Companies spent 20 years trying to keep bots off their websites. Now they’re figuring out how to get bots to buy something. 😂
What interests me more is what happens to customer loyalty. I might stick with a bank for ten years simply because I’m too lazy to switch. My AI agent won’t have that problem. It’ll compare fees, rates and services in seconds and probably find a better deal before I’ve finished my coffee.
That’s a pretty uncomfortable future for businesses that make money from customers not bothering to shop around.
It's super exciting, and maybe a little bit daunting, to consider the ramifications of the agentic economy. I'm building in this space and one of the unanswered questions for me is how to "get the attention" of agents to visit the websites (would they even be called "websites" anymore?) of small or niche businesses. SEO and AEO/GEO still matter in terms of surfacing on their searches, but my sense is that getting chosen will still consist of playing a long game of building brand and domain authority over time. Fun times.
I use claude and codex daily to write the code, so I watch agents choose tools every day. What I have seen matches your coding case study. An agent skips the homepage and reads the docs. It tries the first command, and when that fails it reads the error message. If the error says what to do next, the agent stays. If it does not, the agent moves to the next option in seconds and never tells anyone why. Stripe seeing about 40% of its docs traffic from agents makes sense to me, because for an agent the docs are the sales pitch and the error message is the sales call.
The part I would love to see later in the series is the other side of "on our behalf". A human buyer brings habit and loyalty. An agent brings neither, so it decides again from scratch on every run, and a company can lose a customer who never knew they had picked it. I think the companies that win agents' business will be the ones whose pricing, limits and docs are true and readable by a machine. For twenty years we competed for attention. An agent has all the attention in the world and no patience for a claim it cannot check.
the bit about retailers spending years building walls to keep bots out and now having to fight for their business is the part I keep thinking about, and it's a big reason LangChain doubling to a $3b valuation last week made sense to me, if agents become the customer then whoever runs the layer they're built and managed on gets paid either way. I'd love to know who you think ends up owning the checkout step though because that feels like where most of the money actually sits
As someone actively building an AI application, I look at this ecosystem and see tremendous possibilities for integrating new capabilities. Are we creating technological capabilities faster than we're developing the talent required to integrate them? Where do builders find people with the expertise to bring these technologies together? Are most specialists siloed within their respective domains? And how many people actually have the cross-platform knowledge to architect and deploy advanced stacks?
Great piece. The line that stuck with me: "GEO largely stops at discovery. Agents will take it further, into evaluation and execution." For most categories, execution is a task the agent finishes for you. For entertainment, execution is the experience, an agent can find you a story, but it can't live it for you. That's the bet behind https://prelulu.ai: be the place agents send people when they want a character who actually picks up, live, face to face.
We would like you to try Airtap! We built a phone harness of our own, back in 2020, when there was no GPT and no Agents! We are using that - in addition to coding sandboxes, browsers etc.
If most agent shopping still happens by driving a browser or a throwaway VM, I can see why enterprises are still blocking the traffic over fraud and chargebacks. Does "Sign in with ChatGPT" bind the lent credentials to that one purchase, or can the agent reuse them later for something the user didn't expect?
The distribution point is the one I keep coming back to. In fashion, the product feed is the storefront an agent sees, and most fashion feeds are written for a filter, not a decision. "Black midi dress, 100% viscose" tells an agent nothing about occasion, fit or what it goes with. So agents default to the brands they already know, and mid-tier labels with a better product but thinner data just disappear from the answer. I think the brands that win B2A in fashion won't be the biggest. They'll be the ones whose catalogs read like a stylist wrote them.
If an agent picks a detailing business purely on price and reviews, two years of regulars stops mattering. Or maybe the little guy finally wins because the agent doesn't care about brand. Not sure which one scares me more.
Shopify’s 8x rise in AI traffic alongside nearly 13x more orders is the more interesting number here. Is that a conversion shift, or just a very small base getting bigger?
"Companies spent decades keeping bots out. Now they'll compete for their business."
Olivia and Justine, this made me think about how people actually shop. So much of it is habit and feeling. We buy the pack we know, from the shop where someone remembers our name, or the brand our mother always used. An agent won't care about any of that. It will just compare and pick. I'm curious which brands will feel it first. Do you think loyalty still matters when an agent is doing the buying?
Your line about companies competing for agents as customers flipped my thinking. I have been writing about what agents must never do, and now I am wondering what businesses must start doing to win their business.
Fraud scoring will flag a lot of legit agent checkouts first. The models were trained on human sessions, and an agent filling the whole form in under a second from a cloud IP looks like card testing. If agent orders matter to you, tag them at the edge and watch their decline rate on its own.
