Creative DNA: A First-Principles Framework for Decomposing Direct Response Video Advertisements into Atomic Cognitive Dimensions

Research Thesis — Version 1.0 February 2026

Executive Summary

This document develops a neuroscience-grounded framework for decomposing direct response (DR) video advertisements into measurable atomic dimensions — a "Creative DNA" — enabling data-driven hypothesis testing against performance metrics. The framework is designed for health supplement offers targeting American women aged 50+ on Meta platforms.

Rather than reasoning from advertising conventions, we trace each proposed dimension to a specific cognitive or neural mechanism that influences purchase behavior. The result is a 10-dimension schema that balances explanatory power with statistical viability given a dataset of ~164 creatives.

Core finding: The preliminary 7-dimension framework captures most critical variance but contains two significant gaps (temporal/structural architecture and sensory-channel modality) and one dimension that requires decomposition (Hook Structure conflates formal structure with content, destroying analytical value). The revised framework proposes 10 dimensions organized into three functional tiers based on their position in the cognitive processing pipeline.

Part 1: The Cognitive Processing Pipeline — What Actually Happens in the Brain

Before proposing dimensions, we must map what happens neurologically when a 55-year-old woman scrolling Facebook encounters a video advertisement. Each processing stage has different neural substrates, operates on different timescales, and is influenced by different stimulus properties. This pipeline is the foundational architecture from which all dimensional recommendations derive.

Stage 1: Pre-Attentive Feature Detection (0–200ms)

Neural substrate: Primary visual cortex (V1), superior colliculus, pulvinar nucleus of the thalamus.

Before conscious awareness, the visual system performs massively parallel feature extraction. Itti & Koch's (2001) saliency model demonstrates that three low-level channels drive pre-attentive capture: color contrast, luminance contrast, and orientation contrast. The superior colliculus computes a spatial "priority map" that determines where the eyes move next (Fecteau & Munoz, 2006).

Critical implication for DR ads: The first frame of the video competes against every other element in the Facebook feed. This competition is won or lost by low-level visual properties — not by message content. A text-heavy thumbnail with low color contrast will lose to a high-contrast face regardless of the message's persuasive quality.

Relevant stimulus properties: Visual contrast, presence of faces (fusiform face area activation is pre-attentive; Kanwisher et al., 1997), motion onset (abrupt onset captures attention involuntarily; Yantis & Jonides, 1984), visual complexity.

Stage 2: Attention Capture & Orienting Response (200ms–3s)

Neural substrate: Anterior cingulate cortex (conflict monitoring), ventral attention network (stimulus-driven reorienting), amygdala (threat/novelty detection).

Once pre-attentive features trigger a saccade, the brain performs a rapid threat/novelty assessment. The amygdala processes emotional significance of visual stimuli within 200ms, even before detailed cortical processing (LeDoux, 1996; Vuilleumier et al., 2001). The orienting response — Sokolov's (1963) foundational concept — is triggered by stimuli that are novel, unexpected, or potentially significant.

Two distinct attentional systems compete here (Corbetta & Shulman, 2002): the top-down dorsal network (goal-directed: "I'm looking for a recipe") and the bottom-up ventral network (stimulus-driven: "Something unexpected just appeared"). DR ads must hijack the ventral network to interrupt default scrolling behavior.

Critical implication: The 3-second rule in advertising is not arbitrary — it maps to the approximate duration of the orienting response before habituation occurs (Bradley, 2009). If the stimulus fails to transition from exogenous (involuntary) to endogenous (voluntary) attention within this window, the viewer scrolls past.

Relevant stimulus properties: Novelty/incongruity (prediction error signal; Friston, 2010), emotional arousal intensity, self-relevance detection (medial prefrontal cortex activation; Kelley et al., 2002), information gap (curiosity as a form of deprivation; Loewenstein, 1994).

Stage 3: Relevance Appraisal & Self-Referencing Gate (1–5s)

Neural substrate: Medial prefrontal cortex (self-referencing), temporal-parietal junction (perspective-taking), default mode network (autobiographical retrieval).

This is the most critical and least discussed stage. Once attention is captured, the brain performs a rapid relevance computation: "Is this about me?" The medial prefrontal cortex (mPFC) activates specifically when processing self-relevant information (Northoff et al., 2006). This self-referencing effect dramatically enhances subsequent encoding and memory formation (Rogers et al., 1977; Symons & Johnson, 1997 meta-analysis: d = 0.50 advantage for self-referenced information).

Scherer's (2001) appraisal theory identifies the key questions the brain answers at this stage: (1) Is this relevant to my goals? (2) Does it help or hinder those goals? (3) Can I cope with it? These appraisals determine whether the stimulus generates approach or avoidance motivation.

Critical implication: A video that generates strong attention (Stage 2) but fails the relevance gate (Stage 3) produces the pattern of high 3-second view rates but poor hold rates — a common failure mode in DR ads. The viewer's internal dialogue at this stage is essentially: "Interesting, but not for me." The dimensions that govern this gate are identity-match (avatar), problem recognition (pain point), and perceived applicability.

Relevant stimulus properties: Avatar similarity/identification (Bandura, 1977: observational learning requires perceived similarity), problem specificity and personal resonance, perceived target audience ("people like me"), age/gender/lifestyle cues.

Stage 4: Narrative Engagement & Causal Elaboration (5–60s)

Neural substrate: Temporal pole and anterior temporal lobe (narrative comprehension), dorsolateral prefrontal cortex (causal reasoning), hippocampus (episodic simulation), mirror neuron system (empathetic engagement).

Once past the relevance gate, the brain shifts from evaluative to narrative processing. Hasson et al.'s (2008) neurocinematics research shows that well-structured narratives produce inter-subject neural synchronization — viewers' brains literally align in their processing patterns, indicating shared comprehension and engagement.

Causal reasoning engages the dorsolateral prefrontal cortex and follows predictable structures. Sloman's (2005) "causal model theory" demonstrates that people understand and believe explanations that provide a plausible causal chain from mechanism to outcome. The causal chain need not be scientifically accurate to be persuasive — it must be internally coherent and connect to existing mental models (Fernbach et al., 2013).

Transportation theory (Green & Brock, 2000) provides the key mechanism here: narrative transportation reduces counter-arguing. When a viewer is absorbed in a story, the critical evaluation system (lateral prefrontal cortex) becomes less active, making persuasion attempts more effective. This is not deception — it is the natural consequence of engagement with narrative.

Critical implication: This stage is where the ELM's (Petty & Cacioppo, 1986) central and peripheral routes interact most complexly. The viewer is in a hybrid processing state — following a narrative (peripheral route) while also evaluating causal claims (central route). The quality of the causal mechanism explanation determines whether the viewer builds a mental model of how the product works, which is the foundation for belief formation.

Relevant stimulus properties: Causal mechanism clarity and plausibility, narrative structure (problem → explanation → solution), information novelty (prediction error sustains attention; Berns et al., 2001), evidence type and credibility, emotional arc.

Stage 5: Motivational Activation & Desire Formation (Throughout + Crescendo)

Neural substrate: Ventral striatum/nucleus accumbens (reward anticipation), orbitofrontal cortex (value computation), amygdala-insula circuit (loss aversion).

Motivation is not a discrete stage but a cumulative process. Berridge & Robinson's (2016) incentive salience theory distinguishes between "liking" (hedonic pleasure) and "wanting" (motivational drive). DR ads must generate "wanting" — a future-oriented motivational state that drives action.

