Cited vs. Recommended: The New AI Search Authority Gap

The citation that doesn’t convert

For most of SEO history, getting cited was the goal. You ranked, you earned the click, and the click turned into revenue. AI search breaks that chain in the middle.

A study by SEO director Lily Ray, published June 17, 2026, tracked 100 “best [category] software” queries in Google AI Overviews and Google AI Mode across three checkpoints in April, May, and June. Self-promotional listicles — the “we ranked ourselves number one” articles that became a popular tactic for influencing AI answers — were cited 323 times across those queries. But in 224 of those cases, roughly 69%, the brand that published the listicle was left out of the actual recommendation. Google’s AI surfaces cited the content but pointed buyers somewhere else.

That gap now has a name. “Google appears to have decoupled what it cites from who it recommends,” Lily Ray wrote. The recommendations consistently went to established category leaders — the brands whose entity authority was already strong, regardless of who published the listicle.

For marketplace sellers, this is the uncomfortable truth of AI search: a citation is no longer a recommendation, and the recommendation is what closes the sale.

What Google’s patent reveals about entity profiles

The reason behind the gap is becoming visible. A Google patent — “Data extraction using LLMs” (WO2025063948A1), filed September 2023 by inventors Aarthi Ramachandran and Nidhi Gupta, assignee Google LLC — describes how Google’s AI builds what it calls a “deep, holistic characterization of a particular entity.”

The patent is not about indexing pages. It is about building a profile of a business or brand by collecting information from the entity’s own website and supplementing it with third-party data, then using a large language model to interpret and synthesize everything into an entity summary. That summary takes the form of a hierarchical graph — a weighted network of nodes and edges representing the entity’s attributes and its relationships to other entities.

Search Engine Land published an analysis of this patent on June 22, 2026, framing its core implication for SEO in a single line: the new goal is “teaching AI who you are,” not just optimizing individual pages.

This matters because AI recommendations are not assembled from the page that ranks highest. They are assembled from the entity profile the model has already built. If your entity profile is thin, inconsistent, or contradicted by your own behavior, no amount of on-page optimization will land you in the recommendation.

Citation versus recommendation: the 69% gap

Lily Ray’s study makes the patent tangible. Here is how the gap works in practice.

A seller publishes a “Best Etsy Tools for 2026” listicle and ranks their own product at the top. Google’s AI reads the listicle, finds it useful as a source, and cites it in the AI Overview. But when the model assembles its actual recommendation — the brands it tells the user to buy — it draws on the entity profiles it has already built for each tool. If your brand’s entity profile is weak (few third-party mentions, inconsistent attributes, no corroborating reviews), the model recommends the established competitor instead. Your listicle earned the citation. Your competitor earned the recommendation.

The numbers back this up. Across the three-month tracking period, 74 of the 100 queries returned an AI Overview that cited at least one self-promotional listicle. In 69% of those citations, the self-promoter was excluded from the recommendation. The same study cites research showing that when Google produces an AI summary, users click a link inside that summary in just 1% of visits. So even when you win the citation, almost no one clicks it. The recommendation is where the value concentrates.

How AI engines build your entity profile

The patent specifies which attributes Google’s LLM extracts when it builds an entity characterization:

• Presence — whether the entity exists online, in person, or both
• Age — how long the entity has existed
• Principles — the stated values, positioning, and category
• Items and services referenced — what the entity actually sells or does
• Reputation and social media sentiment — how others describe it

The system then supplements this with external signals: maps data, job listing data, and business information from third parties. Critically, the patent states the LLM can extract this content “without requiring the content of the web page to include specific markup language.” Structured data still helps, but Google’s AI no longer depends on it — the model can read unstructured content and infer the entity attributes on its own.

The patent also describes a feature the AI can flag: a “brand/service mismatch.” If a law firm’s actual business is civil contract services but its brand, reputation, and advertising focus on corporate mergers and acquisitions, the AI notices the gap. For marketplace sellers, this is a direct warning: if your listing copy claims one thing but your reviews, third-party mentions, and product attributes tell a different story, the AI weights the corroborated version — not your marketing.

The shift from markup to meaning

The practical implication is a shift in where effort pays off. Traditional SEO rewarded individual pages: optimize the title tag, fix the schema, build links to that URL. AI search rewards the entity: the consistent, corroborated picture of who you are across every source the model can read.

This is why raw citation counts have become a misleading metric. A high citation count tells you the AI finds your content useful as evidence. It does not tell you the AI considers you authoritative enough to recommend. Bing’s team made this point explicit on June 16, 2026, when Microsoft launched Citation Share in Bing Webmaster Tools — a metric defined as “the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query.” Microsoft explicitly called it “an observational metric — not a ranking system or a competitive scoreboard.” The point is to see how concentrated or fragmented the citation space is for a query, not to treat raw count as a victory.

How to build entity authority that earns recommendations

If citations are the symptom and entity authority is the cause, the playbook changes. Here are the concrete steps.

Audit your entity profile across sources. Search for your brand name on Google, Bing, ChatGPT, and Perplexity and read what comes back. Note the attributes each engine has associated with you: what you sell, who you serve, how you are described. Where the descriptions disagree, that is your first fix.

Eliminate brand and service mismatches. If your listings emphasize features your reviews never mention, or your third-party mentions place you in a different category than your own pages, the AI weights the corroborated story. Make your self-authored content, your reviews, and your third-party mentions tell the same story about what you are and who you are for.

Build corroborating third-party signals. The patent supplements your own pages with maps data, business information, and external sources. Reviews on the marketplaces you sell on, mentions in buyer forums, and consistent category attribution all feed the entity profile. The more independent sources agree about who you are, the higher the model’s confidence in your entity.

Define your attributes explicitly and consistently. State what you make, who it is for, and what category you belong to — in the same terms, across every surface the AI reads. The patent extracts “principles,” “items referenced,” and “services referenced.” When those are stated identically on your store, your marketplace listings, and your third-party profiles, the model’s confidence in your entity rises.

Track citation share, not just citation count. Use Bing Webmaster Tools’ Citation Share to see whether you are gaining or losing share of the citation space for your category. A rising citation count alongside a falling citation share means more sources are being cited alongside you — a sign your entity authority is not translating into recommendation dominance.

How FirstShelf can help

Entity authority is one of the five signals in the FirstShelf Score — alongside semantic density, structure quality, platform fit, and visual proof. The free GEO audit reads your marketplace listing the same way an AI model does, checking whether your title, description, attributes, and images tell a consistent story about what you sell and who it is for. When your listing claims features your reviews do not corroborate, or your category attribution does not match how third parties describe you, the audit flags the mismatch — the same brand and service mismatch Google’s patent describes.

To see how your listings score on entity authority and the other signals AI engines use to decide who to recommend, run a free FirstShelf GEO audit at firstshelf.ai. The dashboard then tracks those scores over time, so you can see whether the consistency work is actually moving the needle on your AI visibility.