> Why this is happening. Two forces are slowing agentic commerce, according to Leigh McKenzie, director of online visibility at Semrush: infrastructure and trust. Real-time catalog normalization across tens of millions of SKUs is a decade-scale problem Google already solved with Merchant Center, and consumers still default to checkout flows they trust — Apple Pay, Google Wallet, and Amazon one-click.
It turns out when you step outside of “hard tech” problems like building GPT6 there are all of these details others have solved already. E-commerce has been optimized to the last decimal point for the last 30 years.
OpenAI is new to it, and if I had to guess, not that interested in getting good at it.
I think they're interested in getting good at it. They just don't want to put in the human time and effort to do so. They expect their many failures and short-comings to be shored up by continuous model training.
But that, of course, means that in the meantime it will suck and nobody will use it.
Also having to wait for ChatGPT for a "thinking" response to search for information that is slower than a Google search loses them lots of money.
I believe that it can still work and I won't claim about being unsurprised about this failure. But this is a great opportunity to execute this problem really well if OpenAI and others are not interested in getting good at this.
Perplexity also attempted this, got sued by Amazon and it appears semi-abandoned.
The only problem is that it must be quicker or just as quick as a Google search, and also compatible with the existing checkout flows.
A lot of the AI companies visibly suffer from the engineer's disease. It's kind of interesting to look at them through that lens, and of course, the claims they make about the future.
Two forces are slowing agentic commerce, according to Leigh McKenzie, director of online visibility at Semrush: infrastructure and trust
Does anybody else just feel the aloof out of touchness just oozing from that sentence? "Trust", as if this is just any old metric they merely have to work to increase.
This is what I want from a purchasing agent: I make a list of items that I repeatedly buy (mostly household supplies), and that I keep roughly updated with my inventory / need. The agent tracks prices and sales across all web stores, making appropriate purchase decisions based on which is the least expensive, combining shipping, taking advantage of sales to stock up, etc. For other one-time purchase items, I input what I am looking for and can create a persistent pan-site shopping cart that once again minimizes costs and shipping fees. Being very explicit here: the main goal of an "agent" should be to represent and carry out my own interests.
And these functionalities have been straightforwardly doable without "AI" for the past few decades, except for the glaring incentives against them! It is in every web store's interest to undermine customers' ability to obtain semantic pricing, shop around, create a cart independent of their site, etc. These incentives are why when you visit any web store these days, the very first thing they do is hassle you with CAPTCHAs (and the bad ones keep doing it throughout the session!) - they want to make sure you're an actual computationally-unassisted human sitting there, wasting your personal time with their bloated pages that take tens of seconds to download and render, so that you don't spend that time being a more efficient market actor.
Now, does "AI" have the capability to go against these trends and enable user-centric algorithmic shopping and purchasing? Perhaps, and I hope so! But it's certainly not going to be led by these popups on web stores nagging me if I want to chat instead of doing the thing I went there to do (which when you think about it, this is just the latest instance of these stores trying to make you waste your time on their site). Rather, it will come from completely third party services (or ideally software) setting out to act in customers' interests, and performing adversarial interoperability to achieve this!
This is all valid (except probably the last sentence), but it also describes so many attempted changes right until they become darn near the default.
This sounds like why I heard Redfin wouldn’t work, or Netflix, or Amazon, or Uber, or PayPal, etc…. There are always these business complexities that make it seem like these spaces have too much friction, but if there’s enough money - if it can be done then people will figure it out.
Shopping research has been pretty funny to me at least, a straightforward way for them to do product placement that people actually want, but implement it so poorly that half of the links it returns are broken.
> E-commerce has been optimized to the last decimal point for the last 30 years.
It certainly hasn't been optimized to anything in 1996. In 1996 it was people clumsily scanning print catalogs, spending 5 hours to upload 10 images on dialup and making a simple HTML page (no DB or any kind of backend) and putting their landline phone on it with a message to "call to checkout"
I know you were exaggerating for effect, but E-commerce and catalog normalization are definitely not "solved" everywhere.
McMaster Carr is a good example of a company that has 90%+ of their stuff ironed out, but most websites and especially small ecommerce isn't like that.
