Google Answered the Question. Nobody Clicked. Now the People Who Wrote the Answer Are in Court.

The open web ran on a trade — you write it, search sends the readers. AI summaries keep the answer and quietly end the trade. What that costs publishers, and what it costs you.

I have spent two years writing about whether machines tell the truth.

Accuracy, hallucination, provenance — is the answer on your screen correct, and does anyone have to say where it came from? I got reasonably good at that question, and so absorbed in it that I never asked the plainer one underneath: who keeps writing the pages the answer is made out of?

That is a professional blind spot, the kind you acquire by specializing. Mine has a very specific shape. Last month I was checking how fast a particular battery chemistry degrades. Google put a summary at the top of the results. I read it, recognized it as broadly right, sanity-checked one figure against something I already knew, and closed the tab. Sites visited: zero.

I had verified the answer, congratulated myself on being careful, and taken — free, without noticing — work somebody paid a writer to do, in a week when I told you to always go read the source.

Here is what I had missed. The open web was never mostly a public good. It was a trade.

The trade nobody ever signed

It went like this. You publish something — a recipe, a repair guide, a court report, a review of a vacuum cleaner. A crawler takes a copy. The search engine shows a fragment and, beside it, a link. The link sends you readers, and readers become subscribers, or ad impressions, or the reason an editor keeps one reporter on the county courthouse beat.

Nobody signed this. No contract, no rate card, no term sheet — just an equilibrium held in place by mutual convenience, which lasted twenty-five years because each side needed the other slightly more than it resented them. An AI summary changes exactly one variable in that arrangement. It keeps the fragment and removes the reason to click.

We know roughly how much that matters. In March 2025, Pew Research Center found users clicked a traditional search result in 8% of visits where an AI summary appeared, against 15% where none did. The sharper number sits inside the box: a link within the summary itself was clicked in just 1% of visits. Pew watched 900 U.S. adults across 68,879 searches — and that fieldwork is now about seventeen months old, a photograph of March 2025 rather than a description of today. The Register, on the same dataset, caught the detail that lodged in my head: with a summary present, users ended the browsing session entirely 26% of the time, against 16% without. Not "went elsewhere." Stopped.

But Pew watched behavior; it did not run an experiment. So let me rank the evidence, strongest first.

The best study anyone has done is a randomized field experiment by Saharsh Agarwal and Ananya Sen, of the Indian School of Business and Carnegie Mellon. A browser extension randomly assigned 1,065 U.S. desktop Chrome users to see the summaries or not. Conditional on appearing, they report, AI Overviews reduce outbound organic clicks by 39.8% and raise zero-click searches by 34.5%. Hold onto "conditional on appearing" — that is the effect on queries where a summary shows, not across all searching. And note what they did not find: any measurable gain in perceived search quality. The results, they write, suggest traffic is diverted without improving the experience. Separate causal work found AI Overview exposure cut traffic to English Wikipedia articles by approximately 15%, with effects the authors call heterogeneous.

Below that sits correlation — not nothing, not proof. Ahrefs, across 300,000 keywords, reported in February 2026 that an AI Overview now correlates with a 58% lower average clickthrough rate for the top-ranking page, up from the 34.5% correlation it measured in April 2025. And at the whole-funnel level, the Reuters Institute reports Google organic traffic to over 2,500 sites down by a third globally between November 2024 and November 2025, and by 38% in the United States.

Different instruments, all pointing the same way.

What is actually alleged, and what a court has actually decided

Penske Media — owner of Rolling Stone, Billboard and Variety — sued Google on Friday, September 12, 2025, the first time a major U.S. publisher had taken the company to court over AI summaries. The 101-page antitrust complaint went to the U.S. District Court for the District of Columbia — the same courthouse where Judge Amit Mehta handled the government's search monopoly case.

Notice that it is an antitrust complaint, not a copyright one. That is the whole argument. Penske claims that because Google holds monopoly power in search, publishers are "forced to acquiesce to this misappropriation of their content" — take the deal or vanish. It alleges that about 20% of Google searches linking to its sites now show an AI Overview, a share it expects to rise, and that its affiliate revenue had fallen by more than a third from its peak by the end of 2024. Those are Penske's numbers about Penske, in a live complaint. Chegg filed a parallel suit in February 2025, with the same lawyers.

Then the case did what cases do. In January 2026 Google moved to dismiss with prejudice, calling the amended complaint "legally defective in every way". Penske answered with a 56-page opposition filed on February 12, 2026. In May 2026 the judge denied the parties' request for oral argument, saying it would set argument "if needed" when ready to address the motions. As of this writing, no ruling has been reported.

Eleven months in, nobody knows whether any of this is unlawful. Meanwhile the News/Media Alliance argues, in its own advocacy, that Google is forcing content creators to give away their content in order to remain in Google Search — a lobby's framing that sits uncomfortably close to a regulator's.

