ARDI™ Research
The State of AI Visibility · Vol. 1
How AI separates trust from recommendation across the buyer journey, and why the overlap between them is far smaller than the market assumes.
of brands are both trusted and recommended by AI. Trust and recommendation are nearly independent signals, and almost no one holds both.
This paper presents findings from the first large-scale observation of how AI models process brand information across the buyer journey. Using the ARDI™ (AI Recommendation and Discovery Intelligence) platform, GOSH AI tracked how leading AI surfaces handle brand citations and recommendations across 70+ industry categories and dozens of geographic markets.
The central finding is that authority and recommendation are nearly independent signals in AI decision-making. In our baseline dataset, the brands AI cites as sources of truth during research and the brands AI ultimately recommends overlap in fewer than 1% of observed cases. We call this divergence the Authority Gap, and classify every brand into one of four Authority States.
AI-powered search and recommendation systems have created a new layer of brand visibility that operates outside traditional search optimization. When a consumer asks ChatGPT, Gemini, Claude, or Perplexity for a recommendation, the answer is shaped by processes that are fundamentally different from how a search engine ranks pages.
Yet the market's understanding of AI visibility remains superficial. The dominant question, "Does AI recommend my brand?", treats AI output as binary: visible or invisible. That framing misses the structural complexity of how AI arrives at a recommendation, and it hides the single most important distinction in AI discovery: the difference between the brands AI trusts and the brands AI recommends.
AI recommendations are not instantaneous judgments. When a model answers a query, it moves through a structured decision sequence, the framework GOSH AI calls UACR™. A brand has to clear all four stages to be chosen.
Trust forms in the first three stages, mostly invisible to the consumer and entirely unmeasured by most brands. Recommendation is the fourth, the only stage the consumer sees, and the only stage most brands attempt to measure.
Recommendation is the visible output. Trust formation is the invisible input. We set out to measure both.
The ARDI™ platform observes AI behavior through two parallel tracks, designed to capture both real-time and training-embedded model behavior. Most existing approaches to AI visibility collapse these two signals into one. We measure them independently, and that separation is central to the findings here.
AI behavior when models have live search access. Current, volatile, retrieval-augmented responses.
AI behavior from training-embedded knowledge. What models have absorbed about brands and categories.
Figures below reflect the Q1-Q2 2026 baseline dataset. The platform now runs on a continuous, twice-monthly cycle, and the dataset grows with every run.
| Parameter | Value |
|---|---|
| AI surfaces observed | 7 (ChatGPT, Gemini, Claude, Perplexity, Grok, Microsoft Copilot, Google AI Overviews) |
| Leading models | 6 |
| Industry categories | 70+ |
| Geographic markets | 50+ |
| Decision stages (UACR™) | 4 (Understanding, Association, Citation, Recommendation) |
| Prompts in testing library | 11,800+ |
| Model executions (baseline) | 22,100+ |
| Citations analyzed (baseline) | 69,500+ |
| Brands identified | 8,000+ |
| Observation period | Q1-Q2 2026 (ongoing) |
All prompts are brand-agnostic: no prompt names a specific brand, so every brand that appears in AI output is model-generated, not prompt-induced. Prompts are categorized by intent (educational, comparative, evaluative, transactional) and mapped to the UACR™ stages, letting us track where a brand appears in the AI's decision process, not just its final answer.
We define authority presence as a brand being cited as a source of truth in the Understanding, Association, and Citation stages, and recommendation presence as a brand being explicitly named in the Recommendation stage. When we measured the overlap between the two across thousands of brands, we expected significant correlation. We found near-complete separation, and it sorts every brand into one of four states.
Recommended, not cited. Visible, but the visibility depends on real-time retrieval, not established trust. Easy to displace.
Cited and recommended. Trusted and chosen. The rarest and most structurally durable position in AI visibility.
Neither cited nor recommended. The state most brands never realize they occupy, absent from the decision entirely.
Cited, not recommended. AI trusts the brand's content but sends consumers to competitors. Recognition without selection.
Percentages reflect brands that appeared in AI output in the baseline dataset. Invisible, brands that appear in neither layer, is the default state for the majority of all brands.
This pattern held in every category we observed. In Pilates studios, one national brand appeared as both a cited authority and a recommended provider across four models, while dozens of competitors with strong recommendation presence had zero upstream citations. In cosmetic dentistry, only one brand out of hundreds bridged both layers. The specific percentages varied by vertical; the structural finding did not.