Prospect theory (Kahneman & Tversky, 1979) provides the key insight: losses are approximately 2x more motivating than equivalent gains. The neural basis for this asymmetry is the amygdala-insula circuit, which responds more strongly to potential losses than potential gains (Tom et al., 2007).

Temporal discounting (McClure et al., 2004) creates a competing force: future benefits are neurally "discounted" relative to immediate costs (effort, money, risk). The beta system (ventral striatum, mPFC) drives immediate gratification while the delta system (lateral prefrontal, parietal cortex) computes long-term value. DR ads must activate the beta system to overcome the delta system's default preference for inaction.

Critical implication: The emotional structure of the ad determines which motivational system is activated. Fear/loss framing engages the amygdala-insula circuit (high urgency, loss aversion). Hope/gain framing engages the ventral striatum (reward anticipation). The optimal strategy depends on the viewer's current state and the product category — but for health supplements targeting pain or weight loss in older adults, loss-framing (losing independence, losing quality of life) typically generates stronger motivational activation than gain-framing.

Relevant stimulus properties: Emotional valence and arousal pattern, specificity of promised outcome, temporal framing (immediate vs. future benefits), social proof as risk reduction, scarcity/urgency cues.

Stage 6: Action Threshold & Decision Execution (CTA Moment)

Neural substrate: Supplementary motor area (action preparation), anterior cingulate cortex (cost-benefit computation), prefrontal cortex (inhibition override).

The final stage requires crossing an action threshold — translating motivation into motor behavior (clicking). Gollwitzer's (1999) implementation intention research demonstrates that the gap between motivation and action is substantial, and is bridged by specificity of the intended action ("I will click and try this" vs. vague positive attitude).

Cognitive load at the moment of decision is critical. Shiv & Fedorikhin (1999) showed that high cognitive load increases reliance on affective (System 1) over deliberative (System 2) processing, making emotionally-primed decisions more likely. The clarity and simplicity of the CTA determines whether the accumulated motivation converts to action.

Critical implication: Everything that happens before the CTA moment is about building sufficient motivational energy to overcome the default state of inaction. The CTA itself must minimize friction and cognitive load while maximizing perceived urgency. The interaction between accumulated emotional state and CTA clarity is multiplicative, not additive.

Relevant stimulus properties: CTA clarity and specificity, perceived friction/cost, urgency framing, risk reversal (guarantees), commitment framing.

Part 2: Stress-Testing the Preliminary 7-Dimension Framework

With the cognitive pipeline established, we can now evaluate each proposed dimension against the question: "Does this dimension correspond to a distinct causal mechanism in the processing pipeline?"

Dimension 1: Problem/Pain Referenced — VALIDATED, NEEDS REFINEMENT

Cognitive role: Governs Stage 3 (Relevance Gate). The problem referenced activates self-referencing circuits in the mPFC when the viewer recognizes their own experience. It also sets the motivational valence — pain conditions activate loss-aversion circuits differently than cosmetic concerns.

Refinement needed: This dimension conflates two separable cognitive phenomena:

(a) Problem category (joint pain, weight, fatigue, digestion) — this determines WHICH population self-selects through the relevance gate. It is essentially an audience-targeting variable embedded in the creative.

(b) Problem severity/framing (mild inconvenience vs. life-altering condition) — this determines the INTENSITY of loss-aversion activation. "Occasional knee discomfort" and "crippling joint pain that's stealing your independence" reference the same problem category but activate completely different neural responses.

Recommendation: Retain as a single dimension but ensure the coding schema captures both category and intensity level. The category drives targeting; the intensity drives motivation.

Dimension 2: Causal Mechanism — VALIDATED, HIGH THEORETICAL IMPORTANCE

Cognitive role: Governs Stage 4 (Narrative Engagement). The causal mechanism is the explanatory backbone that transforms a sales pitch into a story. It engages the dorsolateral prefrontal cortex's causal reasoning circuits and enables the viewer to build a mental model.

Fernbach et al.'s (2013) "illusion of explanatory depth" research is directly relevant: people who are given a causal explanation (even a simplified one) feel they understand the phenomenon deeply, which increases confidence in the solution. A viewer who "understands" that their joint pain is caused by cartilage depletion from a specific enzyme will find the supplement solution more credible than one who is simply told "this works."

Critical nuance: The mechanism need not be scientifically rigorous — it must be narratively coherent and connect to the viewer's existing mental models. "Gut bacteria" works because people have a folk model of gut health. "Mitochondrial uncoupling proteins" requires building a new mental model, which increases cognitive load and reduces persuasion in a low-elaboration context.

Recommendation: Retain as-is. This is one of the highest-leverage dimensions because it directly determines the quality of narrative engagement and belief formation.

Dimension 3: Dominant Emotional Structure — VALIDATED, NEEDS EXPANSION

Cognitive role: Governs Stage 5 (Motivational Activation) and modulates ALL other stages. Emotion is not a single processing stage — it is a continuous modulator of attention, encoding, and decision-making (Pessoa, 2008).

Current values (fear/loss, hope/gain, curiosity/revelation, indignation/injustice) are well-chosen because they map to distinct motivational systems:

• Fear/loss → Amygdala-insula circuit → Avoidance motivation → Loss aversion
• Hope/gain → Ventral striatum → Approach motivation → Reward anticipation
• Curiosity/revelation → Dopaminergic prediction error → Information-seeking → Sustained attention (Gruber et al., 2014: curiosity enhances memory encoding)
• Indignation/injustice → Anterior insula + dorsal anterior cingulate → Moral anger → External attribution (blame) + action tendency

Missing value: Consider adding Shame/Social Comparison as a fifth category. Fessler (2004) demonstrates shame is a distinct motivational state that drives behavior change through social rank computation. Many DR health ads implicitly leverage shame ("you've tried everything and nothing works" / "your friends don't understand"). This is neurally distinct from fear (which is about physical threat) and indignation (which is about external injustice).

However, given the n≈164 constraint, adding a fifth emotional category may dilute statistical power. Recommendation: Retain four categories but add coding notes that distinguish "shame-inflected fear" from "pure threat fear" within the fear/loss category.

Dimension 4: Hook Structure — NEEDS DECOMPOSITION (Critical Issue)

Cognitive role: Governs Stage 2 (Attention Capture). But as currently defined, this dimension conflates two independent cognitive mechanisms:

(a) Rhetorical form (question, statement, conditional, imperative, narrative) — this determines HOW the information is syntactically structured. A question activates the instinctive answer-generation circuit (Meyer et al., 2010: questions create obligatory processing). A statement can be ignored. An "if you have X" conditional triggers relevance-checking.

(b) Content type of the hook (shocking claim, personal story, visual demonstration, social proof) — this determines WHAT information is used to capture attention. A shocking claim exploits prediction error (Friston, 2010). A personal story activates narrative transportation. A visual demonstration exploits the picture superiority effect (Paivio, 1986).

These are orthogonal: you can have a "question + shocking claim" ("Did you know 73% of doctors hide this?") or a "statement + personal story" ("At 57, I was told I'd never walk without pain again") or a "conditional + visual demo" ("If you can see this belly fat, watch what happens when..."). The current schema forces a single classification where two independent codes are needed.

Recommendation: Decompose into two dimensions:

• Hook Rhetorical Device (question, conditional/qualifier, imperative, declarative claim, narrative opening)
• Absorb the content-type element into other existing dimensions (the "what" of the hook is already captured by Problem, Mechanism, Emotional Structure, and Proof Type)

This is the single most important revision to the framework. The conflated dimension is the primary source of the inter-rater reliability problems described in the problem statement.

