The ruling
On March 10, 2026, Amazon won a court order that stopped Perplexity's Comet browser from shopping on Amazon for its users. Amazon's case rested on the Computer Fraud and Abuse Act, the federal anti-hacking law. On August 4, 2026, the Ninth Circuit appeals court threw that order out, reasoning that an agent acting on a user's instruction is the user acting.
Most of what this site argues rests on an idea that no court had tested until this month. When someone tells an AI assistant to go and do something on a website, the visitor that reaches your server is that person, not some scraper in disguise.
A federal appeals court has now said something close to that out loud. It's only one court, the court only threw out a temporary order, and the full case is still undecided. None of this is legal advice, either. Still, it's the first serious sign of how the law is likely to treat a kind of traffic you're about to see a lot more of. Below: what the court decided, why it now matters which kind of bot you're dealing with, and what to change on your own site.
What happened
Amazon sued Perplexity over Comet, Perplexity's agentic browser (a web browser with an AI assistant built in that can click and type for you). Comet could log into a user's Amazon account and finish a purchase for them. Amazon built its case on the Computer Fraud and Abuse Act, the federal anti-hacking law, which turns on whether someone's access to a computer was "authorized." Amazon's terms ban automated purchasing, so Amazon argued that an agent buying things had gone beyond the access it was allowed. On March 10, 2026, a district court agreed enough to issue an order blocking it.
On August 4, 2026, the Ninth Circuit threw that order out (the legal word is "vacated"). The court's reasoning matters well beyond Amazon and Perplexity. The court treated an agent working at a user's direction as a tool in that user's hand, and the user already has permission to be there. If a customer may log in and buy something, the customer's agent logging in and buying that same thing is the same permitted access, carried out a different way.
The order the court threw out was a preliminary one (a temporary order made before the full case is heard), so this isn't a final judgment, and the dispute itself goes on. What matters most is how the court framed it, moving the question from "was this automated?" to "was this the customer?"
Why it now matters legally which kind of bot it is
The most common mistake in this whole area is lumping all non-human traffic into one group. The ruling speaks to only one of two very different things.
Agents working for a user, which identify themselves with names such as
ChatGPT-User, Claude-User, Perplexity-User, and Google-Agent, visit
because one particular person asked for something right now. The agentic
browsers belong in this group too. Someone is sitting there waiting on the
result. Each visit is one request, made for one person, with one
goal. This is the kind of agent the Ninth Circuit was talking about.
Bulk crawlers (programs that visit pages in bulk and copy them), such as
GPTBot, ClaudeBot, and CCBot, work quite differently. They run on the AI
company's own schedule and for its own purposes, like building a training
collection or a search index, and nobody in particular is waiting. No
customer's access rights are being used. Nothing in the Comet decision says you
can't keep these bots out, and all the ways of granting or refusing them
permission that grew up over the years still stand.
If your robots.txt (the file that tells bots which parts of your site they may visit) or your firewall treats those two groups with a single rule, you've made a policy decision you probably didn't mean to make: it either blocks your customers' assistants or lets bulk copiers in on the same terms. We wrote about how OpenAI's own bots split along those lines in what OAI-AdsBot means for your robots.txt.
The detection problem is worse than the legal one
Say you disagree with all of this and want to block agents that work for users anyway. More and more often you can't, because there's nothing to tell them apart by.
Agentic browsers are browsers. Comet is built on Chromium, the open-source
project underneath Google Chrome, and Anthropic's and OpenAI's browser
integrations run inside Chrome itself. Their requests carry an ordinary Chrome
user-agent string (the label every browser sends to say what it is), real TLS
fingerprints (the technical signature of a browser's secure connection), real
cookies, a real logged-in session, and a real screen. Looking at the raw web
traffic, there's often nothing that separates "the user clicked this" from "the
user's agent clicked this for them." Blocking by user-agent string only
catches agents that say who they are, the ones ending in -User. Those are the
well-behaved ones, so a block mostly hits agents playing fair. They're also the
part that's shrinking, as agents move into the browsers people already have. That shift is
the subject of
Atlas is dead, agentic browsing isn't.
