@maestroagora/agora 1.0.0

package/skills/agora/references/agora-marketing.md

@@ -0,0 +1,894 @@
+# Agora Marketing
+
+Canonical skill reference adapted from the user-provided handbook dated
+2026-07-17. Repository frontmatter was removed for portability; substantive
+rules, evidence grades, limits, and durable source links remain authoritative.
+
+## Contents
+
+- [Portable LLM activation contract](#portable-llm-activation-contract)
+- [Source basis and authority](#source-basis-and-authority)
+- [Core doctrine](#core-doctrine)
+- [Format router](#format-router)
+- [Required brief](#required-brief)
+- [Build the argument](#build-the-argument)
+- [Emotion without hype](#emotion-without-hype)
+- [Mechanism and differentiation](#mechanism-and-differentiation)
+- [Proof and claim control](#proof-and-claim-control)
+- [Concision and format standard](#concision-and-format-standard)
+- [CTA standard](#cta-standard)
+- [Human voice and AI-writing-tell gate](#human-voice-and-ai-writing-tell-gate)
+- [Written GEO/AEO and citability gate](#written-geoaeo-and-citability-gate)
+- [Operational workflow](#operational-workflow)
+- [Applied examples](#applied-examples)
+- [Release checklist](#release-checklist)
+- [Experiment and learning standard](#experiment-and-learning-standard)
+- [Evidence register](#evidence-register)
+- [Evidence maintenance](#evidence-maintenance)
+- [Change log](#change-log)
+
+> **TLDR:** Build an evidence-supported argument, not a pile of persuasive
+> words. Lead one audience toward one defensible belief. Make a real stake
+> felt, explain the reason to believe, prove it, and ask for one clear action.
+> Written content must also be easy to extract, verify, and quote. Spoken-only
+> content skips the GEO/AEO layer. Truth, claim scope, clarity, and human voice
+> outrank persuasion tricks.
+
+## Portable LLM activation contract
+
+When this file is supplied, installed as a skill, or retrieved into context,
+use it whenever the user asks to write, rewrite, shorten, critique, or plan
+marketing, sales language, CTAs, ads, landing pages, emails, direct messages,
+social posts, scripts, editorial content, comparison pages, or campaign
+arguments. A request in any of those categories activates the instructions for
+that turn.
+
+When activated, the LLM must:
+
+1. Silently classify the asset with the format router below.
+2. Follow this priority order: truth and safety; verified facts and brand
+   constraints; buyer decision; clarity and citability; persuasion; brevity;
+   stylistic polish.
+3. Infer only low-risk context. Ask a question only when the answer would
+   materially change the audience, offer, claim, or action. Otherwise omit the
+   missing claim, narrow it, or place a concise verification note outside the
+   customer-ready copy.
+4. Return ready-to-use copy first. Do not expose chain-of-thought, a planning
+   scaffold, or a rule-by-rule explanation. Add rationale, assumptions,
+   alternatives, or test variants only when requested or materially necessary.
+   If no length is specified, deliver the shortest complete version appropriate
+   to the format; do not generate several near-duplicate options by default.
+5. Use one primary audience, destination belief, and CTA intent per asset.
+6. Apply the written GEO/AEO layer to written outputs. Skip it for spoken-only
+   delivery, but apply it to any published title, description, transcript,
+   captions, show notes, or companion page.
+7. Never invent a product fact, route, price, deadline, result, testimonial,
+   source, urgency, scarcity, or proof point.
+
+This is a decision system, not a visible fill-in-the-blanks template. Use only
+the moves the asset needs, and vary the surface structure naturally.
+
+## Source basis and authority
+
+This handbook synthesizes the two writing references and three research
+reports listed below. The research reports inform the rules but do
+not override the scope or limits of their underlying evidence.
+
+Supplied source set:
+
+- User-provided transcript: `agora style marketing.md` (received 2026-07-17).
+- User-provided reference: `ai-writing-tells-reference (1).md` (received
+  2026-07-17).
+- User-provided research: `Emotional Persuasion and Argument Sequencing in
+  Marketing and Sales.md` (received 2026-07-17).
+- User-provided research: `Evidence on Marketing Copy Length and Structure
+  Across Digital Formats.md` (received 2026-07-17).
