How Query Fan-Out, Entity Understanding, and AI-Generated Recommendations Are Changing the Relationship Between Search Results and Customer Decisions
For more than two decades, ranking has been one of the primary ways we’ve understood search performance. We tracked where documents appeared, how often they were displayed, whether users clicked them, and what happened after those visits. Those measurements provided a useful picture of how search engines connected people with information, but they were built around a relatively straightforward relationship between queries, search results, and website visits. With search experiences changing, that relationship is now becoming much more complicated.
Google’s AI Mode, Gemini, Ask Maps, ChatGPT, Claude, Perplexity, and other AI search experiences are taking on more and more activities that previously required several searches, website visits, and comparisons. They can interpret complex questions, investigate different aspects of a problem, gather information from multiple sources, and present recommendations based on the user’s requirements. A business might rank highly in a search engine for a relevant keyword without appearing in an AI search response, while another might enter the conversation because its expertise, reputation, or other characteristics align more closely with the user’s circumstances. Even when the same businesses appear, the language used to describe them can influence how they are evaluated.
Search is evolving beyond ranking relevant information to helping users determine which options deserve consideration. Ranking remains an important component of information retrieval, but it doesn’t fully explain how organizations become part of a recommendation, how supporting information is evaluated, or why certain alternatives are presented as more appropriate than others. Understanding this change requires examining how search systems retrieve information, understand entities, gather evidence, compare alternatives, and construct recommendations.
Search has been moving beyond keywords for years
Google’s Knowledge Graph, introduced in 2012, represented an important move toward understanding entities and their relationships rather than relying exclusively on matching words in documents. Featured snippets, product results, local packs, knowledge panels, and review summaries subsequently brought more information directly into the search experience. These developments gradually expanded the work search engines performed between receiving a query and presenting results. Generative AI extends that progression by making it possible to synthesize information from different sources and respond to more complicated requests conversationally.
Consider someone searching for “managed IT services Worcester.” The query identifies a service category and a location, allowing traditional search results to help the user discover companies that may provide those services. A more detailed request introduces additional requirements: “Our Worcester-area business has outgrown its IT provider. We have about 40 employees, need stronger cybersecurity and Microsoft 365 support, and can’t afford significant disruption when switching providers. Which local firms should we consider?” The user still wants a provider in a particular geographic market, but they also want evidence of experience with similarly sized organizations, technical capabilities, responsiveness, and a manageable transition process.
A useful response needs to identify companies that appear capable of meeting those requirements and explain the basis for that assessment. Information about a company’s migration process, cybersecurity capabilities, customer experience, or industry specialization may contribute to a recommendation even when the documents containing that information don’t rank prominently for the original service keyword. Rather than simply returning businesses associated with a category, search systems can combine different kinds of information to help users evaluate which organizations may be suitable for their circumstances.
Query fan-out expands search beyond the original question
One of the most consequential technical developments is query fan-out, which allows search systems to investigate multiple aspects of a complex question. Instead of relying exclusively on the language of the original query, a system can conduct additional searches related to the underlying information needs, gathering material that addresses different components of the question. This makes it possible for documents that are not particularly relevant to the original keyword to contribute useful information to the final response.
In 2024, I collaborated with Laurence O’Toole at Authoritas on a study of how Google AI Overviews select supporting links. At the time, much of the research into AI Overviews compared cited URLs against organic results for the exact keyword entered by the user. I believed that approach overlooked an important part of Google’s underlying retrieval process because Google’s patent on generative summaries described the possibility of expanding retrieval beyond documents responsive to the original query, including material related to additional or reformulated queries.
Our research examined 11,163 queries, of which 2,358 triggered AI Overviews. We compared the URLs cited in those summaries against first-page organic results for the original query, related queries, and reformulated versions of the query. Only 46.3% of cited URLs could be accounted for through first-page organic results for the original query. Adding related queries increased that figure to 60.4%, while including both related and reformulated queries brought the overlap to 67.3%. Higher-ranking documents were also generally more likely to be cited across the different query sets, indicating that conventional organic rankings continued to have an important relationship with source selection.
Our results didn’t establish the exact retrieval operations Google performed internally. The related and reformulated queries were research approximations rather than records of Google’s internal searches. They did, however, demonstrate that examining a broader set of relevant queries explained substantially more of the documents appearing in AI-generated summaries than looking exclusively at the original keyword.
