Summary

Google’s commercial-search position remains intact for now, while AI is capturing the free queries that support its broader habit loop and could eventually reroute its “tax on GDP.” Alex Rampell sees search volume falling for informational queries while Google’s financials still rise, implying users are moving non-monetizable questions—not monetizable shopping—to ChatGPT. The risk arrives when agents own purchase initiation and the “tax might just shift elsewhere.”
AI’s commerce sweet spot is the broad middle between unplanned impulse buys and large, highly considered purchases that often still demand physical experience or human reassurance. A TikTok-triggered shirt requires no research, while a house, car, or wedding venue is unlikely to become fully agentic. Handbags, detergent, bikes, couches, and laptops offer the stronger opening: better research, continuous price scanning, and eventually execution.
Once a buyer has selected a SKU or UPC, an agent could automate the entire money-versus-time optimization problem. It can search prices, delivery terms, coupons, cashback programs, affiliate rebates, and even the best credit card, then buy when a threshold is met—CamelCamelCamel with the action loop closed. Rampell’s test becomes an “IQ test”: “Do you want to pay less for something or more?”
Attribution—not product discovery—is the load-bearing commercial problem, and AI may make today’s bad incentives worse. Last-click systems already let coupon extensions such as Honey intercept customers at checkout and claim credit for sales they did not cause. Agents could become “the last click of the 21st century,” capturing merchant economics despite being only one influence among Reddit, advertising, creators, stores, and prior brand affinity.
The web’s degraded information supply is a structural constraint on AI shopping, because models cannot turn affiliate-optimized inputs into objective advice. Search spans walled gardens while the open web is saturated with “SEO-optimized crap”; Amazon listings and reviews are similarly gameable. The unsolved question is stark: “You can’t turn shill junk into honest analysis,” so how do platforms “decrapify” the corpus?
AI likely strengthens aggregators while exposing undifferentiated direct-to-consumer brands whose products are made elsewhere and whose traffic must be bought. Commodity sellers such as mattress brands can be copied by the same OEMs and must repeatedly acquire customers from Google or Facebook; fashion brands also cannot own every trend. Moore argues that agents can direct buyers once demand starts there, but she and Rampell note that AI may struggle to inculcate demand before culture makes an item desirable.
The durable opportunities sit in trusted curation, specialized buying agents, and merchant infrastructure—not merely another horizontal chatbot. Costco’s membership-funded refusal to sell bad products makes it unusually “AI-proof,” while startups can build domain experts, agent-readable storefronts, and payment infrastructure for delegated purchasing. Amazon’s high-margin advertising is exposed if AI takes control of the presentation layer before shoppers reach Amazon.

Deep dive

1. Existing behavior already sketches the agentic-shopping future

Rampell’s starting point is both personal and historical: he uses ChatGPT “three orders of magnitude more” than Google, and says his search volume is falling for non-commerce queries rather than commerce. His former company TrialPay was a major affiliate, operating through cookies and invisible 1-by-1 tracking pixels; the open question is whether that primitive attribution system can plausibly power agentic purchasing.
Justine Moore sees online shopping as one of consumer technology’s largest markets, but relatively few AI startups have attacked it. Smart LLMs can theoretically improve decisions or purchase autonomously, so the episode’s inquiry is why commerce remains complex, which purchases are susceptible, and where new products can realistically enter.
Rampell prefers to “observe first” and points to CamelCamelCamel: consumers already specify that they will buy an Amazon item when its price reaches a target. He believes it may be Amazon’s biggest affiliate, though he has no stake in it. Turning its alert into automatic execution is merely “one additional appendix to the present,” because the consumer has already declared both intent and condition.
Moore’s matching observation comes from teenagers uploading celebrity photos to ChatGPT and asking what hair braid or sweater the person is wearing. At its best, the system identifies the $5,000 original and offers attainable lookalikes to a “19-year-old girl in Missouri”; that demographic’s behavior suggests visual research will expand into purchasing.

