The Search vs Learned Divergence

Introduction

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 AI's training-embedded knowledge. The remaining 91.72% are visible to one path and invisible to the other.

The first paper in this series, The Authority Gap (State of AI Visibility, Vol. 1, April 2026), established a structural finding about how AI processes brands across the buyer journey: AI does not treat trust and recommendation as the same signal. The brands AI cites as sources of truth during early-stage research and the brands AI ultimately recommends to consumers are, in the overwhelming majority of cases, different entities.

This paper extends the line of inquiry to a different axis of divergence. Vol. 1 measured how AI distributes brand visibility across the stages of the buyer journey. This paper measures how AI distributes brand visibility across the two observational paths AI uses to form a response: the Search Path, where models retrieve information in real time, and the Learned Path, where models draw on knowledge embedded during training.

The dominant assumption in the current AI visibility market - and in most existing measurement tools - is that these two paths converge on the same answer. A brand AI surfaces in real-time search would, presumably, also be a brand AI has embedded from training. A brand absent from search results would, presumably, also be absent from the model's underlying knowledge.

Our findings suggest the opposite. AI's real-time recommendations and AI's training-embedded knowledge are largely independent signals. The observed overlap between the two paths is structurally low, and the brands present in both are the rarest position observed in any AI visibility framework yet measured.

This paper documents that finding, classifies brands by their Search × Learned position, and proposes a sister framework to the one introduced in Vol. 1. Read together, the two volumes describe a multi-axis problem that no single measurement strategy currently addresses.

Research Framework: The Two-Path Observation Model

Vol. 1 of this series introduced the AI buyer journey as a three-stage decision sequence - Think, Discovery, Decision - and demonstrated that AI behaves differently at each stage. That framework identified where in the journey AI processes brand information differently. This paper introduces a complementary framework that identifies how AI processes brand information, through two structurally distinct mechanisms.

When a consumer asks an AI model for a recommendation, the model assembles its response through two parallel processes operating simultaneously: real-time retrieval of current information from the open web, and recall of brand knowledge embedded during the model's training. These are not interchangeable. Each process has different inputs, different characteristics, and different vulnerabilities. We refer to them as the Search Path and the Learned Path, and the ARDI™ platform observes them independently.

The Search Path

The Search Path captures AI behavior when models have access to real-time search. When a model is permitted to retrieve information from the live web during response generation, the resulting recommendations reflect the current state of search engines, citation networks, and indexed content. A brand well-represented in current search results is more likely to surface in the model's recommendation; a brand absent from current search results is unlikely to be named, regardless of whether the model has prior knowledge of that brand.

The Search Path is volatile. Its outputs change with every retrieval policy update, every search index refresh, and every shift in citation patterns across the open web. A brand can appear in Search Path results one month and disappear the next, without any change to the brand itself.

The Learned Path

The Learned Path captures AI behavior when models rely on training-embedded knowledge. When a model generates a response without real-time retrieval, the resulting recommendations reflect the brand information the model absorbed during pretraining and any subsequent fine-tuning cycles. A brand that was well-represented in the training data corpus, through high-volume mentions, structured citations, or category-level prominence, is more likely to surface; a brand not embedded during training is unlikely to be named, regardless of how visible it is in current search results.

The Learned Path is stable. Its outputs persist across retrieval policy changes and remain consistent until the model is retrained, which happens infrequently and unpredictably. A brand that has earned embedded presence in a model's training data has a form of visibility that no real-time intervention can immediately remove.

Why Independent Measurement Matters

Most AI visibility measurement strategies in current use collapse these two paths into a single observed output. They issue a query to an AI model, capture the recommendation that comes back, and treat that recommendation as a unitary signal of brand visibility. This approach cannot distinguish between a brand visible because of real-time retrieval and a brand visible because of embedded training knowledge.

The distinction matters because the two paths fail in different ways. A brand whose visibility depends on the Search Path is exposed to retrieval-policy risk: a single change in how a model selects sources can erase that brand from its recommendations overnight. A brand whose visibility depends on the Learned Path is exposed to retraining risk: at the next training cycle, embedded prominence can shift in ways the brand has no ability to influence in the short term. A brand present in both paths has a form of resilience neither single-path visibility provides.

Observation Methodology

The ARDI™ platform observes AI behavior on a continuous monthly cadence. Each observation cycle issues a structured library of prompts to six leading AI models, captures every brand mention and citation produced in the resulting responses, and records the result in a unified observation spine that preserves buyer-stage, category, geographic market, and model attribution for every observation.

