Ask Google “what temperature should I roast a chicken at,” and Google answers you. There’s a confident paragraph at the top of the page. There are no clicks. The recipe site that taught Google the answer doesn’t know you exist. Its ad inventory served nothing. Its newsletter signup didn’t fire. The cooking writer who tested four roasting temperatures and wrote about it three years ago — they aren’t compensated, credited, or even visible above the fold.
This is the new search bargain, and we’re naming the wrong thing when we call its byproduct “AI slop.”
“AI slop” — Merriam-Webster’s, the American Dialect Society’s, and Macquarie Dictionary’s word of the year for 2025 — gets used to mean low-effort, AI-generated content polluting the internet. That definition treats the problem as a supply issue: too many machines making too much junk. The implication is that if the supply tightens — better detectors, watermarking, AI labels, content moderation — the problem subsides.
It won’t. The supply side is downstream of a demand-side break that happened first. Search stopped sending users to publishers. Publishers stopped being able to afford humans. AI filled the gap. Search now answers from a corpus increasingly written by AI to be summarized by AI. That’s not a content problem you can solve by writing better content. It’s a search problem.
What the Numbers Actually Say
Let’s get the data straight, because the figure that’s been floating around — 34.5% — is now twelve months out of date.
That number came from an Ahrefs study in April 2025 which compared CTR for top-ranking pages before and after AI Overviews appeared on a query. By February 2026, Ahrefs’ follow-up study using December 2025 data put the click reduction at 58%. Almost double, in eight months.
Pew Research, working independently, came to a similar place from a different angle. Their July 2025 study tracked the actual browsing behavior of 900 U.S. adults — 68,879 unique Google searches in March 2025, of which 12,593 surfaced an AI summary.
- Users who saw an AI summary clicked a traditional search result in 8% of visits.
- Users who didn’t see one clicked in 15% of visits.
- Clicks on links inside the AI summary itself: 1% of visits.
When an AI summary appears, the chance any single visitor clicks anything at all is around 9%. Without one, it’s 15%. Almost half the would-be clicks evaporate.
The publisher-side data tracks. Chartbeat, monitoring traffic across 2,500+ news sites globally, reported a 33% decline in Google search referrals across 2025. Digital Content Next — the trade association for major publishers — surveyed its members and found most reporting 1–25% traffic losses, with some exceeding 75%. As of early 2026, approximately 58% of Google searches end in zero clicks.
And the response from publishers has stopped being polite. Penske Media — Rolling Stone, Billboard, The Hollywood Reporter — filed a federal antitrust suit against Google in September 2025, with a 56-page opposition to dismissal in February 2026. The argument is creative: not copyright infringement, but anticompetitive coercion. Google’s search monopoly, Penske argues, forces a “forced choice” — let your content train AI Overviews that cannibalize your traffic, or be excluded from search entirely. The European Publishers Council filed a parallel complaint with similar framing. One-third of publishers surveyed in early 2026 said they plan to block AI Overviews the moment tools become available.
Google’s response was to ship “Further Exploration” — a small section of curated links at the bottom of AI Overview answers, designed to send some traffic back. It’s a thermostat on a burning house.
The Supply-Side Reading Misses the Diagnosis
The dominant framing of “AI slop” goes like this: generative AI made content creation effectively free, content farms exploit this, search engines fail to filter it, the open web fills with low-value junk. From this view, the cure is supply control — better classifiers, mandatory disclosure, platform moderation, “AI-free” certifications.
This framing isn’t wrong, but it’s incomplete in a way that matters. It treats the AI-generated content surge as an exogenous shock — something that happened to the web from outside. The supply surge isn’t autonomous. It’s a rational response to a price signal.
Until roughly 2023, the implicit deal was: write content, get ranked, get clicks, monetize via ads or conversions, fund more writing. That’s the bargain that paid for recipe blogs, product reviews, local journalism, and most of the long tail of the web. Each click had a value. Each post had an expected return.
AI Overviews break the third step. Clicks per ranking position are falling — fast, by every measure we have. But the cost of producing content didn’t fall with them, and human writers didn’t suddenly become cheaper. So what happens to a business whose revenue per article is dropping but whose cost per article is constant?
It writes fewer articles, or it writes cheaper articles. Most chose cheaper.
AI lets you produce a thousand SEO-optimized listicles for the cost of one freelance assignment. The margin on each individual page is awful, but it scales. The economics that produce “slop” aren’t the economics of malice or laziness — they’re the economics of a publisher trying to survive a 33% collapse in search referrals (per Chartbeat) without going out of business.