2. Physical immediacy and experience explain e-commerce’s incomplete takeover

The host notes that e-commerce represents 16% of retail sales, far below what an observer 20 years ago might have predicted. Rampell’s explanation is a distinct demand curve for immediacy: toothpaste delivered at 7 a.m. is not a substitute when someone needs to brush their teeth before bed and Walgreens is nearby.
His parallel from Wise is that money received immediately serves a materially different market from money received two days later. E-commerce has expanded as delivery fell from roughly two weeks toward overnight, but it still cannot satisfy every “real-time toothpaste” need.
Retail also sells an outing, not just inventory: a bored consumer visits the mall, while an aspirational buyer repeatedly looks at a Rolex before a bonus turns browsing into purchase. Those experiential and emotional functions can resist complete migration into a transaction engine.
Moore’s pushback—worth keeping—is that 16% understates the internet’s influence. A shopper may research a MacBook online and then visit a store to compare its physical weight, or research clothing before driving five to ten minutes to nearby stores; the transaction is offline, but the decision journey is not.

3. Personalized pricing is economically neat but commercially dangerous

Asked whether agents enable individualized prices, Rampell explains that charging according to willingness to pay is appealing in Econ 101 terms because it captures consumer surplus. A crude version might charge an iPhone owner more than an Android owner, treating device choice as evidence of lower price sensitivity.
His practical conclusion is far less enthusiastic: companies have repeatedly tried dynamic personalization, with Delta cited as a possible example, but it invites regulation and “very, very high levels of unpopularity.” Even when technically possible, “generally it’s hard to get away with.”

4. Last-click attribution rewards interception rather than causation

Rampell calls last-click attribution the internet business model he hates most because it assigns 100% of a sale to the final measurable referral. A MacBook purchase might reflect Reddit research, a Super Bowl ad, Apple’s website, and an in-store trial; dividing credit is nondeterministic, but rewarding only the last click is deterministically wrong.
Honey is his sharpest specimen: it appears when the customer is already checking out, offers a coupon, redirects through an affiliate page to place a cookie, then returns the buyer to the same merchant. It “steals that attribution,” yet marketing teams can misread correlation as causation and report Honey or RetailMeNot as their best or fastest-growing channel.
AI compounds the problem. If Moore researches across Reddit and advertising, asks ChatGPT a question, and then purchases, the model contributed but did not single-handedly drive the sale. Untangling those influences remains “very, very hard,” even if an agent owns the final action.

5. Aggregators capture value because commodity brands cannot defend the SKU

Rampell’s diagnosis of direct-to-consumer weakness begins with one-and-done products made by third parties. Casper did not manufacture its mattress; an OEM could supply thousands of competing labels, while Casper repeatedly bought traffic from Google and Facebook. The traffic platforms became the bigger victors.
A subscription can partially improve the model: Dropcam attached recurring revenue to hardware even as cameras commoditized. A mattress bought five years ago still serves its owner, leaving Casper to find a new customer while the original factory sells similar products through “5,000 other manufacturers.”
Internet 1.0 already started killing off much of the long tail of location-protected commodity retail. Once shoppers could reach every seller, thousands of stores carrying the same Nike shoe became redundant; advantage moved toward the manufacturer or the aggregator with the fastest shipping and best service.
Moore adds that fashion demand moves too quickly for a single-style brand: Allbirds can yield to retro Adidas, On Running shoes, or New Balance. Amazon and Shopify ride demand across whichever SKUs become fashionable; AI may reinforce that aggregation because it can direct existing intent, but Moore says it will be hard for AI to “inculcate demand.” Rampell illustrates the problem: it is hard to know an On shoe is cool until one sees the cultural signal.

6. Google is losing informational searches before monetizable ones

Rampell reconstructs Google’s model: it arrived in 1998 as perhaps the 47th search engine, used links and PageRank to produce much better results, then adopted Overture’s paid-search concept. AdWords built the roughly $2 trillion company because relevant paid results could improve a commerce query rather than merely subsidize it.
Google’s enduring advantage is the default behavior formed by countless free questions, including searches launched through Safari—Apple makes tens of billions of dollars a year by sending those searches to Google. Now users ask ChatGPT questions such as who won an Oscar in 1977, moving the non-monetizable work elsewhere while continuing purchase-intent searches on Google.
Rampell’s evidence is that Google’s revenue still rises even as search volume appears to fall: “They are starting to lose some of the free but not any of the premium.” He estimates ChatGPT has roughly 800 million weekly active users, but says they are not buying directly in ChatGPT; OpenAI’s work to add commerce indicates that loop remains unfinished.
Moore identifies a reason shoppers returned to Google or Amazon after experimenting: product recommendations hallucinated nonexistent items, obsolete versions, or incorrect prices. Natural-language advice for highly specific needs—leggings for a particular hike and weather—is compelling, but only once the underlying inventory and pricing are current.