While the docket sat still, the regulators moved

On December 9, 2025 the European Commission opened a formal antitrust investigation into Google's use of publisher content for AI. Its framing is the cleanest statement of the trade anywhere in this story: the Commission will examine how far AI Overviews and AI Mode are generated from publishers' content "without appropriate compensation for that, and without the possibility for publishers to refuse without losing access to Google Search." Read its hedges with equal care — "if proven," the practices "may breach EU competition rules," and "the opening of a formal investigation does not prejudge its outcome." Opening a case is not finding a violation. In February 2026 the European Publishers Council filed a complaint of its own under Article 102, its chairman Christian Van Thillo insisting it is about a gatekeeper taking content "without consent, without fair compensation."

Britain actually ordered something. In June 2026, the UK's competition regulator required Google to offer publishers a way out of AI products without sacrificing their visibility in ordinary search results. That last clause is the point: previously, a publisher that opted out of AI Overviews was removed from Google's search index entirely. Google shipped an opt-out the same day, June 3, 2026.

Publishers are unimpressed. Paul Bannister of Raptive called a front-end toggle a light switch while keeping the power plant running, and the complaint underneath is a good one: nobody hands back the click data that would let a publisher work out whether flipping the switch helps or hurts. A month earlier Google had announced five updates designed to send more traffic back to websites. And on August 11, 2026, nearly 300 French newspapers complained to France's competition authority, citing the regulator Arcom's estimate of "between 33% and 38% traffic loss attributable to AI-generated summaries."

Now let me make the other side's case, properly

Google's position is not that the studies are inconvenient; it is that they are wrong. Responding to Pew, the company said the work used a "flawed methodology and skewed queryset that is not representative of Search traffic", and that it has "not observed significant drops in aggregate web traffic." In Google's own published account, Liz Reid, the head of Search, wrote that total organic click volume to websites has been "relatively stable year-over-year," with "slightly more" quality clicks than a year earlier. Note the hedges Google itself chose — relatively, slightly.

Two honest problems with that defense. It is dated: Reid's post went up on August 6, 2025, and nothing comparable has replaced it. And, as Engadget observed at the time, the post shared no numbers whatsoever. Google is the only party who can see the aggregate data, and it has chosen to characterize rather than publish it.

Here, though, is the concession the publishers' side rarely makes. AI Overviews are the latest in a long line of search features — Featured Snippets, Local Packs, Top Stories — that resolve queries on the results page, and Ahrefs' own caveat is that clickthrough rates have been falling even for keywords that never trigger an AI Overview. A meaningful share of what publishers blame on AI was already happening. Asked directly, people also say they still click: 42% of those using AI chatbots for news told the Reuters Institute they always or often follow through to the source. Self-reported, and about chatbots rather than AI Overviews — but it cuts against the tidy "nobody clicks" framing, including mine.

Australia already built the machine American publishers are suing to get

Now to the country that ran this experiment first and got a result nobody predicted.

In 2021 Australia wrote the bargain into statute — Act No. 21 of 2021, assented on March 2, 2021 — establishing a mandatory code of conduct for platforms and news businesses bargaining over news content.

Then the twist. The competition regulator's own page states it flatly: no digital platforms have yet been designated under the code. The power to designate belongs to the Treasurer — a politician, not the regulator — and in five years no Treasurer has used it. Five years on the statute book, and the code has never actually been fired. It worked entirely as a threat.

Did the threat work? Treasury's 2022 review concluded, in its own hedged words, that it is "reasonable to conclude that the Code has been a success to date", citing over 30 agreements it called "highly unlikely to have been made without the Code." Rod Sims, who chaired the ACCC when the code was designed, put the deals at "more than A$200 million per year."

Then the threat stopped working, in the most boring way imaginable. In March 2024 Meta simply refused to renew, and two federal ministers called that a "dereliction of its commitment" to the sustainability of Australian news media. Canberra's answer, on December 12, 2024, was a News Bargaining Incentive — a charge platforms can offset by signing deals — with the government stating outright that "Australia does not intend to raise revenue from this policy."

The bills arrived on Thursday, August 13, 2026, three days ago. They would impose a charge of 2.5% of a platform's Australian digital advertising revenue where it fails to reach agreements with at least eight news businesses. Note the tense: not law yet. Michelle Grattan reports the legislation is expected to pass in this parliamentary fortnight. The U.S. tech industry's trade association calls it a "thinly veiled discriminatory tax" on American services — the objection Washington will raise against any bill like it.

And here is the part that stopped me. Treasury's explanatory materials for the draft say services provided solely or primarily by large language models are excluded from the definition of internet search engine service. Read it precisely: as drafted, that writes LLM-only services — a standalone answer chatbot — out of scope, while Google Search with AI Overviews attached is still a search service and stays in. But Grattan puts the practical upshot bluntly: AI platforms will not be covered by the new deal. Caroline Fisher and Sora Park of the University of Canberra write that the government chose not to include AI companies, and that "excluding AI services creates a significant regulatory gap."

