Artificial Intelligence

Affiliate links, personalized ads, and chatbot revenue optimization

By Eric Benjamin Seufert

Generative AI has undeniably penetrated consumer awareness at an incredible clip: this NBER paper finds that “overall genAI adoption (including work and non-work usage) has been faster than both the personal computer and the internet.”

OpenAI’s ChatGPT is the world’s most popular standalone chatbot. ChatGPT chatbot was launched in November 2022 and, as of the company’s developer day early last month, is now used by 800MM users per week. ChatGPT is empowered by a Generative Pre-trained Transformer (GPT) model, which is an autoregressive LLM model for unidirectional next-token prediction. GPT is “generative” in that it receives input instructions and uses learned pattern representations to produce new data. This contrasts with Google’s BERT, or Bidirectional Encoder Representations from Transformers, a masked language model that uses bidirectional attention to reconstruct information from corrupted inputs.

ChatGPT is currently monetized through subscriptions, an enterprise product, and a nascent eCommerce initiative, Instant Checkout, that allows partner retailers to sell their products directly within the chatbot interface. Instant Checkout generates revenue with an affiliate-like scheme, where retailers pay a fixed percentage fee on all transactions facilitated by ChatGPT. OpenAI has recruited Etsy, Shopify, and Walmart as retail partners: these companies provide ChatGPT with product feeds, and ChatGPT recommends their products to users contextually, based on the content of their queries.

ChatGPT is not OpenAI’s only consumer-facing product: the company recently launched Sora, an app that allows users to create and share generated video content in a social feed. But ChatGPT is the company’s principal commercial engine: Sam Altman, OpenAI’s CEO, claimed last week that OpenAI expects to “end this year above $20 billion in annualized revenue run rate and grow to hundreds of billion by 2030.” But OpenAI is currently wildly unprofitable: per recent reporting from the FT, the company lost $8BN in the first half of 2025. Further, OpenAI has committed to roughly $1.4TN in infrastructure expenditure over the next eight years. Last month, OpenAI raised money at a $500BN valuation; reporting from Reuters suggests that the company intends to go public at a valuation of roughly $1TN.

To support that valuation, OpenAI almost certainly needs to accelerate its revenue growth. Per the FT’s reporting, only 5% of ChatGPT’s users pay for its subscription; the remainder utilize its free tier. OpenAI’s eCommerce initiative has been identified as a potential source of revenue scale: OpenAI could generate a significant volume of revenue if it can position itself as a meaningful conduit for eCommerce transactions. eCommerce sales in the US alone amounted to more than $300BN in Q2 2025; the International Trade Administration estimates that eCommerce will generate $4.9TN in global sales this year.

But for reasons that I expound upon in Agentic commerce is a mirage, I don’t believe that the affiliate-like eCommerce recommendation system implemented by OpenAI is the optimal business model for either ChatGPT or the broader chatbot product category. In this piece, I’ll argue that personalized, conversion-optimized advertising is a superior business model for chatbots than affiliate distribution: it is more scalable, better aligns with consumers’ interests, and has more positive-sum characteristics for society as a whole.

In particular, I believe that personalized advertising will be the model that OpenAI ultimately adopts as the primary monetization scheme for ChatGPT. This is important to consider precisely because OpenAI serves as an avatar for the commercial prospects and general sustainability of the artificial intelligence market, to the extent that the company’s infrastructure commitments are now viewed as systemic economic vulnerabilities. For this reason, the business model that OpenAI adopts to achieve sustainable growth is of critical importance, not just for its own prospects, but also for those of the overall consumer-facing AI category.

I’ve written previously that I believe personalized advertising will be the business model that allows OpenAI, specifically, but the chatbot category as a whole, to support sustainability (see Obviously, OpenAI will monetize with ads and OpenAI’s advertising opportunity). But the chatbot engagement model presents idiosyncratic sensitivities relative to other product paradigms, especially Search, that need to be accommodated with advertising, as I note in Perplexity introduces advertising; exploring the AI-search incentive problem:

With a traditional search engine, the user is best served if a single query leads to a single click. In theory, search engine effectiveness is inversely correlated with the number of queries and clicks conducted: if a user is forced to refine their query multiple times or follows multiple links in pursuit of the information they seek, then the search engine isn’t effective. In this way, a search engine must balance the tension of any monetization opportunity (query) with sponsored links against the retentive power of delivering the best, most helpful possible resources in the most prominent position of search results … But for high-intent queries on an AI search engine, the user expects some editorialization: actual insights around product quality, appropriateness, pricing, etc. The user is not merely looking to be nudged in the right direction; they seek context and insight. And that insight could actively undermine the prospects for an ad being clicked on.