7. Bad source material is the central obstacle to trustworthy AI shopping

Rampell calls the World Wide Web “unhealthy”: search was fragmented into walled gardens long before ChatGPT, with real-time information on Twitter or X and friend activity on Facebook. Meanwhile, the remaining open web became commercialized through affiliate content designed to rank rather than inform.
His contrast is Consumer Reports, funded by subscriptions and refusing advertising, with today’s outsourced “top 10 running shoes” pages. The latter often means “top 10 affiliate revenue to me,” and the model cannot repair the incentive failure: “You can’t turn shill junk into honest analysis.” Rampell also connects the decline of traditional media to Craigslist taking local-classified revenue away from newspapers.
Moore finds better evidence in unsponsored YouTube reviews, where creators compare multiple products, disclose sponsorships, and can earn ad revenue from viewing rather than affiliate conversion. The opportunity is to transcribe those high-quality videos so LLMs can absorb them; because video is not skimmable and Google does not automatically transcribe every video, much of that information is not yet readily surfaced in traditional search.
Rampell mentions the New York Times’ purchase of Wirecutter as a possible trusted-review analogue, but Moore remains skeptical because nearly every recommendation seems to carry an affiliate link. His governing challenge applies across the category: if “most of the things on the internet are crap,” summarization alone cannot “decrapify” them.

8. Trust and automation define the investable layers

Amazon illustrates polluted abundance: sellers can buy 400 gadgets on AliExpress for $2 each, add a logo, and sell them for $25 by arbitraging six-week versus next-day delivery. Rampell found roughly “9,000” heated-sock listings sharing OEM origins, while he says a seller can switch a five-star rock listing’s SKU to heated socks and retain the reviews. Amazon should fix this but has little incentive because it wants to sell more goods.
Costco is his counter-model because membership economics make customer trust central. At roughly $100 per membership and more than 50 million members, membership revenue approximates net income; excessive merchandise margin would degrade the membership, so Costco curates aggressively. Moore says Costco keeps the hot dog at $150 and built a chicken farm to protect rotisserie-chicken economics.
Asked whether Costco should extend that trust, Rampell says expansion could endanger the enterprise, though financial services fit: a Costco loan could aim to be the cheapest rather than maximize lending margin. The decades-long belief that “if it’s sold at Costco, it’s good” makes the company unusually resilient and, in Moore’s words, “AI-proof.”
For new entrants, the clearest agentic target is a known SKU or UPC. An agent can continuously compare price, shipping, coupons, cashback, and card rewards; for detergent, it might autonomously buy when it is 30% cheaper on a specific site than usual and delivery remains reasonable. Products without UPCs—Wayfair’s bar stools are Rampell’s example—require more discovery and judgment before execution.
Moore places conversational assistance around bags, bikes, couches, and laptops: an agent can watch TikToks, read Reddit, ask dynamic follow-ups, and understand criteria deeply enough to recommend an item used for years. Impulse purchases and houses, cars, or wedding venues remain harder endpoints because one has no prior research and the other still calls for touch, experience, or a human expert.
Rampell sees specialist agents mainstreaming behaviors now confined to CamelCamelCamel, Ebates bought by Rockutan, Quidco, coupon tools, and technical bargain hunters. By collapsing “18 things” into one instruction, a service could choose the right card and rebate route, becoming “the last click of the 21st century”—potentially a large company even if it is not ChatGPT.
Moore’s second market is merchant infrastructure: storefronts must become legible to browsing agents, while financial systems need to support software using a consumer’s card and completing purchases. Amazon is exposed here because advertising is its “best SKU”—100% gross margin—and that profit pool is imperiled if AI owns the presentation layer.

LinkedIn Co-Founder Opens Up on the Reality of AI Job Loss | EP #194 Is It Time to Rotate Out of Crypto?