The Internet Is Starting to Say No to AI

For much of the internet's history, the relationship between a website and its visitors has been relatively straightforward. A person opens a browser, navigates to a service, and decides what to do next. The website controls the environment. The visitor controls the decisions.

Artificial intelligence is beginning to complicate that arrangement.

A new generation of AI assistants promises to do more than retrieve information or answer questions. These systems can navigate websites, compare products, fill out forms, and attempt transactions on behalf of their users. Their appeal is simple: instead of spending twenty minutes comparing flights, booking a reservation, or shopping for household supplies, a person could delegate the work to software.

But a growing number of websites are refusing to cooperate.

In recent weeks, users of consumer AI agents have encountered restrictions across major online services. According to reporting by TechCrunch and The Verge, businesses including Amazon, Yelp, Delta Air Lines, and major financial institutions have imposed limits on certain automated interactions.

Their reasons are not identical. Some cite security. Others require licensing arrangements. Some object to unauthorized transactions or access to customer information.

The resulting conflict is exposing a question that the first generation of AI assistants rarely had to confront.

When a customer authorizes software to act on their behalf, does the company operating the website have to accept it?

The answer is becoming one of the more consequential uncertainties surrounding the future of artificial intelligence.

The assistant at the locked door

The immediate difficulty is practical.

An AI assistant may be capable of identifying the right product, comparing alternatives, and navigating a checkout process. None of those abilities guarantees that the retailer will permit the assistant to complete a purchase.

The same distinction applies to travel reservations, banking, customer accounts, and other services where access is controlled.

The Verge reported on October 17 that users of agents from Meta, OpenAI, xAI, and other developers had encountered barriers on popular websites. Yelp told TechCrunch that AI agents must participate in its paid data-licensing program. Delta Air Lines was reported to be restricting agent-made bookings. JPMorgan Chase and Charles Schwab confirmed restrictions involving automated account access, citing security concerns.

These are materially different policies, even when the result looks similar to the user.

A bank refusing automated access to a sensitive financial account raises questions about identity verification, fraud, and liability. A review platform requiring a commercial data license raises questions about the economic value of its information. A retailer refusing an independent shopping agent raises questions about control over its customer relationships.

Grouping these decisions together as hostility toward AI would obscure their differences.

Yet they share a practical consequence: an agent can be technically capable of completing a task and still be unable to perform it.

That distinction challenges one of the central assumptions behind consumer AI agents. Their usefulness depends not only on what they can do, but also on what other businesses allow them to do.

Amazon and the question of authorization

The most revealing dispute has already reached federal court.

In November 2025, Amazon sued Perplexity AI over the operation of Comet, Perplexity's AI-enabled browser. Comet included an assistant that could navigate Amazon's website when activated by a user.

Amazon argued that Perplexity was accessing its systems without authorization. The company had informed Perplexity that its AI products were not permitted to access the Amazon store, and a central disagreement concerned whether the agent adequately identified itself.

Perplexity took a different position. It argued that the assistant was acting under the user's authority, much as a person might use other software tools to interact with a website.

A federal district court initially granted Amazon a preliminary injunction in March 2026.

Then the case took an important turn.

On August 4, the U.S. Court of Appeals for the Ninth Circuit vacated the injunction and sent the case back for further proceedings.

The appeals court reasoned that Amazon was unlikely to succeed on the relevant computer-access claims as presented because the user, employing the assistant as a tool, was the party accessing Amazon's computers for purposes of those statutes.

That is a significant distinction. The court's reasoning did not simply treat an independently developed AI assistant as a separate intruder merely because a website disapproved of its operation.

But the ruling was not a universal declaration that AI agents may access any website, under any circumstances. It concerned a preliminary injunction and particular claims under federal and California computer-access laws. It did not settle every possible contractual, privacy, consumer-protection, or security question.

The wider legal dispute continued.

Nevertheless, the case exposed a fundamental disagreement about the nature of automated assistance.

Amazon wants meaningful control over how third-party software interacts with its service. Perplexity argues that users should be able to choose tools that act on their behalf.

Both positions have implications beyond online shopping.

A ruling or industry standard that broadly favors website operators could leave consumers dependent on whichever AI assistants those operators approve. A system that gives outside agents unrestricted access could impose serious security and operational costs on the businesses responsible for users' accounts.

