Mike King’s recent article, How Do We Move Forward if the SERP Is Generative?, raises an important question about the future of search. If search results are increasingly assembled dynamically, rather than presented as a relatively stable set of ranked links, how should marketers think about visibility, optimization, and performance measurement?
King’s exploration of generative search experiences challenges the assumptions that have shaped SEO for decades. When a search engine can construct comparisons, recommendations, and interactive experiences from retrieved information, the familiar concept of a fixed ranking becomes a less reliable measure of success. His argument for thinking about visibility probabilistically, rather than relying on individual ranking observations, is a logical response to that change.
As I read his article, however, I kept coming back to another implication of generative search, one that connects directly to my work over the past several years on decision models and what I call eligibility gates.
If search systems are increasingly assembling answers that help people evaluate alternatives and make decisions, then measuring the probability of appearing in those answers is only part of the challenge. We also need to understand what makes a brand eligible for consideration, what evidence supports its suitability, and under which conditions it is likely to be recommended.
In other words, if the SERP is becoming generative, GEO needs to become decision-driven.
This is not an argument for abandoning traditional SEO or dismissing the importance of AI visibility. Retrieval, relevance, authority, and accessible content remain fundamental. Rather, it is an argument that generative search creates an additional layer of competition, one that is not adequately explained by rankings, citations, or brand mentions alone. The challenge is no longer simply whether an AI system can find information about a company, but whether that information is sufficient to establish that the company is an appropriate answer to the decision the user is trying to make.
Generative Search Changes The Nature Of The Competition
One of the most interesting aspects of King’s article is his exploration of what happens when the search results page itself becomes a generated experience. His Genraggle concept illustrates how a search engine might move beyond presenting links and instead construct an interface suited to the user’s task, whether that involves comparing products, evaluating services, or navigating a complex decision.
The significance of that possibility extends beyond the presentation of search results. In a traditional search experience, the search engine primarily helps the user find relevant information, leaving much of the evaluation to take place after the click. In a generative experience, some of that evaluation can occur within the search system itself, drawing on information from multiple sources to generate comparisons and recommendations.
The distinction matters because a company can be highly visible within the information being retrieved without necessarily becoming a recommended solution. Its website might be used to establish facts about a category, explain terminology, or provide technical information, yet the final answer may favor another vendor when the user’s requirements are applied.
This is where King’s argument about probabilistic visibility becomes especially relevant. If the results are dynamic, we need repeated observations to understand how frequently a brand appears. But if those results increasingly involve evaluation and recommendation, we also need to understand the conditions that determine whether the brand moves from being a source of information to being a viable option.
That distinction has become increasingly apparent in our own GEO research.
When 98% Citation Visibility Produces Only 17% Recommendations
In a recent client project involving a specialized software category, we examined how AI systems represented and recommended vendors under different decision conditions. One company was cited in 98% of the observed responses, a level of visibility that would appear exceptional in many conventional GEO reporting frameworks.
Yet when we introduced more refined decision criteria, the company was recommended only 17% of the time. In a separate condition where supporting evidence was absent, its recommendation rate was just 8%.
These results represented different observed conditions rather than successive stages of a single conversion funnel, but the contrast was revealing. The company was highly visible and readily associated with the category, yet that visibility did not consistently translate into recommendations when the evaluation required more specific evidence.
The findings illustrate an important limitation of treating citations as a primary measure of GEO success. Being cited indicates that a source or brand has appeared in an AI-generated response, but it does not necessarily mean that the system has established the brand’s suitability for the user’s requirements. A company can be recognized as a participant in a category without being selected as an appropriate solution for a particular buyer.
The difference becomes especially important in business-to-business markets, where decisions are rarely based on general category relevance alone. Buyers have technical requirements, operational constraints, compliance obligations, budget considerations, and implementation needs that can materially change which solution is most appropriate.
A recommendation that makes sense for one organization may be unsuitable for another, even when both users begin with the same broad question.
This is why I believe GEO must distinguish between visibility, eligibility, and preference. Visibility establishes whether the brand is present within the information environment. Eligibility concerns whether it satisfies the requirements of a particular decision. Preference concerns why it might be selected over other eligible alternatives.
These are related outcomes, but they are not interchangeable, and optimizing for one does not guarantee success in the others.
The AI Wasn’t Just Looking For Vendors. It Was Researching How To Decide.