In March 2025, Google formally introduced AI Mode and described its use of query fan-out. The system could issue multiple searches across related subtopics and information sources, then combine those findings into a response. Google expanded on that explanation during Google I/O in May 2025, describing how AI Mode breaks a question into subtopics and conducts multiple related searches simultaneously. Google’s documentation established that a complex search experience can involve a collection of retrieval operations rather than a single query and its associated results, reinforcing the broader retrieval behavior our earlier research had been investigating.
A document may contribute to an answer because it satisfies one part of a larger information need. A page explaining how an IT provider handles client migrations might be more useful to a particular recommendation than its general managed services page, while another company’s cybersecurity documentation could become important when the user expresses specific security concerns. The original keyword remains relevant, but it no longer defines the full set of information involved in answering the question. Retrieval can extend across multiple topics, requirements, and sources that collectively contribute to a more complete response.
Document relevance and decision relevance
Document relevance concerns how well a particular document satisfies an information need, while decision relevance concerns how useful that information is when evaluating a specific choice. A business might have a highly relevant service page while offering little evidence about the qualifications that matter most to the customer. Another business might provide extensive documentation about a particular specialization that becomes important when the search includes additional requirements. Query fan-out allows these different types of information to participate in the same response, even when individual documents address only one part of the customer’s larger question.
The information retrieved also needs to be associated with the correct organization and interpreted in the context of the user’s requirements. Finding a page that describes cybersecurity capabilities doesn’t automatically establish that its publisher is the best option for a company concerned about security. The system must identify what the information establishes, determine which organization it describes, and incorporate those findings into an answer that addresses the user’s situation. This creates a more complicated relationship between the documents retrieved and the entities ultimately recommended.
Entity understanding connects information to the organizations it describes
Search engines have long worked to identify entities and distinguish them from the words used to describe them. Generative systems introduce additional ways to combine information about entities into more complete representations, drawing on descriptions, attributes, relationships, and other information distributed across different sources. The challenge becomes particularly interesting when the system needs to understand an organization’s capabilities rather than simply recognize its name, location, or service category.
In June 2026, I examined a Google patent in Google’s LLM patent suggests a new goal for SEO: Teaching AI who you are. The patent describes a method through which a language-model-based system could examine information from multiple sources, extract attributes and relationships, and synthesize a broader characterization of an entity. Although a patent describes a potential technical implementation rather than confirming what Google currently deploys in production Search, the proposed architecture illustrates one possible approach to developing an understanding of an organization from information distributed across the web.
Consider a professional services firm operating in several markets. Its website may describe its service offerings and office locations, while individual staff profiles document professional qualifications. Industry directories might establish affiliations or specializations, published articles could provide evidence of expertise, and customer reviews or case studies might describe the kinds of problems the firm has addressed. These sources contain different information, but together they can contribute to a more complete representation of the organization and the circumstances in which it may be an appropriate choice.
A system attempting to recommend firms for a particular assignment needs to connect relevant facts to the correct organizations while distinguishing general characteristics from specific qualifications. It might correctly identify a company as an accounting firm without recognizing that it specializes in complex multi-state tax obligations. Similarly, it might recognize an IT provider’s cybersecurity services without finding evidence that the company works with organizations subject to particular regulatory requirements. In both situations, the organization can be recognized while its suitability for a specific customer remains uncertain.
I explored this problem further in my July 2026 Search Engine Land article, How to audit your AI entity footprint. The analysis focused on how AI systems describe organizations, including their identities, differentiators, expertise, and relationships. One exercise involved asking different systems to explain a business without initially providing its website as the primary source. Sometimes the resulting descriptions accurately captured the organization’s specializations and distinguishing characteristics, while other responses were generic, incomplete, or inconsistent with how the business understood itself.
These observations demonstrate that AI systems can produce different representations of the same organization, with varying levels of specificity and accuracy. A system attempting to compare organizations must work with the information it can access and interpret, and incomplete representations can limit its ability to explain why a particular business may be appropriate for a specific need. We don’t yet have sufficient evidence to establish a direct causal relationship between richer entity representations and increased recommendation frequency, but the quality and completeness of those representations are important subjects for further investigation.
Recommendations change what it means to appear in search
The movement from retrieval to recommendation is particularly visible in local search, where a customer’s needs frequently involve a combination of location, service requirements, availability, reputation, and other considerations. Traditional local search results have long incorporated some of these factors through business listings, ratings, reviews, and local packs. Conversational interfaces can bring more of that information into a generated explanation that attempts to identify appropriate businesses for the customer’s circumstances.
