If we use AI to make decisions for us, where do ads fit in?

Building frontier AI models is expensive. As is serving them to hundreds of millions of customers. So far, a small percentage of users are paying $20 a month to use them; back of the envelope math suggests ~5% of ChatGPT’s ~700 million users are doing so today (8% on the high-end, 3% on the low).

Nick Turley, the person in charge of ChatGPT, was recently interviewed on Decoder, where he said:

We will build other products, and those other products can have different dimensions to them, and maybe ChatGPT just isn’t an ads-y product because it’s just so deeply accountable to your goals. But it doesn’t mean that we wouldn’t build other things in the future, too. I think it’s good to preserve optionality, but I also really do want to emphasize how incredible the subscription model is, how fast it’s growing, and how untapped a lot of the opportunities are.

Emphasis mine. I want to zoom in on that bit, that ChatGPT isn’t “ads-y” because it’s “so deeply accountable to your goals.”

I’ve been thinking about this tension for over a year.

AI Will Disrupt the Attention Economy

AI, and I felt this during the deep learning era as well, is an important bit of technology because it allows you to project your decisions.

Gunpowder changed the nature of fighting and war because it allowed combatants to project their force, magnitudes farther than a spear or sword allows. The printing press, telegraph, and the internet changed the world because they allowed people to project their communication beyond their audible reach. AI, née deep learning, allows you to encode your decisions (not all of them, but many) into portable packages of perception and discernment that can sort through mountains of content in moments.

This decision projection will change our information ecosystem. Our digital and media economy is a zero-sum battle to earn and sell your attention. With decision projection our attention is effectively limitless1.

Given most advertising is sold in units of attention, this presents a challenge.

Search Ads Work Because Search Presents Options

Google Adwords (now just “Google Ads”) is perhaps the best ad model for a given product, ever.

When someone searches, a real-time auction begins. Eligible ads bid for the given query, with the winner paying the next best competitor’s bid. The winner’s ad would appear similar to a search result, among the search results. Users perused the search results, including the ad, and would select a link to click.

Today, Google handles ~90% of all searches.

Google Adwords is perfect because:

  1. Users state what they’re looking for
  2. Interested parties compete for that bid, yielding relevant ads
  3. Users select their result from a range of options

That selection is key. Google puts options on the page, ads included, and the user decides.

But there is one way to keep Google from serving you an ad. Start your search from the Google homepage, not your browser’s address bar, and instead of hitting “Search”, click “I’m Feeling Lucky.” Google will skip the results, the ads, the selection, and take you directly to the first result. You’ve ceded the selection decision to Google, hence no ads are shown.

“I’m Feeling Lucky,” is an anachronism. While writing this, I was surprised to see it’s still there. Initially, it was a bit of swagger, confidence manifested as UI. “We are so good at web search,” Google seemed to say, “you can skip the results.” Few ever used it, and dramatically fewer use it today, but oddly it presaged a pattern picked up by ChatGPT.

Chatbots Have Few Good Ad Choices

ChatGPT – and Claude, Gemini, DeepSeek, and all other chatbots – don’t deliver a set of options to peruse, they deliver answers. As Turley says, they are “deeply accountable to your goals.”

Unlike search, there is no obvious play to insert ads. And the options that do exist feel either bolted-on or undermine the chatbot’s core function. These options include:

  1. Display Ads: Advertising placed in or around the response. These could be text or images. This is the dominant ad model for web pages, and not integrated into the content.
  2. Text Integrated Ads: Advertising integrated into the text response. The chatbot would search for or be provided relevant product information that would inform the response. The integrated ad would be noted as an ad, but otherwise naturally integrated into the reply.
  3. Widget Integrated Ads: In responses, product listings could be broken out in rich, carousels. OpenAI is experimenting with this format, Perplexity kind of does this, and Google already presents a carousel of only sponsored options atop your search.
  4. Interstitial Ads: Advertising that is presented in between user interactions. An ad could be displayed for a short time after you submit your query, before you see your result.
  5. Sponsored Prompts: Advertisers could sponsor suggested prompts, either on the landing page (as a suggested query, “Explore sandwich ideas with Kraft”) or as a suggested follow up after a response has arrived (“Would you like to learn more about Product X)?

Off the bat, we can remove display ads as an option. To build an ad product that delivers value at a scale similar to their product, ChatGPT cannot adopt standard ad units and ad targeting. Display ads would be valued the same way display ads on the New York Times or stray blogs are (in terms of page views and clicks), undercutting the special nature of ChatGPT. Adopting display ads devalues their product, creates bad incentives, and won’t generate the returns needed to support OpenAI’s goals. For a deep dive on why this is, read my explanation of why media metrics matter

Interstitial ads, though a natural fit for slow reasoning models, is likely an imperfect fit for the reasons display ads fail. They’re bolted on, not tied to the core query, and outside of the main user flow.

