AI search has moved fast enough this year that a session written in April already reads like an earlier chapter. OpenAI is now testing Sponsored Agents inside ChatGPT, letting brands pay for a labelled conversation with their own AI representative once someone clicks an ad. Amazon just blocked Meta's new Muse shopping agent from its site entirely, over access it says was never authorized. And AI tools recommending coupon codes are turning out to be wrong more often than they're right, since a majority of AI-suggested codes don't actually work. That is the landscape Nick Lafferty, founding marketing engineer at Profound, walked into at Affiliate Summit East, and his session gave the room a framework for understanding exactly why all three of those stories are connected. 

The Session: AI Search Has a Second Dimension 

Lafferty's core argument was that most brands are still only tracking half the problem. The first half is simple visibility, whether your brand shows up at all when someone asks ChatGPT, Gemini, or Perplexity a question in your category. That's the layer most SEO and content teams have spent the last year optimizing for. 

The second half, which he called the real second dimension of AI search, is what the model actually says once you do show up. He used Southwest Airlines as an example. Southwest dropped its unassigned seating policy earlier this year, yet AI models can still tell users to arrive early because Southwest does unassigned seating, a policy that no longer exists. He also pointed to Hoka, where asking whether Hoka makes a good running shoe can return an answer that says yes, then immediately suggests a competitor is better. Showing up is not the same as being described correctly, and most brands have no way of knowing which one is happening to them. 

What the Data Actually Showed 

Profound analysed 50,000 LLM responses across categories including consumer, fintech, and auto. A few findings from that data set stood out. ChatGPT's average response runs around 3,000 characters, roughly the length of eleven tweets, and about half of that length is what Profound calls unsolicited content, meaning rationale, comparisons, and editorial framing nobody explicitly asked for. That unsolicited half is where brand perception actually gets shaped. 

On accuracy specifically, Profound's own fact-checking product found that around 11% of AI responses about one major wearable brand contained factual errors, wrong prices or wrong features. Multiplied across the millions of prompts a category receives monthly, that is a meaningful reputational exposure most brands aren't currently measuring. 

The most striking number from the session, though, was about where citations actually come from in the first place. Lafferty said only 5 to 10% of citations in any given category come from a brand's own domain. The other 90 to 95% come from third party sources: Reddit, Wikipedia, LinkedIn, competitor sites, and affiliate content. He illustrated this with Run Repeat, a niche affiliate site that literally cuts running shoes in half and weighs them against manufacturer claims. Run Repeat is now the single most cited source across all running shoe related prompts Profound tracks, at 7.1% of citations, ahead of Nike's own site, which sits at just 2.4%. 

Why This Explains September's News 

Seen through that framework, the last two weeks of AI news line up neatly with what Lafferty described. OpenAI building a paid Sponsored Agents product only makes commercial sense if organic presence inside AI conversations is already scarce and contested, exactly the scarcity Lafferty's citation data points to. Amazon's decision to block Meta's Muse agent shows that even having a product listed on a retailer's site is no guarantee an AI agent can actually see or buy it, reinforcing his point that visibility and accessibility are not the same thing. And the unreliability of AI-suggested coupon codes is a direct, consumer-facing example of his accuracy problem playing out in exactly the kind of content affiliates produce every day. 

Takeaways From the Session 

Track citation share, not just rankings. If 90 to 95% of citations in your category come from outside your own domain, your usual SEO scoreboard isn't measuring the thing that actually matters anymore. Find out which third party sites are winning citations in your category and treat that list as your real competitive set. 

Produce content AI cannot replicate. Run Repeat wins because it does something physical and specific that a language model simply cannot do. Lafferty's broader point was that first-hand, hands-on, highly specific content consistently outperforms generic brand copy in AI citations. 

Build real relationships with publishers rather than buying mentions at scale. Lafferty was direct that Google is already cracking down on manipulated citations, and that AI search tends to reward content that reads as genuinely independent. 

Audit accuracy and sentiment on a regular cadence. Don't assume that being cited means being described correctly. An inaccurate but visible citation can do more damage to a brand than no citation at all. 

Watch how platforms are drawing access lines. The Amazon and Meta dispute is a preview of a larger fight over which AI agents get to see, browse, and buy from which retailers. That fight will shape what affiliate and agentic shopping flows are even possible in the months ahead. 

The Bottom Line 

Visibility was only ever step one. Citations from your own domain barely register against the 90 to 95% coming from everywhere else, so the real competition is happening on sites most brands aren't even tracking. Winning now means knowing who is actually getting cited in your category, making sure what gets said about you there is true, and keeping a close eye on the platform fights already deciding who gets into the conversation at all.