In this way, the chatbot product experience cannot be used as a direct surrogate for Search; the integration of advertising must similarly be approached with nuance. And while I do believe that advertising as the dominant monetization model for chatbots is inevitable, I don’t make the case in this piece for how that should be implemented. The functional form of the ad placements that best monetize chatbots is a trillion-dollar problem to be solved by sophisticated product minds.

But on comparisons between the chatbot and Search use cases being fallacious / erroneous, it’s helpful to highlight what I view as flawed assumptions about chatbots. Before progressing into the arguments about monetization, these assumptions should be interrogated, because they help to highlight the distinction between chatbots and Search, and thus why chatbot monetization is likely to evolve away from the precedent set by Search. So before the superiority of personalized advertising to affiliate links can be reasoned about, three misconceptions and flawed assumptions that conflate chatbots with search engines must be dispelled.

Flawed assumption one: in-chat context is the only input available for non-subscription monetization.

If one sees the chatbot experience as fundamentally similar in terms of purpose and outcome to Search, then it’s natural to assume that any form of non-subscription monetization — be it affiliate links or advertising — may only be informed by in-chat context, since Search ads are principally (although certainly not singularly) targeted by keywords. But this assumption ignores the basis for most display advertising, especially social media display advertising, which is targeted primarily through behavioral profiling (see The App Tracking Transparency Recession for more background). There is no constraint on the scope of data that a chatbot may use for targeting ads, assuming it can collect the quality and volume of data necessary to institute personalized advertising. I make the point in Digital Advertising, Demand Routing, and the Millionaires’ Mall that, when a sufficient quantity of that data is available, personalized advertising should outperform contextual advertising:

The degree to which one of these approaches outperforms any of the others is dependent on the data available to it: for instance, if little or no behavioral data is available for specific users, contextual targeting could outperform behavioral targeting. Behavioral targeting flourishes when ads for specific products can be exposed to individuals with known affinities for them, as assessed through historical conversion data … One challenge with optimizing for conversions is that, in an ecosystem as vast and mostly heterogeneous as the internet — even in the context of specific, scaled products — the presence at any given moment of a user that 1) has an interest in some product and 2) possesses the disposable income to purchase that product is rare. Digital advertising doesn’t create consumer demand; rather, digital advertising should seek to route existing demand to the products that best serve it. An efficient digital advertising channel matches consumer demand for a product with the most satisfying and fulfilling variant of that product. An advertising channel is not a demand factory but a demand highway: the more efficiently an ad channel can route consumer demand to products, the more economic value it produces.

This doesn’t have to be the case as a rule; certainly, it’s conceivable that contextual targeting might outperform personalized advertising for eg., specialized publications. But consumer-facing products that have humanity-spanning appeal — which, at 800MM WAU, ChatGPT does — are not limited to contextual signals as inputs for advertising targeting. And as I alluded to above, neither is Search, so even when the chatbot use case is seen as qualitatively similar to that of Search, the imposition that only contextual data may be used for advertising targeting is an unnecessary restriction.

Flawed assumption two: Chatbots will always primarily produce text responses.

The digital affiliate advertising model was pioneered in the mid-1990s by CDNow, an eCommerce retailer specializing in CDs and other music-oriented products, although Amazon arguably popularized the model. This model was designed to allow website owners to earn commissions on sales they catalyzed through links or banner ads; it predated modern programmatic digital advertising and existed as a sort of stepping stone that alleviated the need to manage direct relationships with advertisers but didn’t achieve the efficiencies of personalized advertising. The affiliate model gave way to personalized advertising for the same reasons that contextual targeting is generally inferior to behavioral targeting: individuals have idiosyncratic preferences that aren’t captured when they are targeted in groups.

