over. The structural stage is what the surviving cohort is willing to do that the lookup cohort never had to.
]]>The argument is technically defensible. The argument is also, almost line for line, the same one regional sports network executives offered in 2018 about the cable bundle.
I want to write about why those two arguments belong in the same essay.
Start with the numbers, because the numbers are not in dispute. The disagreement is about what they mean.
A field study published last year tracked 42% of organic queries triggering AI Overviews and measured outbound clicks at 0.38 per search; with AI Overviews removed, the same set of queries produced 0.61 clicks per search. That is a 38% reduction in clicks attributable to the Overview. Other studies have reported organic CTR drops between 34% and 61% on AIO-triggered keywords.
Publisher-level data tells the same story at higher resolution. Wired’s Google traffic is down 62% over the AI Overviews rollout window. The Verge, HowToGeek, and ZDNet have each lost more than 85%. Mashable is down 30%. Digital Trends went from 8.5 million monthly Google clicks in March 2024 to 264,861 in January 2026, which is a number that takes a moment to process. Chartbeat reported that small publishers — sites in the 1,000 to 10,000 daily-pageview range — saw Google referral traffic drop by roughly 60% over two years. The IAB Tech Lab estimated AI summaries reduce publisher traffic 20%–60% on average and translate to roughly $2 billion in annual ad revenue losses across the publishing sector.
These are not edge cases. These are not contested numbers. Even Google does not really dispute them in aggregate; what Google disputes is the claim that the clicks lost were the kind of clicks that ever mattered.
Hold that disagreement in mind, because it is the entire game.
The Google argument, in its strongest form, runs like this.
Search has always had a long tail of low-value clicks. A query like “what time does the Super Bowl start” used to send the user to a page that took eight seconds to load, served a banner ad, and answered the question in one sentence the user could have read in a knowledge panel. The user got nothing valuable from the visit. The publisher got a click that was not going to convert, was not going to retain, was not going to do anything for the business beyond produce one CPM impression at the bottom of the page. AI Overviews remove that click. The publisher loses the impression. The user saves five seconds. Net welfare goes up.
The clicks that survive, in this telling, are the clicks the publisher actually wanted — readers who want to spend time on the page, sign up for the newsletter, return for the next article, become a subscriber. Those readers are the publisher’s real audience. The bounce traffic was always noise.
Google has separately argued that the methodologies used in the public studies are flawed and that the company “consistently directs billions of clicks to websites daily” without observing significant drops in aggregate web traffic. More recently, the company has pointed to AI Overviews CTR rising from 1.3% in late 2025 to 2.4% by February 2026 as evidence that the click-through rate is recovering as the product matures.
I want to be careful here. The defense is not nothing. There genuinely is a category of click that the publisher never monetized well. There genuinely is a recovery underway in raw AIO click-through rates. There genuinely are publishers — direct-traffic newsletters, paid Substacks, vertical specialists — who never depended on the bounce-click economy and are not seeing the decline.
The defense is true. That is the part of the defense that is worth saying out loud, because if it were obviously false, this essay would not be necessary.
The problem with the defense is not that it is false. The problem is that it is the same shape, almost word for word, as a defense I have already watched run its course in my own industry.
"The new product is cannibalizing some of the old product, but the old product still has revenue, and the cannibalized portion was the marginal portion." That sentence describes the AI Overviews defense. It also describes every regional sports network deck I saw between 2017 and 2022.
I have spent enough of my career in sports media to know what the conversation around regional sports networks sounded like in the back half of the last decade. By 2018, cable subscriptions had been declining for six straight years. The DTC narrative was already in motion. ESPN+ had launched. DAZN was building out its U.S. footprint. Cord-cutting headlines ran weekly.
What the RSN side said, repeatedly, was a structurally specific version of the bounce-click argument.
The cable bundle was still, by an enormous margin, the largest revenue source for the franchise. Local rights inside the RSN system paid more per fan than any direct-to-consumer alternative could plausibly replicate. The fans who were leaving cable were, the argument went, the marginal fans — the ones who did not watch enough games to justify the bundle anyway. The fans who actually loved the team would always pay for the bundle, because the bundle was where the games lived. The cord-cutters were leaving for reasons that had nothing to do with sports — they were leaving for Game of Thrones and House of Cards and Netflix originals. Sports was the anchor of the bundle. The anchor would hold.