In April 2026, I published Google Ask Maps is moving from listings to recommendations, based on hands-on testing across plumbing, electrical, and HVAC searches. The research examined how Ask Maps responded as prompts moved from broad service categories toward more specific customer problems and decision criteria. Even relatively straightforward requests produced more than conventional business listings, with Ask Maps summarizing options, interpreting aspects of user intent, and describing businesses according to characteristics it considered relevant.
As the prompts became more situational, the system increasingly positioned businesses around particular needs. Responses discussed characteristics such as responsiveness, specialized experience, honesty, and whether a provider might prioritize repairs over replacements. With more complex questions, Ask Maps sometimes provided guidance about the problem itself before recommending businesses. Portions of the evaluation customers previously conducted through reviews, websites, and multiple searches were being incorporated into the search response.
Google’s March 2026 introduction of Ask Maps describes a conversational recommendation experience drawing on information about more than 300 million places and contributions from over 500 million Maps contributors. In May, I followed up with Google Ask Maps: How to optimize for visibility, examining how businesses were represented in responses to different customer needs. One of the more interesting observations was how the explanations accompanying recommendations contributed to the positioning of individual businesses.
A business might appear because of its proximity, but its recommendation could emphasize specialized experience, reputation, or suitability for a particular problem. Another might be included as a general alternative without any clear differentiation. Both businesses appeared in the answer, but the system presented them differently, giving the user different information about why each organization might deserve consideration.
Inclusion, positioning, and preference
An organization can be included in an AI-generated response without receiving particular emphasis. It can also be presented as a strong candidate for a specific requirement, compared with alternatives, or described with qualifications and limitations. The language accompanying a recommendation can influence how the user understands the available choices, and the same collection of businesses may create different impressions depending on how the system explains each option.
Measuring whether a business was mentioned captures only one aspect of its presence. We also need to examine how it was characterized, which customer needs it was associated with, what evidence supported the characterization, and how it was presented relative to competitors. Identifying consistent relationships between customer requirements and recommended organizations could help explain how selection-oriented search experiences differ from conventional ranked results.
Evidence gathering is becoming part of the recommendation process
A search system attempting to recommend businesses faces a more complicated problem than identifying providers within a service category. Finding IT companies in Worcester is relatively straightforward compared with establishing which ones have demonstrated experience handling complicated provider transitions. A company can claim to provide cybersecurity services without establishing the depth of its capabilities, while a website might promote fast response times without providing evidence of what customers can reasonably expect. Similarly, a business may describe itself as experienced with regulated organizations without documenting the industries or requirements involved.
Different sources can help substantiate different kinds of claims. A service page can describe an offering, while a case study might document an implementation. A staff profile can establish professional qualifications, a partner directory may confirm a relationship, and reviews can provide customer perspectives. Each source has limitations, including the possibility that information is incomplete, subjective, outdated, or insufficiently specific to the customer’s circumstances.
A business with extensive online documentation may be easier for a search system to describe than a similarly qualified competitor with relatively little public information. That doesn’t establish that the first business is better. It means the system may have access to more information from which to construct an explanation of its qualifications. AI-generated recommendations may therefore partly reflect differences in publicly accessible documentation rather than comprehensive assessments of the organizations being compared.
There is also no guarantee that retrieved information will be interpreted correctly. In March 2025, Columbia University’s Tow Center for Digital Journalism published AI Search Has a Citation Problem, an investigation into the ability of eight generative search systems to identify and cite news sources accurately. The researchers documented fabricated links, incorrect source attributions, and failures to acknowledge uncertainty when the systems couldn’t confidently establish an answer.
The study examined news retrieval rather than business recommendations, so its findings shouldn’t be generalized as a measured failure rate for commercial search. It does demonstrate that retrieving information, citing a source, and accurately verifying a claim are separate activities. A citation may point to a real document without that document supporting everything stated in the generated response. Recommendations can consequently appear well-supported while relying on incomplete information, mistaken interpretations, or claims the system hasn’t independently verified.
As users delegate more of their research to AI systems, the reliability of those evaluations becomes increasingly consequential. A generated recommendation may save someone considerable time investigating alternatives, but its apparent confidence shouldn’t be confused with evidence that every relevant qualification has been independently confirmed.
Comparison moves part of the evaluation process into search
A recommendation becomes more useful when it addresses the circumstances behind the user’s decision. Someone researching managed IT services may initially want to identify available providers, but subsequent questions might focus on organizational size, cybersecurity experience, Microsoft 365 support, migration procedures, cost, responsiveness, and risk. Traditionally, the user might investigate those questions separately, gathering information from company websites, reviews, directories, and conversations with potential vendors.