Dimension 5: Proof Type — VALIDATED, MINOR REFINEMENT

Cognitive role: Governs the transition from Stage 4 (Engagement) to Stage 5 (Motivation) by modulating BELIEF STRENGTH. Without proof, narrative engagement generates interest but not conviction.

The neuroscience of credibility is well-studied. Source credibility effects operate through the medial prefrontal cortex (Klucharev et al., 2008), and different proof types activate different trust mechanisms:

• Testimonial → Social proof (Cialdini, 2006) + Mirror neuron system (empathetic simulation of the testifier's experience; Iacoboni, 2009)
• Authority/study citation → Expertise heuristic (Petty & Cacioppo, 1986) + Reduced need for personal evaluation
• Visual demonstration → Sensory evidence, the most phylogenetically ancient and trusted form of proof (seeing is believing — direct ventral stream processing bypasses verbal skepticism)
• Before/after → Counterfactual reasoning (dorsolateral PFC) + Emotional contrast (amygdala response to transformation)
• Mechanism explanation (pseudo-scientific) → Illusion of explanatory depth (Fernbach et al., 2013) + Fluency-as-credibility heuristic

Refinement: Add "social proof/numbers" as a distinct category from testimonial. "Over 2 million women have tried this" activates a different neural mechanism (bandwagon heuristic, numerosity processing in the intraparietal sulcus) than a single person's testimony. Also add "none/assertion only" as an explicit value — the absence of proof is analytically important.

Recommendation: Retain with expanded value set: testimonial, authority/study citation, visual demonstration, before/after transformation, social proof/numerical, mechanism-as-proof, none.

Dimension 6: Avatar Identity — VALIDATED, HIGH THEORETICAL IMPORTANCE

Cognitive role: Governs Stage 3 (Relevance Gate) through identification and Stage 4 (Engagement) through parasocial processing.

This is more important than the preliminary framework suggests. Identification theory (Cohen, 2001) and parasocial interaction theory (Horton & Wohl, 1956) demonstrate that the perceived identity of the on-screen person determines:

• Whether the viewer enters the relevance gate (demographic similarity triggers self-referencing)
• The processing route used (identified-with characters enable narrative transportation; observed experts enable authority-based peripheral processing)
• The motivational valence (a "peer who found a solution" activates approach motivation; an "authority warning about danger" activates avoidance motivation)

The current framework captures age, gender, and role — but role is the critical variable and needs richer enumeration:

• Peer/fellow sufferer → "This person is like me" → Self-referencing, empathy, narrative transportation
• Discoverer/whistleblower → "This person found something hidden" → Curiosity, conspiratorial bonding, epistemic trust
• Expert/authority → "This person knows things I don't" → Expertise heuristic, credibility transfer
• Caregiver/advocate → "This person cares about me" → Warmth heuristic, reduced threat response
• Narrator/faceless → No parasocial identification → Relies on message quality alone

Recommendation: Retain but formalize role categories as above. Demographic similarity (age bracket, perceived ethnicity, body type) should be coded as secondary attributes rather than primary codes, since these moderate the role's effectiveness but are not independently testable with n≈164.

Dimension 7: Implicit Promise — VALIDATED, RENAME FOR CLARITY

Cognitive role: Governs Stage 5 (Motivational Activation) by specifying the REWARD that the ventral striatum is anticipating. The specificity and vividness of the promise determine the strength of "wanting" (Berridge & Robinson, 2016).

Rename recommendation: "Promised Transformation" is more precise than "Implicit Promise" because it captures both the end-state and the implied journey. It also avoids the problematic word "implicit" — the promise is often quite explicit in DR ads.

Values should be organized by Maslow's hierarchy as these map to different neural circuits:

• Physical relief (pain reduction, symptom elimination) → Basic safety/pain circuits (PAG, insula)
• Functional restoration (mobility, energy, sleep) → Competence/autonomy circuits
• Aesthetic transformation (weight loss, appearance) → Social comparison circuits (mPFC, temporal-parietal junction)
• Independence/autonomy (not being a burden, freedom) → Self-determination circuits
• Identity restoration ("feeling like myself again") → Self-concept circuits (mPFC, posterior cingulate)

Recommendation: Retain with rename and richer value taxonomy.

Part 3: Identifying Missing Dimensions

The cognitive pipeline analysis reveals three dimensions missing from the preliminary framework that have substantial theoretical support.

MISSING Dimension A: Narrative Architecture (Temporal Structure)

Cognitive basis: The SEQUENCE in which information is presented matters independently of WHAT information is presented. This is supported by three bodies of research:

1. Primacy and recency effects (Murdock, 1962): The first and last items in a sequence are disproportionately encoded into memory. In a DR ad, what appears in the first 3 seconds and the last 5 seconds has outsized influence.

2. Narrative arc and neural coupling (Hasson et al., 2008; Zak, 2015): Stories that follow a tension-release arc (Freytag's pyramid) produce oxytocin release, which increases trust and prosocial behavior. Zak specifically found that narrative transportation in ads predicts donation behavior.

3. Information sequencing effects on persuasion (Hovland et al., 1957; Haugtvedt & Wegener, 1994): Whether the strongest argument comes first (primacy) or last (recency) interacts with involvement level. For low-involvement processing (social media scrolling), primacy effects dominate — the strongest hook must come first.

Why this dimension is missing and critical: Two ads can use identical hooks, mechanisms, proof types, and promises but arrange them in completely different sequences — and the sequence alone will produce different performance. A "problem → mechanism → proof → promise → CTA" sequence activates different processing dynamics than "promise → proof → mechanism → problem → CTA."

Proposed values:

• Problem-first (pain → explanation → solution → proof → CTA) — Classic DR structure. Activates loss aversion, then resolves.
• Discovery-first (revelation → mechanism → proof → application → CTA) — Curiosity-driven. Leads with prediction error.
• Testimonial-first (story → problem → solution → proof → CTA) — Narrative transportation first, then elaboration.
• Demonstration-first (visual result → explanation → mechanism → CTA) — Sensory evidence first, explanation after.
• Authority-first (expert/credential → problem → mechanism → solution → CTA) — Expertise heuristic established first.

MISSING Dimension B: Specificity/Concreteness Level

Cognitive basis: Paivio's (1986) dual-coding theory demonstrates that concrete, imageable information is encoded through both verbal and visual channels simultaneously, producing substantially stronger memory traces than abstract information. Concrete language activates sensory-motor cortex in addition to language areas (Binder et al., 2005), creating embodied representations.

This dimension captures a variable that crosscuts all other dimensions: the level of concrete detail vs. abstract generality. Consider two versions of the same promise:

• Abstract: "You'll feel better and have more energy"
• Concrete: "You'll wake up at 6am without an alarm and walk to the mailbox without stopping to rest"

Both reference the same promised transformation (functional restoration), but the concrete version activates richer mental simulation (episodic simulation in the hippocampus; Schacter et al., 2012), stronger emotional response, and more durable memory encoding.

Similarly, mechanisms can be abstract ("natural inflammation support") or concrete ("targets the TNF-alpha protein that triggers your morning joint stiffness"). Proof can be abstract ("clinically tested") or concrete ("127 women aged 50-70 in a Duke University double-blind trial").

Why this dimension is independent: Specificity/concreteness is not captured by any existing dimension. Problem, mechanism, proof, and promise each have their own content — but the LEVEL OF DETAIL at which that content is presented is an independent variable with its own cognitive mechanism.

Proposed coding: A 3-level ordinal scale: (1) Abstract/vague, (2) Moderate specificity, (3) Highly concrete/vivid. This should be coded as a global property of the creative, not per-dimension, to maintain parsimony with n≈164.

