Search is changing, transactions are changing, attribution is changing—AI is rewriting the foundational logic of e-commerce. The relationship between consumers and platforms has been completely transformed, shifting from “people searching for goods” to “agents understanding, recommending, comparing prices, and placing orders for you.” Over the past two decades, the three pillars of internet commerce have been: advertising, subscriptions, and e-commerce. Advertising once served as the sturdiest pivot—whether it was Google’s AdWords, Meta’s news feed, or ByteDance’s recommendation algorithm, they were essentially attention-based businesses revolving around “precision targeting.”
However, in the age of AI, the logic of advertising is faltering. Users no longer need to be “fed” content and products; instead, they directly communicate their needs to intelligent agents. This means the traditional “advertisement → click → purchase” loop, which the internet has relied on for survival, could be disrupted. AI must find new monetization paths, and e-commerce—particularly e-commerce where AI participates in decision-making—has become the new critical entry point.
In this discussion, a16z partner Alex Rampell and investor Justine Moore pinpointed the crux: what AI is disrupting is not impulse buying but “research-driven purchases.” For mid-to-high-priced items like computers, travel bags, or bicycles, AI can read thousands of reviews, summarize video content, integrate price trends, and automatically place orders for you at the right moment. It fully automates “consumer research,” leaving “advertising reach” with nowhere to gain a foothold. When the intelligent agent itself becomes the most trusted “shopping assistant,” the marketing logic between brands and traffic platforms loses its grip.
Simultaneously, the old system of affiliate marketing—pixel tracking—last-click attribution is collapsing. In the past, we could trace a transaction back to its source channel, specific ad, or user click. But when AI makes purchasing decisions on behalf of people, this path becomes blurred or even invisible. The failure of attribution signifies the disintegration of the traditional advertising model—because advertising relies on “seeing causality,” while the behavior of intelligent agents more closely resembles a black box. In the future, traffic may no longer be priced based on “exposure” but rather on “closed-loop transactions.”
Another insight Rampell raised is even more enlightening: AI does not create demand; it merely executes demand with extreme efficiency. Humans are still driven by culture and social interactions to “get interested” in products, but when it comes to the final step of “making the purchase”—selecting the brand, calculating discounts, deciding which credit card to use for cashback—AI will become the smartest intermediary. This “executive agency” blurs the line between advertising and e-commerce. Advertising may no longer exist in the form of displays but rather as direct transaction commissions.
On the other end, the Costco-style membership trust model has become one of the few moats that cannot be replaced by algorithms. As e-commerce becomes polluted with affiliate kickbacks and fake reviews, and “top ten recommendations” devolve into SEO garbage, users are once again seeking trust-based relationships centered on “helping me choose rather than tricking me into buying.” Rampell describes this trust as sacred—it cannot be acquired through algorithmic bidding nor replicated by large language models.
Ultimately, AI may deliver the “final blow” in the e-commerce world—it doesn’t generate traffic, but it can decide transactions. When consumer research, price comparison, and execution are fully automated, a platform’s moat will no longer be its advertising algorithm but its ability to accommodate the transactional logic of intelligent agents. Whoever masters the “research–compare–execute” closed loop will hold the key to monetizing the next generation of the internet.
I. AI Will Rewrite the Rules of Traffic and Transactions in E-Commerce
Erik Torenberg: You’ve both been focused on the e-commerce space. Alex, can you talk about where this idea came from and how it started?
Alex Rampell: I founded a company called Trial Pay a long time ago, and I was actually trying to sell things online even before the internet became widespread. I was pondering a few questions back then.
First: What will happen to Google? Many people are paying attention to changes in search volume—is it increasing or decreasing?
Personally, I am using Google less for search, though this decline isn’t specifically in shopping scenarios but rather in other non-shopping aspects. Second, Trial Pay, which I founded, was one of the largest affiliate marketing platforms globally at the time. The essence of affiliate marketing is that if you refer a user and facilitate a transaction, you earn a commission. This was actually one of the earliest business models on the internet, even predating AdWords and AdSense.