The supply of AI content isn’t a content problem. It’s a response to the search problem.
The Spiral
Once you see the loop, it’s hard to unsee:
- Answer engines (Google AI Overviews, Bing Copilot, Perplexity, the chat surface in every major LLM product) answer queries in-place, using publisher content as substrate.
- Publishers’ click-driven revenue collapses. Pew’s 8%-vs-15% means the median page is clicked roughly half as often when an AI summary appears. Top-ranking pages, per Ahrefs, are clicked at 42% of their pre-AIO rate.
- Publishers respond by lowering the unit cost of content. Some lay off staff. Some outsource. Many shift to AI-assisted production — a human edits a generated draft instead of researching and writing one.
- The corpus answer engines train on and summarize from increasingly consists of AI-assisted content optimized to be ranked and summarized.
- The summaries get blander, more derivative, more confidently wrong. Users notice. Trust erodes.
- Search platforms respond by amplifying the AI surface further, since their own AI answer feels more authoritative than the slop substrate beneath it.
- Loop.
Each cycle of this loop makes the next cycle cheaper for the platform, more expensive for the publisher, and worse for the user. It’s not a stable equilibrium. It’s a tightening spiral.
The Cloudflare data point that drives this home: per Cloudflare’s own analysis, Anthropic’s ClaudeBot crawls 20,583 pages for every single referral it returns to a publisher. That is the ratio of extraction to acknowledgement in the current AI/web relationship. It’s not a small asymmetry. It is, functionally, a one-way valve.
The Reuters Institute’s Journalism, Media, and Technology Trends and Predictions 2026 puts it tactfully: publishers plan to focus on “investigative journalism, analysis, and distinctive reporting” while “reducing investment in more routine content.” Translated: humans will write the things only humans can defensibly write, and everything else becomes machine work that nobody pays for, that nobody clicks, and that everyone produces anyway because the alternative is producing nothing and dying faster.
That’s the slop, and it isn’t a moral failing of content producers. It’s the equilibrium of a market whose price signal got rewired.
Why Standards Can’t Fix This
I wrote about /llms.txt last September and concluded it was an elegant solution nobody is using. I want to be specific about why it isn’t used, because the same reasoning will apply to every voluntary standard proposed to “fix” AI search.
/llms.txt assumes a cooperative relationship between publisher and platform. The publisher creates a curated, AI-friendly version of their content. The platform reads it and respects the curation. Both win.
That assumption is dead. Platforms have no reason to honor a publisher’s curation, because the platform’s incentive is not to send users to publishers — the platform’s incentive is to keep users inside the platform. Reading /llms.txt to surface a publisher’s preferred summary, then linking out to that publisher, would directly reduce the platform’s most important metric (time-on-Google, in Google’s case). Why would they?
The same argument disqualifies most variants. AI-bot-only robots.txt directives? Honored selectively — most major bots respect them, some don’t, and the ones that don’t are unaccountable. Schema.org annotations specific to AI consumption? Same incentive problem. A new HTTP header that signals “compensate me to summarize me”? The platform would need a reason to read it. There isn’t one.
The Well-Known URIs standard — /.well-known/security.txt, /.well-known/openid-configuration, the IETF’s elegant little namespace — works because the parties on both sides want the discovery to succeed. A security researcher and a website owner both benefit from a working security.txt. An OAuth client and an identity provider both benefit from openid-configuration. Coordination problems get solved when incentives align.
Publisher and AI platform incentives don’t align. There is no standard you can write that fixes a structural conflict between two parties who would prefer the other to disappear.
So the AI Is Poisoning the Well It Drinks From
This is where the second-order effect gets interesting. AI Overviews are trained on, and summarize from, a corpus increasingly produced to be summarized. The model is reading text that was written by a model to optimize for being read by a model. There’s a name for that — model collapse, technically — but you don’t need a paper to see it. Search results have gotten genuinely worse over the last two years, and the worsening isn’t subtle.
Cloudflare’s Q1 2026 robots.txt analysis found that 89.4% of AI crawler traffic serves training or mixed purposes, not search. That asymmetry matters. The web isn’t being read to be indexed and referred to. It’s being read to be ingested, distilled, and returned without attribution. The Reuters Institute’s tracking of AI-generated content in fact-checked claims rose from 7% of cases in 2024 to 16% in 2025 — and that’s just the cases where someone bothered to file a fact-check. The actual prevalence is higher, because most slop doesn’t trigger a check; it just sits in the substrate, doing search-engine-optimization work, training the next generation of summaries.
