AI AND THE OPEN WEB
Publisher traffic is down about a third — and small sites lost the most.
The part that decides who survives
The headline number hides the finding that actually matters: the loss is not distributed evenly, and it is inverted. The smaller the publisher, the harder the hit.
Search referral decline by publisher size. Source: Chartbeat, reported July 2026. Single-source figure; treat the pattern as firmer than the precise values.
Large organisations absorb a 22% loss, sign a licensing agreement, and continue. An independent site losing 60% of its referrals does not have a legal department, a licensing negotiation, or a subscription base to fall back on. The mechanism that concentrates the web is not censorship. It is arithmetic.
What the numbers do not say
This page would be dishonest without these, and most coverage of this topic omits them.
Zero-click rates have not risen monotonically. One tracker measured the zero-click rate on AI-Overview keywords falling from above 45% to around 38% across 2025 — the opposite of the expected direction. Users may be adapting rather than surrendering.
AI referrals are growing fast. ChatGPT's outgoing referrals to publishers rose roughly 52% year over year. The absolute number is still tiny, but a small number compounding quickly is different from a dead channel.
Licensing money is real. Publishers now report meaningful revenue from AI licensing deals — a line item that did not exist three years ago. It is unevenly distributed, but it is not nothing.
Why this is not just a publisher problem
It is tempting to file this as an industry squabble. It is not, for a reason that closes a loop with the rest of this site.
AI answers are made of pages. Every summary an answer engine produces is assembled from material somebody researched, wrote, fact-checked and paid for. If the economics of producing that material collapse, the input degrades — and the answers degrade with it, for everyone, including the people who never click a link.
Add the other measured trend: roughly half of new articles are now primarily machine-generated, per the figures in the quarterly record. Human-made pages are getting less economically viable at the same moment machine-made pages are getting cheaper. That is not a conspiracy; it is two incentives pointing the same way.
If answering a question no longer requires visiting the page that knows the answer, who pays the person who found it out? Licensing deals answer this for a few dozen large organisations. Nothing currently answers it for the independent site, the specialist blog, the forum thread or the documentation page — which together are most of what makes the web worth searching.
The courts, and the distinction that actually emerged
The other half of this economy is being decided in litigation. More than 35 training-data copyright cases were active in major Western jurisdictions by mid-2026, and the early rulings did not land where either side expected.
The question turned out not to be "is training copying?" It turned out to be how the material was obtained.
In the most detailed US ruling so far, a judge held that training a model on lawfully acquired books was transformative fair use — while retaining a corpus the company knew to be pirated was not, and exposed it to statutory damages. That case settled for approximately $1.5 billion, the largest copyright settlement on record in the United States, working out to roughly $3,100 per book. By the March 2026 claim deadline, 440,490 of 482,460 eligible works had been claimed — a 91.3% rate.Bartz v. Anthropic (N.D. Cal.), merits ruling June 2025, settlement announced August 2025, claim data March 2026. Final approval was still pending after a May 2026 fairness hearing at the time of writing.
Around it, the picture is genuinely mixed rather than settled. Another judge granted summary judgment to a different lab on fair use, finding the training highly transformative but the plaintiffs unable to show market harm. A third court found against fair use where an AI product competed directly with the database it had been trained on — the first final judgment in this area, now under the first appellate review, argued in June 2026. A German court rejected the equivalent defence outright under its own law. The US Copyright Office has stated that fair use cannot simply be presumed for AI training.
Provenance became the asset. Labs that can document a clean licensing chain for their training data are surviving motions and settling on better terms; those with poorly documented or piracy-tainted pipelines are not. That is why the licensing market appeared so quickly — nine-figure deals between major publishers and the largest labs, plus a new layer of intermediaries aggregating smaller publishers to license collectively.
And there is a feedback loop worth noticing: every disclosed licensing deal makes the fair-use defence harder for the next defendant, because it demonstrates that a licensing market exists and could have been used. The industry's own deal-making is narrowing its own legal argument.
Which connects directly to everything above. The traffic collapse and the licensing settlements are the same question asked twice — once by the market and once by the courts: if a publisher's work ends up inside an answer, what is owed, and to whom? Neither forum has finished answering, and the outcomes will decide whether independent publishing has an economic basis at all.This is a summary of a fast-moving and jurisdiction-dependent area, recorded August 2026, and it is not legal advice.
AI answer engines are genuinely useful, and this site is built to be readable by them — its own llms.txt exists for exactly that reason. The problem is not summarisation; it is that the value moved without the payment following. The fix most likely to work is not blocking crawlers, which mostly harms the blocker, but attribution with weight: citations that carry traffic, licensing that reaches beyond the largest players, and readers who deliberately visit sources. That last one is free, and it is the only part you personally control.