For much of the past two years, the AI search conversation has focused on visibility. Teams want to know whether their brand is mentioned, where it appears in recommendations, how often competitors surface, and whether they are being cited across ChatGPT, Gemini, Claude, Perplexity, Google, and other AI surfaces.
Those are useful metrics, but I have increasingly questioned whether they tell us enough about the corresponding business outcome.
Nearly two years ago, I began using the term “eligibility gates” to describe the detailed decision criteria that determine whether a company or product remains qualified as a buyer moves through the decision process. I later expanded on that idea when I wrote Decision Coverage: whether an organization has exposed enough information and evidence for AI to evaluate, compare, qualify, and confidently recommend.
That led to two related arguments tied to previous articles. In my article Beyond Content Parity, I argued that using tools and processes that simply match competitor content may create parity without creating any decision advantage. In Beyond Brand Sovereignty, I argued that the knowledge AI needs often exists beyond the marketing site, scattered across product, sales, support, engineering, legal, customer success, and external
A recent project allowed me to test those ideas against an actual business problem.
The Problem Wasn’t Visibility, It Was A Shrinking Pipeline
The company did not have an AI visibility problem. It frequently appeared in AI-generated consideration sets. The concern was what happened afterward, when prospective buyers continued the conversation and began asking more specific questions about requirements, workflows, integrations, outcomes, evidence, and competitive differences.
The question was not “How do we get seen in AI?” It was, “Why, with all this visibility, is our lead pipeline shrinking?” The reality we found was that as AI took a prospective buyer further down the consideration set, they became less competitive.
To prove my thesis and identify the problem, we created 38 detailed, criteria-based prompts across several product categories and buyer segments, and evaluated them on five AI surfaces, yielding 190 responses.
These were intentionally different from broad “best vendor” prompts. They reflected the kinds of requirements buyers use to narrow a field: workflow fit, integrations, regulatory requirements, interoperability, operational conditions, analytics, platform breadth, and other decision criteria. I had to create a detailed explanation of their purpose as diagnostic-only because the brand and their SEO agency argued that no human would ever use that complex a prompt. We actually found that some prompts were even longer after talking to actual procurement managers.
The initial numbers looked excellent – even better than their current visibility scorecard. Our research found the company was mentioned in 98.4% of evaluated responses and appeared as either a primary recommendation or shortlisted option 93.2% of the time. For many GEO programs, that would be the headline of the week and champagne for everyone. The brand was almost always visible and almost always treated as a serious option. But visibility was not the problem we were investigating.
When we evaluated those same responses against the decision criteria in the prompts, 73.7% had weak or missing evidence supporting important claims or recommendations about the company. In 15.8% of runs, an important association, decision criterion, competitive distinction, or discovery signal contained what we classified as a Significant or Critical Gap. The more we refined our evidence-scoring criteria, the higher this number rose.
The critical finding was that AI systems generally knew the company belonged in the conversation. However, they were far less consistent at substantiating why it should remain there once the buyer began applying specific eligibility criteria. That is the distinction that matters and explains their declining lead pipeline.
Visibility Does Not Equal Decision Readiness
There is a meaningful difference between asking which vendors operate in a market and asking which vendors can satisfy a specific combination of technical, operational, regulatory, and business requirements.
A model may know a company is a legitimate vendor, understand its main products, and have enough general information to put it on a shortlist. But once the buyer asks how an integration works, whether a workflow is supported, what happens under a particular operating condition, what limitations apply, or what evidence supports an outcome claim, the information requirements become more demanding. Those are not merely additional keywords. They are evidence requirements.
That explains how a brand can achieve a 98.4% mention rate and 93.2% primary-or-shortlist rate while still having weak or missing evidence in nearly three-quarters of the responses. The numbers are not contradictory. They measure different stages of the decision process.
No one can dispute that AI visibility gets the company into the consideration set. Decision readiness determines whether AI has enough information to keep it there.
The Weakness Was Often Evidence, Not Capability
One of the more important findings was that weak evidence did not necessarily mean the company lacked the capability. In many cases, the capability appeared to exist, but the available information did not explain or substantiate it well enough to satisfy the AI surface’s confidence level.
A marketing bullet might claim an end-to-end process without ever showing the actual process, connections, or how the data flows between systems. We found efficiency claims but did not have any “proof” from internal tests or customer testimonials of the efficiency gain.
Sometimes the information was simply fragmented. A model might need to connect a product page to an integration page, then to another interoperability page, before it could infer how the pieces satisfied the buyer’s requirement. While an AI system could or should be able to make those multi-part connections, each additional inference creates another opportunity for incomplete representation or for a competitor with clearer evidence to become the safer recommendation.
We also found that the strongest content was not necessarily the longest or most buzzword-integrated, but rather content that made the relationships among capability, workflow, buyer requirements, and outcome explicit.
An interesting finding was that only first-party evidence was not inherently a weakness. Many claims lacked third-party validation, but this was not necessarily flagged as a defect. We found that authoritative product documentation may be the best evidence that a feature or integration exists. However, every claim involving superiority, effectiveness, adoption, scale, measurable outcomes, or competitive differentiation did benefit from independent corroboration by customers, industry, government, analysts, or other parties.