MISSING Dimension C: Visual-Verbal Modality Dominance

Cognitive basis: Dual-coding theory (Paivio, 1986) and multimedia learning theory (Mayer, 2009) demonstrate that information presented through both visual and verbal channels simultaneously creates stronger representations than either channel alone — but only when the channels are coherent rather than competing.

DR video ads vary dramatically in their channel strategy:

• Text-dominant (text overlays with background music, talking-head reading script) — Relies on verbal processing (left hemisphere, Broca's and Wernicke's areas)
• Visual-dominant (demonstrations, before/after images, B-roll with voiceover) — Relies on imagistic processing (ventral visual stream, bilateral engagement)
• Integrated (visual demonstrations synchronized with explanatory narration) — Maximum dual-coding benefit

Mayer's (2009) modality effect demonstrates that combining spoken narration with visual demonstrations produces superior learning (and by extension, persuasion) compared to text + visual or narration alone. However, Mayer's redundancy principle shows that adding on-screen text to narration + visual actually REDUCES processing efficacy due to split attention.

Why this matters for the 50+ demographic specifically: Age-related changes in cognitive processing (Salthouse, 2010) include reduced working memory capacity and slower processing speed. This means the 50+ audience is MORE sensitive to cognitive overload from competing channels. Visual-dominant or well-integrated ads should outperform text-heavy ads in this demographic — a testable hypothesis.

Proposed coding: 3-level categorical: Text-dominant, Visual-dominant, Integrated (narration + visual).

Evaluation of Additional Candidate Dimensions (EXCLUDED)

Several other dimensions were considered and excluded to maintain statistical viability:

Production quality — While potentially influential (heuristic credibility cue), it is (a) extremely difficult to code reliably by LLM from script + description, (b) confounded with budget/scaling status, and (c) has contradictory theoretical predictions (high production can signal "corporate" = distrust, while low production can signal "authentic" = trust, in the DR context).

Ad length — Important but is a continuous metadata variable, not a creative dimension. Should be included as a covariate in analysis, not as a coded dimension.

CTA specificity — Important (Stage 6) but is typically standardized within an offer/product, reducing variance in the dataset. Can be included as a binary control variable (specific vs. generic) rather than a full dimension.

Music/audio emotional tone — Theoretically important (music activates the limbic system independently of verbal content; Koelsch, 2014) but cannot be reliably coded from script + visual description by an LLM. This is a known limitation of the LLM-extraction methodology.

Part 4: The Recommended 10-Dimension Framework

Tier 1: Gateway Dimensions (Govern attention capture and relevance — Stages 1-3)

These dimensions determine WHETHER the viewer processes the ad at all. They operate in the first 1-5 seconds and function as sequential gates. If any gate fails, subsequent dimensions become irrelevant.

Tier 2: Elaboration Dimensions (Govern engagement depth and belief formation — Stage 4)

These dimensions determine HOW DEEPLY the viewer processes the ad and whether they form beliefs strong enough to motivate action. They operate in the 5-60 second window and interact multiplicatively.

Tier 3: Conversion Dimensions (Govern motivational intensity and action — Stages 5-6)

These dimensions determine WHETHER the viewer converts. They operate throughout the ad (building cumulatively) and reach peak influence at the CTA moment.

Part 5: Relative Importance Hierarchy

The following hierarchy is derived from the cognitive pipeline's sequential gating structure. A dimension that operates at an earlier gate has higher theoretical leverage because failure at that gate renders all subsequent dimensions irrelevant.

Theoretical Importance Ranking

Rank 1 — D3: Problem/Pain Referenced Confidence: HIGH

Rationale: This is the primary relevance gate. The problem referenced determines which population self-selects into the ad. A perfectly crafted ad with the wrong problem reference for the audience will have zero conversion. This dimension has the highest variance in impact because it functions as a binary filter: either the viewer recognizes their condition or they don't. This maps to the mPFC self-referencing circuit — the most powerful determinant of whether information proceeds to deeper processing.

Additionally, in the context of Meta's algorithmic targeting, the problem referenced is the primary signal through which the algorithm learns which users to serve the ad to. It therefore has a double effect: direct (cognitive relevance gate) and indirect (algorithmic audience optimization).

Rank 2 — D8: Dominant Emotional Structure Confidence: HIGH

Rationale: Emotion is not a single-stage variable — it modulates every processing stage simultaneously (Pessoa, 2008). The emotional structure determines: arousal level (which governs attention persistence), valence (which governs approach/avoidance motivation), and processing mode (fear narrows attention to focal cues; curiosity broadens exploratory processing; Fredrickson, 2001). In the health supplement context for older adults, the emotional structure is likely the strongest predictor of CPI because it directly determines the strength of the motivational signal at the CTA moment.

Rank 3 — D4: Causal Mechanism Confidence: HIGH

Rationale: The causal mechanism is the cognitive anchor of the entire narrative. It provides the "why" that makes the solution believable and the purchase rational. Without a coherent causal mechanism, the ad is reduced to assertion + emotion, which can generate clicks but not sustained purchase intent (high CTR, low conversion rate). The mechanism is the bridge between attention and action.

In the health supplement niche specifically, mechanism is particularly important because the audience (women 50+) has typically tried multiple solutions and failed. Their prior belief is "nothing works for me." The mechanism's role is to explain WHY previous approaches failed and why THIS approach is different — a causal reframing that bypasses accumulated skepticism.

Rank 4 — D2: Avatar Role Identity Confidence: MEDIUM-HIGH

Rationale: The avatar determines the processing route. A peer activates narrative transportation (reduced counter-arguing). An expert activates authority-based acceptance. A discoverer activates conspiratorial in-group identification. The avatar choice interacts strongly with emotional structure and proof type — an expert delivering a fear message with study citations is a completely different cognitive experience than a peer delivering a hope message with a personal testimonial.

Rank 5 — D5: Narrative Architecture Confidence: MEDIUM

Rationale: The sequence of information presentation determines the initial processing frame and the emotional arc. Problem-first structures prime loss-aversion, making subsequent solution framing more motivating. Discovery-first structures prime curiosity, sustaining attention but potentially delaying motivational activation. The architecture interacts with emotional structure to create the overall "shape" of the persuasive experience.

Rank 6 — D1: Hook Rhetorical Device Confidence: MEDIUM

Rationale: The hook's rhetorical form determines whether the 3-second attention window is leveraged effectively. However, its influence is largely confined to Stage 2 — once the viewer is past the first 3 seconds, the hook's rhetorical form is no longer active. Its theoretical importance is high for predicting 3-second view rates and hook rates but lower for predicting downstream conversion metrics like CPI.

Rank 7 — D6: Proof Type Confidence: MEDIUM

Rationale: Proof modulates belief strength, but its importance varies dramatically based on the viewer's elaboration level. Viewers in low elaboration (most social media contexts) rely on proof as a peripheral cue ("a doctor said it, must be true"). Viewers in high elaboration (seriously considering a purchase) evaluate proof quality. The proof type's importance therefore increases later in the funnel — it matters more for conversion rate than for CTR.

Rank 8 — D7: Specificity/Concreteness Level Confidence: MEDIUM-LOW

Rationale: Concreteness enhances encoding and persuasion across all stages but is a modulator rather than a primary driver. Its effect is most pronounced in the elaboration stage, where concrete details create vivid mental simulations. However, it is harder to isolate empirically because it is partially correlated with ad length and mechanism complexity.

