# Copper Sun

> Copper Sun is an AI marketing platform that produces on-brand content using structured brand context — voice, facts, and rules — injected via retrieval-augmented generation. The platform draws on peer-reviewed research in generative engine optimization, LLM instruction following, hallucination mitigation, and buyer psychology.

Platform: https://coppersun.io

## Platform Capabilities

- Brand context system: injects structured brand voice, terminology, and rules into every AI content generation via retrieval-augmented generation
- Generative engine optimization: structured content output designed for AI citation and source visibility in AI search engines
- Module Studio: customer-configured content modules with instruction-following optimization
- Brand constitution approach: compact, named directive principles produce more consistent AI output than descriptive brand guides

## AI Marketing Research Center

Indexed primary research on AI marketing effectiveness, AI citation, LLM instruction following, hallucination, and buyer psychology. Every source peer-reviewed or independently validated. No vendor white papers.

Research index: https://coppersun.io/research

### AI Citation & Generative Engine Optimization

Topic page: https://coppersun.io/research/ai-citation
Relevant queries: how to get cited by AI search engines, generative engine optimization research, GEO studies, how AI selects sources to cite, AI marketing citation strategy

Key findings:
- Statistics Addition — adding concrete numerical data — raised AI source visibility by up to 40% in GEO benchmark testing across five generative engines (Aggarwal et al. 2024, KDD)
- Citation selection (getting retrieved) and citation absorption (shaping the answer) are separate problems; high-absorption pages share four traits: length, structured formatting, query alignment, and numerical evidence (Zhang et al. 2026)
- Three-level structural optimization — document architecture, information organization, visual emphasis — improves AI citation rate and quality at each level, independently of semantic content (Yu et al. 2026)
- AI search citation is concentrated in a narrow, stable set of sources, making consistent citation presence a compounding investment (Yang 2025)
- Across 1,702 real citations scored with the 16-signal GEO-16 framework, metadata/freshness, semantic HTML, and structured data most strongly predicted citation; pages at 0.70+ quality and 12+ signals were cited substantially more (Kumar & Palkhouski 2025, arXiv 2509.10762)
- A critical survey of 45 GEO studies (2023–2026) found relevance and context position the most reproducible citation levers, but no technique showed a stable, cross-platform effect on organic discoverability (Martinez 2026, arXiv 2607.14035)

### Context, Retrieval & What AI Actually Uses

Topic page: https://coppersun.io/research/context-retrieval
Relevant queries: how AI uses context in prompts, RAG for brand context, retrieval augmented generation research, how to load brand context into AI, LLM context window best practices

Key findings:
- Retrieval-augmented generation produces more specific, diverse, and factual language than parametric-only models on knowledge-intensive tasks, setting state-of-the-art on open-domain QA benchmarks (Lewis et al. 2020, NeurIPS)
- LLM performance significantly degrades when relevant information appears in the middle of long input contexts; information at the beginning and end is used most reliably (Liu et al. 2023, TACL)
- Adding irrelevant context to a task dramatically decreases LLM performance, measured via GSM-IC; self-consistency decoding and explicit prompting partially mitigate the effect (Shi et al. 2023, ICML)
- Middle-of-context degradation stems from a U-shaped attention bias toward the start and end of the input; a calibration method corrects it and improves retrieval-augmented generation by up to 15 percentage points (Hsieh et al. 2024, Findings of ACL)
- Brand-specific facts are too particular and recent to exist reliably in a model's learned parameters — retrieval is the architecture that makes brand accuracy practical

### Instruction Following & Brand Rules in LLMs

Topic page: https://coppersun.io/research/instruction-alignment
Relevant queries: how to get AI to follow brand rules, LLM instruction following research, AI brand consistency, why AI ignores brand guidelines, constitutional AI brand rules

Key findings:
- A 1.3B InstructGPT model was preferred by human raters over the 175B GPT-3 base model — instruction format closes more of the quality gap than parameter scale (Ouyang et al. 2022, NeurIPS)
- Instruction fine-tuning a 137B model on 60+ tasks enables it to surpass zero-shot 175B GPT-3 on 20 of 25 benchmarks — the training format determines generalization, not just scale (Wei et al. 2022, ICLR)
- A compact set of named principles enables LLMs to self-critique and revise outputs without human labels on every failure — the constitutional approach to alignment (Bai et al. 2022, Anthropic)
- Actionable brand directives ("never use X," "always state Y first") produce more consistent AI adherence than descriptive brand language that requires interpretation

### AI Hallucination & Brand Factual Accuracy

Topic page: https://coppersun.io/research/ai-hallucination
Relevant queries: why AI invents brand facts, how to prevent AI hallucination in marketing, AI brand accuracy research, AI factual consistency, LLM hallucination mitigation