Every line in that paragraph was true at the time it was spoken. The cable bundle was the largest revenue source. RSN economics did pay more per fan than DTC could match in the late 2010s. Many of the cord-cutters were leaving for reasons unrelated to sports. The structural arguments were directionally correct. The math worked, in the particular accounting frame the RSN executives were operating in.
Diamond Sports Group, the entity that held the largest portfolio of regional sports networks in the country, filed for Chapter 11 bankruptcy in March 2023. It re-emerged as FanDuel Sports Network in early 2025 after a rebrand via FanDuel naming rights, and is now in a second wind-down. Thirteen NBA teams and seven NHL teams and most of the affected MLB clubs spent late 2025 and early 2026 watching their primary local-rights infrastructure dissolve in real time, with most of them now scrambling to assemble local-rights stacks outside the cable bundle.
That is not what an industry that is losing the marginal fan looks like. That is what an industry that priced its model against the marginal fan and found out, too late, that the marginal fan was load-bearing looks like.
The RSN argument’s logical error was not in the math. The math was correct. The error was in the assumption that the categories were stable.
“Marginal fan” is a categorization that depends on what you compare it against. A fan who watches twenty games a year, in the era when the cable bundle is the only delivery system, is a marginal fan — relative to the fan who watches eighty. In a world where DTC streaming exists, that same twenty-game fan becomes the core of a different product. They are someone who would happily pay $15 a month for twenty games, who would not pay $90 a month for the bundle to access the same twenty games. The category of fan did not change. The product changed underneath the category.
When the product changes, the category re-prices.
The RSN executives who used the “marginal fan” frame were not wrong about the marginal fan. They were wrong about what would happen when a different product taught the marginal fan that they were actually a core customer of that product. Every cord-cutter who picked up an NBA League Pass, an MLB.tv subscription, an ESPN+ trial, a YouTube TV vMVPD package, was a fan who had been mis-categorized as marginal in the bundle frame. They were not low-value; they were waiting for a product priced for the way they actually wanted to consume.
The cable bundle did not lose its marginal fans. It lost the category — and discovered, in the loss, that the category had been propping up the entire pricing structure.
Now read the AI Overviews defense again with that pattern in mind.
“The clicks AI Overviews replaced were bounce clicks. The user did not want them. The publisher did not monetize them. The clicks that survive are the high-value ones — the readers who actually wanted the article.”
The mathematical part is, as with the RSN defense, correct. There genuinely is a category of click that bounces. There genuinely is a category of reader that converts. The two categories are real, and they are different.
The structural error is in the assumption that the categories are stable.
What a “bounce click” is depends on what the alternative is. A click that bounces in a world where Google is the entry point to the entire web is a click that the publisher failed to retain. The publisher had the user on the page; the page did not deliver; the user left. That is a publisher problem. A click that bounces in a world where the user can get the answer without ever leaving Google is a click that the publisher never had access to. The user was never going to be on the page in the first place. That is a platform problem. The publisher’s job has changed without anyone telling the publisher.
The “high-value clicks remain” frame is structurally a “core fan stays in the bundle” frame. It is true at the moment it is spoken. It is true the year after. The thing it does not survive is the way the surviving cohort gets retrained.
A reader who used to find Wired through Google now finds the answer to their immediate question in the AI Overview, and never opens Wired again unless they happen to know it exists by name. The reader who does know Wired exists by name and goes there directly is the reader the bounce-click defense says is the real audience. That reader is also a much smaller cohort than the one Wired used to have, and that cohort is not generative — it does not get replenished by the next college student who Googled their way into the Wired universe in 2014, because that pathway is now answered by Google before it reaches the publisher.
The publishers that hold up against this — and there are some — share a pattern with the sports franchises that adapted earliest to the post-RSN world. They built a direct relationship with the audience that did not depend on the platform’s referral.
For the franchises, that has meant team-owned streaming products, free-streaming partnerships outside the cable bundle, and first-party data collection that the RSN never permitted. I wrote a piece a few weeks ago on the parallel mechanism in minor-league baseball, where free local streaming has quietly become the load-bearing layer. The structural shift is from “rented audience through a third-party pipe” to “owned audience through a first-party surface.”
For the publishers, the same shift looks like newsletters that subscribers actually open, paid memberships with retention metrics that make sense, podcasts whose listeners came from the host’s own audience, communities that exist in the publisher’s environment rather than someone else’s feed. None of that is novel. What is new is the realization that the publishers that did not start building the direct relationship five years ago are running out of time to start now.