Rank 9 — D9: Promised Transformation Confidence: MEDIUM-LOW

Rationale: Counter-intuitively, the specific promised transformation may matter less than its emotional framing. A promise of "pain relief" framed with high emotional intensity may outperform a promise of "complete independence" framed blandly. The promise provides the content; the emotional structure provides the motivational energy. Additionally, the promised transformation is partially determined by the product/offer itself, reducing the creative strategist's degree of freedom.

Rank 10 — D10: Visual-Verbal Modality Confidence: LOW-MEDIUM

Rationale: Important in theory (dual-coding, multimedia learning) but likely acts as a hygiene factor rather than a differentiator. Most DR ads in the dataset probably cluster in 2-3 modality patterns, reducing variance. Additionally, this dimension is harder to isolate from production decisions that may be correlated with other factors (budget, team, offer stage).

Part 6: Critical Interaction Effects

The 10 dimensions do not operate independently. The cognitive pipeline analysis predicts several theoretically important interaction effects that should be prioritized in analysis.

Interaction 1: Emotional Structure × Proof Type (HIGH PRIORITY)

Mechanism: The emotional state determines HOW proof is processed. Fear-primed viewers process authority/study citations more persuasively (because fear narrows attention to credibility cues; Eysenck, 1992). Curiosity-primed viewers process mechanism explanations more deeply (because curiosity enhances hippocampal-dependent learning; Gruber et al., 2014). Hope-primed viewers respond more to testimonials (because positive affect increases receptivity to narrative/social proof; Isen, 2008).

Testable hypothesis: "Authority citation as proof type will decrease CPI more when paired with fear/loss emotional structure than when paired with hope/gain emotional structure."

Interaction 2: Avatar Role × Emotional Structure (HIGH PRIORITY)

Mechanism: Avatar-emotion congruence determines the believability of the emotional appeal. A peer delivering fear-based content (personal fear about their condition) is congruent. An expert delivering fear-based content (clinical warning) is congruent. A peer delivering expert-level mechanism explanation is incongruent and triggers credibility skepticism.

Testable hypothesis: "Peer avatar with fear/loss emotional structure will outperform expert avatar with fear/loss emotional structure on CPI, because fear from a peer activates empathetic identification while fear from an expert can activate reactance."

Interaction 3: Causal Mechanism × Specificity Level (MEDIUM PRIORITY)

Mechanism: Abstract mechanisms ("natural inflammation support") create weak mental models. Concrete mechanisms ("the Boswellia extract binds to the 5-LOX enzyme that triggers your morning stiffness") create strong mental models. However, excessively concrete mechanisms can increase cognitive load beyond the processing capacity of the low-elaboration social media context. There is likely an inverted-U relationship between mechanism specificity and persuasion.

Testable hypothesis: "Moderate specificity in causal mechanism description will produce lower CPI than either abstract or highly technical mechanism descriptions."

Interaction 4: Narrative Architecture × Hook Rhetorical Device (MEDIUM PRIORITY)

Mechanism: The hook's rhetorical form sets an expectation for what follows. A question hook ("Did you know...?") creates an information-gap that is best resolved by a discovery-first architecture. A conditional hook ("If you're over 50 and...") sets a relevance frame that is best followed by a problem-first architecture. Mismatch between hook and architecture creates cognitive dissonance and viewer disengagement.

Testable hypothesis: "Question hooks paired with discovery-first architecture will produce lower CPI than question hooks paired with problem-first architecture."

Interaction 5: Problem Severity × Promised Transformation (MEDIUM PRIORITY)

Mechanism: Construal Level Theory (Trope & Liberman, 2010) predicts that psychologically "close" problems (severe, present, concrete) are processed at a low construal level and require low-construal solutions (specific, immediate relief). Psychologically "distant" problems (mild, future, abstract) allow high-construal solutions (general wellness, vitality). Matching the construal level of problem and promise enhances persuasive coherence.

Testable hypothesis: "Severe/specific problem framing paired with functional-restoration promise will outperform severe/specific problem framing paired with identity-restoration promise."

Part 7: Operational Specifications for LLM-Based Extraction

For each dimension, the following specifies coding rules optimized for LLM extraction with maximum inter-rater reliability.

D1: Hook Rhetorical Device

Definition: The syntactic/rhetorical structure of the FIRST complete utterance in the ad (first 1-2 sentences, or first 5 seconds if primarily visual).

Enumerated values:

1. Question — Opens with an interrogative sentence. E.g., "Have you ever wondered why...?"
2. Conditional/Qualifier — Opens with "If you..." or a qualifying statement that segments the audience. E.g., "If you're a woman over 50 dealing with..."
3. Declarative Claim — Opens with a factual or provocative assertion. E.g., "73% of doctors will never tell you this."
4. Imperative/Command — Opens with a direct instruction. E.g., "Stop taking this supplement immediately."
5. Narrative Opening — Opens with a personal or third-person story. E.g., "Last March, I could barely get out of bed..."

LLM coding instruction: "Read the first 1-2 sentences of the script. Classify ONLY the grammatical/rhetorical structure, ignoring the content. A sentence is a Question if and only if it ends with a question mark or uses interrogative syntax. A Conditional opens with 'if,' 'when,' or a qualifying clause. Default to Declarative Claim if ambiguous."

Failure modes: LLM may be confused by rhetorical questions that function as declarations ("Isn't it time you tried something different?" — this is a Question, not an Imperative). Solution: code by syntactic form, not pragmatic function.

D2: Avatar Role Identity

Definition: The primary social role that the on-screen speaker/character occupies in the ad. If multiple speakers appear, code the DOMINANT speaker (most screen time or most prominent in the narrative).

Enumerated values:

1. Peer/Fellow Sufferer — A person of similar demographic who has experienced the problem. Identified by personal testimony language: "I used to..." / "When my doctor told me..."
2. Discoverer/Insider — A person who has found hidden or suppressed information. Identified by revelation language: "I discovered..." / "They don't want you to know..."
3. Expert/Authority — A credentialed professional (doctor, researcher, pharmacist). Identified by credentials or clinical language.
4. Caregiver/Advocate — A person speaking on behalf of the sufferer (spouse, child, friend, health coach). Identified by third-person care language: "I watched my mother..." / "My patients were..."
5. Narrator/Faceless — No identified on-screen persona. Text-over-video, voiceover without persona attribution, or slideshow format.

LLM coding instruction: "Identify the primary speaker. Code based on their STATED or IMPLIED role in the ad, not their actual identity. If a doctor shares their personal joint pain story, code as Peer/Fellow Sufferer (they are performing the peer role, not the expert role). The role is defined by the function they serve in the narrative, not their credentials."

Failure modes: LLM may default to coding credentials (doctor = expert) rather than narrative function (doctor sharing personal story = peer). Solution: explicit instruction to code narrative role, not biographical identity.

D3: Problem/Pain Referenced

Definition: The specific health condition, symptom, or concern that the ad identifies as the viewer's problem.

