And why your marketing dashboard cannot see the evaluation happening
AI systems do not use one universal recommendation formula, but they can only surface a business when they can understand what it is, connect it to the buyer’s question, locate credible information about it, and use that information to construct an answer.
That evaluation may happen before the buyer visits your website, clicks a result, or appears anywhere in your marketing analytics. By the time the buyer reaches you, an AI system may already have formed and communicated a view of who you are, what you do, and whether you belong in the decision.
That is why a business can have healthy traffic, active social channels, strong search rankings, and no clear idea whether it is being understood or recommended when buyers ask AI systems for guidance.
TL;DR
- Understand, verify, synthesize, and surface provide a practical model for the work behind an AI-generated recommendation not a universal description of every platform’s proprietary process.
- Five conditions improve a business’s ability to be represented accurately: entity clarity, question alignment, credible evidence, technical accessibility, and a consistent public record.
- Most conventional analytics measure what happens after visibility or a visit. They do not fully show how an AI system evaluated the business before the click.
- Appearing once is a useful observation. Reliable recommendation must be evaluated across questions, systems, contexts, and time.
A Practical Model of What Happens Before the Answer
A reader asked an important question after the first article in this series: “What signals do you think AI will rely on most when deciding what’s worth recommending?” The most useful answer is not a secret list of ranking factors. No single list could describe every model, search product, retrieval system, or prompt context.
A better way to understand the problem is through four operations that a useful recommendation typically requires:
Understand: interpret the buyer’s question and recognize the businesses, people, services, and problems connected to it.
Verify: look for information that supports, qualifies, or conflicts with the available claims when retrieval and sources are available.
Synthesize: combine relevant information into an answer that responds to the buyer’s situation rather than merely repeating a page.
Surface: include, describe, cite, or recommend the businesses the system determines are relevant enough to mention.
Different platforms perform these operations differently, and some answers depend more heavily on model training while others use live web retrieval. The model is useful because it exposes the practical failure points. A system cannot represent a business accurately when the available information is ambiguous, disconnected from the question, unsupported, inaccessible, or inconsistent.
Five Conditions That Help a Business Become Recommendable
1. Entity clarity
Before an AI system can connect a business to a buyer’s question, it must have enough consistent information to distinguish that business from other organizations, products, and people. This is more difficult than it sounds. Companies often use different names across platforms, describe their services differently from page to page, or fail to state clearly whom they serve and what problems they address.
Entity clarity means creating a coherent public identity: a primary website, consistent naming, accurate profiles, clear descriptions, identifiable leadership, and corroborating references where they legitimately exist. Structured data can help search systems interpret visible information, but it must match what the page actually says. Google is explicit that no special AI schema is required for its AI search features.
2. Question alignment
Article 3 made the distinction between a keyword and a question: a keyword identifies a subject, while a buyer question reveals the decision behind it. That distinction matters because an AI-generated answer must respond to the question the buyer actually asked.
A page about “B2B marketing strategy” may establish topical relevance. It may not answer whether a founder should hire an agency, build an internal team, or change the company’s positioning before investing in distribution. Content becomes more useful when it addresses the full decision with specific reasoning, evidence, limitations, and a clear answer.
3. Credible evidence
After Article 3, David Crowder offered a useful three-part description of the buyer’s test: Does the evidence hold up? Does the provider understand my specific context? Will someone stand behind the recommendation when the decision becomes difficult?
AI systems do not experience trust as humans do, and we should not pretend they conduct the same judgment. But systems that retrieve and cite information can compare claims with other accessible sources, connect content to identifiable authors and organizations, and present corroborating or conflicting information to the buyer. Clear authorship, relevant credentials, first-party evidence, independent references, and consistent claims give the system—and ultimately the buyer—more to evaluate.
That distinction is central to Recommendation Marketing. Confidence is easy to publish. Deserved confidence requires proof.
4. Technical accessibility
Good information cannot contribute to an answer if the relevant system cannot access or interpret it. The exact requirements vary by platform, so the technical audit should begin with official documentation rather than a universal checklist sold as certainty.
The practical review should include:
Crawler access: Confirm that robots.txt, hosting, firewall, and CDN settings do not unintentionally block the crawlers relevant to the platforms you want to reach.
Visible textual content: Keep important expertise, service descriptions, and evidence available in a form crawlers can retrieve and interpret reliably.
Structured data: Use appropriate markup when it accurately describes the visible page; do not treat schema as a guarantee of AI inclusion.
Canonical signals: Reduce avoidable ambiguity about which version of a page should be treated as primary.
Platform-specific controls: Distinguish search retrieval, user-requested fetching, and model-training crawlers because the same provider may use different bots for different purposes.
