AI Knows Your Brand But Won’t Recommend It: The AIVI Index and Six-Dimension Visibility Measurement

In the pilot I asked three different language models about six observability products. No product was named in any question; they were open-ended prompts such as "which platform should a mid-sized engineering team adopt". Every brand name that appeared was the model’s own choice.

One result stood apart. Two of the three models recognized a product called Coroot when asked directly, correctly describing what it is and who builds it. Yet it appeared in none of the eighteen open-ended responses. Not one.

Being known is not being recommended

Classical visibility metrics are one-dimensional: you rank or you do not, traffic exists or it does not. In generative AI, visibility is a chain. The model must first recognize that you exist, then retrieve information about you, then judge it relevant, then decide to include you, then actually write you into the answer. Break any link and you are absent from the response.

The Coroot case shows a chain whose first link is intact and whose later links are broken. A single "AI visibility score" cannot make that distinction; it just hands you a low number without telling you why.

The six dimensions of AIVI

This is why the index is split into six separate measures. Each captures a different link in the chain, and each is normalized to the interval between 0 and 1.

• Entity Recognition Rate (ERR) — does the model know you exist? The precondition for everything else.
• Visibility Recall (VR) — do you actually appear when the question is relevant to you?
• Citation Consistency (CC) — does it lean on the same sources each time it supports a claim about you?
• Recommendation Confidence (RC) — are you merely mentioned, or actually recommended? The gap between those is large.
• Prompt Robustness (PR) — does your visibility hold when the same intent is phrased differently?
• Cross-Model Stability (CMS) — do you exist in one model and vanish in another?

The six can be weighted into a single AIVI score. But the paper deliberately avoids prescribing fixed weights; it argues weights should be estimated from data. Absent such data, treating all six equally (1/6 each) is the most honest starting assumption.

Pilot results

Another detail is worth noting. Of the five products scoring full marks on recognition, SigNoz achieves roughly half the visibility of the others. And that average conceals something: one model mentioned SigNoz in two thirds of its answers while another never mentioned it at all. Measuring on a single model would have produced a completely wrong conclusion.

Two boundary conditions found in practice

The real value of the pilot was exposing two weaknesses in the framework that were invisible on paper.

First: Prompt Robustness and Cross-Model Stability reward low variance. For a brand that appears nowhere, variance is zero — meaning a completely invisible brand scores full marks on both. Coroot was exactly that case. I added a condition to the framework: when visibility is zero, these two dimensions are not computed and are reported as undefined.

Second: Citation Consistency could not be computed for any product, because the models were run without web access and no response carried sources. That is a finding rather than a fault: citation consistency is measurable only in retrieval-enabled deployments. A complete profile requires measuring the same brand in both configurations.

From five layers to six dimensions

In my earlier work I presented AIVI as a five-layer framework: technical accessibility, information quality, machine readability, semantic trust, and citation authority. That model was a diagnostic instrument classifying the factors that shape visibility — it told you what to fix.

This study extends and completes that framework. The five layers describe the inputs; the six dimensions measure the output. One tells you what to do on your site, the other tells you whether what you did registered in the model’s answer. Two ends of the same problem: the to-do list and the thermometer.

What it does and does not tell you

AIVI is not a quality measure. A high score does not mean your business is good; it means generative systems are likely to find and recommend you. The whole point of my earlier work on manufactured authority was that these two can come apart.

The limits of the pilot deserve stating plainly. All three models come from a single provider, so what I measured is consistency across versions of one family rather than genuine cross-provider consistency. One domain, one language, one day. Coding reliability was tested with two annotators (agreement κ = 0.87) but the second annotator was a language model, not a human. All of this is stated in the paper, because a study proposing a measurement standard cannot start by overstating its own measurement.