Enumerated values (for health supplement niche):

1. Joint/mobility pain (knees, back, arthritis, stiffness)
2. Excess weight/belly fat (obesity, stubborn fat, metabolic)
3. Low energy/fatigue (tiredness, brain fog, sluggishness)
4. Digestive issues (bloating, gut health, irregularity)
5. Sleep problems (insomnia, restless sleep, daytime drowsiness)
6. Cognitive decline (memory, focus, mental sharpness)
7. Appearance/aging (skin, hair, wrinkles, looking old)
8. Blood sugar/metabolic (diabetes risk, A1C, glucose)
9. Heart/cardiovascular (blood pressure, cholesterol, circulation)
10. General wellness/vitality (nonspecific "feeling better")

Secondary code — Problem Intensity (ordinal):

• 1 — Mild/inconvenience: "Occasional discomfort," "a few extra pounds"
• 2 — Moderate/limiting: "Keeps me from enjoying activities," "I've tried everything"
• 3 — Severe/life-altering: "Crippling pain," "I was afraid I'd end up in a wheelchair"

LLM coding instruction: "Identify the PRIMARY health problem mentioned in the ad. If multiple problems are referenced, code the one that receives the most screen time or is positioned as the root cause. Code intensity based on the most extreme language used to describe the problem, not the average severity implied."

Failure modes: LLMs tend to code every mentioned condition. Solution: enforce single primary code with explicit "most prominent" instruction.

D4: Causal Mechanism

Definition: The explanatory model provided for WHY the problem exists and/or WHY the solution works. This is the "science" or "secret" of the ad.

Enumerated values:

1. Gut/microbiome (bacteria, leaky gut, gut-brain axis)
2. Inflammation/immune (chronic inflammation, cytokines, immune response)
3. Hormonal/endocrine (cortisol, thyroid, estrogen, testosterone)
4. Metabolic/cellular (mitochondria, metabolism, fat cells, insulin)
5. Nutritional deficiency (missing nutrient, depleted vitamin/mineral)
6. Toxic accumulation (toxins, heavy metals, environmental exposure)
7. Structural/mechanical (cartilage, collagen, joint fluid, nerve compression)
8. Genetic/epigenetic (gene expression, DNA, inherited tendency)
9. Proprietary/branded mechanism (named process unique to the product: "The 7-Second Ritual")
10. None/unspecified (no mechanism provided, pure assertion)

LLM coding instruction: "Identify the causal explanation provided for either (a) why the problem occurs or (b) why the solution works. Code the DOMINANT mechanism — the one around which the narrative is structured. If a branded mechanism is built on top of a scientific mechanism (e.g., 'The 7-Second Trick' that works by 'reducing cortisol'), code the underlying scientific category (Hormonal/endocrine). Only code as Proprietary/branded when no underlying mechanism is explained."

Failure modes: LLMs may conflate the mechanism with the product ingredient. Solution: explicit instruction to code the explanatory model, not the product feature.

D5: Narrative Architecture

Definition: The overall structural sequence of the ad's persuasive argument. Determined by which element comes FIRST in prominence.

Enumerated values:

1. Problem-first — Opens by establishing the viewer's pain/problem, then explains cause, then presents solution.
2. Discovery-first — Opens with a revelation, secret, or news, then connects to the viewer's problem, then presents solution.
3. Testimonial-first — Opens with a personal transformation story, then extracts the lesson/product, then generalizes.
4. Demonstration-first — Opens with a visual result or product demonstration, then explains how/why.
5. Authority-first — Opens by establishing expert credentials or scientific context, then presents findings.

LLM coding instruction: "Examine the first 25% of the script. Which persuasive element is established FIRST and MOST PROMINENTLY? The architecture is defined by the ad's opening strategy, not its overall content. An ad that mentions a problem in the first sentence but immediately pivots to a revelation should be coded as Discovery-first if the revelation dominates the opening."

Failure modes: Most DR ads mention problems early, so LLMs may default to Problem-first. Solution: instruction to identify the DOMINANT opening strategy, with examples of each architecture.

D6: Proof Type

Definition: The primary form of evidence presented to support the product's efficacy claim. Code the STRONGEST proof element, not the most frequent.

Enumerated values:

1. Personal testimonial — A named or unnamed individual shares their experience.
2. Authority/study citation — A doctor, researcher, or university study is referenced.
3. Visual demonstration — The product's effect is shown visually (demo, imagery, animation).
4. Before/after transformation — Explicit comparison of before and after states (visual or verbal).
5. Social proof/numerical — Numbers, statistics, or crowd behavior ("2 million women," "sold out 3 times").
6. Mechanism-as-proof — The explanation itself serves as the evidence ("Now that you understand the science, you can see why...").
7. None/pure assertion — Claims are made without supporting evidence.

LLM coding instruction: "Identify the single strongest piece of evidence used to support the product claim. 'Strongest' means the proof element that would be most convincing to a skeptical viewer. If a doctor gives a testimonial, code as Authority/study citation (the authority credential is the stronger proof element). If a regular person gives a testimonial with before/after photos, code as Before/after transformation."

Failure modes: Multiple proof types often co-occur. Solution: forced-choice for primary proof with option for secondary proof code.

D7: Specificity/Concreteness Level

Definition: The overall level of concrete, sensory-rich, specific detail used across the ad's claims, mechanisms, and promises. This is a GLOBAL property, not tied to any single element.

Ordinal scale:

1. Abstract/Vague — Generalized claims without specific details. "Supports healthy joints." "Helps with weight management." "Clinically tested."
2. Moderately Specific — Some concrete details but mixed with generalities. "Reduces morning stiffness in as little as 2 weeks." "Contains 12 research-backed ingredients."
3. Highly Concrete/Vivid — Rich specific details, numbers, sensory descriptions, named entities. "127 women in a 90-day Duke University trial saw a 63% reduction in CRP inflammation markers." "You'll pick up your grandchild without wincing."

LLM coding instruction: "Evaluate the overall level of specific, concrete detail across the ENTIRE script. Score based on the PROPORTION of claims that include specific numbers, named entities, sensory descriptions, or measurable outcomes. A script with one concrete statistic among mostly vague claims = Moderately Specific. A script where most claims are vivid and detailed = Highly Concrete."

Failure modes: LLMs may anchor on a single vivid detail and rate the whole ad as concrete. Solution: instruction to evaluate proportion, not peak concreteness.

D8: Dominant Emotional Structure

Definition: The primary emotional register that the ad sustains throughout its duration. Not the fleeting emotions triggered by individual moments, but the overarching emotional experience designed for the viewer.

Enumerated values:

1. Fear/Loss — Dominated by anxiety, urgency, threat of worsening condition, loss of function/independence. Core motivation: ESCAPE from negative state.
2. Hope/Gain — Dominated by optimism, possibility, positive transformation, future benefits. Core motivation: APPROACH toward positive state.
3. Curiosity/Revelation — Dominated by intrigue, surprise, hidden information, epistemic desire. Core motivation: RESOLVE information gap.
4. Indignation/Injustice — Dominated by anger at system, blame on external actors (pharma, doctors, industry), feeling of being deceived. Core motivation: REBEL against the source of injustice.

LLM coding instruction: "Read the complete script and identify which emotional response the ad is DESIGNED to sustain in the viewer. Code based on the ad's dominant emotional trajectory, not individual sentences. An ad that opens with fear but transitions to hope should be coded based on which emotion occupies more of the ad's duration and narrative energy. If truly 50/50, code the emotion present at the CTA moment, as this is the emotion intended to drive action."

Failure modes: LLMs tend to identify multiple emotions and struggle to select one. Solution: force a single primary code and provide the tiebreaker rule (emotion at CTA moment).

D9: Promised Transformation

Definition: The primary end-state that the viewer is led to believe the product will enable. Code what the viewer GAINS, not what the product does.

Enumerated values:

1. Physical relief — Freedom from pain, discomfort, or physical symptoms.
2. Functional restoration — Regained ability to perform activities (walking, sleeping, playing with grandchildren).
3. Aesthetic transformation — Visible change in appearance (weight loss, younger-looking skin, flatter belly).
4. Metabolic/health marker improvement — Improved measurable health outcomes (blood sugar, cholesterol, blood pressure).
5. Independence/autonomy — Reduced dependence on medications, doctors, or caregivers.
6. Identity/vitality restoration — "Feeling like yourself again," restored sense of self, regained confidence.