For this paper, the analysis compares Decision-stage observations from both paths. The Search Path executes its prompt cycles against AI models with real-time retrieval enabled. The Learned Path executes the same Decision-stage prompts against the same AI models with retrieval disabled, capturing responses generated from the model's training-embedded knowledge alone. The same buyer-stage filter, the same prompts, the same models, the same categories, the same markets - observed independently in each path.

A note on what the Learned Path measures. The Learned Path observes model behavior with real-time retrieval disabled. The resulting responses are an observable proxy for the model's training-embedded or parametric brand knowledge. We do not claim direct access to any model's training corpus. We measure what the model produces when it is constrained to respond from learned knowledge alone.

A note on the analytical unit. The unit of analysis throughout this paper is the observed brand-category pair. A brand operating in multiple categories appears multiple times in the dataset, once per category in which it was observed. Aggregate percentages and position classifications refer to brand-category pairs, not unique brands. Narrative passages occasionally use the shorthand "brands" where context makes the analytical unit clear; statistical claims always refer to brand-category pairs.

Observation Parameters

*Microsoft Copilot tracking uses Azure OpenAI GPT-4o as the underlying proxy. Live Bing-grounded retrieval is not captured in the current pipeline. See Section 7 for full methodology limitations.

All prompts are brand-agnostic. No prompt references a specific brand by name. Every brand mention or citation in the observed AI output is model-generated, not prompt-induced. This ensures observed brand presence reflects the model's behavior rather than the wording of the prompt.

Classification Methodology

The unit of analysis is the (brand, category) pair. For each such pair, the observation dataset answers three questions:

1. Did this brand appear in Decision-stage Search Path responses for this category?
2. Did this brand appear in Decision-stage Learned Path responses for this category?
3. One path, both paths, or neither?

Brands present in both paths are classified as Anchored Authority. Brands present in the Search Path only are classified as Surfacing Authority. Brands present in the Learned Path only are classified as Latent Authority. Brands present in neither path are, by definition, not represented in observed AI output. We refer to this fourth position as Invisible: a residual position that exists conceptually but is not directly measurable from this dataset. The implications of this measurement boundary are discussed in Section 7.

Scope of the Primary Analysis

Observation depth varies across categories. Some categories have substantial observation cycles in both paths; others have substantial observation in one path but not yet in the other. The aggregate findings in Section 4 are computed across all (brand, category) pairs in the observation dataset. The category-level analysis is scoped to the three categories where both paths have run to comparable depth: Pilates Studios, Massage / Wellness Studios, and Med Spas. Combined, these three categories account for approximately 29,000 brand mentions across approximately 6,750 distinct AI calls, providing sufficient depth for the per-category findings reported.

Categories with substantial single-path observation are addressed in Section 6 as examples of structural coverage asymmetry rather than as primary findings. As Learned Path coverage continues to expand across the full observation library on the standard monthly cadence, future ARDI™ releases will extend the per-category analysis to additional categories.

Key Finding: The Search/Learned Divergence

The central finding of this paper is that AI's real-time recommendations and AI's training-embedded knowledge are largely independent signals. When measured at the (brand, category) level across the observation dataset, the brands AI surfaces from real-time search and the brands AI knows from training data are, in the overwhelming majority of cases, different entities.

We define Search Path presence as a brand appearing at least once in Decision-stage AI responses generated with real-time retrieval enabled. We define Learned Path presence as a brand appearing at least once in Decision-stage AI responses generated with retrieval disabled, drawing solely on the model's training-embedded knowledge.

When we measured the overlap between these two signals across the observation dataset, we did not expect zero correlation. We expected meaningful overlap, perhaps 30 to 40 percent, reflecting the fact that brands prominent enough to be searched today are typically brands prominent enough to be embedded during training. We observed almost the opposite.

Why the Divergence Exists

The two paths reach different conclusions because they operate on different inputs. The Search Path is shaped by what the open web is currently surfacing: citation patterns, local search index behavior, content recency, and the particular signals the underlying retrieval system uses to rank sources at the moment of the query. A brand whose local-search presence is strong this month appears in the Search Path; a brand whose presence weakens next month drops out. Real-time retrieval is, by design, a present-tense system.

The Learned Path is shaped by what the AI model absorbed during training: the brand mentions, structured citations, encyclopedic references, and long-form content present in the training corpus at the time it was assembled. Training data is historical, meaning a brand that was prominent two years ago but has since faded may still be embedded; a brand that has emerged in the last twelve months may not yet be present at all. Training is, by design, a past-tense system.

