The Recommendation Economy

1. Introduction

For nearly three decades, the internet rewarded visibility. Businesses competed for higher search rankings because rankings generated clicks, clicks generated traffic, and traffic generated customers. An entire industry grew up around being easier to find.

AI changes that model at the root. When a consumer asks ChatGPT, Gemini, Claude, or Perplexity for the best running shoe, protein powder, or CRM, they are rarely handed ten competing websites. They get a recommendation.

The scale of the shift is no longer speculative. ChatGPT passed 800 million weekly users in late 2025. Shopping queries roughly doubled as a share of ChatGPT usage in the first half of 2025 alone. Adobe measured a 1,200% jump in AI-referred traffic to US retail sites over eight months, a doubling roughly every two months, and found that 39% of consumers had already used generative AI to shop, nearly half of them for product recommendations specifically.

Meanwhile the old model is breaking. When Pew Research tracked real browsing across 68,879 Google searches, users clicked a traditional result on only 8% of the searches where an AI summary appeared, about half the normal rate. They clicked a source cited inside the AI answer only 1% of the time.

For the first time since Google’s rise, businesses are no longer competing mainly to be discovered. They are competing to be chosen. Traditional search rewarded the website that best satisfied an algorithm built to rank pages. Generative AI rewards the brand an algorithm believes deserves trust.

So the optimization problem changes with it. The question is no longer “How do I rank higher?” It is “Why should an AI recommend my brand over the alternatives?”

That question is the foundation of what people now call Generative Engine Optimization. Most of what brands are told to do about it answers a different question entirely.

2. The problem with today’s GEO advice

GEO is one of the fastest-growing topics in digital marketing. Since ChatGPT went mainstream, hundreds of articles have tried to explain how brands should prepare for AI-powered search. Much of that guidance shares three problems.

First, recommendations get recycled between agencies without validation. The same technical checklists show up everywhere, on agency blogs, LinkedIn posts, and consultant decks: llms.txt, robots directives, schema markup, AI-friendly formatting. They are presented with confidence, and they cite each other rather than original research.

Second, the advice rarely separates AI accessibility from AI recommendation. Making sure a model can read your content is a different problem from understanding why that content gets cited or recommended. Conflating the two is how brands spend technical budget in places that pay diminishing returns.

Third, there is almost no reproducible, large-scale research on what the brands AI recommends have in common. The studies that exist are scattered across vendor blogs and academic preprints, and they often contradict one another. One meta-analysis found that nearly every headline claim about AI search, from conversion quality to zero-click rates to citation patterns, has at least one credible study on each side.

The result is a body of best practices built mostly on assumption. This report assembles what the record actually shows.

3. An initial observation

Our own investigation started with a simple exercise. We looked at a set of brands that leading AI systems kept recommending across product and service categories, including brands whose research the AI quoted and cited, and whose services it actively recommended.

Few of them had implemented the technical optimizations now sold as GEO best practice. Several had no AI-specific crawler configuration at all, and almost none appeared to use specialized GEO tooling.

What they shared was something else: heavy third-party validation. They had been discussed on Reddit, reviewed across major marketplaces, featured in editorial publications, referenced by creators on YouTube, searched for by name, and mentioned repeatedly across the wider web, usually several of these at once.

This is anecdotal, and we treat it that way. But it pointed at a hypothesis: perhaps AI recommendation systems reward brands that have already earned wide digital authority, more than websites that are technically tuned.

4. The checklist

llms.txt. In June 2026, Ahrefs analyzed server logs from 137,210 domains. Of the roughly 38,000 that had published a valid llms.txt file, close to a third of the sample, 97% received zero requests for it all month. The retrieval crawlers that actually power AI answers fetched the file only a few hundred times across the entire sample. Of the files that got any traffic at all, 96% of requests came from SEO audit tools. The GEO industry is auditing itself. Google’s own documentation says the file is not needed to appear in AI search features, and Google’s John Mueller had already compared it to the keywords meta tag, the canonical example of a self-declared signal that search engines learned to ignore.

Schema markup. Reports claim AI-cited pages are roughly three times more likely to carry structured data than uncited ones. But when Ahrefs ran a natural experiment, tracking 1,885 pages that added JSON-LD against 4,000 control pages, the schema produced no citation uplift on any AI platform. The correlation exists because well-run, authoritative sites tend to use schema, not because AI systems reward it. In follow-up tests, AI systems read the visible HTML during retrieval and ignored the structured data entirely.

Crawler directives. A March 2026 study of 4 million AI citations found that most news sites that explicitly block OpenAI’s crawlers still show up in ChatGPT’s citations. Roughly 70% of ChatGPT’s citation volume went to sites blocking its retrieval bots. AI engines can cite a page from search-index metadata without ever fetching it. So if blocking a crawler cannot remove you from AI answers, permitting one will not, on its own, put you in them.

Content tricks. The original academic GEO research, from Princeton, found that keyword stuffing performed at or below baseline in generative engines. A follow-up benchmark at NeurIPS found that most content tricks designed for AI answers were ineffective and often backfired, lowering the page’s ranking in LLM responses. A 2026 survey of the field found that across 54 tactic-and-domain combinations tested, only three produced a statistically significant gain.