Companies spent 20 years trying to keep bots off their websites. Now they’re figuring out how to get bots to buy something. 😂
What interests me more is what happens to customer loyalty. I might stick with a bank for ten years simply because I’m too lazy to switch. My AI agent won’t have that problem. It’ll compare fees, rates and services in seconds and probably find a better deal before I’ve finished my coffee.
That’s a pretty uncomfortable future for businesses that make money from customers not bothering to shop around.
Disrupting of moats for major companies will happen. New ways to make moat will appear as well.
It's super exciting, and maybe a little bit daunting, to consider the ramifications of the agentic economy. I'm building in this space and one of the unanswered questions for me is how to "get the attention" of agents to visit the websites (would they even be called "websites" anymore?) of small or niche businesses. SEO and AEO/GEO still matter in terms of surfacing on their searches, but my sense is that getting chosen will still consist of playing a long game of building brand and domain authority over time. Fun times.
I use claude and codex daily to write the code, so I watch agents choose tools every day. What I have seen matches your coding case study. An agent skips the homepage and reads the docs. It tries the first command, and when that fails it reads the error message. If the error says what to do next, the agent stays. If it does not, the agent moves to the next option in seconds and never tells anyone why. Stripe seeing about 40% of its docs traffic from agents makes sense to me, because for an agent the docs are the sales pitch and the error message is the sales call.
The part I would love to see later in the series is the other side of "on our behalf". A human buyer brings habit and loyalty. An agent brings neither, so it decides again from scratch on every run, and a company can lose a customer who never knew they had picked it. I think the companies that win agents' business will be the ones whose pricing, limits and docs are true and readable by a machine. For twenty years we competed for attention. An agent has all the attention in the world and no patience for a claim it cannot check.
the bit about retailers spending years building walls to keep bots out and now having to fight for their business is the part I keep thinking about, and it's a big reason LangChain doubling to a $3b valuation last week made sense to me, if agents become the customer then whoever runs the layer they're built and managed on gets paid either way. I'd love to know who you think ends up owning the checkout step though because that feels like where most of the money actually sits
As someone actively building an AI application, I look at this ecosystem and see tremendous possibilities for integrating new capabilities. Are we creating technological capabilities faster than we're developing the talent required to integrate them? Where do builders find people with the expertise to bring these technologies together? Are most specialists siloed within their respective domains? And how many people actually have the cross-platform knowledge to architect and deploy advanced stacks?
Great piece. The line that stuck with me: "GEO largely stops at discovery. Agents will take it further, into evaluation and execution." For most categories, execution is a task the agent finishes for you. For entertainment, execution is the experience, an agent can find you a story, but it can't live it for you. That's the bet behind https://prelulu.ai: be the place agents send people when they want a character who actually picks up, live, face to face.
We would like you to try Airtap! We built a phone harness of our own, back in 2020, when there was no GPT and no Agents! We are using that - in addition to coding sandboxes, browsers etc.
If most agent shopping still happens by driving a browser or a throwaway VM, I can see why enterprises are still blocking the traffic over fraud and chargebacks. Does "Sign in with ChatGPT" bind the lent credentials to that one purchase, or can the agent reuse them later for something the user didn't expect?
The distribution point is the one I keep coming back to. In fashion, the product feed is the storefront an agent sees, and most fashion feeds are written for a filter, not a decision. "Black midi dress, 100% viscose" tells an agent nothing about occasion, fit or what it goes with. So agents default to the brands they already know, and mid-tier labels with a better product but thinner data just disappear from the answer. I think the brands that win B2A in fashion won't be the biggest. They'll be the ones whose catalogs read like a stylist wrote them.
If an agent picks a detailing business purely on price and reviews, two years of regulars stops mattering. Or maybe the little guy finally wins because the agent doesn't care about brand. Not sure which one scares me more.
Shopify’s 8x rise in AI traffic alongside nearly 13x more orders is the more interesting number here. Is that a conversion shift, or just a very small base getting bigger?
"Companies spent decades keeping bots out. Now they'll compete for their business."
Olivia and Justine, this made me think about how people actually shop. So much of it is habit and feeling. We buy the pack we know, from the shop where someone remembers our name, or the brand our mother always used. An agent won't care about any of that. It will just compare and pick. I'm curious which brands will feel it first. Do you think loyalty still matters when an agent is doing the buying?
Agreed. We’re moving from stage 3 to stage 4.
AI Tools → Task-Executing Agents → Delegated Autonomous Agents → Accountable Economic Agents → Life-like Autonomous Entities.
Oh my goodness thanks so much for this I am just getting my head round this!
Your line about companies competing for agents as customers flipped my thinking. I have been writing about what agents must never do, and now I am wondering what businesses must start doing to win their business.
Fraud scoring will flag a lot of legit agent checkouts first. The models were trained on human sessions, and an agent filling the whole form in under a second from a cloud IP looks like card testing. If agent orders matter to you, tag them at the edge and watch their decline rate on its own.