The slop isn’t sitting in a separate quarantine the AI can ignore. It is the AI’s input.
The Exits
If standards can’t fix it and the spiral is self-reinforcing, what’s actually left?
A few things, and none of them are search.
Direct subscription. Newsletters. RSS. Bookmarks. The relationships where the reader decides what they read, and the publisher knows their reader exists. These don’t scale the way search did — and that’s precisely what makes them defensible. Algorithms can’t disintermediate a relationship the user formed directly.
Paid relationships. Substack, Ghost, Patreon, individual paid newsletters, the whole micropayment-adjacent ecosystem. Click-driven advertising was always a fragile foundation; AI Overviews just clarified how fragile. The publishers most insulated from this collapse are the ones whose revenue comes from a reader’s deliberate decision, not from incidental ad impressions during a search journey.
Infrastructure pushback. Cloudflare flipped its default in mid-2025 — new customers get AI crawlers blocked unless they opt in, and “pay-per-crawl” exists for those who want a compensation channel. Cloudflare hosts roughly 20% of the web. When the substrate provider changes the default, the negotiation changes. This is the first piece of structural leverage publishers have had since AI Overviews shipped.
Trust signals where discovery used to be. This is where standards still matter — not as a way to be discovered, but as a way to be verified once a user has found you. /.well-known/security.txt, /.well-known/openid-configuration, signed RSS feeds, verified author identity, ATProto handles, Keyoxide profiles. The post-search internet still needs trust infrastructure. It just doesn’t need it to function as a search funnel.
What This Means If You’re Building Something on the Web
Three implications I’d take seriously if I were starting a project right now:
Design for zero referral. Assume the search referral to your project is going to keep dropping. Build something that survives at 10% of today’s discovery traffic. If your business model only works at 100%, your business model isn’t a business model — it’s an artifact of a market structure that’s being dismantled in real time.
Treat the front page as the relationship. Newsletter signups, RSS subscribe buttons, follow-on-Mastodon/Bluesky links — these aren’t 2010 decorations. They’re how someone who finds you once continues to find you. The home page of your site should optimize for the conversion from anonymous visit to known reader with the same seriousness that 2015 sites optimized for visit to page view.
Build for the AI surfaces, but don’t depend on them. Yes, your content will be ingested. Yes, summaries will appear without your link. You can’t opt out and expect to remain visible, and you can’t opt in and expect to be compensated. Optimize for being recognizable inside a summary — distinctive voice, real expertise, claims that are hard to compress — so that the readers who care about provenance will look for the source. That’s a tiny fraction of readers. It’s the fraction you can actually keep.
A Note on Optimism
There’s a version of this post that ends on a confident note about the open web reclaiming itself, RSS triumphant, search dethroned by trust-based discovery. I don’t believe that version. The platforms that built the click-based bargain are larger, better-capitalized, and more entrenched than ever, and the AI surfaces are still in early innings. Things will get worse before they get different.
But “different” is the operative word. The web didn’t go away when Google Reader died, when Twitter/X tilted, when Facebook hid links. It rerouted. Each rerouting cost something — visibility, breadth, frictionless discovery — and produced something else: smaller, slower, more deliberately chosen networks of readers and writers.
The current rerouting will produce another one of those. It will be smaller than the search-driven web ever was. It will be unrecognizable as a “market” by 2015 standards. And if you’re a person who writes things you’d rather not see compressed, paraphrased, and served at the top of someone else’s search results — it might also be the only web worth being on.
Coda
The “AI slop” framing is a comfortable diagnosis. It locates the problem in bad actors making bad content, which means there’s someone to blame and something to filter. The real diagnosis is less comfortable: the search infrastructure the web monetized itself through changed its function from router to answerer, and everything downstream of that change — including the slop — is a market responding rationally.
You can’t fix that with standards. You can’t fix it with quality controls on content production. You can only route around it.
So route.
Sources and further reading:
- Pew Research Center: Google users are less likely to click on links when an AI summary appears (July 2025, n=900, 68,879 searches)
- Ahrefs: AI Overviews reduce clicks by 58% (February 2026 update)
- Search Engine Journal: Antitrust filing says Google cannibalizes publisher traffic (Penske Media v. Google)
- Reuters Institute: Journalism, Media, and Technology Trends and Predictions 2026
- Transparency Coalition: Cloudflare blocks AI crawlers by default
- The Next Web: Google updates AI Overviews with “Further Exploration” as 58% click decline triggers antitrust suits
Companion essays in this trilogy:
Was this helpful?
Have questions about this article?
Ask can help explain concepts, provide context, or point you to related content.