LLM coding instruction: "Identify the MOST PROMINENTLY promised outcome — the one that receives the most emotional emphasis or screen time. Distinguish between the product's claimed FUNCTION (what it does biologically) and the promised TRANSFORMATION (how the viewer's life changes). Code the transformation, not the function. 'Reduces inflammation' is a function. 'You'll dance at your daughter's wedding' is a transformation."

Failure modes: LLMs may code the product claim rather than the viewer-centric transformation. Solution: explicit instruction with examples distinguishing function from transformation.

D10: Visual-Verbal Modality Dominance

Definition: The primary information channel through which the ad's persuasive content is delivered.

Enumerated values:

1. Text-dominant — Primary persuasive content delivered through on-screen text overlays, often with background music or minimal voiceover. Reading is required to receive the message.
2. Verbal-dominant — Primary content delivered through spoken narration (talking head or voiceover). Visual elements are supplementary (lifestyle B-roll, product shots).
3. Visual-dominant — Primary content delivered through visual demonstrations, before/after imagery, animations, or visual storytelling. Audio provides context but visuals carry the argument.
4. Integrated — Deliberate synchronization of visual demonstrations with explanatory narration. Neither channel is interpretable alone; meaning emerges from their combination.

LLM coding instruction: "From the script and visual description, determine which channel carries the PRIMARY persuasive argument. If the ad would still make sense with the video removed (audio only), it is Verbal-dominant. If it would still make sense with audio removed (visual only), it is Visual-dominant or Text-dominant. If removing either channel destroys the argument, it is Integrated."

Failure modes: Most ads have both audio and video, making it tempting to always code Integrated. Solution: the "channel removal" test forces identification of the primary channel.

Part 8: Temporal Dynamics — How Dimension Importance Shifts Across the Video

The 10 dimensions do not maintain constant importance. Their influence follows a predictable pattern mapped to the cognitive processing pipeline:

First 3 Seconds (Pre-Attention → Attention Capture)

Dominant dimensions: D1 (Hook Rhetorical Device), D10 (Visual-Verbal Modality), D2 (Avatar — initial identification)

Cognitive state: Exogenous attention, pre-conscious feature detection, orienting response. The viewer has NOT yet made any decisions. The brain is answering: "Is this worth attending to?"

Analytical implication: These dimensions should correlate most strongly with HOOK RATE (3-second view rate) but weakly with CPI. High hook rate + poor CPI suggests strong D1/D10 but weak D3-D9.

3-10 Seconds (Attention → Relevance Gate)

Dominant dimensions: D3 (Problem/Pain), D2 (Avatar Role — full identification), D8 (Emotional Structure — initial valence established)

Cognitive state: Conscious evaluation, self-referencing, approach/avoidance appraisal. The viewer is answering: "Is this about me? Is this worth my time?"

Analytical implication: These dimensions should correlate with HOLD RATE (percentage of viewers who watch past 10 seconds). High hook rate but low hold rate suggests strong D1 but weak D3/D2 relevance matching.

10-30 Seconds (Engagement → Elaboration)

Dominant dimensions: D4 (Causal Mechanism), D5 (Narrative Architecture), D6 (Proof Type), D7 (Specificity Level)

Cognitive state: Narrative processing, causal reasoning, belief formation. The viewer is answering: "Does this make sense? Can I believe this?"

Analytical implication: These dimensions should correlate with AVERAGE VIEW DURATION and VIDEO COMPLETION RATE. They predict the depth of engagement and the strength of belief formation that precedes conversion.

Final 10 Seconds / CTA Moment (Motivation → Action)

Dominant dimensions: D8 (Emotional Structure — peak intensity), D9 (Promised Transformation — crystallization of desire), D6 (Proof Type — final credibility boost)

Cognitive state: Motivational activation, cost-benefit computation, action threshold. The viewer is answering: "Should I do this? Is it worth the risk?"

Analytical implication: These dimensions should correlate most strongly with CTR and CPI — the metrics that capture actual conversion behavior.

Part 9: Statistical Design Considerations

The Granularity-Power Tradeoff

With n≈164 creatives and 10 dimensions (each with 3-7 values), the parameter space is enormous relative to the sample size. Naive multivariate regression will be grossly underpowered and unreliable.

Recommended analytical approach:

Phase 1: Univariate screening (all 10 dimensions) For each dimension independently, compute: mean CPI by value, win rate by value, and hit rate by value. Use Kruskal-Wallis nonparametric tests (appropriate for small, non-normal groups) to identify dimensions with statistically suggestive differences (p < 0.20 threshold for screening, not p < 0.05). This requires NO multiple-regression and works with n≈164.

Phase 2: Bivariate interaction analysis (top 4-5 dimensions from Phase 1) For the dimensions that show univariate signal, test the 5 predicted interaction effects using stratified analysis (e.g., "What is the mean CPI for Fear + Testimonial vs. Fear + Authority vs. Hope + Testimonial vs. Hope + Authority?"). Visualize with heatmaps. With 164 creatives and 2 dimensions of 4 values each, you get approximately 10 creatives per cell — sufficient for directional signal, not for statistical significance.

Phase 3: Hypothesis generation, not hypothesis testing With n≈164, the goal is to GENERATE specific testable hypotheses ("Fear + Authority + Problem-first architecture produces the lowest CPI"), not to prove them statistically. The hypotheses are then tested prospectively in the next creative cycle. This is the hypothesis-driven entrepreneurship framework (Camuffo et al., 2020) applied to creative strategy.

Dimensionality reduction strategy: If the dataset proves too sparse for 10 dimensions, the priority reduction order is:

1. Drop D10 (Visual-Verbal Modality) — hardest to code reliably from scripts
2. Drop D7 (Specificity Level) — most subjective ordinal rating
3. Merge D1 (Hook Device) into D5 (Narrative Architecture) — sacrifice hook granularity for architecture
4. Core irreducible set: D2, D3, D4, D5/D1, D6, D8, D9 — 7 dimensions

Sample Size Requirements by Analysis Type

Recommended: The "Creative DNA + Outcome" Matrix

Rather than regression, create a visual matrix where each creative is a row, each dimension is a column, and the final columns are performance metrics. This enables PATTERN RECOGNITION — identifying clusters of dimensional configurations that co-occur with high or low performance. This is the content analysis methodology (Krippendorff, 1980) applied to creative performance analysis.

Sort the matrix by CPI (best to worst). Visually scan for patterns: "The top 20 creatives by CPI are disproportionately Fear/Loss + Peer + Problem-first + Authority proof." This pattern becomes a hypothesis for the next cycle.

Part 10: Implementation Roadmap

Step 1: Codebook Development (Week 1)

Create a formal codebook document with the 10 dimensions, each containing: definition, enumerated values with examples from actual ads in the dataset, decision rules for ambiguous cases, and the specific LLM prompt for extraction.

Step 2: LLM Extraction Pipeline (Week 1-2)

For each of the 164 creatives, prepare: (a) the full script text, (b) a structured visual description (if available), (c) metadata (offer, cycle, format). Run through an LLM with the codebook as system prompt, extracting all 10 dimensions. Use forced-choice output format (JSON with enumerated values only) to prevent free-text drift.