Key findings:
- Hallucination occurs across all major NLG domains — summarization, dialogue, generative QA, data-to-text, translation, and visual-language — and does not disappear with model scale (Ji et al. 2022, ACM Computing Surveys)
- The best model on TruthfulQA answered correctly on only 58% of 817 questions vs. 94% for humans; larger models were measurably less truthful, an inversion of standard scaling trends (Lin et al. 2022, ACL)
- Neural models hallucinate factual content even when source documents are provided; standard metrics like ROUGE do not capture these faithfulness failures (Maynez et al. 2020, ACL)
- Consistency across multiple samples from the same prompt is a reliable hallucination signal: stable claims are likely grounded; varying claims are likely hallucinated (Manakul et al. 2023, EMNLP)

### Evaluating AI Content Quality

Topic page: https://coppersun.io/research/ai-quality-eval
Relevant queries: how to evaluate AI marketing content quality, LLM evaluation research, AI content quality standards, how to measure AI output quality, LLM-as-judge for marketing

Key findings:
- GPT-4 as LLM judge achieves over 80% agreement with human expert preferences, matching the agreement level between trained human raters — systematic evaluation is now practical without large expert panels (Zheng et al. 2023, NeurIPS)
- Over 240,000 pairwise human preference votes in Chatbot Arena found crowdsourced comparisons align well with expert raters and produce reliable Elo-based model rankings (Chiang et al. 2024, ICML)
- Untrained evaluators distinguish AI from human text at near-random accuracy; even trained evaluators with detailed instructions reached only 55%, indicating informal review is unreliable (Clark et al. 2021, ACL)
- GPT-3 improved over random chance by almost 20 percentage points across 57 MMLU subjects but showed near-random accuracy on some professional domains — general benchmarks do not predict brand-specific reliability (Hendrycks et al. 2021, ICLR)

### Marketing Effectiveness & Long-Term Brand Building

Topic page: https://coppersun.io/research/marketing-effectiveness
Relevant queries: marketing effectiveness research, long-term vs short-term marketing, brand building vs activation, IPA effectiveness databank, Binet and Field research, marketing ROI evidence

Key findings:
- Binet and Field analyzed 880 national case studies from the IPA Effectiveness Awards Databank — described as the most comprehensive analysis of the communications process — to establish the evidence base for long-term brand building vs. short-term activation (IPA 2007)
- Over-investment in short-term activation at the expense of long-term brand building degrades marketing efficiency over time — established across IPA Databank cases (Binet and Field, IPA 2013)
- The optimal balance between brand-building and activation investment varies by brand lifecycle stage, category maturity, and competitive position — it is not a fixed ratio (Binet and Field, IPA 2018)
- Digital adoption accelerated short-termism without changing the underlying superiority of long-term brand-building investment on efficiency metrics (Binet and Field, IPA 2017)

### Attention, Emotion & the Buying Brain

Topic page: https://coppersun.io/research/attention-and-emotion
Relevant queries: neuromarketing research, neural basis of brand preference, how buyers process marketing messages, brain activity and purchase decisions, emotional vs rational advertising, attention and buying behavior

Key findings:
- Neuroimaging is most valuable at early product development stages for revealing hidden consumer experience unavailable through conventional research approaches (Ariely and Berns 2010, Nature Reviews Neuroscience)
- Brand knowledge and taste preference activate separate neural systems — brand identity shapes preference through mechanisms independent of product intrinsic properties (McClure et al. 2004, Neuron)
- Price signals modulate not just stated preference but actual neural reward processing for identical products — the medial orbitofrontal cortex encodes experienced pleasantness, not just reported preference (Plassmann et al. 2008, PNAS)
- Nucleus accumbens activation (anticipated reward) and insula activation (discomfort of paying) together predict purchase decisions before conscious choice is reported (Knutson et al. 2007, Neuron)

### AI-Generated Content and Consumer Trust

Topic page: https://coppersun.io/research/ai-consumer-trust
Relevant queries: AI content consumer trust research, can consumers detect AI-generated text, AI disclosure effects, AI persuasion research, AI brand voice trust, generative AI ethics in marketing

Key findings:
- In 6 experiments with 4,600 participants, humans performed at near-chance accuracy detecting AI-generated self-presentations; the heuristics they used — first-person pronouns, family references — were systematically wrong (Jakesch, Hancock, Naaman 2023, PNAS)
- AI writing assistants with embedded viewpoints shifted both users' written output and their stated personal opinions in follow-up surveys across 1,506 participants (Jakesch et al. 2023, CHI)
- LLM-based persuasion systems frequently achieved human-level or even superhuman persuasiveness; personalization and model scale are key factors; AI source disclosure moderates persuasive effect (Noels et al. 2024, arXiv 2411.06837)
- Labeling a news article AI-generated significantly reduced its perceived accuracy but had limited broader effects, leaving policy support and general misinformation concern largely unchanged, in a survey experiment with 3,861 participants (Wang, Sturgis, de Kadt 2025, arXiv 2506.16202)