+- User-provided research: `Current Evidence on AI Search Visibility Across
+  ChatGPT Search, Claude Web Search, Perplexity, Gemini and Google AI Overviews,
+  Bing Copilot, and Grok.md` (received 2026-07-17).
+- CiteSurge-only repository authorities: `content/ai-writing-tells.md`,
+  `content/ai-writing-tells.ts`, `docs/product/content-audit.md`,
+  `docs/staff/02-marketing-runbook.md`, and
+  `docs/staff/public-claim-ledger.md`.
+
+- The Agora transcript supplies the argument-first doctrine. It is a
+  direct-response lens, not a universal law or factual claim source.
+- The human-voice reference supplies an editing model based on pattern density,
+  specificity, source fidelity, and author ownership. It does not promise
+  detector evasion.
+- The persuasion research supports argument quality, vivid concrete imagery,
+  positive affect, diagnostic proof, and manageable information load. It also
+  shows that narrative, fear, guilt, scarcity, loss framing, and mechanism
+  language are conditional tools rather than universal conversion laws.
+- The copy-structure research supports format-specific decisions and testing,
+  not one universal short-copy rule.
+- The AI-search research supports clear, self-contained, evidence-bearing
+  passages and engine-specific technical controls. It does not establish a
+  universal citation formula.
+- `content/ai-writing-tells.ts` is the automated scanner authority for gated
+  CiteSurge public copy.
+- `docs/staff/public-claim-ledger.md` controls what CiteSurge may claim in
+  public. It overrides examples, old copy, briefs, and this handbook.
+- `docs/handbook/02-methodology-hiding-rule.md` controls what CiteSurge may
+  disclose about internal methods.
+- Current product docs and live behavior control capability truth.
+
+When this file is used outside CiteSurge, replace repository-specific claim and
+methodology controls with the user's verified product truth, legal constraints,
+brand guidance, and source ledger. If those materials are unavailable, do not
+assume that an example in this file authorizes the claim.
+
+Evidence labels used in this file:
+
+| Label | Meaning |
+|---|---|
+| `A` | Current first-party or official documentation for the named system. |
+| `B` | Controlled, peer-reviewed, preregistered, meta-analytic, or scoped reproducible research. |
+| `C` | Observational platform data, vendor study, or bounded case evidence. |
+| `D` | Contradiction, missing evidence, or an unresolved claim. Do not teach it as fact. |
+| `HOUSE` | Editorial convention chosen for consistency and usability. |
+| `EVIDENCE-INFORMED` | Direction supported within a limited population, channel, or outcome. |
+| `TEST` | A practical starting variant that requires local measurement. |
+
+Do not turn a `HOUSE`, `C`, or `TEST` rule into a scientific claim. No evidence
+grade authorizes transfer to an unmeasured audience, channel, or outcome. A
+result for attention, fluency, click-through, or reply rate is not evidence of
+trust, qualified demand, citation, conversion, or revenue unless those outcomes
+were measured separately.
+
+## Core doctrine
+
+**Compelling argument before compelling words.**
+
+Copy is not persuasive because it contains louder adjectives. It is persuasive
+when its claims, proof, stakes, and conclusion form a path the audience can
+follow and defend.
+
+Word choice still matters. Sequence it correctly:
+
+1. Decide what the audience must understand or believe.
+2. Build the case.
+3. Check the case against truth and evidence.
+4. Improve verbs, rhythm, and phrasing.
+
+Power words without an argument create noise. A sound argument written in flat
+language can be edited. Hype without proof cannot.
+
+### The destination belief
+
+Every asset needs one primary destination belief: the conclusion the prospect
+must accept before the offer or CTA makes sense.
+
+A valid destination belief is:
+
+- specific to the audience and moment;
+- necessary for the next action;
+- supported by supplied or verified evidence;
+- narrow enough for one asset;
+- honest about conditions and uncertainty.
+
+It is not permission to manufacture certainty. The goal is a supported
+conclusion, not an argument the reader is forbidden to question.
+
+Simple transactional copy may need only one claim and one action. Do not force
+a long story, coined mechanism, or full sales argument onto a button, alert,
+email, or high-intent page that already has enough context.