Text integration ads hit on the tension Turly describes: ChatGPT is “deeply accountable to our goals,” so taking time to not deliver a single answer to our question, given the context, undermines its core function. Turly elaborates:

If we ever [added advertising to ChatGPT] I’d want to be very, very careful and deliberate because I really think that the thing that makes ChatGPT magical is the fact that you get the best answer for you and there’s no other stakeholder in the middle. It’s personalized as to your needs and tastes, etc. But we’re not trying to upsell you on something like that or to boost some pay-to-play provider or product. And maybe there are ways of doing ads that preserve that and that preserve the incentive structure, but I think that would be a novel concept and we’d have to be very deliberate.

OpenAI and others could try to identify when users are asking for options and use these moments to serve ads. This brings us to widget ads. In April, OpenAI announced the addition of product carousels to their search mode, similar to Google.

Ads naturally fit in this interface, as it presents a selection. But, for now, this functionality is hidden in ChatGPT’s search mode…which itself is hidden (hit the “+” button, select “More”, select “Web Search”). They are clearly being cautious. One gets the feeling search mode is a place to explore these tricky questions without spoiling the core ChatGPT experience.

Thinking through widget ads, you end up landing on affiliate marketing, or affiliate links. Affiliate marketing is when advertisers pay people or companies a commission for leads or sales they generate. This is big business, though smaller than traditional advertising.

And yes, Turly says, OpenAI is thinking about affiliate marketing:

There is actually something that is neither ads nor subscriptions, which is if people buy things in your product after you very independently serve the recommendation. Wirecutter famously does this with expert-selected products.

But then if you buy them through a product like ChatGPT, you could take a cut. That is something we are exploring with our merchant partners. I don’t know if it’s the right model, I don’t even know if it’s the right user experience yet, but I’m really excited about it because it might be a way of preserving the magic of ChatGPT while figuring out a way to make merchants really successful and build a sustainable business.

Affiliate marketing, and the question of whether it consciously or unconsciously influences recommendations, is a fraught topic. We have a hard enough time determining if it affects human reviewers; trying to understand if it affects AI reviewers is another question entirely.

If I were at OpenAI, I would argue strongly against generating affiliate revenue from in-response recommendations if only because it could function as an explanation for why ChatGPT’s results aren’t good. One challenge facing chatbot products is that they are black boxes. How they arrive at their results is largely hidden (with the exception of reasoning chains), and even among researchers at top labs can’t explain why an LLM returns a specific result. This black box nature leaves the door open for users to come up with their own explanations, factual or not, that can take on a life of their own. Adding a visible incentive – affiliate revenue – introduces an easy reason why one chatbot has worse vibes than another. And often, that’s enough to cause real damage.

Further, I have questions about if it’s even technically possible to implement affiliate marketing without influencing the results. If you provide your chatbot with a well designed, tested, and maintained tool for obtaining product specs and features (let’s call these ad prompts), this set of product information will be easier to obtain and consume than an inconsistent or unruly webpage. Simply providing an ad prompt will almost certainly increase your likelihood of recommendation due to the nature of contexts2.

I will be shocked if ChatGPT is the first to pull the trigger on affiliate recommendations. I think they can work, provided they are contextualized within a larger array of options and limited to an “ad slot” amongst the array. But monetized product recommendations integrated into text answers will undermine the core service ChatGPT provides.

If I were forced to pick an ad format for ChatGPT, today, I would pick sponsored prompts. I believe it’s the best, worst option of the formats identified above. It’s relevant to the core chatbot user interaction, isn’t a bolted-on distraction like intersticials and display ads, yet doesn’t influence the response ChatGPT generates. ChatGPT’s stock conclusion to it’s answers, suggestions to users about next step for them to pursue (“Would you like to learn more about X?”), could be broken out of the text response itself. Below the text response there would be a couple buttons representing these entreaties, one of which could be sponsored.

This is where I’d start, but it’s not ideal.

AI Disrupts Advertising Foundations

Ads are designed to influence or perceptions and ultimately our decisions. But as we outsource more decisions to AI tools, and those tools become better at projecting our decisions and discernment…where does that leave advertising? Will the task of advertising be split between appealing to us and appealing to our agents? Are these jobs the same or different?

It’s hard to say at the moment, and I don’t think we’ll get an answer from anyone for a bit. The big labs are blitzscaling and there’s no shortage of funding to pay the bills. The goal is marketshare, and no one wants to be the first tool to compromise their product. But this can’t go on forever; an ad model will emerge. Let’s just hope it fits the chatbot product.

  1. By the way, I suspect this is the reason Meta is spending so aggresively when it comes to AI. If they have a unifying strategy throughout their existence it’s earning and selling attention. Their king KPI is “share of timespent”, aka how much of your waking hours is spent staring at Meta products. 98% of their revenue is from advertising, selling this attention. If AI turns attention from a zero-sum game into, well, anything else, it’s an existential event for Meta. ↩

  2. I tried this myself this week, scraping the product pages from a few bicycle manufactorers and rephrased their content as ad prompt markdown files (here’s one example). I staged these documents behind an MCP armed with simple vector and text search (another great use case for Chroma), and wired it up to Claude with intructions to both browse the web and use the affiliate tool to assemble recommended products for my queries. Over and over, the affiliate listings would be richer, more descriptive, and would appear more often. I suspect this is because the data had been prepped, and that ease delivered better results. ↩