It’s said to have originated in the adult content industry because it was one of the first business models to emerge on the internet. Its implementation relies on cookies and pixel tracking: a cookie is placed in the user’s browser upon their visit, and when an order is completed, a hidden 1×1 pixel image on the confirmation page reads the cookie to confirm, “This user was referred by Eric.”
Trial Pay grew large based on this model, but the question is: Can this mechanism support the future e-commerce ecosystem? Does it still make sense for many scenarios, like impulse purchases? Impulse buys account for a significant portion of e-commerce, and impulse purchases themselves hardly rely on AI. When you grab a Coke at the supermarket checkout on a whim, that behavior won’t be driven by an AI recommendation.
In fact, the Coke bought at the checkout often costs more than those on the shelves—retailers design this psychological trick to make you spend more. Such consumption cannot be driven by AI. However, another category is high-value goods. People spend a lot of time researching these, and AI can play a crucial role in this research and price comparison. But the question here is: Without affiliate links, how will e-commerce transactions be completed? So, the first point I was considering was: What kind of system do different types of shopping behaviors constitute?
The second point was: Does affiliate marketing still make sense in an AI-driven commerce landscape?
The third point is my personal behavioral change: I now use ChatGPT about a thousand times more often than Google, which is interesting in itself.
Erik Torenberg: Justine, why were you interested in this topic? What made you think it was most worth discussing?
Justine Moore: For me, there are several major consumer markets, and one of the largest is probably online shopping. However, so far, there haven’t been many startups truly attempting to enter this market with AI. Although, as Alex said, the opportunity is substantial because we now have powerful large language models and agents that can help consumers make better purchasing decisions than they could themselves, or even place orders directly on their behalf. Logically, this should attract more people to package them into consumer-facing products, but we haven’t seen many doing that yet.
So, part of the purpose of writing our article was to delve into: Why is this system so complex? In which different shopping scenarios can AI play a role? What changes do we hope to see against the backdrop of this larger market? We also hope that by doing this, people exploring this field can share their ideas and approaches with us.
Alex Rampell: Personally, I prefer to “observe” first because observation is objective, and only then “predict.” Predicting the future is very difficult—there’s a famous quote about prediction I can’t recall, but the core idea is: Predicting the future is extremely hard. But actually, there are already many observable behavioral patterns. For example, CamelCamelCamel is one of the best websites in the world (we have no relationship with them, this is not an advertisement). It’s essentially like a price version of Google News Alerts.
I once gave a talk to the Amazon Prime team, and they were very aware of this website because it might be Amazon’s largest source of affiliate marketing. Users say every day: If this product drops to this price, I will buy it. It’s currently at this price, please notify me when it drops to my desired price. And when the price actually hits the target, users do buy immediately. This shows that consumers are already acting as “Agents,” but this Agent is very inefficient.
If this process could be closed-loop—not just notifying about price changes but directly completing the purchase for the user—users would accept it. Because we’ve observed this behavior exists; they just lack a more convenient way to execute it. This is the simplest way to “predict the future”: You’re just recording the current situation and then adding a natural extension.
Justine Moore: Another example I’ve observed comes from some viral use cases. Some are very successful because AI perfectly found the product; others are hilarious because AI completely failed. For instance, some teenagers use ChatGPT, uploading photos from Lana Del Rey concerts or Taylor Swift street snaps, and ask, “What brand is this hair clip she’s wearing?” or “Where can I buy this sweater she’s wearing?”
When AI gets it right, the effect is excellent. For example, it might prompt: “This sweater costs $5,000, and as a 19-year-old from Missouri, you probably can’t afford it, but here are some similar, more affordable alternatives you can buy.” Users in this age group are often early indicators of consumer behavior trends. So, this actually hints that the process from product research to final purchase is increasingly likely to be completed by agents in the future, especially when the price is right, as Alex just mentioned.