Neither outcome is obviously desirable in every situation.

The difficult question is where to establish the boundary.

Security is a real problem

There are legitimate reasons for websites to resist unfamiliar automation.

Modern online services already contend with credential theft, fraudulent transactions, mass scraping, abusive requests, and automated purchasing systems that undermine fair access.

An AI agent can resemble those systems from the website's perspective.

It may navigate unusually quickly, generate sequences of requests, access personal information, or attempt transactions with limited human supervision. A website may struggle to determine whether the activity represents a legitimate customer's instructions or an unauthorized automated process.

The distinction becomes especially important when an agent handles financial information, changes reservations, or accesses private accounts.

Even a well-intentioned agent can make mistakes. It might misunderstand a product's conditions, select the wrong option, submit incorrect information, or continue operating after the user's instructions become ambiguous.

Unlike a search query, some of these actions can create consequences that are expensive or difficult to reverse.

Anti-bot defenses were developed for precisely this kind of uncertain environment. Security tools often assess behavior rather than a program's stated intentions. An agent acting legitimately may therefore trigger safeguards intended to prevent fraud.

As The Verge reported, users have encountered situations in which automated assistance appears to be treated as suspicious activity rather than a permitted method of using a service.

That friction is not necessarily evidence of deliberate discrimination against AI. Sometimes it reflects security infrastructure that was designed before autonomous consumer agents became commonplace.

The challenge for agent developers is to demonstrate that their systems can identify themselves, respect authorization boundaries, protect credentials, and produce reliable records of their actions.

The challenge for website operators is to determine whether those safeguards are sufficient.

There is also a commercial struggle

Security, however, does not fully explain the resistance.

Many internet businesses depend on controlling how customers discover products, compare alternatives, and complete purchases.

A retailer's website is not simply a digital cash register. It is also a place to promote products, display advertisements, collect information about shopping behavior, recommend additional purchases, and develop an ongoing customer relationship.

An independent AI shopping agent could change that relationship.

Instead of browsing multiple pages and encountering the retailer's chosen recommendations, a consumer might ask an assistant to compare products across competing stores and purchase the most suitable option.

The customer could receive a more efficient shopping experience. But the retailer might lose some ability to influence how the decision is made.

This creates a potential conflict between customer convenience and platform economics.

The interests are not necessarily opposed. A retailer might welcome a trustworthy agent that generates additional sales. But a retailer could also be reluctant to surrender the shopping interface to an intermediary that controls product comparisons, customer discovery, and recommendations.

Perplexity has argued that Amazon's opposition reflects, in part, commercial incentives surrounding advertising and customer attention. That is Perplexity's interpretation of the dispute, not an independently established explanation for Amazon's decisions.

Still, the underlying economic question is substantial.

If a consumer increasingly interacts with an AI assistant rather than the businesses fulfilling their purchases, who controls the most valuable part of the relationship?

The answer could influence advertising, distribution, customer acquisition, and platform bargaining power.

The internet has seen related shifts before. Search engines and app stores changed how businesses reached customers, creating new intermediaries between consumers and online services.

AI agents could introduce another layer of intermediation, this time extending beyond discovery into execution.

But that outcome is not inevitable.

An alternative internet is being built

While some companies restrict independent agents, others are developing infrastructure specifically designed to accommodate them.

In January 2026, Shopify and Google announced the Universal Commerce Protocol, an open standard intended to help AI systems interact with merchants through structured commerce processes.

The protocol addresses more than product discovery. Its design includes shopping carts, checkout, order information, payments, and the operational requirements involved in completing a transaction.

In June, Shopify announced that developers could access its agentic commerce infrastructure without the previous approval requirement, broadening the tools available for connecting agents with participating merchants.

The development highlights an important alternative to having AI systems imitate human browser activity.

Rather than sending an agent through a website designed for people, a merchant can provide a defined technical interface through which authorized systems perform permitted actions.

Such an interface can make the interaction more reliable. It can establish what information an agent receives, which actions are supported, and when human confirmation is required.

Google's technical documentation describes the protocol as a common structure connecting AI surfaces, merchants, and payment providers.

This does not eliminate every concern about security, commercial power, or consumer choice. The specific terms of access still matter. Standards can be formally open while practical participation depends on integrations, platform policies, or business relationships.

Nevertheless, the approach suggests a possible direction for the industry.