Step 3: Inter-Rater Reliability Check (Week 2)

Randomly select 20 creatives (~12% of dataset). Code them independently using two different LLM configurations (e.g., different prompt formulations, different temperature settings) or one LLM + one human coder. Compute Cohen's Kappa for each dimension. Target: κ > 0.70 for all dimensions. Any dimension with κ < 0.60 needs codebook revision.

Step 4: Matrix Assembly & Pattern Analysis (Week 2-3)

Assemble the Creative DNA matrix. Merge with performance data (CPI, CTR, Win Rate, Hit Rate by cycle). Conduct Phase 1 univariate screening. Visualize top hypotheses with heatmaps.

Step 5: Hypothesis Formulation (Week 3)

From the pattern analysis, formulate 3-5 specific, falsifiable hypotheses for the next creative cycle. Each hypothesis should take the form: "Creative with [Dimension X = Value A] + [Dimension Y = Value B] will achieve CPI below [threshold], controlling for offer."

Step 6: Prospective Testing (Next Cycle)

Design the next cycle's creative briefs to explicitly test the formulated hypotheses. This means creating matched pairs that differ on only the hypothesized dimension while holding others constant. This is where the framework generates its highest ROI — systematic learning rather than random variation.

Appendix A: Complete Dimension Reference Card

Appendix B: Key References

• Bandura, A. (1977). Social Learning Theory. Prentice Hall.
• Berridge, K.C. & Robinson, T.E. (2016). Liking, wanting, and the incentive-sensitization theory of addiction. American Psychologist, 71(8), 670-679.
• Binder, J.R. et al. (2005). Distinct brain systems for processing concrete and abstract concepts. Journal of Cognitive Neuroscience, 17(6), 905-917.
• Bradley, M.M. (2009). Natural selective attention: Orienting and emotion. Psychophysiology, 46(1), 1-11.
• Camuffo, A. et al. (2020). A scientific approach to entrepreneurial decision making. Management Science, 66(2), 564-586.
• Cialdini, R.B. (2006). Influence: The Psychology of Persuasion. Harper Business.
• Cohen, J. (2001). Defining identification: A theoretical look at the identification of audiences with media characters. Mass Communication & Society, 4(3), 245-264.
• Corbetta, M. & Shulman, G.L. (2002). Control of goal-directed and stimulus-driven attention in the brain. Nature Reviews Neuroscience, 3, 201-215.
• Eysenck, M.W. (1992). Anxiety: The Cognitive Perspective. Psychology Press.
• Fecteau, J.H. & Munoz, D.P. (2006). Salience, relevance, and firing: a priority map for target selection. Trends in Cognitive Sciences, 10(8), 382-390.
• Fernbach, P.M. et al. (2013). The illusion of explanatory depth. Psychological Science, 24(4), 510-522.
• Fredrickson, B.L. (2001). The role of positive emotions in positive psychology. American Psychologist, 56(3), 218-226.
• Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11, 127-138.
• Gollwitzer, P.M. (1999). Implementation intentions: Strong effects of simple plans. American Psychologist, 54(7), 493-503.
• Green, M.C. & Brock, T.C. (2000). The role of transportation in the persuasiveness of public narratives. Journal of Personality and Social Psychology, 79(5), 701-721.
• Gruber, M.J. et al. (2014). States of curiosity modulate hippocampus-dependent learning via the dopaminergic circuit. Neuron, 84(2), 486-496.
• Hasson, U. et al. (2008). Neurocinematics: The neuroscience of film. Projections, 2(1), 1-26.
• Horton, D. & Wohl, R. (1956). Mass communication and para-social interaction. Psychiatry, 19(3), 215-229.
• Isen, A.M. (2008). Some ways in which positive affect influences decision making and problem solving. In M. Lewis et al. (Eds.), Handbook of Emotions.
• Itti, L. & Koch, C. (2001). Computational modelling of visual attention. Nature Reviews Neuroscience, 2, 194-203.
• Kahneman, D. & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263-291.
• Kanwisher, N. et al. (1997). The fusiform face area: A module in human extrastriate cortex. Journal of Neuroscience, 17(11), 4302-4311.
• Kelley, W.M. et al. (2002). Finding the self? An event-related fMRI study. Journal of Cognitive Neuroscience, 14(5), 785-794.
• Klucharev, V. et al. (2008). Reinforcement learning signal predicts social conformity. Neuron, 61(1), 140-151.
• Koelsch, S. (2014). Brain correlates of music-evoked emotions. Nature Reviews Neuroscience, 15, 170-180.
• Krippendorff, K. (1980). Content Analysis: An Introduction to Its Methodology. Sage.
• LeDoux, J.E. (1996). The Emotional Brain. Simon & Schuster.
• Loewenstein, G. (1994). The psychology of curiosity: A review and reinterpretation. Psychological Bulletin, 116(1), 75-98.
• Mayer, R.E. (2009). Multimedia Learning (2nd ed.). Cambridge University Press.
• McClure, S.M. et al. (2004). Separate neural systems value immediate and delayed monetary rewards. Science, 306(5695), 503-507.
• Murdock, B.B. (1962). The serial position effect of free recall. Journal of Experimental Psychology, 64(5), 482-488.
• Northoff, G. et al. (2006). Self-referential processing in our brain — A meta-analysis. NeuroImage, 31(1), 440-457.
• Paivio, A. (1986). Mental Representations: A Dual Coding Approach. Oxford University Press.
• Pessoa, L. (2008). On the relationship between emotion and cognition. Nature Reviews Neuroscience, 9, 148-158.
• Petty, R.E. & Cacioppo, J.T. (1986). The Elaboration Likelihood Model of Persuasion. Advances in Experimental Social Psychology, 19, 123-205.
• Rogers, T.B. et al. (1977). Self-reference and the encoding of personal information. Journal of Personality and Social Psychology, 35(9), 677-688.
• Schacter, D.L. et al. (2012). The future of memory: Remembering, imagining, and the brain. Neuron, 76(4), 677-694.
• Scherer, K.R. (2001). Appraisal considered as a process of multilevel sequential checking. In K.R. Scherer et al. (Eds.), Appraisal Processes in Emotion.
• Shiv, B. & Fedorikhin, A. (1999). Heart and mind in conflict: The interplay of affect and cognition. Journal of Consumer Research, 26(3), 278-292.
• Simon, H.A. (1955). A behavioral model of rational choice. Quarterly Journal of Economics, 69(1), 99-118.
• Sloman, S.A. (2005). Causal Models: How People Think About the World and Its Alternatives. Oxford University Press.
• Sokolov, E.N. (1963). Perception and the Conditioned Reflex. Pergamon Press.
• Symons, C.S. & Johnson, B.T. (1997). The self-reference effect in memory: A meta-analysis. Psychological Bulletin, 121(3), 371-394.
• Tom, S.M. et al. (2007). The neural basis of loss aversion in decision-making under risk. Science, 315(5811), 515-518.
• Trope, Y. & Liberman, N. (2010). Construal-level theory of psychological distance. Psychological Review, 117(2), 440-463.
• Vuilleumier, P. et al. (2001). Effects of attention and emotion on face processing in the human brain. Neuron, 30(3), 829-841.
• Yantis, S. & Jonides, J. (1984). Abrupt visual onsets and selective attention. Journal of Experimental Psychology: Human Perception and Performance, 10(5), 601-621.
• Zak, P.J. (2015). Why inspiring stories make us react: The neuroscience of narrative. Cerebrum, 2015, 2.

This document is a working thesis intended to guide empirical validation. All theoretical predictions are falsifiable hypotheses, not established facts. The framework should be iteratively refined based on empirical results from creative performance data.