Erik Torenberg: Alex, can you imagine an extreme scenario of “dynamic customized pricing”? For example, we look at the same product on Amazon, but it shows you a higher price than me, perhaps because I’m more “stingy” than you, or because you’re wealthier. Do you think such a world will emerge?
Alex Rampell: Many have already tried similar approaches. From an Economics 101 perspective, it’s actually a smart logic: Extract as much “consumer surplus” as possible. Consumer surplus benefits buyers but represents a loss for sellers. I’ve heard Delta Airlines is trying something like this.
Actually, there are some “crude versions” of dynamic pricing. For example, if you browse with an iPhone, you should pay a higher price than someone using an Android phone, because iPhones are more expensive, implying your demand elasticity is different from a user with less money. However, this will likely encounter regulatory issues and, more probably, trigger significant consumer backlash. Although some have tried, it’s usually hard to implement effectively.
II. The E-Commerce Attribution Problem: What Exactly Causes the “Final Blow”?
Erik Torenberg: If we look back at past technological platform shifts, we can see similar patterns. E-commerce now accounts for about 16% of total retail sales. But if you asked people 20 years ago, “What will be the future share of e-commerce?” everyone would probably estimate higher. Why hasn’t reality reached that level?
Alex Rampell: The reason lies in the significant difference in the demand curve between “immediacy” and “non-immediacy.”
For example, “next-day delivery” is already cool, but it can never match the demand intensity of “instant purchase.” Like when I’m going to bed at night and find I’m out of toothpaste, I’ll immediately go to the nearby Walgreens to buy some. Amazon is great, but if it delivers at 7 AM the next morning, it’s useless to me. There’s a market curve for “real-time toothpaste demand.”
Another scenario is “browsing”—I’m bored today, so I go mall shopping. This is an experience in itself. Then there are “long-cycle, deliberative purchases,” even with a bit of a “wish list” nature, like eyeing a Rolex for a long time and buying it impulsively after getting a bonus. The shopping process itself is part of the experience.
I serve on the board of a company called Wise, which deals with cross-border remittances. They found that the market demand for “real-time arrival” is completely different from “arrival in two days.” The demand for the latter is much smaller. The development of Amazon’s logistics tells a similar story. In the early days of e-commerce, items might take two weeks to arrive. Looking at the demand curve today, it’s still expanding, so the increase in e-commerce share is reasonable.
Justine Moore: You said e-commerce only accounts for 16%; that number seems low. I’m not questioning the data, but many user behaviors aren’t captured. For example: People often research online but complete purchases offline. When buying big-ticket items, I might read various reviews on Reddit, Instagram, or Apple’s website, but ultimately go to a physical store to personally compare the weight difference between a MacBook Pro and a MacBook Air.
Another example is buying clothes. In San Francisco, many people order a bunch of clothes online, try them on, and return half, because there aren’t many large department stores locally. But in my hometown in Oregon, there’s almost no need for that because you can drive 5–10 minutes to various clothing stores, making the cost of returns much lower. But indeed, many people research online first—which store to go to, what specific product to buy, what style to look for. So, while only about 16% of transactions currently occur entirely online, the internet plays a research role even in many offline purchases.
Alex Rampell: This leads to the trickiest problem: attribution. Attribution is almost everyone’s nightmare. For example: How should I assign attribution for Justine’s MacBook purchase? The most corrosive business model on the internet is so-called “last-click attribution.” It assigns 100% of the credit to the last-click channel. But the reality might be: Partly inspired by a post I saw on Reddit, partly by a Super Bowl ad I saw—these factors collectively drove the purchase.
A more reasonable approach would be “fractional attribution,” but it’s not entirely deterministic. The seemingly “deterministic” method—treating the last-click channel as the sole driver—is actually wrong. Many people fall into this trap without distinguishing between correlation and causation. The business model I hate the most is this kind, like Honey.
It pops up when you’re already on the checkout page, asking if you want a discount code. Who would refuse? Click, get 10% off. Its operational logic is: redirect you to an affiliate link page, place a cookie on your computer, then send you back to the original page. This way, it “steals” the transaction attribution.