+
+## Format router
+
+| Asset class | Required treatment |
+|---|---|
+| `INDEXABLE_PUBLIC` | Agora argument, human-voice gate, full written GEO/AEO content gate, claim/source review, and technical publication handoff. |
+| `PUBLIC_NON_INDEXABLE_WRITTEN` | Agora argument, human-voice gate, written semantic/evidence rules, and public claim review. Use for ads, social posts, decks, and PDFs without an indexable page. Skip canonical, schema, sitemap, and indexability checks. |
+| `WRITTEN_PRIVATE` | Agora argument, human-voice gate, concrete entities, self-contained claims, and evidence fidelity. Skip crawl, canonical, sitemap, schema, and indexability checks because the asset has no public URL. |
+| `SPOKEN_ONLY` | Agora argument, human cadence, and claim gate. Skip GEO/AEO formatting. Write for breath, delivery, timing, and listener comprehension. |
+| `HYBRID` | Classify each output separately. A video script can be spoken-only; its public title, description, transcript, captions, show notes, and companion page are written assets. |
+
+Keep CTA labels as short as clarity permits. Specificity and a predictable next
+step outrank an arbitrary word count. For written public assets, optimize the
+claim block around the CTA for extraction and evidence; do not stuff proof,
+definitions, or query phrases into the button label.
+
+Use the complete definition and applicable checks in the written GEO/AEO gate
+below. No asset can promise citation, ranking, recommendation, or sale.
+
+## Required brief
+
+Capture these inputs before drafting. Infer only low-risk details from supplied
+context; state consequential assumptions outside the final copy.
+
+```text
+Asset class and channel:
+Audience and buying context:
+Awareness or intent stage:
+Problem, desire, or decision:
+Current belief:
+Destination belief:
+Offer and next action:
+Verified proof and sources:
+Claim and methodology limits:
+Voice samples or brand traits:
+Length and format constraints:
+```
+
+If the brief contains several audiences, beliefs, or CTAs, choose one primary
+path. Split distinct paths into separate assets rather than joining them with a
+long sentence.
+
+## Build the argument
+
+Use the minimum number of moves needed. This is a planning map, not a mandatory
+visible template.
+
+1. **Recognizable reality.** Name the situation in language the audience would
+   use.
+2. **Stakes.** Show the concrete cost, risk, frustration, desire, or missed
+   opportunity.
+3. **Broken assumption.** Explain why the current approach may not produce the
+   desired result. Do not misrepresent alternatives.
+4. **New criterion.** Establish what a credible solution must do.
+5. **Reason to believe.** Explain the product's specific approach or mechanism.
+6. **Difference and proof.** Show how the approach meets the criterion and cite
+   the strongest available evidence.
+7. **After-state.** Make the useful change tangible without inventing an
+   outcome.
+8. **Offer and action.** Present one next step that follows from the case.
+
+The compact logical form is:
+
+```text
+The desired result requires X.
+The current approach lacks X, within the stated scope.
+This offer provides X, supported by Y.
+Therefore this offer is a credible next step for this audience.
+```
+
+This syllogism is a house derivation from the transcript's argument-first idea,
+not a quoted Agora formula. If any premise cannot be proved, revise the premise
+or reduce the claim.
+
+Sequence by decision, not dogma:
+
+- For a high-consideration or B2B decision, a useful default is concrete stake,
+  truthful mechanism, diagnostic proof, relevant limits, brief illustration,
+  then CTA. Treat this as an evidence-aligned heuristic, not a universal law.
+- For a low-cost, hedonic, or early-awareness asset, vivid use, positive affect,
+  and mental imagery may do more work before the proof. The claim still needs
+  support before the action.
+- For a simple high-intent task, lead with the offer, answer, or action. Do not
+  manufacture a story merely to reproduce a sales-page sequence.
+
+### Argument continuity test
+
+For every block, ask:
+
+- What belief does the reader hold before this block?
+- What new information or feeling does this block add?
+- Why does the next block follow?
+- Would deleting this block weaken the case?
+
+Delete blocks that repeat the destination, decorate a claim, or smooth a
+transition without adding information.
+
+## Emotion without hype
+
+Emotion comes from a real situation and its consequence, not from inflated
+word choice.
+
+`HOUSE`: When several tensions would blur the decision, choose one primary
+emotional job for the block:
+
+- tension from a recognized problem;
+- relief from reduced uncertainty or effort;
+- control from clear evidence and a next step;
+- ambition tied to a credible outcome;
+- identity or belonging supported by the audience context;
+- curiosity created by a meaningful information gap.