During another phase of the project, we examined the fan-out questions associated with a prompt asking an AI system to identify the best solution in the software category. The original request was straightforward, but the observable research activity was considerably more complex.
Using IPullRank’s Qforia Fan Out tool, we captured the fan-outs for our prompts. We found that 62% of the fan-out questions focused on understanding the evaluation checklist, establishing knowledge of the product category, and applying that information to comparisons between vendors.
This was particularly interesting because the user had not explicitly asked for a checklist or a detailed explanation of the category. The request was simply to identify the best solution. Yet a substantial proportion of the system’s research activity concerned gathering the information needed to understand and evaluate the available options.
That does not mean every AI system follows a fixed sequence in which it establishes criteria, applies eligibility gates, and then selects a winner. Fan-out questions provide a window into observable retrieval behavior, not a complete explanation of a model’s internal reasoning. Nevertheless, the pattern is instructive because it demonstrates that the information needed to construct a recommendation can extend well beyond the wording of the original prompt.
The system was investigating what characteristics mattered, how the category should be understood, and how competing vendors compared against relevant criteria. In practical terms, it involved gathering information to support a decision.
This has significant implications for content strategy. If an AI system is researching the criteria for a suitable solution, then a vendor needs more than content that merely describes its products or establishes general category relevance. It needs accessible, credible evidence that addresses the criteria the system is investigating.
Content that explains what a buyer should look for may be as consequential as content describing an individual vendor. Technical documentation, implementation guidance, industry standards, product specifications, and independent evidence can all contribute to the information available for evaluation.
This is the connection I see between King’s vision of generative search and the need for decision-driven GEO. As search experiences become more capable of constructing answers and comparisons, the competitive challenge increasingly involves the information used to support those evaluations.
The Visibility Trap In Today’s GEO Strategies
Despite this shift, much of the GEO activity I see in the market remains focused on improving visibility through approaches inherited from traditional SEO. Marketers examine the content competitors have published, identify missing topics, and create similar material in the hope of achieving comparable coverage. AI visibility programs frequently rely on collections of simple informational prompts, often resembling People Also Ask questions, to measure whether brands are mentioned or cited.
These approaches are not without value. Informational content helps establish category understanding, and tracking broad prompts can reveal whether a brand is discoverable within a particular subject area. Competitive content analysis can also identify genuine gaps in the information available to prospective customers.
The problem arises when content parity is mistaken for decision relevance, or when visibility across simple prompts is treated as evidence that a brand is well positioned to be recommended.
Consider the difference between asking an AI system what a customer relationship management platform does and asking which CRM is best suited to a regulated multinational organization with specific integration requirements, regional data controls, and complex implementation constraints. Both questions concern the same product category, but the information required to answer them is fundamentally different.
The first is primarily an informational task. The second requires evaluating alternatives against a set of criteria, some of which may be mandatory and others negotiable. A vendor could appear consistently in responses to the first question while rarely being recommended for the second.
Yet conventional GEO dashboards may combine these very different interactions into a single measure of visibility.
Content parity creates a similar problem. If every vendor publishes articles describing the same ten important features of a software category, the resulting material may contribute to understanding of the category without providing the evidence needed to distinguish one solution from another. The content may answer common questions while leaving the most consequential buying criteria insufficiently addressed.
Our finding that 62% of the observed fan-out questions concerned evaluation checklists, category understanding, and vendor comparisons illustrates the potential gap. A GEO strategy organized primarily around common informational questions may miss a substantial portion of the research required to make a recommendation.
The issue is not that marketers need to stop answering common questions or producing educational content. It is that they need to stop treating topic coverage as a substitute for decision coverage.
The relevant question is no longer simply whether the website contains content about what the customer is searching for. It is whether the company has provided the evidence needed to establish its suitability for what the customer is trying to decide.
Eligibility Gates And The Evidence Required To Be Considered
For the past two years, I have used the concept of eligibility gates to explain an important aspect of AI-driven discovery. An eligibility gate is a requirement that a solution must satisfy to be considered appropriate for a particular decision.
These gates are not necessarily explicit rules embedded within a model. They are a practical way to represent the criteria and constraints that influence whether a recommendation can be justified.
A buyer looking for enterprise software may require compatibility with an existing technology environment, support for particular operational workflows, compliance with specific standards, or integration with systems already in use. Some of these requirements may be preferences, while others are conditions that a vendor must satisfy to be a viable option.
If a vendor genuinely lacks a mandatory capability, then increased visibility cannot make it the preferred recommendation. The solution is simply not appropriate for that decision.