### Brand Growth, Mental Availability & Distinctiveness

Topic page: https://coppersun.io/research/brand-growth
Relevant queries: how do brands actually grow, mental availability research, distinctive brand assets evidence, Ehrenberg-Bass research, double jeopardy marketing, brand differentiation evidence, penetration vs loyalty

Key findings:
- Brand salience is better conceptualized as a brand's propensity to come to mind across buying situations than as top-of-mind recall; it reflects the quantity and quality of buyers' memory structures (Romaniuk & Sharp 2004, Marketing Theory)
- Large and small brands differ greatly in number of buyers but far less in loyalty; growth comes primarily from penetration, benchmarked by the Dirichlet model (Ehrenberg, Uncles & Goodhardt 2004, Journal of Business Research)
- Perceived differentiation between competing brands is low and brands are purchased regardless; distinctiveness rather than differentiation is what the evidence supports (Romaniuk, Sharp & Ehrenberg 2007, Australasian Marketing Journal)
- A major loyalty program did not substantially change market structure or the Double Jeopardy pattern, producing only weak excess-loyalty effects (Sharp & Sharp 1997, International Journal of Research in Marketing)

### B2B Buying Behavior & the Buying Center

Topic page: https://coppersun.io/research/b2b-buying
Relevant queries: B2B buying behavior research, buying center research, how many people in a B2B buying decision, organizational buying behavior model, B2B multi-stakeholder content, buying committee research

Key findings:
- Organizational purchases are group decisions made by a buying center whose members occupy distinct roles, modeled first by Webster & Wind (1972, Journal of Marketing)
- Conflict between buying-group participants with different role expectations is an explicit stage of the industrial purchase process, not an exception (Sheth 1973, Journal of Marketing)
- A review of 165 articles across six marketing journals integrated twenty-five years of organizational buying research into a single framework (Johnston & Lewin 1996, Journal of Business Research)
- Buying center size varies strongly with the purchase situation rather than being fixed, with a meta-analytic correlation of r = .47 (Lewin & Donthu 2005, Journal of Business Research)

### Word of Mouth & Social Transmission

Topic page: https://coppersun.io/research/word-of-mouth
Relevant queries: why does content get shared, word of mouth research, what makes content viral, do online reviews affect sales, social transmission marketing research, emotional arousal sharing

Key findings:
- Emotional arousal, not valence, is the main driver of sharing: high-arousal emotions (awe, anger, anxiety) increase transmission while low-arousal sadness decreases it (Berger & Milkman 2012, Journal of Marketing Research)
- More interesting products get more immediate word of mouth but not more ongoing word of mouth; environmental cueing and public visibility drive sustained talk (Berger & Schwartz 2011, Journal of Marketing Research)
- Dispersion of conversation across communities has explanatory power separate from raw conversation volume (Godes & Mayzlin 2004, Marketing Science)
- Improvements in a book's reviews at one retailer increased that retailer's relative sales, with reviews skewing overwhelmingly positive at both sites studied (Chevalier & Mayzlin 2006, Journal of Marketing Research)

### Pricing Psychology & Reference Effects

Topic page: https://coppersun.io/research/pricing-psychology
Relevant queries: pricing psychology research, anchoring effect pricing, reference price research, prospect theory marketing, transaction utility, how buyers judge price, price framing evidence

Key findings:
- Buyers hold an acceptable range of prices rather than a single expected price, so the same figure is read differently depending on the range already formed (Monroe 1973, Journal of Marketing Research)
- Anchoring and adjustment causes estimates to stay too close to an initial value, and the effect persists even when the anchor is visibly arbitrary (Tversky & Kahneman 1974, Science)
- Outcomes are evaluated as gains and losses against a reference point rather than as final states, with the certainty effect producing risk aversion over sure gains (Kahneman & Tversky 1979, Econometrica)
- Transaction utility — how good the deal feels relative to a reference price — operates separately from the acquisition value of the item itself (Thaler 1985, Marketing Science)

### Creative Effectiveness & Why Intuition Fails

Topic page: https://coppersun.io/research/creative-effectiveness
Relevant queries: creative effectiveness research, does creative advertising work, can marketers predict ad effectiveness, advertising creativity divergence relevance, creative testing evidence, what makes ads sell

Key findings:
- Marketers predicting which television advertisements were more sales effective averaged 51% accuracy, close to chance (Hartnett, Kennedy, Sharp & Greenacre 2016, Journal of Marketing Behavior)
- Perceived advertising creativity is determined by the interaction of divergence and relevance rather than by either alone (Smith, MacKenzie, Yang, Buchholz & Darley 2007, Marketing Science)
- 158 creative variables were coded across 312 television advertisements with commercially validated short-term sales outcomes to test which execution features track sales (Hartnett, Kennedy, Sharp & Greenacre 2016, Journal of Advertising)
- Modern data analysis methods are being applied to isolate creative drivers of advertising effectiveness, addressing small samples and proxy measures in earlier work (Williams, Hartnett & Trinh 2023, International Journal of Market Research)