Recommendation Science: The Emerging Battle to Become the Name AI Suggests

AI-generated recommendations raise a question beyond search visibility: what evidence makes a person or organization worth considering, and how accurately do systems represent that evidence?

Publisher note: Dr. Trudy Beerman founded Influence Media News and PSI TV and created REACHology®. This article discusses concepts she is developing and topics related to her commercial authority-building and visibility services.

Businesses have long sought visibility in search results. AI-assisted discovery adds another question: when someone asks whom to hire, interview, book or consider, which names appear in the answer? A generated shortlist can introduce candidates to a user, but inclusion should not be mistaken for verified competence or suitability.

Some generative search systems retrieve information and combine sources into a conversational response. Research on Generative Engine Optimization examines content visibility within those responses, while established recommender-system research examines how systems match users with items and information. These areas provide useful foundations for inquiry, but neither supplies a universal explanation of why a public AI assistant recommends a particular professional. Aggarwal et al.: GEO: Generative Engine Optimization

In this analysis, I use Recommendation Science as a proposed area of inquiry into how evidence, reputation, relationships, context and technological systems influence who or what becomes a suggested choice. The term describes the direction of my developing work. It does not claim a newly established academic discipline, ownership of existing recommender-system research or a validated formula for securing AI recommendations.

Recommendation Research Already Exists

Recommender systems are an established research area concerned with helping users navigate information and identify relevant options. Large language models have become part of that research, with proposed roles in interpreting information, representing preferences, ranking options and supporting conversational interaction. A survey by Jianghao Lin and colleagues examines their use across different stages of the recommendation pipeline. Lin et al.: How Can Recommender Systems Benefit from Large Language Models

A separate survey by Likang Wu and colleagues organizes research on language models for recommendation into discriminative and generative approaches. These surveys establish that recommendation is an active technical research subject. They do not demonstrate that publishing a particular number of biographies, interviews or articles causes an AI assistant to recommend their author. Wu et al.: A Survey on Large Language Models for Recommendation

Retrieval can also be incorporated into recommendation systems. For example, Jian Xu and colleagues’ RALLRec paper combines textual and collaborative information and introduces a reranking method to address changing user preferences. Its experiments concern the system and datasets described in the paper; they should not be generalized into a claim about how every commercial chatbot selects consultants, speakers or founders. Xu et al.: RALLRec

Visibility Is Different from Suitability

A website can appear in search without demonstrating that its owner is qualified for a particular assignment. A professional can have substantial experience while providing little accessible documentation of it. These examples illustrate a practical distinction: being discoverable and being appropriate for a user’s needs are different questions.

Consider a hypothetical comparison between two founders with similar professional ability. One has a brief biography and a basic service page. The other provides relevant qualifications, published work, documented results and independent coverage that readers can examine. The second founder gives an outside observer more material to evaluate, but that difference alone does not prove superior competence or a higher likelihood of recommendation by any particular AI system.

This distinction matters because a visibility strategy can improve documentation without improving the underlying service. Responsible authority-building should connect accurate evidence to actual capability. A recommendation that overlooks relevance, availability or the user’s constraints may be unhelpful even when the recommended person is well known.

The “Density of Digital Dots” as an Analytical Metaphor

I use the phrase density of digital dots to describe the collection of discoverable evidence associated with a person, organization or body of work. A “dot” might be a credential, article, interview, event record or documented result. This is an analytical metaphor, not a standardized metric or a proven ranking factor.

The usefulness of a dot depends on what it establishes. A license may support a claim about professional authorization. An interview may demonstrate how a person explains a subject. A case study may document a result, provided its methods and limitations are clear. None should automatically be treated as evidence for every claim the person makes.

Quantity also needs to be distinguished from independence. Ten copies of the same promotional biography do not constitute ten separate confirmations of expertise. A paid feature, an affiliated publication and independent reporting have different source relationships, which readers should be able to identify.

The practical objective is therefore a coherent and verifiable body of evidence. Whether greater evidence density changes AI recommendation frequency, under which conditions and by how much remains a research question. It should be tested rather than asserted.