Ironically, many marketers at large e-commerce companies actually fall for it. They say, “Oh, our best channel is Honey, it’s growing so fast!” There’s also RetailMeNot from earlier times, which even went public with a high valuation, but essentially it was “stealing attribution.” The real reason remains: attribution is too hard to disentangle.
In the future AI world, this will only get more complex. For example, Justine researched on Reddit, saw a Super Bowl ad, asked a question on ChatGPT, and finally clicked “buy.” If Apple attributes all the credit to ChatGPT, saying “it drove this transaction,” that would be wrong. It’s just one link, not the sole driver. How to decompose and assign attribution will become increasingly difficult.
III. Aggregator Platforms Like Google and Facebook Are the Winners in the E-Commerce Race
Erik Torenberg: Let’s return to the broader industry landscape. So far, the biggest winners have been aggregator platforms like Shopify and Amazon, along with a few individual brands like Allbirds and Casper. These companies quickly achieved large-scale revenue but failed to grow into long-term sustainable businesses; they didn’t become stronger as they scaled. Why?
Alex Rampell: Because if the transaction is a “one-off deal,” you’re not actually manufacturing the product. Take Casper, for example; they didn’t produce their own mattresses, probably just sourced from Chinese OEM factories and slapped the Casper logo on them. Their model essentially involved buying traffic from Google and Facebook. So the real winners were Google and Facebook, not Casper itself. When people saw Casper’s success, they’d say, “Wow, mattresses are a good category, I should do that too.” Then they’d go to Shenzhen to find factories, put their own logo on products, and undercut Casper on price. This was bound to happen.
One solution to this problem is bundling with a “subscription model.” Like Dropcam. It’s an e-commerce product but paired with a subscription service. Although there are now billions of cameras on the market capable of the same things, and competition is fiercer, because Google acquired Dropcam and integrated it into Nest, it still makes decent money in this category.
In contrast, Casper has it tough. If I bought a Casper mattress five years ago and am still using it, they must constantly seek new customers and cannot generate recurring revenue from existing ones. Meanwhile, the original factory that produced for Casper has now sold the same mattress to about 5,000 other brands. This is clearly not a good business model.
Overall, if you’re just a reseller of commodities, it’s a big problem. Many would say, “Casper has its own mattress, Allbirds has its own shoes.” But the fact is, they usually don’t actually manufacture these goods.
This is similar to what happened during the Internet 1.0 era: long-tail commodity retailers gradually disappeared because the “advantage of geographic location” no longer mattered. In the past, retail was driven by, say, Justinestown (a place in Oregon) having only one store, so of course you’d shop there. You could drive to another town, but it was too far. But on the internet, you can buy from any store anytime. So if 5,000 stores don’t manufacture shoes themselves but all sell the same Nike shoes, why not go directly to Nike’s website or a store with faster delivery and better service? Thus, the long-tail retailer model gradually died out, a trend we’ve already seen.
But actually, the direct-to-consumer e-commerce experience isn’t much better. Because there are almost no barriers to entry, no barriers mean extremely fierce competition. For consumers, this is a benefit of capitalism; but for those “any one of thousands of sellers,” it’s completely bad news.
Justine Moore: Especially consumer goods, like shoes (e.g., Allbirds) or cosmetics, are highly dependent on “trends.” Mattresses might be more like utilitarian goods, but shoes or cosmetics are entirely trend-driven. And on the internet, trends never last. For example, Allbirds was once hugely popular, but soon everyone was into retro Adidas, then this year in TikTok’s BamaRush videos, every girl was wearing On Running shoes; last year they were wearing Japanese-style New Balance.
This is a huge problem for Allbirds because it can’t capture all trends. Its SKUs might only cover one or a few styles. Whereas platforms like Shopify or Amazon can easily accommodate any trend and allocate demand to corresponding SKUs.
This will be even more challenging in the AI era because AI agents can directly guide users to make purchases (if users start the buying process through them). This could be both an opportunity and a risk for single-SKU retailers. But my intuition is that aggregator platforms will ultimately benefit.
