Meta's Muse was downloaded 5 million times in its first month. Between August and September, agentic web traffic tripled. Most websites responded the only way their architecture allowed: they closed the door.
A recent piece on MIT Technology Review laid out what happens when millions of personal agents arrive on a web built for humans. The reporting is worth reading in full. Here is how I read it from the commercial side.
Two categories were enough for twenty years
Every analytics setup, fraud filter and ad auction I have worked with since 2008 rests on one assumption. A visitor is either a person or a bot. People get the experience, bots get the rules in robots.txt, and anything in between gets a captcha.
Personal agents break that split. An agent sent by a customer to book a table or compare insurance is not a scraper. It carries real purchase intent, it just moves at machine speed and never sees your banner.
So the agent gets treated as a bot. TechCrunch reported people failing to get agents through sites from Pizza Hut to Yelp. Amazon blocked Muse and is fighting Perplexity in court to keep its agents out, appealing a ruling that went Perplexity's way.
Amazon's real argument is about ads
The most revealing detail in the story is Amazon's reasoning. In its appeal, it argues that AI agents threaten its advertising business because it becomes harder to tell ads seen by humans from ads served to bots.
Read that as an operator, not as a lawyer. A retail media business sells human attention.
If a growing share of the shoppers on the page are software, the impression count stops meaning what advertisers are paying for. The block is not about security. It is about protecting the unit of sale.
Every business with an ad model or an attention model has the same exposure. Publishers, marketplaces, comparison sites. The question is not whether agents are good or bad. The question is whether your revenue depends on the visitor being human.
The counter-example is Shopify, which is giving agents their own checkout lanes that skip the pages designed for humans. That is a very different bet. It says: if the agent brings a buyer, take the order, and stop pretending the visitor has to scroll a product gallery first.
The economics explain the split. According to the same reporting, the vast majority of agentic traffic goes to e-commerce sites. If your money comes from the transaction, an agent is a customer with a faster cursor. If your money comes from the attention around the transaction, an agent is a leak.
I have seen the same divide play out in affiliate marketing for years. Whoever owns the conversion welcomes any traffic that converts. Whoever monetizes the path to conversion fights anything that shortens it.
The customer service problem nobody budgeted for
Ethan Mollick from Wharton made a prediction in the piece that most operators should take seriously. Once people learn that agents can request refunds and negotiate prices, they will send agents to do exactly that, at volume. Customer service queues sized for humans will not survive it.
Think about what that means in practice. A refund policy that relied on friction, a phone call, a form, a wait, was never really a policy. It was a filter. Agents remove the friction, so the real policy becomes visible, and so does its cost.
The same goes for pricing. If different agents start seeing different prices, every inconsistency in your pricing architecture gets discovered, logged and compared within days. Hidden discounts stop being hidden. Negotiable prices get negotiated every single time.
Your conversion data is about to get strange
There is a quieter second-order effect. An agent does not browse the way a person does. It lands on the right page, checks three facts and either buys or leaves, all in seconds.
That wrecks the behavioural signals most marketing teams optimize against. Time on page drops, bounce rate climbs, funnels look broken. Teams that do not separate agent sessions from human ones will spend the next year fixing pages that are working perfectly well for the visitor actually using them.
What I would decide this quarter
First, find out how much agent traffic you already have. Most analytics tools still file it under bots or drop it entirely. Ask your developers to segment by user agent and behaviour, and look at what those sessions were trying to do.
Second, write an explicit agent policy, the way you once wrote a cookie policy. Which actions can an agent take on a customer's behalf? Browse, compare, buy, cancel, request a refund? Decide each one on purpose, instead of letting a default firewall rule decide for you.
Third, make the parts you want agents to use legible. I wrote earlier about how to make your site legible to machines, and the logic is the same here: plain prices, clear stock status, structured policies. And if paid acquisition is part of your model, the shift I described in paid search being rebuilt for agents is already showing up in this traffic.
Finally, stress-test your policies against a visitor with infinite patience. Any rule that only worked because humans give up is now a liability. Fix it before an agent finds it for you.
The web had two kinds of visitors for a long time. It has three now, and the third one carries a wallet.