Fragmented Identity Signals and the Person Behind the Brand

In my developing framework, Fragmented Identity Signals™ describes a documentation problem: relevant information about a person exists in separate places without clear connections. For example, a book may use one version of a name, an event biography another, and a current company page may omit the relationship between them. The concept is presented here as part of my own work, rather than a diagnosis validated by the cited technical papers.

Fragmentation can make verification more laborious for readers. It can also create practical identity questions, such as whether two records refer to the same person or whether an old role is still current. Correcting names, dates, affiliations and biographical links improves the documentation available for evaluation without guaranteeing how a search engine or AI system will respond.

For a founder-led business, this review should consider the person and the organization separately. A company may explain its services clearly while providing little evidence about the professional delivering them. Google’s people-first content guidance encourages authorship information and background that help readers evaluate expertise, making substantive author pages useful for publishers. Google Search Central: Creating Helpful, Reliable, People-First Content

Cited, Known and Recommended: Three Different Observations

My proposed framework distinguishes three observable outcomes: cited, known and recommended. These are categories for examining responses, rather than a sequence every system follows. An assistant may recommend a name without citing a source, or cite an article without recommending its author.

Cited

A response cites a source when it identifies or links that source in connection with an answer. The review must then ask whether the source supports the associated statement. A citation is observable attribution; it does not establish that the system has verified the author’s broader qualifications or that the source influenced every part of the answer.

Known

“Known” is shorthand here for a system producing an accurate description of an entity without receiving an extensive description from the user. It should not be interpreted as proof of durable model knowledge or inclusion in training data. The response may depend on retrieval, conversation context or other features of the system being tested.

Recommended

A response recommends an entity when it presents that person or organization as an option suited to the user’s request. The assessment should record whether the user supplied the name, what constraints were stated and whether the supporting explanation was accurate. Being offered as an option remains different from being selected, contacted or hired.

Keeping these outcomes separate makes measurement more informative. A change in citation frequency does not necessarily show a change in recognition, recommendation or customer demand. Each outcome needs its own definition and evidence.

An Unnamed Provider Is Not Proof of an Unclaimed Market

An AI response that offers general advice without naming a provider should be interpreted cautiously. It could reflect an underspecified question, limited available information or the behavior of the system in that interaction. The response alone does not establish that a category is commercially unclaimed or that the first visible founder will become its preferred recommendation.

A more defensible investigation repeats clearly documented questions and examines how responses change across systems and conditions. It also compares AI output with the actual market. An established provider’s absence from one answer does not mean the provider lacks expertise, customers or a strong reputation.

Trust Requires More Than a Name in an Answer

The National Institute of Standards and Technology’s AI Risk Management Framework identifies characteristics of trustworthy AI, including validity and reliability, accountability and transparency, privacy, and fairness with harmful bias managed. Its Core organizes risk-management activities around govern, map, measure and manage. These resources concern responsible AI development and use, not a marketing checklist for earning recommendations. NIST: AI Risks and Trustworthiness; NIST: AI RMF Core

Google’s guidance addresses helpful and reliable search content. It explains experience, expertise, authoritativeness and trustworthiness, while clarifying that E-E-A-T is not itself a specific ranking factor. Quality raters evaluate search performance; their judgments do not directly determine the rankings of individual pages. This guidance should not be presented as a universal rulebook for recommendations across AI products. Google Search Central: Creating Helpful, Reliable, People-First Content

A biblical reflection adds an ethical perspective on corroboration. Proverbs 27:2 counsels allowing others to offer praise rather than relying on self-praise. Applied here as a principle, it encourages humility about our claims and attention to what others can substantiate. It is not empirical evidence about how recommendation systems operate. Proverbs 27:2, ESV

What Recommendation Science Could Investigate

The value of the proposed inquiry lies in questions that can be investigated, including results that might challenge the framework. Rather than assume that more visibility produces more trust, a study should define the outcome, compare observations and report limitations.

• How do relevant credentials and documented results affect human evaluation?
• Does correcting inconsistent identity information change the accuracy of AI descriptions?
• How do independent coverage, paid coverage and owned content differ in citation patterns?
• How sensitive are recommendations to wording, geography, language and user constraints?
• Are qualified but less visible candidates systematically overlooked?
• Do changes in documentation correspond with changes in citation or recommendation frequency?
• When does an AI recommendation lead to an inquiry or a purchase?