> E-commerce has been optimized to the last decimal point for the last 30 years
Sort of, but there's a ton of middlemen between "resources in the ground" and "product in my hand". For example, how much utility is there in a "store" to thee consumer at this point? Let me buy from the manufacturer.
I am behind schedule on developing a "summer phase" [1] for my foxographer costume and was chatting with Gemini about a crash priority "spring phase" [2] and asked it for suggestions and it gave me a 10-pack of results that had one good thing in it at rank #8, a similar query run against a normal search engine actually got something better at #1. Now sure I am talking w/ Gemini with big words like "supergraphic" whereas a normal search would be heavy on 3-letter and 5-letter words used in the product descriptions.
It makes think though of expert system based product configurators back in the 1980s
thing is that kind of product configurator is based on an ontology, constraints and rules as opposed to embeddings which might capture the "feel" of things like clothing.
[1] Busytown meets Arknights
[2] supergraphic shirt + camera gets resonance with my promotional system and people keep approaching me (e.g. laugh but every KPI in the system has an extra zero on the left)
30 years? E-commerce hasn't been around that long - try 5 years of optimization MAX.
FWIW OpenAI is desperately trying to monetize and they think e-commerce is a "simple" problem to solve. I mean they do need to convert their funnel without alienating their users. I assume they are going to have some big payouts for agentic purchases gone awry or leave merchants on the hook.
ChatGPT recommended me some good hard drives for price per TB, and one particularly cheap one had direct checkout with Walmart, so I tried it, because why not? It let me get all the way to the payment step before it told me it was out of stock. Walmart's website told me it was out of stock when I decided to click on the link. This is probably part of why it doesn't convert.
You can either have AI be honest or AI become a marketing tool. The two are fundamentally incompatible.
You won't get it to push your products when users ask what's the best XYZ - either because it'll be too honest to lie or because it'll be too expensive for you.
The idea that AI will suddenly solve e-commerce demonstrates a lack of understanding on everything that has happened in this space over the last 25 years.
There’s a lot of this going on in AI at the moment. New folks come in thinking they have a magic solution and then produce a total train wreck as it turns out domain expertise is still a thing.
Walmart does not, over 10 years after they were released, even accept the contactless payment systems in common use. Instead, they push their in-house version in part so they can capture the relevant customer data.
And we're meant to believe that Walmart planned to outsource the entire series of touchpoints represented by the discovery & checkout process? Yeah, okay.
This was never going to be more than an experiment for Walmart.
Wow the sceptics really came out in force for this one.
I’m currently using Gemini to research components for a remote controlled plane. I have the frame of the plane and now need to buy correctly specced servo motors, an engine, battery, etc etc. It has saved me so much time and educated me tremendously on how the different components interact and the options available.
If I could just press “buy” from within Gemini and pay via Google Pay (or better still, Apple Pay) I’d do it in a heartbeat.
How many people tried for the novalty with no intention of purchasing? It being a thousand times worse conversion wouldn't matter if they are additional sales???
It's probably stuff like this along with investor pressure that will make AI companies slowly make their AIs more profit maximizing (and the long term reason ilya etc was so against even going down that path)
Last year they couldn't draw fingers on hands properly, this year they can't convert at checkouts, I wonder what they'll be failing to do a year from now.
The shift to AI is currently a boon to consumer. Penny’s has obviously done this, as they have had a $119 Man U jersey ring up at $19 for a week now, with many of my mates having bought one. It’s unbelievable that anyone thinks gutting human oversight builds a better company.
They don't want complete shopping automation because if it really worked, people would buy LESS. Walmarts in person and online storefront is engineered to get people to buy things they didn't come in for. Upselling is programmed at every step.
That doesn't seem terribly surprising, a human can quickly look through a grid of shirts to find one they like. ChatGPT would be guessing what they might want and the human would probably get a bad experience there with some regularity.
Personally, I cancelled my subscription the day that they announced there would be ads in ChatGPT. Not that I was surprised on that day.
I find it kind of fascinating seeing people now debate the ins and outs of just exactly how useful the ads/affiliate schemes inside the chatbot are.