The bounce-click defense is exactly the kind of argument that delays the start. If a publisher believes the clicks they are losing are the bad ones, the publisher does not feel the urgency to rebuild the audience model. The publisher waits for the AIO CTR recovery curve. The publisher trusts Google’s directional language about quality of clicks improving. The publisher reads the 2.4% AIO CTR figure as a turn rather than a ceiling. And the publisher loses another two years.
I do not think most publishers have those years.
I am not saying AI Overviews are a bad product for users. They are an excellent product for users. The compression of “look up a fact” into a single result is a real welfare gain, and the welfare gain is going to keep showing up in usage.
I am not saying Google is lying. The bounce-click argument is internally consistent. The CTR recovery is real. The category of valuable click that survives the Overview is a real category. None of that is in question.
I am saying that “true and operationally fatal” is a category that exists, and the AI Overviews defense — like the RSN defense before it — sits squarely inside that category. The defense is correct in its premises and wrong in its conclusion, and the wrongness shows up not in the next quarter but in the cohort math three to five years out.
The most useful thing I can offer, having watched the same shape of argument play out once already, is the observation that the inside of the cliff and the bottom of the cliff look the same in the accounting until they do not.
I would not be optimizing for AI Overviews citations. I would not be trying to game the new SERP. I would not be spending the next twelve months in the SEO conference circuit waiting for a CTR recovery that is not coming back to its old shape.
I would be doing what the franchises that survived the RSN collapse did. I would be building the direct surface. I would be moving the audience off Google’s rails and onto a relationship the platform cannot disintermediate. I would be giving up some short-term traffic to invest in a model the next product cycle cannot collapse.
That is the part of the answer that none of the defenses, mine or Google’s, makes easier. Building a direct audience is harder than running an SEO factory. Running an SEO factory was supposed to be the business. The business changed.
The bounce-click defense is the sound of an industry that has not yet noticed that the business changed.
That sound is familiar. I have heard it before.
]]>In November, Apple announced that MLS Season Pass would no longer be sold as a separate product. Every match in the league, on every Saturday and Wednesday and Sunday, is now part of the regular Apple TV subscription — same price, more inventory, no upcharge, no separate authentication flow. The standalone MLS subscription was the front of a 10-year, $2.5 billion deal Apple signed in 2022. Three seasons in, Apple folded the wrapper.
The same Apple TV bundle now also carries Friday Night Baseball — two MLB games each Friday, weekly doubleheaders through the regular season, exclusive to the platform. F1 races. Eventually, more.
Read those two moves together and the strategy gets obvious.
Apple is not buying sports rights to make money on sports. Apple is buying sports rights to lower churn and raise sign-ups for the rest of the Services bucket. That is the whole game. Every other major sports-rights buyer in 2026 is running a P&L on the rights themselves — selling ads against the inventory, building subscription tiers around the live games, splitting cable affiliate fees with the partner pipes. Disney, NBC, Amazon, ESPN, Paramount: each one of them has to make the rights deal pay for itself in the same fiscal year, on the same line item, against the same partners.
Apple is not running that model.
Apple TV+ exists, in its parent company’s framing, to reduce churn and justify Apple One bundle pricing. Sports rights, inside that framing, are not a P&L line. They are a customer-acquisition cost. The metric that matters is not whether the MLB or MLS rights pay back per game; it is whether the new subscriber stays in the ecosystem — keeps using iCloud, keeps paying for Apple Music, eventually upgrades the iPhone, eventually buys the Watch.
The deal structure tells the strategy. Two games a week, on a Friday night, with no stand-alone subscription option, in a bundle that also includes a streaming service most subscribers signed up for to watch Severance — that is not how you build a P&L. That is how you build a habit.
It is also why the MLS folding move is so telling. A traditional rights buyer would have run the math on MLS Season Pass standalone — how many subscribers, what ARPU, what marginal contribution to rights amortization — and either kept the wrapper or renegotiated the deal. Apple ran a different calculation: how many people will start paying for Apple TV who weren’t, how many will stop canceling, how many will move from monthly to annual. The standalone MLS revenue was leaving a larger Services number on the table.
This is what it looks like when sports rights are priced against a different P&L than the partners use.
The other buyers cannot run this model. The other buyers do not have an iPhone to amortize against. That, more than anything else, is why Apple’s sports strategy looks idiosyncratic from the outside. From the inside, it is the only model that makes sense for the company actually buying the rights.
]]>