So the contrast is not the one you expect. Australia built the machinery American publishers are litigating for, learned it only works as a threat, rebuilt it after a platform called the bluff — and still drew the boundary just short of the thing that broke the trade.

Just imagine the web in 2031

Push the current trend out five years and the interesting failure isn't dramatic. It's quiet.

Start with the split. A few dozen large publishers sign licensing deals; their work is quoted, attributed, paid for. Everyone else is still crawled, still summarized, and paid nothing — not out of malice, but because negotiating with two million websites is impossible.

Then the incentives invert. If almost nobody arrives at your page, you stop writing for a person and start writing for the extractor. Pages get shorter, flatter, more machine-legible, stripped of the photographs and digressions that only ever existed because a human was going to see them. The recipe loses the story about the grandmother — and eventually stops being tested, because testing costs money and nobody is watching.

Now the part that should worry you more than the money. Models need new facts, and new facts come from someone physically present: at the zoning hearing, in the courtroom, on the phone with a third source. In 2031, the county that lost its last reporter in 2029 still gets confident answers about its school board — synthesized from a 2027 archive, fluent as ever, and wrong in ways nobody local can check. The loop closes there: an answer assembled from a page written to become an answer. Nobody lied anywhere in that chain. Nobody verified anything either.

The optimistic version is genuinely possible. Just watch where the money lands.

The people who agree about the problem and about nothing else

The critics of the obvious fix are not shills. They are ideologically scattered, and they keep arriving at the same objection.

From the free-market right, Paul Matzko of the Cato Institute argues that a link tax modeled on Australia's could have "serious negative consequences" for both newspapers and consumers, noting that of roughly $200 million paid over within months of the 2021 regime, "90 percent" went "to the three largest conglomerates." That is Cato's characterization, and Treasury and Rod Sims read the same period more favorably — but the distributional worry is not crazy. At R Street, Josh Withrow makes the narrower version: the smallest and newest outlets are entirely excluded from the benefits of an American bargaining bill.

From the digital-rights left, the Electronic Frontier Foundation gets there by another road: competition doesn't flourish when a group is allowed to form a cartel, even one of small newsrooms, and paying for links implies "a sort of property right in links" that would be poisonous for anyone who is not a large publisher.

Against them stand the News/Media Alliance and the European Publishers Council. And here is the honest bit: none of them is arguing about the same thing. The critics are attacking a bargaining code. The publishers are describing an extraction problem. Australia's drafters just demonstrated the gap by writing a code that does not reach a pure answer engine. Nobody here has a good answer for the genuinely new fact: a summary is not a link at all.

So what do you actually do about it?

Policy will take years. Your behavior changes in a week, and it is the input the whole system runs on.

• Treat the summary as a claim, not a conclusion. It is a synthesis of pages you have not seen, delivered in the confident register of a reference work. For anything medical, legal, financial or contested, open one underlying source before you act.

• Click through once a day, on purpose. One deliberate visit to a site that did the work. Traffic is the only vote most publishers can count, and it costs you eleven seconds.

• Pay one publication directly. Not five. One you would genuinely notice closing. Direct money is the only revenue a summary cannot intercept.

• Get specific about local. Ask whether anyone still physically attends your council, court and school board meetings. If the answer is one part-time reporter, that is your information supply chain — one resignation from zero.

• If you run a site, ask for the numbers. Opt-outs are secondary; data is the demand. How often does your material appear in a summary, and what does that do to your clicks?

• Distrust any traffic claim without a denominator. "Billions of clicks" and "traffic collapsed" are both true-sounding sentences that survive because neither side publishes the arithmetic.

What I should have been asking all along

My blind spot had a comfortable shape: I interrogated the output and took the input for granted. Two years asking whether the machine's answer was accurate, never once asking who funds the material that makes accuracy possible. Accuracy is downstream of somebody bothering. Somebody bothering is downstream of getting paid.

And let me be fair about where the record stands. No court has ruled any of this unlawful. Europe has opened an investigation, not delivered a verdict. Britain ordered a door that publishers say they cannot afford to walk through. Australia legislated harder than anyone and still stopped short of the AI. Google says aggregate traffic is fine and has not shown its work.

The lesson, as I see it, is that the trade did not need anyone's bad intent to break. It only needed to become unnecessary to one side. Which means the remedy was never really a lawsuit — a lawsuit is what you file once the remedy didn't happen. It is duller: publish the traffic data, price the input, make the payment reach the outlets doing the reporting rather than the ones with lobbyists, and write rules that follow the answer rather than whichever technology happens to be delivering it.

And in the meantime, on the days when the summary at the top of the page is right and complete and saves you four minutes: click anyway. Somebody has to.

Working out how automated systems quietly rewrite the deals we never noticed we were part of is exactly why the HAIA Foundation exists. If that is your kind of thing, subscribe — and then, in the spirit of the argument, go read something all the way to the end.