Pretty much in the same way that I'm fascinated about people debating the quality of ads shown in their browser while I've been using ad blockers for forever now and cannot imagine seeing the web without them.
So they are comparing to the conversion rate of people who click on a link in the chat and go to Walmart's website to view the product? Wouldn't that be a really strong intent-to-buy signal?
The better comparison might be conversion rate for those who searched on Walmart.com vs those who searched within ChatGPT. Or maybe that is what they're comparing and I misunderstood?
The experience is a lot like when you are talking with a friend, then they decide to ask siri or google a question using voice. The result is always imprecise. Meaning they either have to repeat their query, or end up typing it anyway.
If you want to buy a Walmart product, the easiest way is to go to Walmart. Why add an imprecise middle man in between?
278 comments
> Why this is happening. Two forces are slowing agentic commerce, according to Leigh McKenzie, director of online visibility at Semrush: infrastructure and trust. Real-time catalog normalization across tens of millions of SKUs is a decade-scale problem Google already solved with Merchant Center, and consumers still default to checkout flows they trust — Apple Pay, Google Wallet, and Amazon one-click.
It turns out when you step outside of “hard tech” problems like building GPT6 there are all of these details others have solved already. E-commerce has been optimized to the last decimal point for the last 30 years.
OpenAI is new to it, and if I had to guess, not that interested in getting good at it.
> not that interested in getting good at it
I think they're interested in getting good at it. They just don't want to put in the human time and effort to do so. They expect their many failures and short-comings to be shored up by continuous model training.
But that, of course, means that in the meantime it will suck and nobody will use it.
I believe that it can still work and I won't claim about being unsurprised about this failure. But this is a great opportunity to execute this problem really well if OpenAI and others are not interested in getting good at this.
Perplexity also attempted this, got sued by Amazon and it appears semi-abandoned.
The only problem is that it must be quicker or just as quick as a Google search, and also compatible with the existing checkout flows.
>
Two forces are slowing agentic commerce, according to Leigh McKenzie, director of online visibility at Semrush: infrastructure and trustDoes anybody else just feel the aloof out of touchness just oozing from that sentence? "Trust", as if this is just any old metric they merely have to work to increase.
This is what I want from a purchasing agent: I make a list of items that I repeatedly buy (mostly household supplies), and that I keep roughly updated with my inventory / need. The agent tracks prices and sales across all web stores, making appropriate purchase decisions based on which is the least expensive, combining shipping, taking advantage of sales to stock up, etc. For other one-time purchase items, I input what I am looking for and can create a persistent pan-site shopping cart that once again minimizes costs and shipping fees. Being very explicit here: the main goal of an "agent" should be to represent and carry out my own interests.
And these functionalities have been straightforwardly doable without "AI" for the past few decades, except for the glaring incentives against them! It is in every web store's interest to undermine customers' ability to obtain semantic pricing, shop around, create a cart independent of their site, etc. These incentives are why when you visit any web store these days, the very first thing they do is hassle you with CAPTCHAs (and the bad ones keep doing it throughout the session!) - they want to make sure you're an actual computationally-unassisted human sitting there, wasting your personal time with their bloated pages that take tens of seconds to download and render, so that you don't spend that time being a more efficient market actor.
Now, does "AI" have the capability to go against these trends and enable user-centric algorithmic shopping and purchasing? Perhaps, and I hope so! But it's certainly not going to be led by these popups on web stores nagging me if I want to chat instead of doing the thing I went there to do (which when you think about it, this is just the latest instance of these stores trying to make you waste your time on their site). Rather, it will come from completely third party services (or ideally software) setting out to act in customers' interests, and performing adversarial interoperability to achieve this!
This sounds like why I heard Redfin wouldn’t work, or Netflix, or Amazon, or Uber, or PayPal, etc…. There are always these business complexities that make it seem like these spaces have too much friction, but if there’s enough money - if it can be done then people will figure it out.
> E-commerce has been optimized to the last decimal point for the last 30 years.