Notably, the affiliate model took root when the internet was primarily comprised of text-based, browser-rendered websites. In the same way that the consumer internet experience evolved into a diverse set of formats, so too can the chatbot experience be expected to evolve. There is no reason to assume that chatbot interactions will always be dominated by text; most chatbots are already multi-modal, across both input reception and output generation. It’s important to note that a chatbot itself is an agent, and that agent can choose from a range of models to render output.

It’s not difficult to imagine that a large proportion of chatbot output might be rendered as video in the near-term future; for example, the response to “How to change a flat tire” is almost certainly better served as a video tutorial than a series of text instructions. If that is the case, a consumer’s chatbot engagement may more resemble YouTube than Google Search, and so should the underlying business model.

Flawed assumption three: Chatbot commerce will always have access to the same qualitative review data that it does now.

This assumption is anchored to the view that chatbots can serve as high-purchase-intent research agents, surfacing the best product options for a given set of preferences and requirements. This is true, but that research capability presumes that the chatbot’s underlying model can be trained on high-quality, insightful product reviews, and that the chatbot’s output can a) synthesize those into a small number of product recommendations and, by extension, b) monetize those recommendations with affiliate links.

But the reviews that a chatbot ingests are themselves likely monetized through affiliate links. If a chatbot simply absorbs the knowledge contained in expert reviews and pre-empts their monetization, then those reviewers are starved of income and are unlikely to write more reviews. Similarly, if a chatbot relies on ratings and reviews from eCommerce retailers in implementing its own checkout mechanism — as OpenAI has done — then in capturing the transaction, it might deprive the platform of the ability to aggregate more reviews.

In other words, if a chatbot’s ability to provide relevant product recommendations is predicated on synthesizing expert reviews from blogs, and its own eCommerce solution prevents those blogs from monetizing, then those blogs will stop being written, and the chatbot will lose those reviews as an input. And if chatbots prevent reviews or ratings from being submitted to eCommerce platforms, then it will lose access to a valuable input to its own recommendation mechanism. In either case, the chatbot loses access to a critical input that its own personalization engine needs for recommending relevant items, and its recommendations may end up being less helpful than those of retail platforms, undermining its value as a research agent.

With these assumptions addressed, below, I present three arguments for why I believe the affiliate model is inferior to personalized advertising as the monetization engine for chatbots:

Personalized advertising maximizes the value of user attention

While Instant Checkout’s commercial terms are not public, product recommendation systems tend to prioritize (rank) candidates based on conversion probability, which is the likelihood that an exposed product is purchased. Digital advertising, however, utilizes an auction mechanism that correlates an ad’s rank with its expected value: this is the advertiser’s bid times the calculated probability of the ad leading to some outcome (like a purchase), with some quality conditions applied. For more background, see Understanding conversion optimization in digital advertising, The economics of advertising auctions, and this presentation I gave earlier this month, which explores recent research in ML applications for both ad ranking and product recommendations.

The two dominant auction mechanisms for digital advertising are Generalized Second Price (GSP) and Vickrey-Clarke-Groves (VCG). These auction designs differ in fundamental ways, most notably that in VCG auctions, bidders pay the externality costs imposed on other bidders (ie., how much other bidders are hurt by the presence of their bids) and that in GSP, bidders pay just above the next-highest bid below their own. Thus, VCG auctions are generally characterized as “truth-telling,” in that the dominant strategy for a bidder is to bid the actual value of the impression to them, whereas GSP auctions incentivize “bid shading,” or bidding a lower amount than what the impression is worth to them. These ideas are well established in Hal Varian’s Position auctions (2006) and Online ad auctions (2009), although a wealth of literature exists on this topic. While VCG serves as a canonical reference design, most scaled advertising platforms implement a modified second-price-like auction mechanism that institutes various quality constraints (like pacing, perceived ad quality, etc.).