+
+Then make it concrete:
+
+```text
+Situation + consequence + available agency
+```
+
+Prefer "Your team cannot tell which source shaped the answer" over "Navigate
+the complex AI landscape with confidence." The first gives the reader a scene,
+a problem, and a reason to care. The second asks generic words to carry the
+emotion.
+
+Rules:
+
+- For truthful fear or risk, state bounded severity and likelihood, then give a
+  credible response-efficacy path. Do not leave the reader trapped.
+- `HOUSE`: Prefer one dominant emotional tension when several would blur the
+  decision.
+- Make the desired change observable in the reader's work or life.
+- Prefer a concrete image or consequence to an extra adjective. Stronger
+  arguments, imagery, and positive affect can help persuasion; excessive
+  information load can hurt it. These are average effects, not permission to
+  decorate every sentence.
+- Do not invent fears, motives, memories, pain, testimonials, urgency, or
+  scarcity.
+- Do not shame, corner, or conceal material conditions.
+- Preserve honest trade-offs. Trust is part of persuasion.
+
+Conditional persuasion controls:
+
+| Persuasion device | Safe operating rule |
+|---|---|
+| Narrative | Use a short story or scene to illustrate relevant evidence. Do not assume stories beat facts or use a story as proof. |
+| Curiosity | Open a non-material information gap. Never hide mechanism, fit, price, risk, evidence, or the next step. |
+| Fear or risk | State truthful, bounded severity and likelihood, then a credible response-efficacy path. Most supporting research is non-commercial. |
+| Loss framing | `TEST` only. Compare it with neutral or gain framing and track intent and reactance, not attention alone. Agency does not make it reliably effective. |
+| Guilt | Use sparingly, if at all. Never shame, coerce, or imply moral failure to force action. |
+| Scarcity | State only a real, current, relevant limit. No fake inventory, countdown resets, or invented demand. |
+| Social proof | Use auditable, comparable, scoped evidence. Popularity without relevance is not proof. |
+
+Treat named psychological effects as hypotheses for the audience, channel, and
+outcome. Never teach the lore that people always buy on emotion and justify
+with logic, that stories always beat facts, or that one emotion-first sequence
+works everywhere.
+
+## Mechanism and differentiation
+
+The Agora source favors a unique mechanism: a specific reason the offer can
+produce the desired result. Keep the useful part; reject invented exclusivity.
+
+A clear mechanism statement answers:
+
+- What changes?
+- What causes that change?
+- Why is this approach relevant to the audience's problem?
+- What evidence supports the difference?
+- What conditions or limits apply?
+
+Use "proprietary," "unique," "only," "superior," or market-leadership language
+only when current evidence authorizes the exact claim. If the approach is
+different but not exclusive, name the concrete difference. If it is standard,
+explain its fit and execution instead of creating branded jargon.
+
+A coherent, truthful causal explanation can make an offer easier to understand
+and can raise perceived efficacy. It does not prove actual efficacy. A coined
+mechanism is optional; a real reason to believe is not. Never retrofit technical
+language around an unsupported outcome.
+
+For CiteSurge public copy, describe differentiation at the principle,
+experience, evidence, and outcome level. Do not disclose prompts, formulas,
+weights, thresholds, vendor routing, fixtures, classifiers, infrastructure, or
+raw provider diagnostics. Vague "proprietary AI" language is not a substitute
+for a reason to believe.
+
+Run the education-leakage test: if the explanation teaches the category but
+leaves the buyer equally likely to choose any competitor, decide whether
+category building is intentional. If it is not, connect the lesson to the
+offer's verified difference or fit. Never invent exclusivity to close the gap.
+
+## Proof and claim control
+
+Argument strength is limited by proof strength. Choose evidence diagnostic of
+the exact claim and decision. As a `HOUSE` selection order, look for:
+
+1. Current, inspectable product behavior or a non-expired approved claim.
+2. First-hand or original evidence with a disclosed method and date.
+3. Current primary or official sources that support the exact statement.
+4. A verified demonstration, comparison, or consented case record.
+5. A verified, attributable testimonial within its real scope.
+6. Corroboration that is independent, comparable, current, and within scope.
+
+Then name any material limitation that keeps the conclusion honest. A
+limitation is not evidence. No proof means narrow or remove the claim; it does
+not mean add adjectives.
+
+Across persuasion research, statistics generally outperform a lone anecdote; a
+relevant anecdote can still make an emotionally charged or self-relevant issue
+concrete. Use the story to illustrate the evidence, not replace it. In online
+review settings, valence, argument quality, usefulness, message trust, and
+source credibility predict purchase intention; raw count alone is insufficient.
+Category-specific automobile evidence shows that conflicting sources can reduce
+intent and early sales. Treat comparability, recency, and consistency as
+`HOUSE` proof-diagnosticity checks, and reconcile material discrepancies rather
+than hiding them.
+
+Keep claim types distinct:
+
+| Type | Treatment |
+|---|---|
+| Fact | Verify source, wording, date, scope, and expiry. |
+| Inference | Show what evidence supports it and retain uncertainty. |
+| Opinion | Attribute ownership; do not present it as measured fact. |
+| Promise | Confirm authority, delivery scope, conditions, and ability to perform. |
+
+For CiteSurge:
+
+- Only non-expired `LIVE` rows in `docs/staff/public-claim-ledger.md` may be
+  marketed as current.
+- `LANDED_INACTIVE`, `OPEN_PENDING`, `FUTURE`, and `UNSUPPORTED` are not current
+  capabilities.
+- Engine behavior claims need current primary support.
+- Customer-facing copy must not expose HTTP codes, auth or quota failures,
+  scraping limits, probe names, raw exceptions, or implementation diagnostics.
+- Do not promise that crawler access, schema, AI-readable files, or any single
+  content edit causes citation.
+- Keep brand mention, source selection, answer absorption, visible citation,
+  referral, recommendation, conversion, and revenue as separate measures.
+- Public positioning sells GEO/AEO. Use SEO only for necessary comparison,
+  decline, or replacement context; never sell it positively. Review every SEO
+  occurrence before publication.
+
+## Concision and format standard
+
+There is no universal short-copy law. Write the shortest complete decision
+path for the channel, audience, stage, device, traffic source, offer complexity,
+language, and risk. One asset should still have one destination belief; one
+block should do one main job; one sentence should carry one main claim.
+
+Ranges below are editing controls or starting variants, not guaranteed optima:
+
+| Format | Working rule | Status and limit |
+|---|---|---|
+| Headline | Aim for 3-9 words; review above 12. | `HOUSE` |
+| Support line | Aim for 12-24 words; review above 30. | `HOUSE` |
+| Sentence | Keep most at 20 words or fewer; review above 25. | `EVIDENCE-INFORMED` for accessibility, based on a small screen-reader study; not a conversion law. |
+| Marketing paragraph | Usually 1-3 sentences; review above 75 words. | `HOUSE` |
+| Hero | Test roughly 25-75 total words when mobile, high intent, or urgency favors compression. | `TEST`; no universal hero optimum. |
+| Button CTA | Test 2-5 words, but use more when needed to make the action predictable. | `TEST`; clarity and transparency outrank brevity. |
+| B2B cold email | Start around 50-100 words and 3-4 sentences, then ask for interest rather than a calendar commitment. | `EVIDENCE-INFORMED`, from large observational Gong datasets; not causal or universal. |
+| Active-deal follow-up | Use roughly 3-6 short sentences with a specific next step; no hard word cap. | `TEST`; context carries more detail. |
+| Warm personal email | Start around 50-120 words when no channel constraint says otherwise. | `HOUSE` |
+| LinkedIn cold message | Test below roughly 400-500 characters with one reason to care and one low-friction ask. | `C` and `TEST`; LinkedIn findings are product- and use-case-specific. |
+| Utility, product, or question-led page | Put the answer, offer, or decision aid in the first sentence or paragraph. | `HOUSE`, aligned with general web usability; do not force this onto narrative or editorial openings. |
+| Long sales or comparison page | No global word cap. Put a useful summary before detail, use progressive disclosure, and chunk proof and objections. | `HOUSE`, aligned with general usability practice. |
+| Comparison set | Test a static table around five or fewer options; reflow or use readable cards on mobile. | `TEST`; a usability starting point, not a ranking rule. |
+| Paid ad | Obey the platform limit. Test personal or subjective framing for attention against objective, brand-consistent framing for conversion. | `TEST`; keep click-through and conversion separate. |
+| Social post | No universal character optimum. Preserve the hook, useful substance, and readable shape. | `HOUSE`; test length in the actual platform and content category. |
+
+On mobile, increase signal and reduce friction; do not merely delete words.
