For a publisher, the appeal of the old search deal was simple: let a search engine find the work, and it might send readers back. No audience was promised, but the chance to reach readers gave publishers a reason to take part.
AI search changes that arrangement when it uses the work to answer the reader's question on another company's page. A source becomes part of the answer, and the reader may have no further business with the site that produced it.
Google's new publisher controls improve the choice about taking part. But the ability to refuse is only part of a workable exchange. Publishers also need a return they can judge when sending readers onward is no longer central to search.
The old exchange was useful and imperfect
Early search was never a gift of free traffic to grateful writers. Search engines needed useful material. Publishers wanted to be found. Both could earn money from the exchange even when their goals differed.
The 1998 paper on Google's early search engine explained how it used the web's links to help find pages that matched a search. The material lived elsewhere. A search could send the user to a newspaper, a university or someone writing about a narrow personal interest.
For a site owner, the important part was what happened after arrival. A visitor might read several pages, buy something or return a week later. The publisher had a chance to introduce its work in its own setting, with its own advertising and reasons to stay.
That chance came with dependence. Letting a search engine find a page gave its owner no right to a high rank. The operator could change how it selected and presented results. A publisher could do valuable work and still struggle to be found. Calling this a bargain describes what each side hoped to gain, not a contract every website had signed.
Nor did the reader always leave the search page. Scores, quick facts and other direct answers were familiar well before generative AI. Few people would want a search engine to make them open a separate page for every simple fact it could accurately display.
The useful part of the arrangement was its ability to bring people to sources they would otherwise never find. That gave many publishers a reason to keep their sites open to search, even though they could not know how many readers would arrive.
An answer changes what is being exchanged
A generated answer can carry a much larger share of the work than a short description of a link. It can compare options, combine explanations and respond to a follow-up question. The reader may leave satisfied after using information from several sites without visiting any of them.
A useful answer can spare the reader time and effort. Someone new to a subject may learn more when the terms are explained together. Readers should not have to click more just to keep the old search model intact.
A publisher that once gave people a place to read can now provide the work they read on another company's site. The search service gets something closer to a full answer to offer its user. The source may still be discovered, but a reader who has the information has less reason to visit it.
A July 2026 preprint by Qiaoni Shi, Kai Zhu and Kai Gu examines that change using U.S. desktop browsing data. It finds that ChatGPT chats lead to few outward clicks, which go to a different mix of sites from those reached through traditional search. It does not measure what Google's AI features caused or what every AI product does. The findings give publishers reason to question whether AI referrals will serve them as traditional search did.
A website owner therefore has more to decide than whether a crawler may enter. A page can be indexed so readers can find it, used as a source for an answer or used to train a model. Each use does a different job. Consent to one should not be treated as consent to all of them.
A publisher might want its pages to be found, agree to some uses in answers and oppose training. Another might be willing to sell access for a use that produces few visitors. Those uses need clear terms so publishers know what they are agreeing to.
What the new opt-out gives publishers
By August 31, 2026, Google had rolled out its Search generative AI control worldwide. A publisher can exclude its links and content from AI Overviews, AI Mode and generative AI features in Discover, including use to ground those answers. Google says this setting does not affect ranking or inclusion elsewhere in Search. It also removes visits from the covered AI features. Training is separate, with Google-Extended identified as the relevant control. Other sites can still supply similar material. This is a meaningful improvement over treating ordinary search and those AI uses as inseparable, even though it cannot guarantee that opting out will improve a site's business.
A workable exchange needs a return
Google's Mrinalini Loew points to new ways for publishers to reach readers. These include more useful links in answers, ways to feature preferred sources and support for the subscriptions readers already have. The company's view is that richer answers can introduce people to more material and help them find sites worth visiting.
That deserves a fair hearing. Publishers benefit when a service makes their work easier to understand and find. A reader who discovers a specialist through an answer may become exactly the kind of regular visitor that specialist needs. Better controls and ways to display sources can help publishers, even when Google gains from them too.
A publisher weighing Google's offer has to consider where its readers now spend time. Opting out protects its material from the covered AI uses but also excludes it from those features. Accepting it can preserve a chance to be seen while helping the service answer questions that once brought people to the site. The same choice will not suit every site.
When Reuters Digital's Phil Andraos discussed the new opt-out with Digiday in June, he said the choice would need careful study and testing. He wanted enough information to judge the tradeoff. He had more to offer than a rush to sign up or a demand that every publisher leave at once.
Google has since made its separate AI visibility reports available worldwide. Its published description includes impressions and which pages appear. That helps an owner see where a site's work is used. It does not show the full effect of accepting or refusing the new terms.
A fair exchange should let both sides see what they give and what they get. For some publishers, useful referrals may remain enough. Others may seek:
- Payment for particular uses.
- Better routes into subscriptions.
- Limits that preserve the value of work sold directly to readers.
None of these options requires every piece of public information to be hidden away.
Both sides need room to disagree about what is enough. A source should be able to say that an arrangement does not work for its business. The fact that a platform has readers does not settle whether taking part is worth the cost. A search company should also be free to improve its answers without pretending every old habit of use can be kept.
There will be no single replacement bargain that suits a retailer, a reference site and a small publication. The next useful arrangement will be the one each can understand well enough to accept, revise or leave. A publisher should be able to make that decision before learning, from the following month's results, what it had agreed to.