By utilizing an advertiser’s bid to evaluate candidate ads based on expected value, an advertising platform optimizes its own revenue subject to quality and other idiosyncratic scoring constraints: it ranks ads according to the amount of money they are predicted to generate for the platform. And because the bid is submitted by the advertiser, based on their own proprietary calculations of revenue, an ad can be seen as the content that best monetizes the user’s attention — not in terms of counts of clicks or purchases but of total transaction volume. I explore this idea in The best and highest use of customer attention.

In an ad auction, each candidate’s rank encapsulates the advertiser’s private estimate of an impression’s value: its bid. So the winning ad tends to be the highest expected monetization of the impression (again, conditioned on quality constraints) and not merely the highest conversion probability. Those bids produce a larger expected surplus and dictate how it is divided among the platform (auction revenue), the advertiser (profit on incremental conversions, assuming bids are appropriately priced), and the consumer (higher match quality if the ad improves utility).

An affiliate-monetized recommendation model, on the other hand, is more likely to disproportionately rank on price, given the lack of private bid information. This is susceptible to defaulting to the lowest-price option. In this case, the affiliate model is worse for the platform than the advertising model, since its compensation is applied as a fixed commission: it does not optimize for transaction volume but conversion rate, so its revenue may become a function of the price of the products it recommends. While on-site product recommendation systems also optimize to expected value by considering a product’s price (see research from Alibaba and Criteo), an affiliate-based recommender system can’t use the advertiser’s private estimate of value to optimize total surplus and thus optimally monetize its own inventory.

And in this way, platform revenue and total welfare are positively correlated under the advertising model, because when a winning bid increases, the platform makes more money. But an affiliate model can mis-order candidates based on price if conversion probabilities overwhelm the value equation without being moderated by a bid, resulting in lower-priced products being favored that reduce the platform’s commission fee and generate lower aggregate welfare and platform revenue. Thus, the auction mechanism not only maximizes the platform’s expected revenue but also achieves an allocatively efficient outcome: impressions are awarded to the advertisers who value them most.

As a brief formalization:

Affiliate links only monetize queries with a commercial locus

One benefit of personalized advertising being detached from context is that it allows ad load to remain relatively stable across user sessions. With an affiliate model, only commercially relevant content can be monetized, since targeting is anchored to context. In a 2020 blog post, Google notes (emphasis mine):

Nearly all of the ads you see are on searches with commercial intent, such as searches for “sneakers,” “t-shirt” or “plumber.” We’ve long said that we don’t show ads–or make money–on the vast majority of searches. In fact, on average over the past four years, 80 percent of searches on Google haven’t had any ads at the top of search results. Further, only a small fraction of searches–less than 5 percent–currently have four top text ads. 

If a chatbot borrows its monetization model from Search, then it must contend with the limitations of that model: the overwhelming majority of engagement can’t be monetized. This doesn’t, in and of itself, necessitate lower aggregate revenues if affiliate links monetize at a higher rate. But as argued above, the affiliate model is susceptible to overemphasizing conversion rates through its ranking process such that price becomes the dominant determining factor in selection. If that’s the case, given a fixed commission rate, the affiliate model may deliver lower revenue per transaction than the ad model for a smaller number of impressions.

Beyond the smaller monetization surface area, a separate but related problem arises from this commercial intent issue. The limited universe of commercial queries offers retailers fewer opportunities to reach potential customers. If chatbot product discoverability is limited to the commercial query set, and the recommended products are principally determined by historical conversion rates with no bid to moderate the expected value, then it becomes challenging for new products to reach customers.

This incumbency bias, which is attenuated through the bid in the auction model, may depress the overall revenue capacity of a chatbot by reducing the scope of products it recommends. This, in turn, could diminish the value of chatbots for discovery, leading consumers to other avenues. Product variety has been found to correlate with digital advertising spend; if the affiliate model proliferates across chatbots, it could lead to in-category concentration, which could result in reduced levels of eCommerce economic activity.

The affiliate model engenders incentive misalignment between chatbots and retail platforms

The central argument of my recent piece, Agentic commerce is a mirage, is that independent agents making autonomous purchase decisions on behalf of users is incongruent with both retail platform and end-consumer incentives. The first point is mostly moot now: what OpenAI calls “agentic commerce” doesn’t involve autonomous purchasing, as is implied by the phrase, but rather in-agent product discovery and integrated checkout. That’s a semantic quibble I’m willing to live with.

