New Permissions And Growth Statistics Support A Vision For Agentic Shopping

As we saw last week, after taking an “ask forgiveness, not permission” approach to ingesting the entire internet, AI is being given retroactive approval to train on web content, both open and paywalled. In the US and India, there’s now weight on the side of the scale that says training AI models is considered transformative enough to constitute fair use, or fair dealing.

This week, another boon fell on the AI labs’ shoulders, as the Ninth Circuit Court of Appeals, in the USA, overturned a previous injunction that had effectively banned Perplexity’s Comet from operating on Amazon. 

In the lower courts, Amazon had successfully argued that Comet, which you might remember was one of the wave of “agentic browsers” that were all the rage towards the end of last year, was effectively hacking the marketplace. That sounds like a glib summary, but it’s fully accurate: Amazon went after Perplexity under the CFAA (Computer Fraud and Abuse Act), which was drafted in the mid-1980s as a way to give prosecutors a way to pursue crimes conducted on computers and involving “intangible property”.

The Eastern Herald has the best analysis of this case that The Interline found, and we encourage readers to go through it, but the fulcrum around which the reversal turned was, effectively, the idea that the agentic browser is doing the user’s bidding, and Perplexity consequently receives content from the user, not directly from Amazon. Perplexity is not, therefore, hacking Amazon – at least not of its own volition.

This looks like a weird technicality, and it kind of is, but it’s also the sort of small hole that giant trucks later get driven through. 

Most websites receive a tremendous amount of bot traffic today, and while a good amount of it is of the variety that web properties and readers are already used to (search crawlers and the like) AI bot traffic surged 300% between 2025 and 2026, and today nearly one in every thirty web visits is an AI bot. A lot of websites go to great lengths to try and block at least some of that traffic, since it doesn’t benefit them, and since it carries bandwidth and infrastructure overheads with it.

The Interline doesn’t block anything but the most obviously malicious automated traffic, so we can see that it comes at a cost. Every time we have hit concurrent memory allocation ceilings on our server infrastructure, bots have been to blame. And the bandwidth cost of serving up open-access PDF reports of 150mb and upwards has become progressively harder to justify over the last year or so.

We are not, to be clear, anything like the scale of Amazon or any other big retailer or brand website you care to name. The volume of bot traffic arriving at those websites is gigantic. As we’ve established, a lot of that traffic is effectively zero-value if it doesn’t translate into browsing or purchasing intent. And, Amazon originally argued, some of it is also net-negative value, because it believed that Perplexity was essentially conducting a man-in-the-middle campaign and using the convenience of automated browsing and checking-out to capture a lot of information about what Amazon sells, and how.

This is where the new ruling to overturn that injunction has a bigger scope than it might seem on the surface. In theory at least, Perplexity could still be receiving that information, depending on its policies for data retention and privacy where agentic browsing is concerned. The difference is that it’s being sent screenshots (which is largely how agentic browsing works) from the user’s viewport rather than directly from Amazon.

You can now apply this same standard to basically any agentic behaviour online, which makes this the third ruling in roughly the last twelve months to say that AI has, for most of the relevant purposes, open access to web properties.

Now, The Interline doesn’t believe that many retailers or direct to consumer brands have been especially concerned about agentic browsers over the last few months. While Comet is still technically available, Perplexity has shifted a lot of its efforts to Computer, which is, as the name suggests, a general purpose computer-use agent that can use a browser (headless or by clicking around on a running window) as part of a broad set of permissions to use local applications. 

OpenAI went further, and actually retired its browser, Atlas, after less than a year of availability – again in favour of general agents that perform a broader horizon of local work, in its Codex application. Claude Code and Cowork, of course, offer the same, and even though the OpenClaw hype has thankfully vanished, there are still no small number of people using Hermes Agent and similar harnesses to perform a suite of agentic tasks.

This move away from the browser as the agentic surface has also coincided with the integration of plugins, apps, and connectors that bring retailers’ catalogues directly into conversational AI interfaces. 

All of which has laid the groundwork for one of this week’s other major stories: Shopify’s Q2 earnings report, which placed a pretty serious emphasis on AI and agentic shopping (albeit not through AI-native browsers) as the next growth avenue for eCommerce. According to president Harley Finkelstein, AI traffic, and AI-influenced orders on Shopify storefronts have tripled year over year, and this growth is additive to traditional search, which is, apparently, itself growing 1.3x over a twenty-four month period.

This runs directly counter to two things: Amazon’s argument that agentic browsing cannibalises online sales and represents a security risk to retailers; and the broad understand amongst webmasters that interactions with AI are directly substituting for traditional web visits. We’ve always been open, here at The Interline, about the composition of our organic traffic, and while we haven’t seen a decline in Google referrals just yet, we have seen a steady uptick in AI traffic to where it makes up around 7% of our monthly new visits.

Shopify’s new data does include some suggestion of where this additional traffic comes from: according to the report, three quarters of “AI-attributed” purchases that happened in Q2 of this year took place outside the top 100 product categories. 

In practice, this means exactly what you think it means. Traditional search still rules the roost when it comes to the kinds of products that brands and retailers build their brands around, and for which traditional marketing and established channels still drive engagement and action. AI shopping is more active outside that purview, and in categories where pure quantitative analysis is more useful, such as the Shopify example of someone looking for baby car seats and the AI proactively evaluating them based on whether or not they’ll fit the user’s vehicle.

The important part for our readers to remember is that these AI interactions are not coming from browsers that are automatically clicking through existing websites. Instead they’re being measured in agent API calls made to Shopify’s catalogues, where agents are “working with richer structured data to match products with the buyer’s specific intent, rather than just keywords”.

This all sounds promising for anyone selling online, obviously. But it also sounds suspiciously like the original marketplace promise that Amazon has been fighting to try and keep out of the jaws of AI agents. 

Marketplaces were originally billed as an additive layer – a source of sales you wouldn’t get through your own channels, or through traditional retail partnerships. And for a while they did provide that for many brands, but then household names chose to withdraw from them as they began to realise that they’d ceded control to obtain reach, they’d lost access to critical consumer data, and they’d begun to see private label knock-offs competing against them in the same shared spaces.

So this week’s final headline, the funding round of a new eCommerce startup from the team that built Spotify’s algorithmic recommendation infrastructure, is probably well-timed, even if the announcement was accidental. Rather than needing to align themselves with Shopify to go after agentic traffic, Malachyte is aiming to deliver ‘proactive personalisation’ that’s aimed at human visitors first, but that very obviously lends itself to the kind of high-intent handoff and long-tail catalogue engagement that constitutes so much of the promise of agentic commerce. And it’s aiming to deliver it by giving retailers a way to control their own agentic surfaces, rather than either signing on to Shopify’s, or mounting a losing case to keep the agents out.

In an interview for The Interline Podcast at the beginning of the year, Jonathan Arena of New Generation seemed to see this coming: “You need to own first-party data, or else you risk being compressed out and just being a carousel competing with other carousels in a giant AI mode shopping mall.” 

Based on the stories from the last fortnight, that shopping mall now has the legal framework and the market justification it needs to exist. Which brings up the timeline for brands and retailers who want to stand apart from it.