There is a seller on Amazon with roughly $20–30 million in annual revenue. Paladone is its name — they sell lamps, novelty goods, the category of objects that populate wish lists and arrive in gift-wrapped boxes. In January 2026, Paladone's performance on Amazon looked like this: 500,000 clicks from traditional sponsored product ads. And 88 clicks from the advertising format Amazon has positioned as the future of its $68 billion ad business.
Eighty-eight. Not a typo. Not a rounding error.
The format is Rufus. Amazon's AI shopping assistant, launched in February 2024 beta and fully deployed by July of that year, now has 250–300 million active users. It answers questions, surfaces products, and increasingly mediates — or in some cases, replaces — the search-and-browse behavior that has generated Amazon's advertising revenue at scale. When Paladone's marketplace advertising lead, Stefan Jordev, described the results to The Information, he framed them with the particular resignation of someone who has done everything right and found that the game has changed the rules without informing the players.
Amazon's advertising business reached approximately $68.64 billion in trailing twelve-month revenue as of early 2026 — nearly double the combined size of Publicis and WPP, growing at 21–22% year-over-year. The entire model rests on a premise: humans, at the moment of purchase consideration, can be shown a thing and persuaded to want it. The sponsored product, the headline search ad, the video spot between episodes — these are instruments calibrated to the human nervous system, optimized for attention, desire, the click.
Rufus does not have a nervous system. Rufus has a product graph and a language model.
When a shopper asks Rufus "what's a good gift for a 9-year-old who likes space," the assistant returns a ranked set of products. Those products may include sponsored items — Amazon transitioned Sponsored Prompts from free beta to cost-per-click paid placement on March 25, 2026 — but the sponsored item must still answer the question. It must be genuinely relevant to a 9-year-old space enthusiast. The ad does not interrupt the answer. It must become part of the answer.
This is the structural inversion that Paladone's 88 clicks makes legible: when the AI assistant is the shopping interface, the ad stops being a persuasion vehicle and becomes a data object. Its job is no longer to change your mind. Its job is to be included in the consideration set before the question is fully asked.
The sponsored item must still answer the question. It must be genuinely relevant. The ad does not interrupt the answer. It must become part of the answer.
Here is the detail that makes the Paladone case not just interesting but necessary: Rufus clicks are cheaper. Substantially. The cost-per-click on Rufus's Sponsored Prompts runs approximately $0.31, compared to $0.50–$0.70 for traditional Amazon sponsored product formats. On pure unit economics, Rufus advertising looks efficient. It is not. Paladone spent less per click and received approximately 5,680 times fewer of them.
The lower cost is not a feature. It is the market's signal that the format does not work as intended.
When a format delivers fewer conversions at lower cost, the efficient interpretation is that the mechanism is misaligned, not that you've found a bargain. Paladone was buying clicks from an audience that had already received an answer to their question — from Rufus — before the sponsored result could register as relevant. The CPC discount was compensation for participating in an auction that was settled before the bids were placed.
Research from MIT's ABxLab (ICLR 2026) found that LLM agents are 3–10x more susceptible to standard persuasion cues — authority signals, social proof, scarcity, pricing heuristics — than human baselines under controlled conditions. The implication, widely discussed in the advertising trade press, was that AI shopping agents would be highly manipulable by existing advertising techniques. Persuade the agent, redirect the human.
This research describes a different entity than Rufus.
ABxLab's framework tests agents that are making choices within a presented set. Rufus is not choosing within a set — Rufus is generating the set. Its 60% higher purchase completion rate (Amazon's own figure, not independently verified) reflects its function as a matchmaker between shopper intent and product relevance. The purchases it drives do not flow from persuasion. They flow from accurate identification of what the shopper already wanted.
This distinction — between an agent that can be persuaded and an agent that generates the context in which persuasion is possible — is the axis around which Amazon's advertising future turns. If Rufus's recommendations determine which products are visible at the moment of purchase consideration, then the value of advertising shifts from influencing the recommendation to qualifying for it.
Rufus does not choose within a set. Rufus generates the set. The purchases it drives do not flow from persuasion. They flow from accurate identification of what the shopper already wanted.
The Mars Agency conducted an analysis comparing Amazon's traditional search results with Rufus recommendations for identical queries. The findings were not marginal: 22% of products ranking on page one of Amazon search appeared in Rufus results for those same queries. 36% of Rufus recommendations were not on page one of search at all. Rufus is not echoing the search algorithm — it is building its own relevance model, optimized for conversational query understanding rather than keyword bidding.
This is the finding that should be receiving more attention in the marketing trade press than it has. Amazon's sponsored product auction rewards advertisers who bid on keywords. Rufus's recommendation engine rewards products that are semantically coherent with the questions shoppers ask in natural language. These are different optimization targets. An advertiser can win the keyword auction and be absent from the conversation. The $68B question is which mechanism determines the next trillion dollars of commerce.
For now, Amazon is managing both systems simultaneously. Sponsored Prompts uses the same campaign infrastructure, bid amounts, and CPC bidding as traditional Sponsored Products. Advertisers are auto-enrolled. The transition from free beta (November 2025) to paid CPC (March 2026) came without fanfare, with Amazon generating the prompts from existing product detail page content, reviews, and Q&A — the advertiser's only option is to opt out.
The competitive picture is not contained within Amazon. Google announced its Universal Commerce Protocol in early 2026, enabling AI agents to complete purchases directly within Gemini and Search without redirecting to retailer sites — a direct architectural challenge to the "search, click, convert" model. Walmart partnered with Google Gemini for AI shopping integration. Shopify's Sidekick and agentic storefront initiatives are designed to make products discoverable inside AI platforms. TikTok's Seller Assistant AI and GMV Max tools auto-allocate budgets across AI-optimized shopping formats.
Every major platform is building the same thing: an AI shopping interface that surfaces products based on intent analysis rather than bid position. Bain & Company estimates that approximately 65% of retail media spending currently flows to on-site sponsored search and product listings — the formats most directly threatened by AI-mediated purchase decisions. An industry forecast from 2025 projected that as much as 40% of current search ad spend could be at risk by 2027 as consumers shift to agent-based product discovery.
Amazon is not naive about this. Its positioning of Rufus as a "research tool" — an characterization reported by advertisers including Paladone — is revealing. The value of Rufus advertising is not direct response. It is understanding what signals Amazon's AI uses to rank products in conversational contexts, building data assets around the prompts that generate visibility, and participating in the infrastructure before the rules are finalized.
This is a different game than the one Amazon's $68B ad business was built to play. It is not performance marketing. It is more like institutional positioning — the advertising equivalent of securing a domain name before the brand launches.
The Paladone data point — 88 clicks vs. 500,000 — should not be read as evidence that Rufus advertising is failing. It should be read as evidence that the metric framework is wrong. Clicks are a meaningful unit for a system where a human types a query and receives a result. They are a less meaningful unit for a system where an AI assistant generates a ranked consideration set before the human has finished formulating what they want.
The question Amazon's advertisers are beginning to ask — and that the platforms have not yet answered with satisfactory measurement infrastructure — is: what does it mean to advertise to something that has already decided?
The $68 billion answer is still being written.
· · ·References