These are research questions, not established conclusions. Observing that a recommendation changed after a new publication appeared does not prove the publication caused the change. The platform, available sources, prompt or user context may also have changed.

Measuring Without Overclaiming

A transparent assessment should record the platform, available model information, date, exact prompt and whether browsing or retrieval was enabled. It should also document relevant conversation context and preserve the responses being evaluated. Repeated observations are preferable to treating one favorable answer as a permanent result.

Reporting should explain how many tests were conducted and what counted as a citation, accurate identification or recommendation. It should include errors and absences as well as favorable mentions. Comparisons are more useful when questions and evaluation criteria remain consistent.

Google notes that conventional SEO practices remain relevant to AI Overviews and AI Mode, with no special optimization or additional eligibility requirements necessary. This is a useful reminder that an experimental visibility framework should not be marketed as a mandatory technical requirement for appearing in Google’s AI features. Google Search Central: AI Features and Your Website

PSI TV as an Operating Context for Inquiry

My work with PSI TV provides an operating context for questions about how experts present their knowledge and distribute supporting material. Interviews and published content can provide records of what a person says, how they explain a subject and what experience they claim. Those records can then be examined for accuracy, clarity and usefulness.

That practical setting is not equivalent to an independent research laboratory. Because I own PSI TV and provide related services, any evaluation involving my business requires clear disclosure and methods that allow others to assess the findings. This article reports no verified citation ranking or measured recommendation improvement for PSI TV.

Operational experience can help generate questions, but commercial interests should not determine the answers. A credible study must be able to report that an intervention had no apparent effect, that a competitor performed better or that the available evidence was insufficient for a conclusion.

Founder Focus Is a Strategy, Not a Guaranteed Advantage

A founder-led business can concentrate its documentation around a clearly defined area of expertise. Accurate biographies, useful explanations and verifiable examples may make that expertise easier for a prospective customer to assess. This is a practical strategic option, not proof that smaller businesses outperform larger organizations in AI recommendations.

The aim should be to document legitimate value within a relevant category. Claims that narrow positioning, content volume or consistent descriptions will overcome every visibility disadvantage require evidence. A useful strategy remains accountable to its results.

The Question Beyond Visibility

Recommendation reaches beyond artificial intelligence. Human referrals, editorial source selection and institutional decisions also involve judgments about relevance, evidence and trust. A broader inquiry can compare those environments while respecting their differences.

The central question is therefore larger than whether a machine says a name: what makes that name a defensible choice for this particular need? For founders and experts, that means making claims that can withstand examination, providing supporting evidence and correcting inaccuracies wherever they appear.

Becoming visible creates an opportunity to be considered. Becoming recommended creates an opportunity to be evaluated more closely. The responsibility that follows is to ensure the evidence supports the introduction.

---

Reporting Basis

This editorial analysis draws on research about language models and recommender systems, Google Search Central documentation and NIST’s AI Risk Management Framework. Recommendation Science, the density of digital dots, Fragmented Identity Signals™ and the Cited/Known/Recommended distinctions are presented as the author’s developing analytical concepts. The cited sources do not independently validate those concepts.

The founder comparison is hypothetical. No original experimental results, verified PSI TV citation ranking or client performance claim is reported. The biblical reference is an ethical reflection rather than technical evidence.

Sources

• Lin, Jianghao, et al. How Can Recommender Systems Benefit from Large Language Models: A Survey . First submitted in 2023; revised in 2024. The arXiv record reports acceptance by ACM Transactions on Information Systems.
• Wu, Likang, et al. A Survey on Large Language Models for Recommendation . First submitted in 2023; revised in 2024.
• Xu, Jian, et al. RALLRec: Improving Retrieval Augmented Large Language Model Recommendation with Representation Learning . 2025. The arXiv record reports acceptance as a short paper at The Web Conference 2025.
• Aggarwal, Pranjal, et al. GEO: Generative Engine Optimization . First submitted in 2023; accepted to KDD 2024.
• National Institute of Standards and Technology. AI Risks and Trustworthiness .
• National Institute of Standards and Technology. AI Risk Management Framework Core .
• Google Search Central. Creating Helpful, Reliable, People-First Content .
• Google Search Central. AI Features and Your Website .
• Proverbs 27:2, English Standard Version .