These two systems also tend to favor different brand profiles. National franchises with sustained press coverage, structured encyclopedic entries, and long histories of reference content are more likely to be embedded in training. Local and regional brands, especially those whose visibility comes through current local-search optimization, are more likely to surface in real-time retrieval. Specialized verticals where consumer queries route through dedicated platforms before reaching general AI (automotive marketplaces, advisor-matching tools) tend to develop substantial Learned Path presence without comparable Search Path activation. The two paths are not noisy versions of the same signal. They are structurally different systems, and brands that look strong in one have no guarantee of being visible in the other.

The pattern was consistent across categories with deep dual-path observation, but the specific distribution varied. If the divergence were isolated to a single category it could be dismissed as noise. It is not.

Across the three categories, the Anchored position remains the rarest in every case, while Surfacing and Latent vary in dominance:

Pilates Studios shows the closest distribution to the global pattern. 14.4% Anchored, 27.2% Surfacing, 58.4% Latent. Recognizable national brands cluster in the Anchored position. Club Pilates, the most widely observed brand in the category, appears in both paths with near-identical average recommendation positions (2.38 in the Search Path, 2.17 in the Learned Path). Most independent and regional Pilates studios fall into Latent or Surfacing rather than Anchored. The competitive structure of the category, read through this distribution, is two-tiered. National franchises hold the Learned Path through years of accumulated reference content, while local studios occupy the Search Path through current local-search optimization. Being the brand AI knows and being the brand AI surfaces are, in this category, two distinct positions held by different operators.

Massage / Wellness Studios shows the most balanced distribution observed. 24.0% Anchored, 39.6% Surfacing, 36.4% Latent. The category produces the highest Anchored rate of the three, nearly three times the global average, driven by national franchise brands like Massage Envy and Elements Massage that appear in both paths. Surfacing exceeds Latent here, which is unusual. The pattern suggests that for this category, AI's training- embedded brand knowledge is comparatively shallower than its real-time search visibility. For operators inside the Latent population in this category, the divergence reads as a recoverable signal: the brand-recognition asset already exists in the model. What is missing is the present-tense surfacing layer that converts recognition into citation at the moment of the consumer query.

Med Spas shows the most extreme Latent dominance. 10.0% Anchored, 10.9% Surfacing, 79.1% Latent. Nearly four out of five observed Med Spa brand-category pairs are present in AI's training-embedded knowledge but not surfaced by real-time retrieval. SkinSpirit, a national medical aesthetics brand, illustrates the rare Anchored position with near-identical recommendation positions in both paths (2.20 in each). The dominant pattern, however, is structural. AI has substantial latent knowledge of Med Spa brands that is not being activated in current Decision-stage real-time recommendations. Neither path in this category is dominated by a stable set of recognizable names. The Learned Path reflects a long tail of clinical and aesthetic brands with inconsistent encyclopedic presence; the Search Path reflects a constantly shifting set of locally optimized practices. Med spas, more than the other two verticals, have not yet produced a stable AI-visible brand hierarchy in either path. The competitive opening is correspondingly wide.

The Anchored position remains the rarest in every category observed, even where the overall Surfacing-versus-Latent balance shifts dramatically. The pattern is not a feature of any single category. It is a feature of how AI itself processes brand information across paths.

The Authority Gap Framework, Extended

Based on these findings, we extend the four-position classification system introduced in Vol. 1 to the Search × Learned axis. A brand's position in any category is defined by its presence in each path:

The framework's predictive hypothesis - currently under longitudinal validation - is that Anchored brands are structurally more resilient to AI model and retrieval changes than brands present in only one path. Anchored brands carry visibility that no single retrieval-policy update can erase and no single training cycle can easily move. Surfacing and Latent brands, by contrast, are exposed in different ways to events outside the brand's control.

Named Exemplars by Position

Anchored Authority exemplars include national franchise brands with substantial dual-path presence. Club Pilates leads the Pilates Studios category with 992 Search Path mentions and 682 Learned Path mentions across multiple AI models and observation cycles, with average recommendation positions of 2.38 and 2.17 respectively. Massage Envy holds the equivalent position in Massage / Wellness Studios (361 search, 632 training; positions 2.54 and 2.61). SkinSpirit, a national medical aesthetics chain, shows the cleanest Anchored signal observed: 45 search and 40 training mentions with average recommendation positions identical at 2.20 in each path. AI's two paths agree on both presence and rank.

Surfacing Authority exemplars are typically locally prominent brands whose visibility is driven by current real-time citation behavior. In Pilates Studios, Page One Pilates, Atlas Pilates, and Breathe Pilates Studio appear in Search Path responses but not in Learned Path responses. In Med Spas, Elase Med Spa and Metamorph Med Spa show the same pattern. These brands have demonstrated visibility through real-time retrieval but have not yet earned embedded presence in the model's training data. Their visibility is genuine, but exposed.