That means some evidence gaps are not content-writing problems at all, but does anyone else agree with the evidence?
This Is Where Content Parity Reaches Its Limit
The findings reinforced the concern I raised in my earlier work about how easy-button AI content-generation systems create “me-too content parity.” As I explained to this brand, most of the GEO efforts to date were what I would call SEO-adjacent GEO. It is a newly rebranded SEO process that leverages AI speed for traditional content-gap analysis. Teams look at competitor coverage, People Also Ask questions, common prompts, and pages receiving citations, then create content to close those gaps. There is nothing inherently wrong with that. In this case, it appears to have worked extremely well at establishing this company’s category presence. The limitation appears when everyone says roughly the same thing, and you don’t offer details.
If every vendor claims to be end-to-end, highly integrated, AI-enabled, scalable, flexible, and focused on efficiency, those statements may help establish that each belongs in the category. They do much less when the model has to determine what “end-to-end” covers, what “integration” connects, what the AI actually does, or what evidence supports the efficiency claim.
Parity can help a company qualify for the conversation without providing enough information to survive deeper evaluation.
The Hardest Part Was Building The Evidence Scorecard
Running the prompts was relatively straightforward. There wasn’t a tool or off-the-shelf framework for what we were trying to measure. We needed to build a multi-variable evidence scorecard capable of determining whether the model had enough support to evaluate the company against the criteria that actually mattered in the buying decision.
That meant defining what counted as strong evidence, what first-party documentation was enough, and when independent corroboration became important. The bigger challenge was thinking beyond the “we are great” organizational mindset to determine when an evidence gap or omission was serious enough to affect consideration or selection. That turned out to be even harder than getting SEOs to agree on ranking factors.
Those judgments then had to be translated into a repeatable scoring system and applied consistently across the prompts and five AI surfaces. It also required conversations beyond content. Sales had to help identify criteria and objections that affect deals. Product had to validate capabilities and limitations. Customer-facing teams helped identify real outcomes. In some cases, customers themselves were the best
This is why I do not believe there will be one universal GEO evidence scorecard that every organization can download. The methodology can be reusable if you can honestly define the decision criteria and what constitutes sufficient evidence, then evaluate whether AI can connect the brand to those criteria.
The Content Team Cannot Invent The Missing Knowledge
As we evaluated eligibility criteria and evidence requirements, many findings no longer looked like conventional content gaps. Looking at this massive evidence gap matrix as one manager exclaimed, “that no human will ever read,” it was obvious this would become a team effort to make the changes. Content teams cannot make up content, nor will they necessarily know how everything works.
Some of the knowledge required to close these gaps is organizational in nature. We had to pry it out of product, engineering, and sales, and then run it through marketing and compliance. In some cases, we had to loop in external and partner relations when stronger third-party or attributable proof was required beyond what we could create.
The content team still has a critical role: organize that knowledge, make relationships explicit, create coherent evidence paths, and expose information in ways humans and AI systems can understand.
GEO Measurement Has To Follow The Buyer Deeper
Most importantly, this project validated and reinforced my belief that measuring only the first AI response is insufficient. True, those broad PAA-like prompts are valuable for generating a list of options worth considering. The problem is the current tools cannot maintain state like a human workflow and aggregate the sequence of prompt turns most users leverage. Each subsequent prompt digs deeper into the criteria for consideration, maintaining preferences, and memory, and culminating in the multivariate result that I affectionately call prompt math.
This means that GEO measurement needs to become more sophisticated and account for both visibility and decision readiness. Many tools now have answer accuracy monitoring, which is a step forward. I argue that we need to put more effort into the element that is much closer to the business problem organizations should actually care about.
Do Not Turn Evidence Into The Next Content Tactic
There is an obvious danger here because once marketers conclude that AI systems need evidence, we will probably see an explosion of “evidence-shaped content”: longer pages, superficial diagrams, weak surveys, synthetic research, and AI-generated case studies created primarily because someone believes those formats will influence an LLM, just like the explosion of Reddit submission services. That misses the point. The objective is not to manufacture artifacts that appear persuasive to AI, but to provide real content that reduces uncertainty and risk for the buyer.
If the information would not help a knowledgeable customer make a better decision, it probably should not exist merely to improve a GEO score.
Visibility Is Only The Starting Point
As noted, the company had already achieved what many GEO programs would consider an amazing success: 98.4% visibility and 93.2% primary-or-shortlist consideration, yet 73.7% of those same responses had weak or missing evidence. Unfortunately, our analysis could not prove that this evidence gap caused every lost sale or every subsequent AI exclusion. What it does show is that a visibility-only measurement model would have missed the weakness we were asked to investigate.
Any company can succeed in entering the conversation but fail to offer enough credible evidence to remain qualified as the conversation becomes more specific.
- Brands Are Tracking AI Visibility, But Are They Measuring The Right Things?
- How To Measure AI Search Visibility
- Beyond Content Parity: Building A Validated Content Workflow For AI Search
CategorySearch Visibility & ValueAI Search