Alex Rampell: Additionally, AI finds it hard to “create demand.” For example, why do I know On shoes are cool? Because I saw them in Bama Rush videos on TikTok, every girl had a pair. So “I should buy a pair too, I’m a sorority member, I need to have them.” This demand is formed through social and cultural dissemination, which AI struggles to achieve. AI is better suited for scenarios where “I already know what I want to buy,” like helping me buy those shoes. This is actually similar to Google’s positioning.
I greatly respect Google, but it’s essentially a “GDP tax.” A large portion of GDP comes from consumer spending, and much consumption activity starts from that little Google search box. It takes a cut from every consumption activity through pay-per-click, pay-per-impression, or pay-per-action charges. However, this “taxation” model might be threatened in the future; the tax might shift elsewhere.
IV. AI Will Erode Free Search, But Google’s Commercial Search Remains Unshaken
Erik Torenberg: Let’s delve into this: What will be taken away from Google? What will remain? Also, the relationship between different types of consumer spending and e-commerce. Maybe start with you?
Alex Rampell: Google has always been the “classic premium business model.” They built a better search engine. Everyone knows it started in 1998, probably the 47th search engine on the market at the time, and many thought “it couldn’t succeed.” But it was much better because its linking approach was different, essentially similar to the h-index in academic research but applied to web search. For example, if you search for “bagel,” and everyone links to the same website, its PageRank would be high, and Google would rank it first.
When Google first appeared, the internet wasn’t highly commercialized; most searches were free and non-commercial. Back then, using Google felt much better than Hotbot or other search engines. Later, they basically adopted the business model of Overture, a company founded by Bill Gross at Idealab, which was later acquired by Yahoo. This is why Yahoo held Google stock early on. If you know this history, you’ll understand that the key to Google’s rise was AdWords.
Interestingly, many “premium models” make users think: I don’t want to pay. But in Google’s model, ads not only didn’t degrade the experience but made search better. For example, if I search for “tennis racket,” and no one has done SEO for that term, the results might be poor. But advertisers appear alongside the results, and these ads only show as “relevant” if users click. If no one clicks, they’re irrelevant. This combination of “organic ranking + ads” makes search results higher quality.
So Google was premium from the start and remains so today. Often, what we search for isn’t related to purchasing, like just looking up information—this is the default behavior: if I want to know something, I Google it. Sometimes it’s not even directly on Google but through Safari, because Apple receives tens of billions from Google annually just for directing traffic there.
The change happening now is: Google is losing some free search, but its premium part remains. For example, “Who won the Oscar in 1977?”—that’s a non-monetizable search, people are more inclined to ask ChatGPT directly. And this is already happening. ChatGPT has about 800 million weekly active users, a huge number. People use it for information but haven’t started shopping directly on it yet. We know OpenAI is trying e-commerce, but clearly, it’s not built yet.
The premium part, the searches that actually make money, still happen on Google. We can see this from Google’s financial reports: revenue is still growing, but search volume is actually declining. So what Google is truly losing is not revenue but some non-commercial “free search.”
Maybe they’re directing some traffic to Gemini, but unlikely. The current situation seems more like: People still conduct paid searches on Google, business as usual; but for the free part, they’re turning to AI.
V. Fake Recommendations Weaken Trust; AI Must Solve Authenticity to Enter E-Commerce
Justine Moore: Like all large language models, ChatGPT has a major problem with “hallucinations” in product recommendations. Almost everyone who has tried it has encountered this.
For example, you want to buy leggings. Searching on Google or Amazon, you find the highest-ranked results. But what you really want is personalized recommendations combining your specific needs (like what type of hiking you’ll do, the weather), not the “overall best.” Many people, especially young women, think: I’ll ask ChatGPT, it understands my natural language input, it’ll recommend suitable products. But often, many recommended products simply don’t exist, or existed but are discontinued, or the price information is completely wrong. This makes many people, after one or two tries, return to Google or Amazon.