Conversational search can bring portions of that research into a single interaction. Google describes AI Mode as particularly useful for complex, multi-part questions and comparisons, with query fan-out allowing different aspects of a request to be investigated. Similar capabilities appear in OpenAI’s shopping research, which supports refining preferences and comparing potential purchases. These capabilities extend beyond local services wherever a decision involves multiple criteria, including software purchases, healthcare services, professional advisors, travel arrangements, and consumer products.
Generated comparisons can simplify research for the user while obscuring aspects of how the alternatives were evaluated. We generally cannot determine from the final response which organizations were considered before the recommendations were generated, how competing criteria were weighted, or whether the system had sufficient evidence to assess every relevant option. An organization might be excluded because it doesn’t meet the user’s requirements, but it might also be absent because the system failed to retrieve relevant information, misunderstood its specialization, or selected other candidates.
A well-organized comparison doesn’t establish that the underlying evaluation was comprehensive. Search systems can produce plausible explanations for why particular options are recommended, but those explanations aren’t necessarily transparent accounts of the complete retrieval and evaluation process. Studying selection-oriented search requires examining both the recommendations presented and the limitations of the evidence available to produce them.
Organic rankings still matter, but they explain less of the complete experience
Developments in AI search are frequently framed as a competition between traditional rankings and generative recommendations, but the evidence suggests a more complicated relationship. Our 2024 Authoritas research found a meaningful association between organic rankings and documents linked in AI Overviews, with higher-ranking documents generally more likely to be cited across the original, related, and reformulated query sets. Ranking remained an important part of understanding source selection, although looking exclusively at the original query missed a substantial portion of the observed overlap.
More recently, Ahrefs analyzed 15,000 prompts in its August 2025 study, Only 12% of AI Cited URLs Rank in Google’s Top 10 for the Original Prompt. The researchers found that, on average, approximately 12% of URLs cited across the AI assistants they examined appeared in Google’s top ten results for the same prompt. Perplexity showed greater overlap than several other assistants. The study also contrasted these findings with separate research showing substantially higher overlap between Google AI Overview citations and Google’s conventional results.
The studies used different systems, samples, and methodologies, so their percentages shouldn’t be treated as directly comparable measurements of one underlying process. They do indicate that the relationship between conventional rankings and AI citations varies across search experiences. A document may be retrieved through a related query, an alternative search system, or another information source, while the cited document may differ from the material used to identify or initially evaluate an entity.
Citation overlap also doesn’t measure recommendations. An organization might be recommended without its website receiving a prominent citation, while another might have several documents cited without receiving any particular endorsement in the response. Ranking, retrieval, citation, entity inclusion, and recommendation are related outcomes, but examining them individually provides a more complete view of how search experiences represent information and organizations.
The customer journey is becoming harder to observe
As more discovery and comparison occur inside search experiences, the traditional relationship between rankings, clicks, and customer behavior becomes more difficult to interpret. In January 2025, I examined this problem in Measuring the Impact of Google AI Overviews on SEO Performance. The analysis explored how Google Search Console combines conventional search and AI Overview performance data, making it difficult to isolate the effects of appearing in AI-generated summaries.
The measurement problem becomes more substantial when search systems help users compare alternatives before visiting a website. Someone might receive a list of recommended providers, investigate two through follow-up questions, and then search directly for the organization they prefer. Another user might encounter a negative or incomplete description of a business and remove it from consideration without clicking anything. In both cases, the search experience could influence the decision without generating an attributable visit from the original query.
Independent behavioral research suggests that AI-generated answers are already associated with different patterns of search activity. In July 2025, Pew Research Center reported that users clicked traditional search results in 8% of visits involving an AI summary, compared with 15% of visits without one. Links appearing within AI summaries themselves received clicks in just 1% of visits, and users were more likely to end their browsing session after encountering an AI summary.
The study was observational and didn’t establish that AI summaries independently caused these differences. Searches generating AI summaries may differ in important ways from those that do not, but the findings illustrate how conventional search-result clicks capture only part of the interaction occurring within AI-supported search experiences. For researchers, the measurement challenge extends beyond whether AI search sends more or less website traffic to include how organizations enter a user’s consideration set, how they’re described, and whether those descriptions influence subsequent research or action.
