When a user asks ChatGPT for a product recommendation, the model gives an answer. A brand’s ad appears below it, clearly labeled as sponsored, visually separated. OpenAI is explicit: the ad doesn't influence the answer. The systems are structurally independent.

But the user sees both on the same screen. A May 2026 study from the University of Waterloo and UCL, published in Communications Psychology, found that people consistently perceive AI systems as more confident in their responses than humans, even when the content is identical. The researchers call it the "illusion of confidence." Users attribute authority that may not be warranted.

That matters commercially. If ChatGPT's organic answer discusses your category without mentioning your brand, the user has anchored to a high-confidence opinion before they scroll to your ad. The model's answer functions as an implicit endorsement the user over-trusts. Your ad is working against a trust signal the user has already formed, and formed quickly, and over-weighted.

When the model's answer already positions you favorably, the dynamic flips. The user's over-attribution of AI confidence amplifies your credibility. The ad reinforces authority the model already assigned. Paid spend accelerates trust rather than fighting it.

In search, ads compete with other ads. In an AI assistant, your ad competes with the model's own answer. That's a different commercial environment than anything the industry has operated in before.

What happens if you buy ads in ChatGPT without investing in how the model sees your brand? 

The ads still function. OpenAI's system matches ads to conversations using "context hints" from advertisers, landing page content, ad copy, and title. If the targeting aligns, you get impressions and clicks whether or not the model has heard of your brand.

But the ceiling is lower.

Your web content partly determines the matching signals that control where your ad appears. If that content is thin or poorly structured, context matching weakens. You show up in fewer relevant conversations, or in conversations where the fit is loose enough that the user doesn't engage.

ChatGPT also lets users tap "Ask ChatGPT" on any ad to get the model's organic take on the product. At that point, what the model knows about you shapes whether the follow-up reinforces the ad or works against it. If the model has strong, well-sourced information across the web, it builds on the ad's message. If it doesn't, the follow-up may surface a competitor or return something generic.

The measurement challenge compounds this. Users in AI environments ask multiple questions across sessions, sometimes across devices. The assistant summarizes sources without sending a click. Last-touch attribution breaks. The industry is already moving toward incrementality as the baseline, but most measurement stacks weren't built for a channel where the user never leaves the assistant's interface. Customer acquisition costs will fragment across query types, AI platforms, and product categories in ways that are hard to model with current tools.

Structured data, third-party validation, and web content now directly affect ad performance

Google's Universal Commerce Protocol, launched in January and now backed by Amazon, Meta, Microsoft, Salesforce, Stripe, Shopify, Target, and Wayfair, is building the standard for how AI agents interact with commerce. Agents evaluate brands based on structured data, pricing signals, fulfillment reliability, and the consistency of product information across the web.

PwC calls this Generative Engine Optimization. Google Cloud Next was more direct: the biggest ad spend in the world can't compensate for missing catalog attributes. If your data isn't structured for agent parsing, you don't make the consideration set.

This is work that has historically been filed under "brand marketing" or "PR." Consistent product claims across the web. Structured data readable by AI systems. High-quality reviews and third-party validation. But a growing share of it is much more operational. How products are described, categorized, and merchandised is becoming a critical input for AI environments. Product titles, attributes, descriptions, and taxonomy all feed the data layer that AI agents and platforms read when deciding what to surface. Retailers and brands that treated product content as a catalog management task are finding it's now a performance lever. If the AI can't parse your product data cleanly, it can't recommend you confidently, and the ad you're paying for below that answer inherits that weakness.In an AI-mediated environment, this work directly affects ad performance. It's the foundation your paid strategy sits on.

The funnel in AI-assisted commerce looks less like awareness → consideration → conversion and more like legitimacy → eligibility → recommendation → conversion. Paid can accelerate any stage. It can't create legitimacy from nothing.

This is all very early

ChatGPT ads have been live for less than four months. They hit $100 million in annualized revenue within six weeks with 600+ advertisers enrolled, which is significant. But the ad formats, targeting signals, and pricing models will look different six months from now. OpenAI is still transitioning from $60 CPMs to $3-5 CPC pricing. Google's UCP is weeks old. Self-serve tools are just rolling out.

The brands entering now have a window to shape how they show up in these environments before the playbook solidifies. But only if they pair the media buy with the upstream work.

Six months from now, the mechanics will have changed. How much and in what direction is hard to say. What we're reasonably confident about is the underlying dynamic: in an AI-mediated environment, the model's existing knowledge of your brand shapes the effectiveness of every dollar you spend in it. The brands that invested in both legitimacy and visibility will be better positioned when the market matures. How much better is something we'll be watching closely.

That's our read on where things stand today. The ground is shifting fast, and the right answer six months from now may look nothing like the right answer now.