Once ChatGPT truly solves the e-commerce problem, people will continue trying in the future because OpenAI is indeed working on e-commerce integration, aiming to make recommendations more relevant and data more real-time. By then, Google might lose some search traffic. But I completely agree with Alex: We haven’t seen this behavior at scale yet.
Alex Rampell: However, zooming out to the entire internet, the biggest trouble now is: the internet itself is “unhealthy.” I spoke with John Lilly (he was the CEO of Firefox back then), who is somewhat a guardian of the early internet. His view is: The internet was open in the past, originating from DARPANET → ARPANET → Internet, mainly used by researchers initially, everything was on the open web, no so-called walls.
Today, search has long been fragmented, not starting with ChatGPT. For example: For real-time search, go to Twitter/X; to check on friends, go to Facebook. Google can’t search these; they’re locked within their respective platforms. This is the first aspect of “unhealthiness.”
The second unhealthy aspect is commercialization. Commercialization itself isn’t wrong; I like capitalism too. But today, much content is polluted by commerce. For example, you search for “best running shoes.” In 1995, someone might have blogged about their experience, hosting their own server, just out of interest. Later, affiliate marketing emerged, originally a monetization method but gradually polluted the open internet. So “top ten recommendations” are everywhere, essentially “top ten affiliate kickbacks,” articles are outsourced, then疯狂SEO to make money.
In contrast, the traditional Consumer Reports model was completely different. They refused ads, relied entirely on subscription fees, so users could truly trust the review results. For example, they’d write: This blender is dangerous, cuts hands, don’t buy it; that one is reliable, recommended.
This model has almost disappeared today. The rise of Craigslist almost killed traditional media. Newspapers used to make money mainly from two sources: ads and local classifieds. Craigslist directly cut off the latter, so newspapers declined. Newspapers occasionally did some “public service” reviews, like comparing various blenders, ensuring they didn’t recommend dangerous products. These things are mostly gone now.
So, the problem now is: The share of the open internet is shrinking, much content is behind walls, and the remaining open part is flooded with spam. Hence we often say: You can’t treat sponsored content as genuine reviews. This is the difficulty. Even if AI no longer “hallucinates,” even if its summarization is strong, if the underlying content is garbage, the results won’t be good. SEO-optimized garbage articles filled with affiliate kickbacks, and summarizing this garbage doesn’t turn it into valuable information. So, the question is: How to “de-spam”? This is actually very difficult.
Justine Moore: Honestly, the channel least polluted by spam currently is video. The reason: With the decline of traditional media, many creators now review products themselves, like testing 10 pairs of running shoes, and clearly state in videos if it’s brand-sponsored. Better creators take no sponsorship at all, relying only on YouTube or Google ad revenue. So, when I want to see “someone’s real experience with five different hair dryers,” especially for a specific hair type, I watch non-sponsored videos. Such videos often have high view counts because many people have similar questions.
But the problem is, Google doesn’t utilize this content well. Because videos aren’t easy to “skim,” and they haven’t automatically transcribed all videos into text, this information doesn’t appear in traditional search results. Some companies are now trying to convert quality videos into text, letting LLMs analyze and give recommendations, but this hasn’t truly entered Google’s mainstream search system yet.
Erik Torenberg: For example, The New York Times’ acquisition of Wirecutter is an example.
Alex Rampell: But there’s a problem here too: Wirecutter’s content almost always includes affiliate links. Isn’t that a bit odd? Does it mean the recommendations themselves are biased? I’m quite skeptical of this model. In contrast, the reviews from the Consumer Reports era were more trustworthy; although they might have had subjective biases, like an editor just disliking a certain company, they were independent in principle.
You might think, with algorithms and various technologies now, if we could get truly objective feedback, that would be great, right? Amazon itself is a huge search engine, but unfortunately, this platform is also heavily polluted. For instance, many sellers source 400 small gadgets from AliExpress at $2 each, delivered in six weeks. Then they put their logo on them and sell them on Amazon for $25. This goes back to the delay issue we discussed earlier: How many people are willing to wait 6 weeks for delivery? How many just want “delivery tomorrow”? Amazon essentially arbitrages this time difference.