The emerging llms.txt convention may also be worth testing as a low-cost way to provide a curated map of important content. It remains a voluntary proposal rather than a broadly guaranteed discovery or ranking mechanism, so it belongs in the experimental layer—not at the center of the strategy.
5. A consistent public record
AI answers can vary because the systems, models, retrieved sources, prompts, and available information vary. A business cannot eliminate that variability, but it can reduce unnecessary ambiguity by maintaining a focused and consistent public record.
That does not mean publishing for the sake of activity. It means continuing to answer the questions the business has earned the authority to address, updating information that has materially changed, and keeping descriptions of the company aligned across its website, leadership profiles, articles, case studies, and relevant external listings.
Consistency makes the business easier to classify. Recency helps prevent an outdated version of the business from becoming the clearest available version. Neither guarantees recommendation, but both improve the quality of information available for evaluation.
Why Your Marketing Dashboard Cannot See the Whole Evaluation
Most marketing analytics were designed around observable interactions: impressions, rankings, clicks, sessions, form submissions, and conversions. Those measures remain useful, but they begin at or after exposure.
AI-mediated evaluation can begin earlier. A buyer may ask a system to define the problem, identify possible providers, compare approaches, and build a shortlist without visiting any of those providers. If the buyer later calls one company directly, the company’s analytics may record the call while missing the research process that produced it.
This does not mean the process is completely unmeasurable. Google includes traffic from its AI features within Search Console’s Web reporting, and referral analytics may identify some visits from answer platforms. But those measures largely capture traffic after a link is shown or clicked. They do not provide a complete view of which companies were considered, how they were described, which sources shaped the answer, or how often the answer changed without producing a visit.
A Five-Step AI Recommendation Audit
Article 3 offered the first step: ask an AI assistant the question your best prospective buyer would ask and examine what appears, what does not, and what evidence sits behind the answer. To make that exercise more useful, continue with four additional steps:
Repeat the test. Ask the same question on different days and across more than one relevant AI system. Record what remains stable and what changes.
Vary the buyer context. Change the company size, industry, location, problem, or decision criteria. Note when your company enters or leaves the answer.
Inspect the evidence. Open the cited sources. Determine whether they support the description, whether your strongest proof is visible, and whether outdated or inaccurate information is shaping the response.
Audit the technical path. Review crawler access, visible content, canonical signals, structured data, and consistency across the digital properties an answer system may encounter.
The objective is not to celebrate one favorable answer or panic over one omission. It is to identify patterns across the questions that matter and determine what those patterns reveal about the way the business is being understood.
From Visibility to Recommendation
Visibility gets a business noticed. Clarity helps the system and the buyer understand it. Credibility and evidence make the recommendation defensible. Engagement begins when that recommendation creates a question, a visit, a reply, or a conversation.
The five conditions are therefore not tricks for manipulating an AI answer. They are the operating conditions for making a business easier for buyers and AI systems to understand, verify, and recommend.
If you want to see how your business currently appears across the questions that matter, Right Angle combines Answer Engine Optimization with Sightline to diagnose visibility, representation, citations, and competitive presence. Explore AEO with Sightline.
Next issue: the proof problem. Trust may decide the outcome, but most businesses still cannot measure whether they are earning it. I will examine what honest AI-recommendation measurement requires, and what it cannot responsibly claim.
Frequently Asked Questions
Do all AI systems use the same signals to recommend businesses?
No. Platforms use different models, indexes, retrieval methods, data sources, and policies. The five conditions in this article are a practical framework for improving the information available to those systems, not a universal ranking-factor list.
Do I need to rebuild my website?
Not necessarily. Some technical gaps can be addressed through focused changes to crawler access, page structure, visible content, canonical signals, or structured data. The larger work may be editorial: clearly answering the buyer questions for which the business can provide credible evidence.
Is this just SEO with a new name?
No, although the disciplines overlap. Foundational SEO remains important because crawling, indexing, helpful content, and technical accessibility can support inclusion in AI search experiences. Recommendation Marketing extends the question from whether a page can be found to whether the business is understood, credibly represented, and surfaced in response to consequential buyer questions.
Should I add llms.txt to my website?
It may be reasonable to test, but it should not be treated as a guaranteed visibility mechanism. llms.txt is an emerging voluntary convention, and adoption varies. Foundational accessibility, clear content, evidence, and platform-specific crawler controls deserve priority.
If my business appears today, am I done?
No. One appearance is a useful observation, not proof of reliable recommendation. Test across relevant questions, systems, contexts, and dates, then track the pattern rather than a single favorable result.
B Randall Willis
Founder, Right Angle
rightangleagency.com
About this newsletter:
Recommendation Marketing is published by B Randall Willis, Founder of Right Angle. It documents the emergence of a marketing discipline built for a world in which buyers increasingly use AI systems to discover, evaluate, and compare businesses.
