AI Visibility Gap: Why Brand Mentions Are Not Recommendations

Overview

The AI visibility industry is facing a critical reckoning regarding how brand presence is measured in generative search. A new analysis by Telman Gadimov, founder of CueScout, suggests that the standard metric of "mention count" is fundamentally flawed because it treats being named identically to being recommended.

This conflation creates a distorted view of brand health. In many cases, being mentioned by name without a positive endorsement is worse than total obscurity, as it provides potential customers with a reasoned argument against choosing a product. The industry needs to shift from counting names to classifying intent.

The Visibility Gap

Current AI visibility tools predominantly rely on string matching to determine brand presence. If a large language model (LLM) outputs a brand name, it is logged as a win. However, this approach fails to capture the nuance of consumer decision-making journeys.

When Gadimox reviewed actual AI responses, he found a common pattern: an answer would name multiple products, compare them briefly, and then decline to pick a winner. In one example, a query about product analytics led to a response that named Mixpanel, Amplitude, and Heap. While three brands were technically "mentioned," none were recommended. The model handed the decision back to the user, leaving them with no clear guidance.

This highlights a stark difference between presence and persuasion. A dashboard registering three wins for such an answer is misleading. For the user seeking a solution, the response is effectively a non-answer, requiring them to conduct further research to resolve the ambiguity.

Four States of Brand Presence

To address this gap, CueScout proposes categorizing brand appearances into four distinct states. This framework moves beyond binary presence/absence metrics to offer a more accurate picture of market positioning.

1. Recommended: The answer steers the reader toward the brand, either directly or by describing a scenario where it is the optimal choice.
2. Listed only: The brand is named, but the model advises selecting a competitor, or offers no specific advice at all.
3. Dismissed: The brand is mentioned, but the content argues against it or highlights significant drawbacks.
4. Absent: The brand does not appear in the response.

The critical insight is that mentioned and dismissed are often grouped together in traditional analytics. Gadimov argues this is a dangerous oversight. Being dismissed by name is actively harmful; it plants a seed of doubt in a potential customer’s mind while the analytics software records it as positive visibility.

This problem is exacerbated by the types of sources LLMs frequently cite. "Best tools" listicles, comparison pages, and category roundups are designed to name as many products as possible. If a brand’s visibility is derived primarily from these formats, it is being summarized and skipped rather than advocated for.

A Technical Solution: Decision Models Over Prose

Initial attempts to solve this involved prompting chat models to identify recommendations. However, this approach proved unreliable for quantitative analysis. The responses were varied prose—phrases like "leans toward," "clear recommendation," or "hedges"—which required complex and brittle natural language processing to parse. A slight prompt tweak could cause widespread parsing errors without any obvious failure signals.

Instead, the team utilized a decision model called Jev via OpenRouter. Unlike chat models, decision models return probabilities across predefined options rather than generating text. This eliminates the need for post-hoc parsing and ensures consistent, countable data.

The system posed four specific questions for each stored answer:

• Which single brand does the answer steer the reader toward? Options included the brand, competitors, none, or other.
• Does it advise a choice, or merely list names?
• What is the classification: recommended, listed, dismissed, or absent?
• What type of source is the model leaning on (e.g., listicle, community thread, docs)?

This method produced four typed answers for every data point, removing ambiguity and enabling precise measurement.

Cost and Feasibility

A major barrier to deeper AI analytics has historically been cost. Running complex models over large corpora of search results can be prohibitively expensive. However, this new approach demonstrated exceptional efficiency.

In a test using a corpus of 63 distinct stored answers across ChatGPT, Perplexity, and Gemini, the judgment process took just 7.28 seconds. The total cost was approximately $0.0033, calculated at roughly $0.000013 per individual judgment.

"At three tenths of a cent it stops being a project and becomes a flag you turn on."

This drastic reduction in cost means that re-reading stored answers with a second layer of intent classification is no longer a speculative research project but a viable, everyday operational tool.

Initial Findings and Caveats

The test results from the demo workspace revealed significant discrepancies between mention counts and actual recommendations. For instance, while certain brands appeared in over half of the responses, their recommendation rates were negligible. Notably, 52.4% of answers recommended nobody at all, merely listing products and leaving the final decision to the user.

Gadimov emphasizes that these figures come from a controlled demo environment and should not be extrapolated to the broader market. The primary value lies in the methodology rather than the specific percentages. The expectation is that in real-world data, the most-mentioned brand will rarely be the most-recommended one, and a substantial portion of buying-intent queries will result in no clear recommendation.

The article concludes by inviting other practitioners to run similar analyses on their own corpora. If the shape of the data holds up—specifically the gap between visibility and endorsement—it could force a reevaluation of how AI marketing success is defined.