The question of what to do about an internet filling with machine-generated filler finally got a price tag in July, and the number that matters is not the volume share. The Trustworthy Accountability Group, the Association of National Advertisers and Fiducia put AI slop at between 1.3% and 2.4% of open web programmatic spend, roughly level with made-for-advertising inventory. Modest, containable, the sort of figure that lets an industry declare the problem tractable.
Then read the quality signals. Slop inventory recorded an invalid traffic rate of 0.05% against 0.32% for clean supply. Viewability of 77.2% against 74.9%. Once measurability was accounted for, it graded as premium more than 70% of the time. It cleared at a true CPM of $7.08 against $6.15 for everything else.
Slop is not being tolerated at a discount. It is being bought at a premium, by systems built specifically to prevent that.
The verification stack was built for a different crime
Ad verification grew up fighting fraud, and fraud is a technical lie. Domain spoofing, pixel stuffing, ad stacking, hijacked residential proxies. Each one leaves a signature because each one requires misrepresenting something a machine can check.
Slop misrepresents nothing. The domain is real. The page renders. A human being genuinely arrives, genuinely scrolls, and genuinely leaves eleven seconds later having gained nothing. Every metric in the stack is a proxy for attention mechanics and not one of them is a proxy for worth. A page engineered to satisfy those proxies at minimum cost will beat a page written to be read, because the second one carries expenses the first one does not and earns no measured credit for them.
This is the part buyers should sit with. The measurement system did not fail to catch slop. It graded slop favourably and paid up accordingly. Any remedy that runs through the same metrics inherits the same blindness.
Zero marginal cost against a per-impression auction
The production cost of a passable article has collapsed from several hundred dollars to something close to the electricity. Distribution was already free. The auction on the other side still pays per impression.
When marginal cost approaches zero, any clearing price above zero justifies unlimited supply. There is no natural volume ceiling and no point at which a rational operator stops publishing. A publisher carrying salaries is bidding against an opponent carrying none, for inventory the buyer cannot distinguish.
The ANA’s Q1 benchmark showed made-for-advertising spend rising from 0.6% to 1.1%, the first meaningful increase since the industry’s 2023 crackdown began. That crackdown worked because it faced a finite adversary. A few thousand operators running a known playbook can be listed and excluded. The current adversary regenerates faster than it can be catalogued.
Detection is the reflex answer and it does not survive contact
Ahrefs found that around 74% of newly created pages carry some AI-generated content, and only a small fraction are pure machine output with no human hand. The rest are blends. The line runs straight through legitimate newsrooms, agency work, and every corporate blog with a subscription.
That distribution breaks classification as a market mechanism. Detectors that perform respectably on raw unedited output degrade sharply on edited hybrids, and hybrids are where the mass of the web now sits. Set the threshold tight enough to catch content farms and it condemns edited human work. Set it loose enough to spare that work and the farms sail through. There is no setting that separates cheap from careful, because the tell being measured is process, and process is not the thing anyone actually objects to.
The evidence base has its own problem. The most quoted figure in this debate, the claim that nine-tenths of online content would be synthetic by now, traces back to a scenario sentence in a 2022 European law enforcement report and has been recycled through secondary sources ever since, acquiring authority through repetition rather than measurement. Estimates of AI’s share of the web currently range from under 40% to over 90% depending on who is selling what. A meaningful share of the published analysis of slop is itself slop. Pricing an asset class off classifier scores means pricing it off that.
What is actually being priced is access
The mechanism attracting real money is not quality scoring. It is permission.
From 15 September, Cloudflare’s defaults block mixed-use crawlers, those serving search and model training and agent retrieval through one pipe, from any page carrying advertising. The setting applies to new customers, new sites, and the entire free tier, which is to say millions of domains flipping at once without anyone opting in. Pay Per Crawl becomes Pay Per Use, paying publishers when content shows up inside an answer rather than merely when a bot fetches it. Stack Overflow has already run the model. Cloudflare’s own data suggests more than half of AI crawl traffic is spent re-fetching pages that have not changed, which is pure deadweight cost currently borne by whoever hosts the page.
None of this removes a single slop page from the web. What it does is put a floor under the cost of the input. Charge for the corpus and generation stops being free at the source, while holders of verified archives acquire a revenue line that does not depend on anyone clicking through. For a publishing business watching search referrals evaporate into answer boxes, that is the more durable asset.
The buy side ends up with inclusion lists, and that is the real cost
Exclusion cannot beat regeneration. Nobody blacklists faster than domains spawn, and the slop network exposed earlier this year ran to more than two hundred coordinated properties on its own. The only stable equilibrium on the demand side is inversion: curated supply paths, direct deals, and a few thousand approved domains that everything else is presumed outside of by default.
That works. It also ends the open web as an advertising market. The long tail loses monetisation entirely, and the long tail is where independent publishing has always lived. The cure for slop and the cure for small publishers are the same procedure. Very few of the parties advocating the first have said out loud that they are performing the second.
Three positions follow from this. Verification vendors gain a new product category in slop avoidance while carrying an awkward record, since their existing metrics graded the stuff premium. Infrastructure sitting between crawler and origin acquires a toll position that neither publishers nor model developers can route around cheaply. And owners of large, dated, human-verified archives hold something that appreciates as the surrounding corpus degrades, which is a strange reversal for assets that were being written down as recently as last year.
Where it moves next
The payoff is not destroyed by any of this. It relocates. Gate display advertising behind inclusion lists and production shifts to whatever surface still converts volume into money: citation inside answer engines, affiliate feeds, app store listings, review corpora, short video. Some of that migration is already visible in the current numbers.
Which is why the question as usually posed has no answer. Slop is not a failure of taste or a shortage of filters. It is arbitrage between a production cost that has gone to nothing and a payout mechanism that still prices volume. It ends surface by surface, as each one stops paying by the unit, and never all at once.
The supply follows the payout. It always has.