What to do about it
- Write your terms about behavior, not technology. "No more than N orders per account per hour," "no scraping for resale," and "no getting around rate limits" (caps on how many requests one visitor can make in a set time) can be enforced and defended. "No automated access" is turning into a rule against your own customers, and the Ninth Circuit just made it a weaker thing to sue on.
- Set rate limits per logged-in account. Limit by the account, the session, or the key a program uses to reach your API (application programming interface, the way software connects to your site). Don't limit by the user-agent string. Limits tied to an account still work when the visitor looks exactly like a browser, and they line up with real harm.
- Keep separate rules for reading and for acting. Letting an agent read a product page and letting it place an order are different risk decisions, so give each one its own policy.
- Keep bulk-crawler rules apart from agent rules. They're different traffic, under different law, with different economics, so a single rule will be wrong for one of them.
Splitting reading from acting is the idea to keep
Reading and acting carry very different risks, and that split applies well beyond this case. Reading is cheap, low-risk, and mostly good for you. An agent that reads your documentation and quotes you accurately is giving you free distribution. Acting is where the risk is. Money moves, stock levels change, records get updated, and if the agent misunderstood what its user wanted, you get a support ticket and maybe a chargeback (a card payment the bank reverses).
If your site has one policy covering both, split it. A single bot policy usually dates from when the only non-human visitor anyone thought about was a search engine's crawler (the bot it sends to read and index pages). The sites that handle the next few years well will write two. One is an open, generous reading policy that gives agents the best chance of finding you and quoting you correctly. The other is a careful acting policy, with sign-in, confirmation steps, limits, and a way to undo things built into every place an agent can change something. Our interaction models material works through what the acting side looks like on a real site.
The rest of the legal picture
The Comet ruling didn't arrive out of nowhere, and it helps to be precise about what the cases around it do and don't settle.
Two other cases are about copyright and training data, not about agents visiting a site for a user. CNN sued Perplexity on May 28, 2026, over how Perplexity's answers use CNN's content. Bartz v. Anthropic, over how Anthropic got its training data, ended in a $1.5 billion settlement that a court gave final approval on July 20, 2026. A settlement doesn't set a rule other courts have to follow.
The 2025 Cloudflare and Perplexity stealth-crawling dispute was never formally resolved. Cloudflare published research claiming Perplexity used undeclared user-agents (bots that hid who they were) to reach content that had blocked its declared ones. No court ruled and no regulator acted. It ended as a fight over reputation, which we covered in detail in the Perplexity-Cloudflare crawling controversy.
Taken together: how AI companies collect training data is contested, reusing content in answers is unresolved and in active litigation, and access by an agent working for a user is, for now and in one appeals court's region, treated as the user's own access.
The caveats, in plain terms
This is one Ninth Circuit decision throwing out a preliminary order. It isn't settled national law. Another appeals court could decide the opposite and create a split between the courts, and the case itself could still end differently. State computer-crime laws, contract law, and consumer-protection rules all work separately from the federal anti-hacking law, and they could reach different results on the same facts. None of this is legal advice. If you're making a decision that depends on it, talk to a lawyer who can look at your situation.
References
- CNBC, "Amazon wins court order to block Perplexity's AI shopping agent", March 10, 2026
- Engadget, "Appeals court vacates the order blocking Perplexity's Comet from Amazon", August 4, 2026
- Cloudflare, "Perplexity is using stealth, undeclared crawlers to evade website no-crawl directives", August 2025
- OpenAI, OpenAI crawler and user-agent documentation
- Anthropic, Claude bot and agent user-agents
See the path on your own site
If an agent acting for someone counts as that someone, your site is already serving those agents. See the path one takes through your pages, and where it gets stuck.