Understanding search performance requires a broader framework
If search systems increasingly participate in the selection process, we need ways to investigate their behavior beyond ranking positions and citation counts. I think six dimensions provide a useful starting point for analyzing the relationship between information retrieval, entity understanding, recommendations, and observable user behavior.
Retrieval concerns what information is available to the system and which documents or data sources it can access in response to a question. Entity recognition examines whether the system correctly identifies an organization, its services, locations, specializations, and relationships. Evidence considers what information is available to substantiate the qualifications relevant to the user’s requirements.
Representation examines how the system describes an organization and how accurately that description reflects the available evidence. Selection concerns which organizations are recommended, how they’re positioned, and what changes when the user’s requirements become more specific. Response considers what observable actions follow, including clicks, subsequent searches, direct visits, and other measurable customer behavior.
These dimensions don’t describe a universal internal architecture used by AI search engines. Different systems employ different methods, and substantial portions of their retrieval and generation processes aren’t publicly observable. The framework is intended to separate questions that often become conflated when evaluating AI search performance.
An organization might be recognized accurately but excluded from a recommendation. Its service offering might be retrieved but described incorrectly, or it could be recommended frequently while receiving relatively few direct citations. Each outcome requires a different kind of investigation, and counting brand mentions or citations alone doesn’t necessarily reveal how an organization is being evaluated or positioned.
What we still need to investigate
Several important questions remain unresolved. We have documented evidence that systems such as Google’s AI Mode can decompose questions and retrieve information across multiple subtopics. We also have observations demonstrating that conversational search experiences can change their recommendations as users supply more detailed requirements. What we don’t have is a reliable, universal account of how retrieved information becomes a final recommendation.
How does a system determine which organizations are plausible candidates? How does it distinguish relevant experience from broad marketing claims? What happens when necessary evidence is unavailable or contradictory? How consistently does it recommend the same organizations when the same question is asked repeatedly? These questions become more complicated when comparing systems because ChatGPT, Gemini, Perplexity, Claude, and Google Maps don’t necessarily retrieve information from the same sources, interpret questions identically, or construct recommendations using equivalent methods.
Differences in their responses may reflect retrieval availability, entity recognition, source selection, system instructions, model behavior, or other factors that aren’t directly visible. Controlled testing can help identify patterns in observable recommendation behavior, particularly when researchers examine how the same category of businesses is represented across increasingly specific prompts. A test might begin with general discovery and progress through qualifications, decision constraints, and evidence verification, recording changes in recommendations, descriptions, and supporting sources across repeated sessions.
Such experiments wouldn’t reveal the complete internal mechanisms of the systems, but they could help distinguish consistent behavior from isolated examples and develop more defensible hypotheses about the factors associated with recommendations. They could also provide a more useful basis for evaluating AI search performance than assuming a single prompt produces a representative or stable result.
The next question is no longer just where something ranks
The evolution of search isn’t a straightforward replacement of ranking with recommendation. Conventional search remains a significant part of how information is discovered and retrieved, rankings still influence which documents are encountered, and the technical foundations of search continue to matter. What has expanded is the amount of interpretation that can occur between retrieving information and presenting an answer.
Query fan-out expands the information considered, while entity understanding connects facts and relationships to identifiable subjects. Evidence gathering provides material that may support descriptions and comparisons, and generative systems can combine those elements to produce recommendations tailored to the user’s needs. Relevance to a search and suitability for a recommendation are related, but they aren’t necessarily determined by the same information.
This is where the progression of my own research has led. The 2024 AI Overview study investigated how sources appearing in generated answers related to traditional search results. Subsequent work on search intent examined why different kinds of questions produce different search experiences. The 2026 Ask Maps research explored how increasingly specific customer needs could change business recommendations, while my work on entity understanding and AI entity footprints examined how systems construct and communicate their understanding of organizations.
Together, these investigations point toward a broader research problem: understanding how search systems move from identifying information to evaluating and presenting choices. The research is still developing, and generated explanations shouldn’t be mistaken for transparent evidence of internal decision-making. We can observe which organizations appear, how they’re described, which sources are cited, and how responses change across questions and systems, but substantial portions of the process remain inaccessible to outside researchers.
A ranking indicates where information appears within a set of results. A recommendation introduces judgments about which alternatives are worth considering and why. As search systems take on more of that evaluative work, understanding search will require studying how information is retrieved, how organizations are represented, how alternatives are compared, and how the resulting recommendations influence the choices available to users.
Leave a Reply