But once you search on Amazon, especially for consumer electronics, the results are chaotic. I once wanted to buy heated socks for skiing, searched, and found over 9,000 seemingly different brands, almost all from the same OEM factory. Worse, they abuse fake reviews.
For example, a seller first sells stones, accumulates perfect reviews, then changes the SKU from “stones” to “heated socks,” continuing to use these five-star reviews to sell the new product. Amazon fully knows about this behavior but has no incentive to fix it because they just want to sell more. Honestly, if you’re willing to wait, buying directly from AliExpress is cheaper than Amazon. But overall, Amazon has become a sea of garbage.
In contrast, my favorite business model is Costco. Costco is the best company in the world. It refuses to sell inferior products and refuses to charge high margins. Why would they give up high margins? It sounds strange in business logic, right? The answer: Because it damages the value of the membership system. Costco’s profit doesn’t mainly come from goods but from membership fees. They charge about $100 annual fee, with over 50 million members globally. Looking at their financial reports, their net profit is almost membership count × membership fee, with the rest basically breaking even.
So, if a piece of clothing has over 50% margin, Costco would say: Too high, cut it. Because that would reduce member trust. They even insist on keeping the hot dog at $1.50 and built their own chicken farm to control the cost of rotisserie chicken.
Their private label Kirkland is also strong: Kirkland wine, beer, shirts are all high quality. They even got sued for making athletic pants better than Lululemon. This is Costco’s uniqueness: It curates products strictly like Consumer Reports but simultaneously has massive scale and extreme user trust.
Justine Moore: Users really trust Costco. Take my mom, for example; she’s been a Costco member forever. Now she even uses Costco for glasses, booking flights, because she firmly believes Costco offers the best price and quality. And事实证明, she’s right most of the time.
Alex Rampell: This “member trust” is almost sacred and inviolable. They could definitely make more money by loosening standards, but they refuse. This contrasts sharply with Amazon’s logic.
Jeff Bezos once said: There are two business models:
First, make the most money possible—typical of Apple, e.g., an iPhone sells for $1600, can it sell for $1700? They pursue extremely high margins.
Second, make the least money possible—typical of Amazon, they offer a huge array of goods, regardless of quality, let consumers filter themselves, make money through scale.
Costco is the third, very unique: Through decades of trust accumulation, it achieves “you just trust me.” As Justine’s mom said: If Costco sells it, it must be good. This model is hard to replicate and hard for AI to break. Because its core isn’t traffic or recommendation algorithms but long-term brand trust.
If you were Costco’s CEO, would you leverage this trust to expand into other businesses? It would endanger the entire company’s foundation. However, they do have room for expansion. One of our partners met with Costco’s board and proposed they develop financial services. Because all banks try to make more money from customers, like raising loan rates or lowering deposit interest. Costco could offer the cheapest loans because they don’t intend to profit from them but from membership fees. So, they indeed have significant potential to expand this business.
VI. Standardized Product Transactions Will Be Fully Automated by AI; Non-Standard Goods Still Rely on Human Experience
Erik Torenberg: Justine, let’s talk about how AI will change e-commerce? You mentioned several different purchase types earlier.
Justine Moore: We roughly divide purchasing behavior into a spectrum: At one end is impulse buying, like Coke on supermarket shelves, or now many people see a video on TikTok Shop and buy a T-shirt directly because it looks cool; at the other end is major purchases, like buying a house, booking a wedding venue, or buying a car, these consume large portions of income, are one-time or few-time消费, and people do a lot of research.
Both ends are not easily disrupted by AI. Impulse buying has little pre-research; you decide the moment you see it. Future algorithms will indeed become more precise, like pushing a T-shirt with your dog’s name, making you more likely to buy than other clothes. But this isn’t the “generative AI” scenario we’re discussing. As for major purchases, you might use ChatGPT, Gemini, or other AI tools for research first, but ultimately, because the amount is too large, you’d still want physical experience: see, touch, talk face-to-face with experts before deciding.
