Yelp wants paid data licenses. Delta won't let bots book seats. Two of the largest US banks have shut agents out of accounts. The promise that an assistant could shop and bank for you now runs into a wall of checkboxes.
What The Verge found in October
The Verge reported on October 17 that users of Meta's Muse, xAI's Grok, OpenAI's Dots, and a startup agent called Instinct have been running into blocks across major websites in recent weeks. Several of the affected companies confirmed the restrictions to reporters.
The pattern is uneven but consistent in direction. Retail, travel, banking, and local listings have all tightened up, and each site has picked a different reason.
Company | What it blocks | Who confirmed it |
|---|---|---|
Yelp | AI agents that don't pay for its data licensing program | Yelp, to TechCrunch |
Delta Air Lines | Agents booking tickets on a user's behalf | Delta, per The Verge |
JPMorgan Chase | Agents accessing customer accounts | JPMorgan, per The Verge |
Charles Schwab | Agents accessing customer accounts | Charles Schwab, via Bloomberg |
Amazon | Third-party shopping agents | Amazon, public statements in litigation |
Users also report that Cars.com and Amtrak stopped working with their agents after functioning before. Neither company has confirmed a policy change, so those two are user reports, not confirmed blocks.
Why Yelp's line is about money and banks' is about safety
Yelp's position is the most explicit commercial stance of the group. The company told TechCrunch it only allows AI agents that pay into its data licensing program. That's a straightforward extension of how Yelp already negotiates with search engines and data aggregators: if an agent reads your listings and transacts on them, you license the feed.
Delta and the banks are telling a different story. Both JPMorgan Chase and Charles Schwab have confirmed to reporters that they block agents from accounts because of security concerns. That framing is easier to defend publicly. An authenticated bot acting on someone's bank account is a fraud-surface question that site operators can point at without sounding anti-competitive.
Delta's reasoning is less clear in public. The airline is stopping agents from booking tickets on behalf of users, and the specific scope of the block isn't settled in reporting. It could be a bot-detection side effect. It could be a deliberate policy. The company didn't spell that out.
The Amazon and Perplexity fight set the template
Amazon has been the loudest of the group, and its dispute with Perplexity laid out the playbook the others are now following.
Amazon banned Perplexity's shopping agent and sued the startup. Perplexity fired back in a blog post arguing that Amazon doesn't want bots on its site because fewer human eyes mean fewer ad impressions. The startup framed its agent as an extension of the user, saying Amazon "does not believe in your right to hire labor, to have an assistant or an employee acting on your behalf."
A judge ordered Perplexity to stop its agents from shopping on Amazon. In August 2026, an appeals court overturned that ruling. Amazon has since filed an amended complaint, so the legal question about whether an agent acting for a customer counts as that customer remains unresolved.
Amazon has applied the same logic to Meta. When Muse kept hitting its site, Amazon said the assistant's "continued access" violated its site policies. The company also published a broader position earlier in the dispute: third parties buying on behalf of customers should "operate openly and respect service provider decisions whether or not to participate."
Meta's counterargument: a blocked agent is a turned-away customer
Meta answered with a line that's likely to be quoted for a while. In a post earlier this month, the company said that "turning away a personal agent means turning away the customer behind it," and that it would rather work with merchants on the underlying concerns than watch them lose the sale.
That's not just rhetoric. Meta has moved toward partnering with retailers rather than fighting them, working with Walmart, Gap, Sephora, and Best Buy on agent-friendly shopping paths. Shopify has opened its store network to more AI agents as well.
The contrast is instructive. Meta is betting that merchants will accept agents once the checkout flow is designed for them, with the merchant's rules intact. Amazon is betting that its own storefront is the product and that third-party agents degrade it.
Both bets can be right at the same time, which is why the fight looks like it's about technology when it's mostly about who owns the customer relationship.
Grok, Dots, and the X posts that made it public
The user reports clustered around Amazon first. People using Grok, OpenAI's Dots, and Instinct described being unable to complete purchases on Amazon. Elon Musk responded on X that maybe he "should start a competitor to Amazon."
There are also reports on Reddit of Google accounts linked to Muse getting flagged for "suspicious activity," and of checkouts stalling at Walmart. Those are community claims. Neither Google nor Walmart has confirmed a policy aimed at agents, and account flags can come from ordinary abuse detection.
The same caution applies to Cars.com and Amtrak. A site breaking for an agent is a fact about that user's session, not proof of a deliberate block.
The bots may be tripping ordinary fraud systems
One of the more useful observations in The Verge's reporting comes from Chris Canal, CEO of the AI safety evaluator EquiStamp. Many sites already run anti-bot systems aimed at scrapers and fraud, and autonomous agents trip those same defenses. A browser session that clicks through checkout at machine speed, from an unusual network, with no mouse movement looks like an attack.
Cloudflare has added to the pressure. It introduced a one-click block for AI crawlers in 2025, then moved against crawlers that scrape for both search indexing and model training. Those tools were built for crawlers, but they catch agents that behave similarly.
Canal's warning is that the defensive layer will get more visible to ordinary people. Expect more "verify you're human" prompts on sites that never needed them, including grocery shopping.
What this means for anyone who bought into agent hype
The practical answer is that agents work well on sites that want them and badly on sites that don't, and the split is not obvious until you try. A shopping agent that sailed through an independent retailer may stall at a major bank or an airline.
Canal put the strategic cost plainly. If these assistants promise to handle anything you'd do online, and companies are actively working to prevent that, "there are just fewer useful things the agents can do."
That has a knock-on effect for the labs selling the agents. A feature that fails on the sites people actually use is a weaker feature, whatever it does on a demo.
There's also a quieter risk for users: agents that route around blocks will look more like scraping, which invites harsher countermeasures and more aggressive account flagging. An assistant that gets your account suspended has cost you more than the task it was going to finish.
Where this goes next
Three things are worth tracking. First, Amazon's amended complaint and the Perplexity appeal, because a ruling that agents are legally distinct from users would reshape the argument for every site in the table. Second, whether Yelp's licensing model spreads, since a paid-feed arrangement is easier for large platforms to sustain than a total block. Third, whether the agent makers standardize on something like an identifying header, which would let sites separate a legitimate assistant from a scraper.
None of those is settled. Until they are, the honest summary is that the open web is getting selective, and the first group to feel it is the people who believed the demos.