It certainly hasn't been optimized to anything in 1996. In 1996 it was people clumsily scanning print catalogs, spending 5 hours to upload 10 images on dialup and making a simple HTML page (no DB or any kind of backend) and putting their landline phone on it with a message to "call to checkout"
I know you were exaggerating for effect, but E-commerce and catalog normalization are definitely not "solved" everywhere.
McMaster Carr is a good example of a company that has 90%+ of their stuff ironed out, but most websites and especially small ecommerce isn't like that.
> OpenAI is new to it, and if I had to guess, not that interested in getting good at it.
Maybe if they burn more tokens the answer will become clear
Already your favourite e commerce site has all your data. You can switch on the "buy this automatically" feature.
> E-commerce has been optimized to the last decimal point for the last 30 years
Sort of, but there's a ton of middlemen between "resources in the ground" and "product in my hand". For example, how much utility is there in a "store" to thee consumer at this point? Let me buy from the manufacturer.
I am behind schedule on developing a "summer phase" [1] for my foxographer costume and was chatting with Gemini about a crash priority "spring phase" [2] and asked it for suggestions and it gave me a 10-pack of results that had one good thing in it at rank #8, a similar query run against a normal search engine actually got something better at #1. Now sure I am talking w/ Gemini with big words like "supergraphic" whereas a normal search would be heavy on 3-letter and 5-letter words used in the product descriptions.
It makes think though of expert system based product configurators back in the 1980s
https://en.wikipedia.org/wiki/Xcon
thing is that kind of product configurator is based on an ontology, constraints and rules as opposed to embeddings which might capture the "feel" of things like clothing.
[1] Busytown meets Arknights
[2] supergraphic shirt + camera gets resonance with my promotional system and people keep approaching me (e.g. laugh but every KPI in the system has an extra zero on the left)
FWIW OpenAI is desperately trying to monetize and they think e-commerce is a "simple" problem to solve. I mean they do need to convert their funnel without alienating their users. I assume they are going to have some big payouts for agentic purchases gone awry or leave merchants on the hook.
Why would anyone have an extra layer of friction too where things could go wrong, where handing over payment details in another chain.
Just let me buy my stuff in peace. Shopping is not the 'killer app' for GenAI.
A chat interface is just fundamentally incompatible with this. The agent makes it too easy to ask questions and comparison shop.
You won't get it to push your products when users ask what's the best XYZ - either because it'll be too honest to lie or because it'll be too expensive for you.
There’s a lot of this going on in AI at the moment. New folks come in thinking they have a magic solution and then produce a total train wreck as it turns out domain expertise is still a thing.
Walmart does not, over 10 years after they were released, even accept the contactless payment systems in common use. Instead, they push their in-house version in part so they can capture the relevant customer data.
And we're meant to believe that Walmart planned to outsource the entire series of touchpoints represented by the discovery & checkout process? Yeah, okay.
This was never going to be more than an experiment for Walmart.
I’m currently using Gemini to research components for a remote controlled plane. I have the frame of the plane and now need to buy correctly specced servo motors, an engine, battery, etc etc. It has saved me so much time and educated me tremendously on how the different components interact and the options available.
If I could just press “buy” from within Gemini and pay via Google Pay (or better still, Apple Pay) I’d do it in a heartbeat.
If ChatGPT can do this today, I need to try it.
"I need mayo, ketchup, mustard and ground beef"
"Here is a list of products with prices ... proceed to pay $25 (yes/no)" Yes
"Your card has been charged. Delivery will knock on your door in 7 minutes"
I'll code that app in one month, what's there to lose?
The latest AI is trained on the average citizens social media output. Iq 90.
That’s why AI seemed smart. The bar will not be raised again. We’re cooked.
I find it kind of fascinating seeing people now debate the ins and outs of just exactly how useful the ads/affiliate schemes inside the chatbot are.
Pretty much in the same way that I'm fascinated about people debating the quality of ads shown in their browser while I've been using ad blockers for forever now and cannot imagine seeing the web without them.
The better comparison might be conversion rate for those who searched on Walmart.com vs those who searched within ChatGPT. Or maybe that is what they're comparing and I misunderstood?
If you want to buy a Walmart product, the easiest way is to go to Walmart. Why add an imprecise middle man in between?