5. What actually predicts recommendation?

Between 2024 and 2026, several independent teams, commercial and academic, measured what correlates with brand visibility in AI answers at scale, using different engines, different prompt sets, and different methods.

The largest study so far examined 75,000 brands across ChatGPT, Google AI Mode, and AI Overviews. The single strongest predictor of AI visibility was not a website metric at all. It was YouTube mentions, at a correlation of roughly 0.74, higher than any other factor tested, and nearly identical on ChatGPT, which has no corporate tie to YouTube. Close behind were third-party web mentions of the brand, at 0.66 to 0.71 depending on the platform.

The number of pages on a site correlated at only 0.19, and raw backlink counts at 0.22. An unlinked mention of your brand on someone else’s website predicts AI visibility roughly three times better than a backlink does. In the same data, the three strongest predictors for Google AI Overviews were brand mentions, branded anchors, and branded search volume, all off-site signals. Every on-site metric trailed them.

Independent work replicates the pattern. Seer Interactive ran 10,000 real consumer questions through GPT-4o and found that while page-one Google presence correlated strongly with being mentioned, around 0.65, backlinks came out “weak or even neutral,” against the researchers’ own expectations.

Academic work goes further. A 2026 study tracking 102 brands across five AI engines found that pre-existing brand stature dominated everything else: household-name brands appeared in 73% of relevant AI answers, mid-market brands in 44%, and niche brands in 11%. The same study found that ranked “best-of” listicles, third-party comparison content, were the single most-cited content format in AI answers, at roughly 21% of all citations. And when researchers analyzed 366,000 citations from real user conversations, they found AI systems route overwhelmingly to trusted, high-reliability sources: 90 to 96% of news citations across OpenAI, Google, and Perplexity models went to outlets with high credibility ratings, concentrated among a small set of established brands.

Visibility is also nonlinear. Brands in the top quartile for web mentions averaged 169 AI Overview appearances; the next quartile down averaged only 14. Authority compounds, and the recommendation economy is shaping up to be winner-take-most.

6. How AI systems decide

The correlations make sense once you see how these systems retrieve information. When a user asks an AI assistant for a recommendation, the model runs many searches in parallel.

Research found that AI systems issue an average of 9 to 11 sub-queries per prompt and fuse the results. Brands that appear across many of those result sets accumulate evidence. Brands optimized for a single page or query get filtered out.

For high-stakes queries, the AI explicitly generates sub-searches for third-party trust signals: reviews, credentials, endorsements. It trusts what the rest of the internet says about you over what you say about yourself.

This is why the most-cited domains in AI answers are not brand websites. Depending on the engine and the study, they are Reddit, YouTube, Wikipedia, LinkedIn, review platforms, and editorial publications: the surfaces where independent people evaluate brands in public. Review sites like G2 and Yelp show up disproportionately in recommendation-style queries.

Search engines have even written this dependency into contracts. Google licenses Reddit’s data for a reported $60 million a year, and OpenAI signed its own Reddit partnership in 2024. Community discussion has become paid infrastructure.

The system is volatile at the platform level. Semrush’s 13-week study of 230,000 prompts watched Reddit citations in ChatGPT collapse from roughly 60% of responses to 10% in about six weeks. Any strategy anchored to a single source is fragile, which is part of why platforms like Reddit work to keep their content trustworthy. But it is stable at the brand level: despite entirely different retrieval systems, ChatGPT, AI Mode, and AI Overviews surface largely the same brands, an output overlap of 0.78.

7. What technical work still matters

None of this says technical execution is irrelevant. It shows technical execution is the entry fee.

OpenAI’s and Anthropic’s crawlers do not execute JavaScript, so content rendered only on the client is invisible to them. AI crawlers waste roughly a third of their requests on broken URLs, so clean structure and accurate sitemaps matter. A brand absent from the search index cannot be reached by AI at all. Being accessible, indexable, and server-rendered is necessary. It is also only the floor.

Site accessibility is mostly binary. Past that threshold, the measured payoff of more technical investment is close to zero, with one exception. The Princeton GEO study found that on-page content changes could raise a source’s visibility in AI answers by up to 40% in controlled settings, by adding credibility markers: quotations, statistics, and citations. The tactics that worked made a page read like a credible primary source.

8. The next evolution of brand building

Together, the evidence points to a conclusion much of the GEO industry has been slow to admit, and a bigger picture it has missed.

Generative Engine Optimization is less a technical discipline with a marketing benefit than a brand discipline with a technical floor.

The signals that predict an AI recommendation are the signs of a brand people vouch for when nobody named it in the prompt. Recommendations cannot be configured or bolted onto a website. They are earned in public, over time, across surfaces the brand does not control.

The practical shift is away from schema audits and llms.txt tweaks, toward earned coverage, community presence, and becoming a brand customers seek out by name. There is a structural reason this is the right path. Content tactics like keyword stuffing are copyable by everyone, and research shows their gains shrink as they spread. Authority is different. A decade of reviews, coverage, and community trust cannot be copied by a competitor in a week, and that is the whole point. It is slower to build than any technical fix, and that is exactly why it holds.