A buyer types a long, detailed prompt into ChatGPT and gets back a recommendation, not ten blue links to sort through. Andy Crestodina of Orbit Media Studios recently mapped the nine steps behind that recommendation, and his fix, rewrite the homepage to mirror the buyer's language, targets the wrong terrain. Before an AI model reaches any brand's site, it has already queried and cited a layer of third-party pages that decided who gets recommended. Call that the AI recommendation layer: it is the real competition now, and almost nobody is being paid to be found there.

  • AI chatbots recommend brands based on a fan-out of third-party citations, not the brand's own homepage.
  • Ahrefs found AI search visitors convert 23 times higher than organic search visitors, though that figure comes from a single company's data, not an industry average.
  • Similarweb found 55.9 percent of AI-influenced visits arrive as branded organic search, invisible to standard referral tracking.
  • Only 36 of more than 1,200 brands tracked by Semrush kept consistent AI visibility across ChatGPT, Gemini, and Google AI every month.
  • No paid channel yet exists for buying into the third-party content AI models actually cite when deciding a recommendation.

What the AI recommendation layer is

The AI recommendation layer is the set of third-party pages an AI model queries and cites while deciding what to recommend, before a buyer ever lands on a brand's own site. When ChatGPT or Gemini expands one prompt into several related searches, a mechanic researchers call query fan-out, it is building that layer in real time. Peec AI's analysis of five million fan-out queries collected in April 2026 found Gemini 3 splitting a single prompt into roughly nine sub-searches on average, compared with about two for ChatGPT. Whichever comparison sites, reviews, or trade articles turn up across those sub-searches quietly become the evidence the model weighs, well before the brand's homepage gets its turn.

From the rank-and-click economy to the cited-and-recommended economy

In the rank-and-click economy, a brand competed for one visible slot on a results page, and winning that slot delivered a click it could convert on its own site. In the cited-and-recommended economy, the AI model has already synthesized an answer from several third-party sources before the buyer sees anything, and the click, if it comes at all, arrives after the decision is functionally made. The skill set carries over: entities, structure, and evidence still matter. What does not carry over is where the value gets captured, since the citation that won the recommendation may sit on a publisher's site the brand never paid, and currently cannot pay, to appear on.

What Orbit Media reported, and what we verified

Crestodina's nine-step breakdown argues that AI-referred visitors convert far better than organic search visitors, and recommends brands mirror the buyer's own prompt language while backing every claim with evidence rather than taglines. The 23-times conversion multiple he cites traces to Ahrefs' June 2025 self-reported analysis of its own traffic, where AI referrals were 0.5 percent of visits but drove 12.1 percent of signups. Semrush's broader, cross-industry figure for the same pattern is a more modest 4.4 times, a discrepancy worth flagging since a single company's case study and an industry average are not interchangeable evidence. Crestodina's own caveat is honest: no tool yet shows how people actually phrase prompts at scale, so his framework is inference from convention, not observed data.

Where "good SEO is still good GEO" gets the economics wrong

The strongest case for treating AI optimization as SEO continuity

Google's own VP of Search argued in June 2026 that the fundamentals of good SEO already cover what AI search rewards: clear structure, named entities, demonstrated expertise. Crestodina's framework fits that camp too, since his advice amounts to doing SEO better with sharper evidence, not inventing a new discipline. The practices genuinely do carry over, and a brand that ignores them will lose visibility in both channels at once.

Why practice continuity isn't payment continuity

Continuity of technique says nothing about continuity of who gets paid. A brand can execute every step of Crestodina's framework perfectly and still lose the value of the recommendation, since Similarweb found 55.9 percent of AI-influenced visits register as plain branded organic traffic with no trace of the AI recommendation that produced them. Meanwhile the comparison articles and review pages the model actually cited to build that recommendation earn nothing for having supplied the evidence. Good SEO may still be good GEO as a craft; it is not yet a business model for anyone but the AI platform itself.

The recommendation layer is the new front door. Nobody owns it, and almost no one is paid to be found there.

— Smalk AI

What this means for brands and for publishers

For CMOs and media buyers: budget for the fan-out layer, not just the homepage

  • Audit which third-party pages surface across your category's likely fan-out queries, not just your own site's rankings.
  • Treat comparison content, reviews, and trade coverage as paid-media targets, since Semrush found only 36 of over 1,200 tracked brands held consistent visibility across ChatGPT, Gemini, and Google AI every month.
  • Stop reporting AI impact solely through referral traffic; the Wall Street Journal's January 2026 investigation found one agency's clients saw AI-sourced referrals rise from near zero to 44 percent of total referrals within a year.

For publishers: your comparison content is the uncompensated answer box

The evidence Crestodina tells brands to add, client counts, case studies, awards, is exactly what publishers already produce in reviews and comparison posts, and NP Digital's May 2026 survey found original research earns AI citation 82 percent of the time, the highest of eleven content types tested, with comparison content close behind at 76 percent. That content is doing the persuasive work inside the AI recommendation layer today, without a payment rail attached to it. Publishers pricing their future on recovered referral traffic are pricing the wrong asset; the asset that matters is being the cited evidence, and right now that has no invoice attached.

Three signals the payment rail still doesn't exist

eMarketer's June 2026 forecast, authored by principal analyst Nate Elliott, projects US AI advertising spend more than doubling from just over 32 billion dollars to over 68 billion dollars by 2030, with more than 80 percent of that spend still sitting beside AI answers rather than inside them. In-chat ad pricing is already under pressure: CPMs for ads inside ChatGPT launched near 60 dollars in February 2026 and fell to about 25 dollars within nine weeks, evidence that the current format, not the underlying demand, is mispriced. A parallel signal sits in Semrush's finding that 45 percent of marketers admit they cannot properly measure their own brand's visibility inside AI-generated answers, the same measurement gap that leaves publishers uncompensated for the citations doing the deciding.

Conclusion

Hold on to this: the AI recommendation layer, not any single brand's homepage, is where AI search decisions actually get made, and right now almost nobody, brand or publisher, can pay to be inside it.

Smalk AI is built for exactly that gap in the AI recommendation layer: a generative engine advertising network that places native ads for AI agents while opening a revenue line for the publishers whose evidence already decides the recommendation. Watch which platform first prices a standardized placement inside that layer, since that is the moment content marketing becomes a media buy.