Latent Authority exemplars are brands AI has retained from training but is not currently surfacing. SLT, the national Pilates and barre chain, appears 33 times in Learned Path observations for Pilates Studios but not in current Search Path responses. Burke Williams Spa shows the same pattern in Massage / Wellness Studios. SKINNEY MedSpa, a New York-based medical aesthetics chain, appears 41 times in Learned Path observations for Med Spas but is absent from real-time Decision-stage recommendations. These brands hold a form of dormant authority: present in AI's underlying knowledge, not active in its current responses.

Cross-Path Ranking Alignment

Within the Anchored quadrant, AI does not always agree with itself on rank order. Some brands sit at near-identical recommendation positions across both paths (Club Pilates, Massage Envy, SkinSpirit), suggesting the brand's authority is not just present in both paths but stable across them. Other Anchored brands show meaningful position drift, present in both paths but at different ranks. Pure Barre, for example, ranks at average position 2.74 in the Search Path but 4.15 in the Learned Path, indicating the brand's recommendation strength is currently retrieval-driven and may weaken if training-embedded prominence does not catch up. Position alignment within Anchored is itself a signal of structural resilience: the closer the two ranks, the more durable the position.

Two Frameworks, One Question

Vol. 1 introduced the Authority Gap framework along the Trust × Recommendation axis (Full Authority, Borrowed Authority, Unconverted Authority, Invisible). This paper introduces a sister framework along the Search × Learned axis (Anchored, Surfacing, Latent, Invisible). Together, the two frameworks describe the same underlying question, where is a brand's AI visibility durable and where is it vulnerable, from two structurally distinct vantage points.

In both frameworks, the most durable position is the rarest. In Vol. 1, fewer than 1% of observed brands occupied Full Authority. In this paper, 8.28% of observed brand-category pairs occupy Anchored Authority. The two positions are not equivalent, but they identify the same kind of brand: one whose AI visibility rests on multiple supporting signals rather than a single fragile one. Brands that occupy both positions simultaneously, cited as a source AND recommended AND embedded in training AND surfaced in real-time, are structurally the safest brands in AI visibility, and they are vanishingly few.

Implications

For the Market

The current generation of AI visibility tools observes a single path. Most send queries to AI models with real-time retrieval enabled, capture the resulting recommendations, and report those as the brand's "AI visibility." This methodology cannot distinguish between visibility driven by real-time retrieval and visibility supported by training-embedded knowledge. A brand whose Search Path performance is strong but whose Learned Path presence is zero will appear, by single-path measurement, to be doing well. Vol. 1 established that AI visibility is a multi-stage problem. This paper establishes that single-path measurement also misses an entire dimension of the picture.

For Brand Strategy

Each position implies a different strategic priority and a different state-transition target. Brands in Latent Authority have already earned the harder asset, embedded presence in AI's underlying knowledge, and need to activate real-time visibility signals to convert that latent presence into active recommendation. Brands in Surfacing Authority have demonstrated current visibility but lack the embedded foundation that protects against retrieval shifts; their priority is to develop the long-cycle authority signals that move a brand from real-time-only to dual-path. Brands in Anchored Authority have the rarest position and need to defend both layers against erosion. Brands in Invisible need to engineer entry into either path, whichever the category's structural pattern most rewards.

For AI Visibility Measurement

Single-path measurement systematically understates AI visibility risk. A brand that appears strong in real-time search alone is exposed to retrieval policy changes the brand cannot anticipate or control. A brand that appears strong in training-embedded knowledge alone is exposed to retraining cycles that may not preserve current rank. Comprehensive measurement requires both paths observed independently, then cross-referenced at the (brand, category) level. Any AI visibility report that does not state which path it measures is, in effect, choosing which half of AI's behavior the buyer can see.

Coverage Asymmetries Across Categories

Not every category in the observation universe has comparable observation depth in both paths. The category-level analysis in Section 4 was scoped to three categories with substantial dual-path coverage. The broader dataset reveals that many categories have substantial observation in one path but not the other, and the asymmetry pattern is itself a structural finding.

Search-only categories tend to be local-service verticals where AI surfaces brands via real-time retrieval against current local search results: Pet Groomers, Chiropractors, Cosmetic Dentists, and similar service-driven categories. The category structure favors Search Path activation: brands appear because they are currently retrievable, not because they have earned embedded knowledge presence in the model.