But there is a more subtle problem that is particularly relevant to GEO. A company may satisfy the requirement in practice while failing to provide sufficient retrievable evidence of that.
This distinction between capability and evidence is critical. A product may support a required integration, but the available documentation may not clearly establish that support. A company may have extensive experience serving a particular industry without publishing information that demonstrates that experience. It may meet an important technical or compliance requirement while making the supporting details available only during sales conversations.
In these situations, the brand’s challenge is not necessarily a product deficiency. It may be an evidence deficiency. And when AI systems gather information to evaluate alternatives, evidence deficiencies can become barriers to making recommendations.
This does not imply that publishing a claim guarantees inclusion or that AI systems consistently verify requirements. The availability of evidence is one of several influences, including retrieval behavior, source credibility, model behavior, and the specific decision context.
Nevertheless, it changes the way marketers should approach content gaps. The objective is not merely to produce more material about the product category, but to identify the information required to demonstrate eligibility and ensure that appropriate evidence is accessible.
“We Have All Of That. We Just Don’t Have It On The Website.”
One of the most revealing moments in our client work came during a discussion of the evidence gaps we had identified. As we reviewed the criteria AI systems appeared to be researching and compared them with the information available on the company’s website, a member of the client’s team made a simple observation: “We have all of that. We just don’t have it on the website.”
The comment captured an organizational challenge that extends well beyond the mechanics of SEO or GEO.
The company was not necessarily missing the capabilities, expertise, or underlying documentation. In many cases, the information existed internally but had not been made publicly available. There were also understandable reasons for those decisions.
Some information had been intentionally reserved for sales conversations, allowing representatives to explain capabilities in the context of a prospect’s specific requirements. Other information was confidential or commercially sensitive. In some cases, details had been omitted to keep the website concise and focused on the messages most likely to engage prospective customers.
Those choices made sense within a model where the website’s primary role was to introduce the company, establish credibility, and encourage the buyer to engage with sales. The deeper evaluation could then take place through conversations, demonstrations, technical reviews, and proposals.
Generative search complicates that model because the evaluation may begin before the company has an opportunity to participate directly. When a user asks an AI system to identify the best solution for a particular set of requirements, the system may assess capabilities, establish criteria, and compare vendors based on the information it can retrieve. If important evidence is available only in a sales presentation, an internal document, or a conversation with an account executive, it may not contribute to the initial evaluation.
The company may have the capability and evidence demonstrating it, yet still be poorly represented because the evidence is not accessible to the system conducting the research.
This creates a tension between traditional sales strategy and AI-mediated discovery. Historically, withholding some detail from the website could help create opportunities for direct engagement. In a generative search environment, that same choice may reduce the information available to establish eligibility before the prospect ever makes contact.
The answer is not to publish confidential information, disclose proprietary methods, or abandon a carefully designed sales process. Rather, companies need to determine what can be made available safely and credibly to support the decisions for which they want to be considered.
A company may be able to demonstrate integration support without exposing proprietary implementation details, establish industry experience without identifying confidential customers, or document compliance without disclosing sensitive security information. Technical documentation, carefully constructed capability statements, public certifications, and appropriately anonymized case studies can all help make legitimate evidence more accessible.
The objective is not maximum disclosure. It is sufficient, credible, retrievable evidence to support an accurate evaluation.
This also makes GEO a more cross-functional discipline. Marketing cannot always resolve evidence gaps independently because the relevant information may be owned by product teams, sales organizations, technical specialists, legal departments, or compliance functions.
The question becomes not simply what content marketing should create, but what information the organization needs to make available so that AI systems can accurately evaluate its capabilities.
That is a much more consequential conversation than identifying another keyword gap.
Why The Decision Model Matters More Than The Prompt List
If eligibility depends on the decision’s requirements, then the decision itself becomes a more useful organizing principle for GEO strategy than an isolated prompt.
Two buyers can ask the same broad question while requiring very different recommendations. Someone asking for the best software platform might prioritize affordability and ease of implementation, while another buyer using the same wording might require advanced integrations, regulatory compliance, or enterprise-scale deployment.
The phrase “best software” does not define the decision. The criteria, constraints, and trade-offs do.
This is the purpose of a decision model. It describes the problem the buyer is trying to solve, the requirements that must be satisfied, the preferences that influence the choice, and the evidence needed to compare available options.