The shape of the gap reveals more. In service verticals like Pilates and med spas, AI over-relies on recommendation without authority, naming local providers it has never cited. In product categories like grocery and beauty, AI builds authority without converting it, citing brands as trusted sources but rarely recommending them. In almost no category does AI do both.
The four states form a classification system for brand visibility in AI, scored on two axes and one prediction: how durable the position is as models change.
| State | Cited (trust) | Recommended | Structural durability |
|---|---|---|---|
| Full Authority | Present | Present | High |
| Unconverted Authority | Present | Absent | Moderate |
| Borrowed Authority | Absent | Present | Low |
| Invisible | Absent | Absent | None |
The framework's predictive hypothesis, under longitudinal validation, is that brands with authority presence are structurally more resilient to model changes than brands with recommendation alone. Authority reflects trust embedded in both the search and learned paths; recommendation alone may reflect only real-time retrieval that shifts with every model update.
What AI recommends today depends on search. What AI recommends six months from now depends on authority.
The AI-visibility conversation is focused almost entirely on the Recommendation stage: "Does AI recommend us?" That question is valid but insufficient. Recommendation status can change with any model update or index refresh. Authority, being treated as a source of truth, is the more durable and defensible position.
Brands in Borrowed Authority may be overestimating the stability of their AI presence. Brands in Unconverted Authority may be underestimating their advantage: they hold the harder-to-build asset (trust) and lack only the conversion layer (recommendation). Brands that are Invisible are not in the conversation at all.
Meaningful AI-visibility measurement requires stage-level observation: tracking what AI cites separately from what AI recommends, then measuring the relationship between them. A single-number "visibility score" cannot capture the Authority Gap.
Sample density. While baseline citation volume exceeds 69,500, per-category density of upstream citations is still developing. The authority signal strengthens as twice-monthly cycles accumulate.
Entity resolution. Mapping citation URLs and source names to normalized brand entities is an ongoing refinement. Overlap percentages are directional and may adjust as resolution improves.
Causality. This documents a correlation pattern between authority and recommendation. It does not yet establish causation. Longitudinal analysis underway examines whether authority predicts future recommendation.
Model coverage. Citation behavior varies across models: some expose structured citations (ChatGPT, Gemini), others do not (Claude). For those, ARDI™ uses entity extraction to identify brand references in unstructured text, at lower confidence than citation-linked observations.
Section 8The Authority Gap is real, measurable, and consistent across industries. AI does not treat trust and recommendation as the same signal. The brands it cites as sources of knowledge and the brands it recommends to consumers are, overwhelmingly, different entities, creating a previously unmeasured dimension of brand visibility.
For brands, the implication is direct: knowing whether AI recommends you is only half the picture. Knowing whether AI trusts you, and measuring the gap between the two, is what determines whether your AI visibility is durable or dependent on conditions outside your control.
The brands that win in AI will not be those most visible at the moment of recommendation, but those most trusted before the decision is made.
What is the Authority Gap in AI?
The Authority Gap is a measurable divergence between the brands AI cites as sources of truth during research and the brands AI recommends to consumers during decisions. GOSH AI's research found fewer than 1% of brands appear in both layers, meaning authority and recommendation are nearly independent signals in AI systems.
What are the four Authority States?
Full Authority (cited and recommended, under 1% of brands), Unconverted Authority (cited but not recommended, ~27%), Borrowed Authority (recommended but not cited, ~72%), and Invisible (neither, the default for most brands). Percentages reflect brands that appeared in AI output; Invisible is the default state for the majority of all brands.
What is UACR™?
UACR™ is GOSH AI's model for how AI decides what to recommend, in four stages: Understanding (AI resolves the brand as an entity), Association (it connects the brand to the right category and intent), Citation (it finds authoritative sources), and Recommendation (it names the brand). Trust forms in the first three stages; recommendation is the fourth.
How does ARDI™ measure AI visibility?
ARDI™ (AI Recommendation and Discovery Intelligence) tracks how six leading AI models across seven surfaces, ChatGPT, Gemini, Claude, Perplexity, Grok, Microsoft Copilot, and Google AI Overviews, cite and recommend brands across 70+ categories, using two parallel paths: the Search Path (real-time retrieval) and the Learned Path (training-embedded knowledge).
APA:
GOSH AI. (2026). The Authority Gap: How AI Separates Trust from Recommendation (Updated ed.). ARDI™ Research, The State of AI Visibility, Vol. 1. https://www.mygosh.ai/the-authority-gap
MLA:
GOSH AI. "The Authority Gap: How AI Separates Trust from Recommendation." The State of AI Visibility, vol. 1, updated July 2026, www.mygosh.ai/the-authority-gap.