Learned-only categories tend to be specialized verticals where consumer queries typically route through dedicated platforms before reaching general AI: Car Dealerships through automotive marketplaces, Wealth Management Firms through advisor-matching tools. AI has substantial embedded knowledge of these brands but is not asked to surface them in general Decision-stage queries. The category structure favors Learned Path retention without current real-time activation.

The implication for brand strategy is that a brand's available positions in the framework depend partly on category-level dynamics outside its control. In the current observation universe, a pet grooming brand operating in a Search-only category does not have a path to Anchored Authority through brand-level work alone, because the category does not currently produce Learned Path observations in our pipeline. A wealth management firm in a Learned-only category does not have a path to Anchored Authority through training-data prominence alone, because the category does not currently produce Search Path observations. Strategy must be calibrated to the category's observable pipeline shape, not only to the brand's individual position.

These single-path designations should be read as observation-stage findings, not as final category laws. The asymmetry may reflect true market structure, where the category genuinely routes consumer attention through one path more than the other, or it may reflect the current maturity of ARDI's observation pipeline in that vertical. As dual-path coverage expands in future research, GOSH AI will distinguish structural single-path categories from those where observation has not yet caught up. The strategic guidance above applies to the category as currently observed.

Limitations and Ongoing Research

This paper presents findings from an ongoing observation program. Several limitations should be noted.

Sample density. The aggregate findings (8.28% Anchored, 28.82% Surfacing, 62.90% Latent) are computed across all observable (brand, category) pairs in the dataset. The category-level percentages in Section 4 reflect the three categories with comparable observation depth in both paths. As additional categories accumulate dual-path observation cycles, aggregate percentages may adjust. Trend stability across cycles is monitored monthly.

Pipeline coverage asymmetry. Observation depth varies by pipeline and by category. Section 6 documents the asymmetry pattern explicitly. For categories with substantial single-path observation, brand-position analysis cannot yet be conducted at the four-position level. Future ARDI™ releases will extend the analysis as Learned Path observation expands across the broader category library on the standard monthly cadence.

Brand canonicalization. Entity normalization for the brand-extraction layer is an ongoing refinement. A small percentage of brand entries in observed AI output reflect generic phrases or category terms misclassified as brand entities. Aggregate percentages and category-level distributions are directional and stable across canonicalization refinements; the structural finding holds whether or not noise entities are filtered. Named brand examples in this paper are hand-verified against the underlying observation data. Future ARDI™ releases will further reduce normalization noise.

Causality. This paper documents a structural correlation pattern between path presence and brand visibility. It does not yet establish a causal relationship between path position and downstream consumer outcomes (clicks, leads, revenue, market share). The longitudinal observation program currently underway is examining whether Anchored Authority is predictive of resilience to AI model and retrieval changes; the linkage between path position and consumer behavior is the subject of the in-development Vol. 4.

Microsoft Copilot proxy. Microsoft Copilot tracking uses Azure OpenAI's GPT-4o as the underlying model proxy. This captures the reasoning layer of consumer Copilot but not its real-time web grounding (Bing Search), which Copilot applies at runtime. Live web grounding via the Azure AI Foundry Grounding with Bing Search integration is scheduled for ARDI™ v3.1. Until then, Copilot results in this dataset reflect the model's training-embedded knowledge layer only, not its real-time retrieval behavior.

Conclusion

The Search/Learned Divergence is real, measurable, and consistent across the categories with deep dual-path observation.

AI does not treat real-time search results and training-embedded knowledge as the same signal. The brands AI surfaces in real-time and the brands AI knows from training are, in the overwhelming majority of cases, different entities. Only 8.28% of observed brand-category pairs are present in both paths. This creates a previously unmeasured dimension of brand visibility, one that single-path measurement strategies cannot capture.

For brands, the immediate implication is structural. A position of strong real-time visibility without embedded knowledge is exposed to retrieval-policy risk. A position of strong embedded knowledge without real-time visibility is exposed to retrieval activation gaps. The Anchored position, presence in both paths, is the only structurally durable form of AI visibility observed in this dataset, and it is rare across every category measured.

This research will continue as a longitudinal observation program. Vol. 1 of this series showed that AI separates trust from recommendation. Vol. 2 shows that AI also separates what it has learned from what it currently retrieves. AI visibility is a four-axis problem covering trust, recommendation, retrieval, and embedding. Brands that measure only one axis are flying blind on three. Subsequent volumes will examine model-level variance (Vol. 3), the linkage between path position and consumer behavior (Vol. 4), and category-specific deep dives within individual verticals.

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.