When applied to GEO, the model becomes a framework for both content strategy and measurement.
Instead of beginning with a list of prompts and asking how frequently a brand appears, we can begin with the decisions customers are trying to make. We can then identify the criteria that influence those decisions, examine the information AI systems retrieve when researching them, and determine whether the brand has sufficient evidence to establish eligibility and preference.
This approach also reveals why aggregate visibility metrics can be misleading.
A vendor might be recommended frequently when ease of implementation is the dominant consideration but rarely when advanced customization is required. Another might perform well for large enterprises but poorly for smaller organizations with limited resources.
Neither pattern is necessarily a GEO failure. The relevant question is whether the brand is being recommended in the circumstances where its capabilities genuinely make it an appropriate choice.
Decision models make it possible to examine those differences rather than treating all prompts and recommendations as equivalent.
They also help distinguish among problems that require different responses. If a brand is rarely considered because the system cannot find relevant information, the issue may be discoverability. If it is considered but the system lacks evidence of a required capability, the issue may be documentation or retrievability. If competitors are preferred despite comparable eligibility, the issue may involve differentiation or the strength of supporting evidence.
And if the brand does not actually satisfy the requirement, the issue may belong to product strategy rather than content strategy.
That diagnostic capability is what makes decision-driven GEO fundamentally different from simply expanding prompt coverage or matching competitors’ content.
The Same Decision Can Produce Different Recommendations
Our research also revealed another important limitation in measuring GEO performance using a fixed set of visibility-centric prompts. Even when the underlying product category remained the same, changes in how the decision was framed could produce meaningfully different recommendations.
We tested variations in how the buyer identified themselves, comparing prompts that specified a particular professional role with prompts that described the decision more collectively, using language such as “we are” or “our department.” The responses were not always equivalent. Identifying a specific role could produce different or more refined answers, suggesting that the system was interpreting the request through the responsibilities and likely priorities associated with that role.
This matters because a purchasing decision rarely belongs to a single, generic buyer. A technical leader, operational manager, procurement specialist, and department head may participate in selecting the same product, but each can approach the evaluation from a different perspective. Their priorities overlap, yet the relative importance of integration, usability, governance, cost, and operational performance may differ.
When a prompt identifies the person making the decision, it can provide additional context that changes how the system interprets what constitutes an appropriate recommendation. Even seemingly minor differences between an individual professional role and a collective departmental perspective can influence the answer.
We observed a similar effect when changing the order in which requirements were presented.
In one test, a company emerged as the primary recommendation. When we reordered the requirements, changing which criteria received emphasis, that company disappeared from the recommended set altogether. In its place was a vendor that had not appeared in the original list but was exceptionally strong against the criterion that had become more prominent.
The underlying category had not changed, nor had the broad objective of finding the best solution. What changed was the framing of the decision and the relative emphasis placed on its requirements.
That distinction is critical.
It suggests that recommendation performance cannot be understood as a single, stable property of a brand. A vendor may be highly competitive when one set of requirements dominates the evaluation but less suitable when another criterion becomes more important. A competitor that is largely invisible in one decision context may become the preferred option in another.
The results do not establish whether the changes were caused specifically by the order of the requirements, the model’s interpretation of their relative importance, or other variability in the generated responses. Establishing that would require repeated controlled tests. Nevertheless, the observed substitutions demonstrate why testing only one formulation of a buying question can produce an incomplete picture of a brand’s competitive position.
This is another reason decision models are more useful than static prompt lists.
A well-constructed decision model identifies not only the criteria associated with a purchase but also the roles involved, the constraints they bring, and the trade-offs that can change which solution is most appropriate. Those dimensions can then be systematically varied to examine how recommendations respond.
Instead of asking whether a brand ranks first for a particular prompt, we can investigate the conditions under which it becomes the preferred recommendation, remains eligible but loses to a competitor, or disappears from consideration.
That creates a much richer understanding of the competitive landscape.
It also changes how marketers should think about content. If a brand’s strengths are particularly relevant to one stakeholder or operational requirement, the supporting evidence needs to be clear enough for AI systems to recognize that relevance. If another criterion represents a weakness, the company needs to understand whether the issue is missing evidence, insufficient differentiation, or a genuine product limitation.
In this sense, the decision model becomes more than a framework for generating prompts. It becomes a way to investigate the boundaries of a brand’s eligibility and preference across the circumstances in which customers make decisions.