The future may not depend on whether AI agents can navigate every existing website. It may depend on whether businesses create reliable ways for agents to interact with their systems directly.

That would represent a change in the internet's operating model.

Websites designed around human visitors would increasingly coexist with interfaces designed around software acting under delegated authority.

The emerging business of permission

Cloudflare is approaching the issue from another direction.

The internet infrastructure company has introduced controls intended to give website operators greater influence over how AI systems access their content.

In 2026, Cloudflare outlined changes allowing site owners to distinguish between search crawling, AI training, and certain forms of agent use. Its September policy changes introduced different default treatment for categories of automated activity on advertising-supported pages.

That distinction matters.

A crawler collecting text to train a model is not performing the same function as an assistant retrieving information for a user. Neither is necessarily equivalent to an agent attempting to purchase a product or access a private account.

The legal, commercial, and technical considerations differ substantially.

Treating every automated system as an ordinary visitor would ignore those differences. Treating every automated system as prohibited would ignore legitimate uses of automation.

Cloudflare's approach points toward an increasingly granular access system in which website operators specify which kinds of automated activity they permit.

The company has also explored monetization and reporting tools that could help publishers understand and potentially charge for certain AI-related access.

These developments suggest that permission could become an important commercial layer of the AI-enabled internet.

Access may depend on combinations of user authorization, technical identity, contractual terms, licensing, and security controls.

Such arrangements might make automation more dependable. They could also create new barriers to entry.

A large AI company may have the resources to negotiate integrations with major retailers, financial institutions, and service providers. A smaller developer may not.

If the most capable agents become those with the broadest negotiated access, competition could depend partly on commercial relationships rather than product quality alone.

That is a possibility worth examining, not an outcome that has already been established.

Whose assistant will it really be?

The deeper issue is whether AI agents will primarily represent their users or operate within commercial ecosystems defined by the companies that control access.

Imagine asking an assistant to find the best flight for a particular journey.

A genuinely useful assistant might consider price, travel time, baggage conditions, cancellation policies, and reliability before recommending an option.

But what if it can transact directly with only a subset of airlines?

It could still provide a valuable service. Its limitations, however, would influence which choices it can actually execute.

A similar problem could emerge in shopping, insurance, banking, and other industries where access is selective.

The agent might be capable of comparing a broad market but permitted to complete transactions with only certain participating businesses.

That creates a question of transparency.

Would users know which services their assistant can access? Would recommendations be influenced by commercial relationships? Could the system clearly distinguish a preferred option from an available option?

These questions do not prove that AI agents will systematically favor their commercial partners. Nor do they establish that closed integrations are inherently harmful.

Authorized partnerships can enable better security, clearer accountability, and more reliable transactions.

But the interests of the user, the AI provider, and the business fulfilling the service may not always align.

The design of these relationships will matter.

The internet's next negotiation

It is tempting to describe the emerging conflict as a battle between progress and resistance.

The picture is more complicated.

AI developers have legitimate reasons to pursue systems capable of acting across the internet. Consumers may benefit from faster, more accessible services and greater independence from individual platforms.

Website operators also have legitimate reasons to control access to private accounts, defend against abuse, and determine the conditions under which outside software interacts with their systems.

And both sides have commercial incentives that can complicate their public arguments.

What is changing is the assumption that a sufficiently capable AI assistant will automatically be able to do everything a person can do online.

The Amazon–Perplexity litigation demonstrates that the legal boundary is still being contested. Shopify's commerce infrastructure demonstrates that cooperation can take a structured technical form. Cloudflare's controls demonstrate that website operators are developing more sophisticated ways to distinguish acceptable and unacceptable automation.

These developments do not point toward a single settled outcome.

Some businesses may embrace independent agents. Others may permit access only through approved interfaces. Certain services, particularly those involving sensitive accounts or significant financial risk, may remain highly restrictive.

The internet is unlikely to adopt one universal rule for every kind of automated interaction.

But the central issue is becoming clearer.

The next stage of AI will not be determined by model intelligence alone. It will also depend on the systems, agreements, and institutions that decide how intelligence is permitted to act.

For years, the internet's most important interfaces were designed around what people could see and click.

Now the industry is beginning to negotiate what software can do on their behalf.

And that negotiation may ultimately shape the future of AI more than another improvement in what a model can say.