So, what AI might truly disrupt is that large middle category of goods. For example: My usual travel bag broke, I want a new one, must fit a laptop, large water bottle, and fit in overhead bins. Doing research myself takes time, but if an AI assistant can watch TikTok videos, read Reddit posts, aggregate real user feedback, then recommend to me, that’s very valuable. Maybe I’ll still click to see different options, but if the shopping process can seamlessly connect with AI, I’d likely order directly through it.
Another scenario is you already know what to buy, like consistently buying a certain brand of laundry detergent. The AI assistant can handle price optimization for you: scan全网 prices daily, if a website suddenly offers 30% cheaper than usual and can deliver in reasonable time, it directly buys an extra box for you.
Further up, are higher-value but not yet “major purchase” goods, like bicycles, sofas, laptops. You’ll use these for years, must ensure they fit and won’t become obsolete quickly. People now often dig deep into Reddit’s “Buy It for Life” forum, or directly trust brands like Apple, willing to pay a premium for reliability.
In the future, I can imagine a deeply understanding AI assistant, you could even call it, dynamically answer its questions, then it researches based on your needs and finally helps you make the optimal decision.
Alex Rampell: Another angle is whether the thing you’re buying has a UPC. UPC is the barcode, equivalent to ISBN for books. If there’s a UPC, you can simply run a “find the lowest price” algorithm. In the pre-AI past, people could do this themselves, often ending up on Amazon. But without a UPC, like why Wayfair succeeded and sells well? Because many goods they sell lack unique codes, like bar stools. You just say “I want a bar stool,” but there’s no UPC to uniquely identify it. So Wayfair can offer different sizes, styles.
Thus, back then, almost all players except Amazon were eliminated, and Amazon survived and grew. Of course, I’m simplifying. If a product has a UPC code, then AI will exponentially enhance the “lowest price algorithm.” In the past, some people valued time more, others money more. If I valued money more, I’d be the “algorithm” myself: find the best coupons, best cashback sites. Such cashback sites abound on the internet, suitable for those who value money over time. In the future, AI will fully automate this for consumers, but前提是你确定了SKU或UPC.
Without a UPC, it’s another story. Like a bicycle, you might not know which model to buy, but once determined to be a specific professional model, it has a UPC, can be handed to AI, which picks the optimal price, shipping, terms. This is done manually now, will be automated in the future.
Erik Torenberg: Combining these, over the past decade, there have been almost no entirely new big winners; gains were mostly taken by “aggregator platforms.” Why do we think the next decade will present new opportunities, capable of birthing new, sustainable giant companies?
Alex Rampell: For example, ChatGPT is a new player, not Amazon or Shopify, but it will certainly play a role in e-commerce. The question is whether specialized companies in niche segments will emerge?
Yes. Like independent sites such as CamelCamelCamel, which took no venture capital, might be highly profitable. Also cashback sites, like Ebates in the US (later acquired by Rakuten), Quidco in the UK. Their logic is “you value saving money over saving time.” Such shopping Agents will become more mainstream in the future, especially in areas not involving complex research, and the scale could be very large.
Returning to the attribution problem, most companies can’t figure out ad attribution, but in the future, AI will become the 21st century’s “final blow”—the crucial step that ultimately facilitates the purchase. It might not be a general AI like ChatGPT but vertical shopping Agents: you give it all your credit card info, it helps choose which card offers the highest cashback, integrates cashback, coupons, rebates, and finally places orders for you.
This might not be good for merchants, but it’s not hard to imagine. Previously, these services were niche, only attracting more tech-savvy users willing to spend time saving money. Like my mom, she’s retired, certainly willing to spend time saving money, but the tools are too complex for her. If AI makes this extremely simple, it’s almost an “IQ test”: Do you want to spend more or less? Of course, everyone wants to spend less, but if it requires 18 steps, downloading a bunch of plugins, most people give up.
So this opportunity likely goes to startups, not Amazon. Because Amazon wants