And it reinforces the central implication of generative search: there may be no single best answer independent of the decision context. There is a best-supported recommendation for a particular combination of requirements, priorities, and constraints.
From Visibility Probability To Recommendation Probability
King’s argument that generative SERPs require probabilistic measurement is an important starting point. If search results vary across observations, a single ranking or response cannot adequately represent performance. Repeated testing across relevant prompts and contexts becomes necessary to understand how consistently a brand appears.
But once we recognize that AI-generated experiences can also evaluate and recommend solutions, probability needs to be considered at multiple levels.
We need to understand not only how frequently a brand is retrieved or cited, but also how often it is included among viable alternatives, how accurately its capabilities are represented, and how often it is recommended under specific decision conditions.
Those measures should not be collapsed into one visibility score because they describe different outcomes.
A brand cited in nearly every response may still show weak recommendation performance when the user’s requirements become more specific, as our client research has shown. Conversely, a brand with less overall visibility may be consistently recommended within a narrower set of commercially valuable decisions.
The latter may represent a stronger business outcome, even if its aggregate citation rate is lower.
A decision-driven measurement framework would therefore evaluate recommendation frequency across defined customer scenarios, using repeated observations to account for variation while preserving the criteria that make each scenario meaningful.
It would also examine the relationship between evidence availability and recommendation behavior. When a brand is excluded from a comparison, the analysis should investigate whether the system lacked relevant evidence, failed to retrieve available information, relied on competing sources, or identified a genuine mismatch between the product and the user’s requirements.
These observations can guide optimization, but they should not be confused with proof of causation. Demonstrating that a specific content improvement changes recommendation outcomes requires controlled testing and careful interpretation.
The larger point is that GEO measurement becomes more useful when it helps explain not just whether a brand appeared, but why it may have been considered suitable or unsuitable for a particular decision.
That is how measurement begins to inform strategy rather than simply report visibility.
The Next Phase Of GEO Is About Earning The Recommendation
Mike King’s exploration of generative SERPs is valuable because it challenges marketers to reconsider the assumptions behind traditional search measurement. If search results are increasingly dynamic and assembled around the user’s task, our methods for evaluating visibility must evolve accordingly. I believe the same logic should apply to how we organize GEO strategy.
When AI systems help people make decisions, the competition is no longer limited to the information they retrieve. It also involves determining whether the available information establishes that a brand meets the decision’s requirements and provides credible reasons to prefer it over alternatives.
Our research illustrates the importance of this distinction. A company with 98% citation visibility was recommended only 17% of the time under more refined decision criteria, while a separate condition without supporting evidence produced an 8% recommendation rate. In our fan-out analysis, 62% of the observed questions associated with a request for the best solution focused on understanding evaluation criteria, establishing category knowledge, and comparing vendors.
The role-and-requirement-order tests revealed another dimension. Changing how the buyer was represented or which criteria received greater emphasis could alter the recommended vendors, in one case replacing the primary recommendation with a company that had not appeared in the original list.
These findings do not establish a universal mechanism governing AI recommendations. They do, however, reinforce the importance of understanding the decision-related information that generative search systems retrieve and use.
They also reveal why the next phase of GEO cannot be solved simply by publishing more content, achieving topic parity with competitors, or increasing visibility across collections of informational prompts.
For years, companies have designed websites to provide prospective customers with enough information to initiate a sales conversation. Generative search increasingly creates situations in which the system may need enough evidence to establish a company’s suitability before that conversation begins.
That requires a different approach to content, one organized around the criteria that determine eligibility and the evidence that supports preference.
Decision models provide a way to understand those criteria. Eligibility gates help identify the requirements that must be satisfied. And credible, accessible content enables AI systems to evaluate a brand against those requirements.
The goal should not be to manipulate systems into recommending a company regardless of whether it is appropriate. It should be to ensure that when a company genuinely meets the requirements of a decision, the evidence necessary to recognize that fit is available.
King’s article makes the case that the future of search visibility is probabilistic. The next step is to recognize that the recommendation is conditional.
A brand’s likelihood of being recommended depends not simply on whether it is visible, but on the decision being made, the perspective of the person making it, the relative importance of their requirements, and the evidence available to support the evaluation.
If the SERP is becoming generative, GEO must evolve from primarily optimizing for appearances to understanding and supporting those decisions.
Because the future of GEO is not about becoming the answer to every question. It is about providing the evidence that makes a brand the right answer for the decisions it is best equipped to satisfy.