But the second point holds: integrating checkout into an independent agent is at odds with the motivations and reward structures of most retail platforms, but, specifically, Amazon. This is demonstrable: as I point out in that piece, Amazon has blocked all agents from accessing its website. And, last week, Amazon sued Perplexity for allowing its AI-empowered browser, Comet, to make purchases on behalf of users. From Bloomberg (emphasis mine):

Amazon.com Inc. is suing Perplexity AI Inc. to try and stop the startup from helping users buy items on the world’s largest online marketplace, setting up a showdown that may have implications for the reach of so-called agentic artificial intelligence … The US online retailer filed a lawsuit Tuesday demanding Perplexity stop allowing its AI browser agent, Comet, to make purchases online for users. The e-commerce giant is accusing Perplexity of committing computer fraud by failing to disclose when Comet is shopping on a real person’s behalf, in violation of Amazon’s terms of service, according to the complaint in San Francisco federal court.

I make the point in the Mirage piece that Amazon’s resistance to third-party agentic commerce is motivated by the value that it derives from its direct relationship with consumers, which empowers its own advertising business. As I’ve noted in my Amazon, the advertising behemoth series (see parts one and two), that business is strategically critical to Amazon: the company generated $17.7BN in advertising revenue in Q3 2025, for 22% year-over-year growth. Critically, Amazon’s Q3 advertising revenue represented 83% of the company’s net income in the quarter.

One might argue that Walmart’s participation in OpenAI’s Instant Checkout program serves as a meaningful counterargument to the notion that third-party agentic commerce contrasts with retailer incentives. But Walmart and Amazon are vastly different companies with fundamentally distinct strategic priorities: Walmart’s eCommerce sales amounted to 18% of net sales for FY 2025 (which ended January 31, 2025), and its advertising business generated $4.4BN in revenue in that time period, representing 22.6% of net income. While Walmart is an omnichannel retailer, it generates a minority share of its revenue from eCommerce (although that share is growing), and its advertising revenue as a proportion of net income is roughly a quarter of Amazon’s. As a result of this distinction, Walmart likely views the incremental transactions delivered by integration with third-party commerce agents as more strategically valuable than Amazon does.

Further, Amazon offers its own, on-platform commerce agent: Rufus. Amazon noted in its Q3 2025 earnings call that Rufus saw “250 million active customers in 2025” and is projected to deliver “over $10 billion in incremental annualized sales.” eMarketer estimates that Amazon commands roughly 40% of the US eCommerce market; if Amazon remains a holdout for chatbot commerce integration, can any third-party agentic commerce initiative reach appreciable scale?

And what might change that could result in Amazon refining its approach or softening its position? Again: Amazon operates a commerce agent. What benefit does Instant Checkout offer to Amazon that isn’t offset by a loss in advertising revenue (which is dominated by sponsored search but also includes that from its DSP, Prime Video, and other sources) and all of the downstream consequences of being intermediated?

I’ve seen it argued that Instant Checkout does provide merchants with a direct relationship with ChatGPT users, but I don’t think that conclusion can be substantiated. The merchant that fulfills an Instant Checkout order is the merchant of record for the transaction: that merchant is named on the customer’s credit card statement and is responsible for chargebacks and refunds, per OpenAI’s Instant Checkout documentation. However, while that merchant receives the end user’s email address and whatever other contact information is required to fulfill the order, OpenAI prohibits merchants from using a customer’s email address for marketing. In this way, I’d argue that Instant Checkout does not provide merchants with a direct relationship with customers — at least, not in the specific ways that benefit them, such as the ability to communicate discounts and promotions via email, or, presumably, to use their email address or other contact information for targeted advertising campaigns.

This is all to say: the choice between an affiliate model or an advertising model may be dictated not by the underlying economics of those structures or consumer tastes, but by Amazon. In the end, any new entrant to the retail landscape must contend with its competitive realities. In this piece, I argue that the affiliate model is economically suboptimal for chatbots relative to digital advertising and that it conflicts with the best interests of consumers. More importantly, it’s unclear how this model adapts to the specific contours of the retail environment.

Comments: