ARDI™ Research
The State of AI Visibility · Vol. 2
Quantifying the gap between the brands AI surfaces from real-time search and the brands embedded in its training, and why most brands sit in only one.
of brand-category pairs are present in both AI's real-time search and its training-embedded knowledge. The other 91.72% live in one path, invisible to the other.
This paper extends The Authority Gap (Vol. 1) to a second axis of AI-visibility divergence: the gap between the brands AI surfaces from real-time search and the brands embedded in its foundational training. Using the ARDI™ platform, GOSH AI measured presence in each path independently for every observed brand-category pair, across six leading AI models.
The central finding: only 8.28% of observed brand-category pairs are present in both paths. 62.90% appear only in training data; 28.82% only in real-time search. We name these positions Anchored, Latent, and Surfacing Authority. Most AI-visibility tools see only one path, and single-path measurement is insufficient.
AI does not have one view of your brand. It has two, and they barely overlap. Only 8.28% of observed brand-category pairs are present in both AI's real-time search results and its training-embedded knowledge. The remaining 91.72% are visible to one path and invisible to the other.
Vol. 1, The Authority Gap, established that AI does not treat trust and recommendation as the same signal: the brands AI cites as sources of truth and the brands it recommends are, overwhelmingly, different entities. This paper extends the inquiry to a different axis, how AI distributes visibility across the two mechanisms it uses to form a response: the Search Path, where models retrieve information in real time, and the Learned Path, where they draw on knowledge embedded during training.
The dominant market assumption is that these two paths converge on the same answer. Our findings suggest the opposite: they are largely independent signals, and the brands present in both are the rarest position in any AI-visibility framework yet measured.
When a consumer asks an AI model for a recommendation, the model assembles its answer through two parallel processes: real-time retrieval from the open web, and recall of brand knowledge embedded during training. They are not interchangeable. ARDI™ observes them independently.
The Search Path is volatile. Its outputs change with every retrieval-policy update and search-index refresh. A brand can appear one month and vanish the next, with no change to the brand itself. The Learned Path is stable. It persists across retrieval changes until the model is retrained, which happens rarely and unpredictably.
The two paths fail in different ways. Search-dependent visibility is exposed to retrieval-policy risk; a single change in how a model selects sources can erase a brand overnight. Learned-dependent visibility is exposed to retraining risk. A brand present in both has a resilience neither single path provides.
The Search Path tells you what AI surfaces today. The Learned Path tells you what AI knows. Most brands are visible in only one.
ARDI™ observes AI behavior on a continuous monthly cadence. For this paper, we compared Decision-stage observations from both paths: the Search Path with real-time retrieval enabled, the Learned Path with retrieval disabled, capturing responses from training-embedded knowledge alone. Same prompts, same models, same categories, same markets, observed independently in each path.
The Learned Path is an observable proxy for a model's parametric brand knowledge. We do not claim direct access to any training corpus; we measure what the model produces when constrained to respond from learned knowledge alone. The unit of analysis is the observed brand-category pair.
| Parameter | Value |
|---|---|
| AI models observed | 6 (ChatGPT, Claude, Gemini, Perplexity, Grok, Microsoft Copilot*) |
| Industry categories | 36 actively observed (57 in library) |
| Geographic markets | 32 actively observed (41 in library) |
| Prompts in testing library | 20,800+ |
| Model executions (baseline) | 23,800+ |
| Citations analyzed (baseline) | 80,300+ |
| Brands identified | 8,600+ |
| Observation period | March-May 2026 (ongoing) |
*Microsoft Copilot tracking uses Azure OpenAI GPT-4o as the underlying proxy. Live Bing-grounded retrieval is not yet captured; Copilot results reflect the training-embedded layer only. See Section 7.
For each brand-category pair, the dataset answers: did the brand appear in Decision-stage Search Path responses, in Learned Path responses, in both, or neither? Present in both is Anchored Authority; Search only is Surfacing Authority; Learned only is Latent Authority; neither is Invisible, a conceptual position not directly measurable from observed output. The category-level analysis is scoped to the three categories with comparable dual-path depth: Pilates Studios, Massage / Wellness Studios, and Med Spas.
We expected meaningful overlap, perhaps 30 to 40 percent, since brands prominent enough to be searched today are usually prominent enough to be embedded in training. We observed almost the opposite: the two signals are largely independent.
Known to the model but not surfaced today. The most commercially misleading position: brands believe they have visibility while consumers are sent elsewhere.
Surfaced in real time but not embedded in training. Genuine but exposed; one retrieval-policy change can erase it.
Named in real-time search and embedded in training. The rarest and most durable position, resistant to policy and training change.
Latent is the most actionable position. AI already knows the brand; re-entering the Search Path is the addressable part. Training-data presence is the expensive part, and it is already paid.
The two paths reach different conclusions because they run on different inputs. Search is present-tense, shaped by what the open web surfaces now. Learned is past-tense, shaped by what the model absorbed during training. They favor different brand profiles: national franchises with sustained reference content dominate the Learned Path; locally optimized brands surface in the Search Path. The pattern held in every category with deep dual-path observation.
Pilates Studios (14.4% / 27.2% / 58.4%) tracks the global pattern. Club Pilates appears in both paths at near-identical rank (2.38 search, 2.17 learned). Massage / Wellness (24.0% / 39.6% / 36.4%) is the most balanced, driven by Massage Envy and Elements Massage. Med Spas (10.0% / 10.9% / 79.1%) shows extreme Latent dominance: nearly four in five brands are known but not surfaced. In every case, Anchored is the rarest position.
AI knows more than it says. AI says more than it knows. The overlap between knowing and saying is where durable visibility lives, and it is just 8% of observed brands.
We extend the four-position classification from Vol. 1 to the Search × Learned axis. A brand's position in any category is defined by its presence in each path.
| Position | Search Path | Learned Path | Durability |
|---|---|---|---|
| Anchored Authority | Present | Present | High |
| Surfacing Authority | Present | Absent | Low |
| Latent Authority | Absent | Present | Moderate |
| Invisible | Absent | Absent | None |
Not represented in observed AI output.
The predictive hypothesis, under longitudinal validation, is that Anchored brands are structurally more resilient to model and retrieval changes than single-path brands. Within the Anchored quadrant, rank alignment is itself a resilience signal: Club Pilates, Massage Envy, and SkinSpirit sit at near-identical ranks across both paths, while Pure Barre drifts (2.74 search vs 4.15 learned), a sign its strength is currently retrieval-driven.
Vol. 1 mapped the Trust × Recommendation axis (Full, Borrowed, Unconverted, Invisible). This paper maps the Search × Learned axis (Anchored, Surfacing, Latent, Invisible). Together they describe the same question, where a brand's AI visibility is durable and where it is vulnerable, from two vantage points. In both, the most durable position is the rarest: under 1% Full Authority in Vol. 1, 8.28% Anchored here. AI visibility is a four-axis problem, trust, recommendation, retrieval, and embedding, and brands that measure one axis are flying blind on three.
The current generation of AI-visibility tools observes a single path, usually real-time retrieval, and reports it as "AI visibility." That cannot distinguish visibility driven by retrieval from visibility supported by embedded knowledge. A brand strong in search but absent from training will look healthy by single-path measurement, and be one policy change from gone.
Each position implies a different move. Latent brands own the hard asset (embedded knowledge) and need to activate real-time signals. Surfacing brands have current visibility but lack the foundation that protects against retrieval shifts. Anchored brands hold the rarest position and must defend both layers. Invisible brands need to engineer entry into whichever path the category rewards.
A brand's available positions depend partly on category dynamics outside its control. Some categories produce deep observation in one path but not the other.
| Coverage pattern | Example categories | Search depth | Learned depth |
|---|---|---|---|
| Dual-coverage | Pilates, Massage / Wellness, Med Spas | Substantial | Substantial |
| Search-only | Pet Groomers, Chiropractors, Cosmetic Dentists | Substantial | None |
| Learned-only | Car Dealerships, Wealth Management | None | Substantial |
A pet-grooming brand in a Search-only category has no current path to Anchored through brand work alone; a wealth-management firm in a Learned-only category has no current path through training prominence alone. Strategy must be calibrated to the category's observable pipeline, and these designations are observation-stage findings, not final category laws.
What AI surfaces depends on retrieval. What AI knows depends on training. Durable AI visibility requires both.
Sample density. The aggregate figures span all observable brand-category pairs; the category-level percentages reflect the three deeply-observed categories. As more categories accumulate dual-path cycles, aggregates may adjust.
Coverage asymmetry. Observation depth varies by category and pipeline. For single-path categories, four-position analysis is not yet possible; future releases extend it as Learned coverage grows.
Brand canonicalization. Entity normalization is an ongoing refinement. Aggregate distributions are directional and stable across refinements; named examples are hand-verified against the underlying data.
Causality. This documents a structural correlation, not causation. Longitudinal work examines whether Anchored predicts resilience; the link to consumer behavior is the subject of the in-development Vol. 4.
Microsoft Copilot proxy. Copilot tracking uses Azure OpenAI GPT-4o and captures the reasoning layer, not Copilot's live Bing grounding. Live grounding is scheduled for ARDI™ v3.1; until then Copilot reflects the training-embedded layer only.
Section 8The Search/Learned Divergence is real, measurable, and consistent across the categories with deep dual-path observation. AI does not treat real-time search and training-embedded knowledge as the same signal. Only 8.28% of observed brand-category pairs are present in both, a previously unmeasured dimension of brand visibility that single-path measurement cannot capture.
Vol. 1 showed AI separates trust from recommendation. Vol. 2 shows it also separates what it has learned from what it currently retrieves. AI visibility is a four-axis problem: trust, recommendation, retrieval, and embedding. Subsequent volumes examine model-level variance (Vol. 3) and the link between path position and consumer behavior (Vol. 4).
The brands that win in AI will not be those most visible in any one path, but those most present across every layer of the system that decides what AI says next.
Yet nearly every AI-visibility tool on the market watches a single path, usually real-time search, and reports it as the whole picture. It is not. Search is one of two independent systems that decide what AI knows about a brand and whether it recommends one.
Most companies are measuring only half of their AI visibility.
The other half, the training-embedded layer, is the more durable one, and it is the half almost no one is watching. Measuring both paths, and the gap between them, is the entire point of ARDI™.
A measurable gap between the brands AI surfaces in real-time search and the brands AI knows from training. GOSH AI found only 8.28% of observed brand-category pairs appear in both paths, meaning AI's real-time recommendations and its training-derived knowledge are largely independent. Most brands are visible to one path but invisible to the other.
Anchored Authority is a brand present in both AI's real-time search responses and its training-embedded knowledge. It's the most durable AI-visibility position, because no single retrieval-policy change or training cycle can easily move it. Only 8.28% of observed brand-category pairs hold it, the rarest of the four Search/Learned positions.
Vol. 1, The Authority Gap, showed AI separates trust from recommendation across the buyer journey. Vol. 2 shows AI also separates real-time search from training-embedded knowledge across observation paths. Together they frame AI visibility as a four-axis problem: trust, recommendation, retrieval, and embedding.
Most tools observe a single path, sending queries with real-time retrieval on and reporting the result. That can't distinguish retrieval-driven visibility from training-embedded visibility. ARDI™ observes both paths independently and cross-references them at the brand-category level, producing the four-position classification: Anchored, Surfacing, Latent, Invisible.
APA:
GOSH AI. (2026). The Search vs Learned Divergence: Quantifying the gap between brands AI surfaces from real-time search and brands embedded in foundational training data (Updated ed.). ARDI™ Research, The State of AI Visibility, Vol. 2. https://www.mygosh.ai/search-learned-divergence
MLA:
GOSH AI. "The Search vs Learned Divergence." The State of AI Visibility, vol. 2, updated July 2026, www.mygosh.ai/search-learned-divergence.
These are independent third-party studies. None are inputs to the ARDI™ dataset reported above, and none were used to derive any finding in this paper. They are listed because they examine the same question from outside our data, and readers evaluating this work should be able to check it against them.