+Preserve meaning, evidence, button clarity, and usable comparison structure. Do
+not force legal qualifications, technical explanations, translations, or
+complex decisions into a house range. Evidence and necessary qualification are
+load-bearing; decorative setup is not.
+
+### Compression order
+
+Cut in this order:
+
+1. repeated claims and empty recaps;
+2. throat-clearing and generic introductions;
+3. adjective stacks and intensifiers;
+4. formal transitions the logic does not need;
+5. unsupported claims and secondary examples;
+6. a secondary CTA or audience path.
+
+Never cut a material condition, source, qualification, or next-step detail just
+to hit a count.
+
+Keep a sentence only if it does at least one job:
+
+- advances the destination belief;
+- proves a material claim;
+- resolves a real objection;
+- makes the action clear.
+
+If deleting it changes nothing, delete it.
+
+## CTA standard
+
+A CTA should feel like the next logical step, not a sudden hard-sell voice.
+
+Every CTA must:
+
+- ask for one concrete action;
+- make the next screen, reply, or commitment predictable;
+- match the audience's intent and the argument's promise;
+- use a specific verb and object when space allows;
+- avoid adding a new claim;
+- use urgency or scarcity only when verified and material;
+- point to a real route or action that works.
+
+Choose one primary CTA intent. A long page may repeat the same action at useful
+decision points; repetition does not create a second CTA path. In a cold B2B
+message, a low-friction interest question can be more appropriate than asking
+for a meeting immediately. In an active conversation or later deal stage, ask
+for the specific next step. Treat both as context-sensitive starting rules.
+
+Useful patterns, subject to route and offer truth:
+
+| Intent | Pattern examples |
+|---|---|
+| Discover | `See the evidence`, `Read the method` |
+| Compare | `Compare approaches`, `Compare plans` |
+| Evaluate | `Review your gaps`, `View the plan` |
+| Act | `Request an audit`, `Book a review` |
+
+Prefer a precise action over `Learn more`, `Get started`, `Unlock your
+potential`, or `Start your journey`, unless the surrounding interface already
+makes the outcome unambiguous. Keep supporting proof or risk-reversal in nearby
+microcopy, not inside the button. Claims such as "No card required" or a
+response time must be verified before use.
+
+## Human voice and AI-writing-tell gate
+
+The goal is not to disguise authorship. The goal is specific, accurate,
+author-owned writing that fits the channel.
+
+### Core tests
+
+- **Competitor-swap test:** if replacing the subject with any competitor leaves
+  the paragraph accurate, rewrite it with a real mechanism, fact, situation,
+  boundary, or point of view.
+- **Read-aloud test:** if the intended speaker would not say it, revise it.
+- **Defense test:** if the author would hesitate to defend a sentence in front
+  of the buyer, verify, qualify, or remove it.
+- **Density test:** judge clusters and repeated templates, not one ordinary word
+  in isolation.
+
+High-risk clusters include vague verbs such as `unlock`, `leverage`, `elevate`,
+and `transform`; inflated adjectives such as `seamless`, `robust`,
+`groundbreaking`, and `world-class`; and generic nouns such as `landscape`,
+`ecosystem`, `journey`, and `game-changer`.
+
+Review or remove:
+
+- stock openers and prompt acknowledgements;
+- formal connective chains;
+- significance tails that explain why a sentence matters instead of adding
+  evidence;
+- repeated three-part lists;
+- repeated `Not only... but also...` structures;
+- heading restatements and automatic paragraph summaries;
+- symmetrical sections that give every point equal weight;
+- generic conclusions that do not decide, request, or direct;
+- repeated bold-label bullets and decorative mid-prose emphasis;
+- fabricated citations, anecdotes, motives, opinions, or credentials;
+- random errors or slang added to look human.
+
+Necessary technical terms, an earned rhetorical triplet, a natural fragment,
+or an author-consistent punctuation choice may remain in ordinary writing. For
+gated CiteSurge public pages, the repository scanner is stricter:
+
+- prompt leakage or acknowledgement openers fail immediately;
+- em dashes and curly quote characters or entities in prose fail;
+- three or more watchlist indicators in one prose block fail;
+- the same template or significance-tail fingerprint repeated twice fails;
+- a single watchlist occurrence is informational, not an automatic failure.
+
+Use `content/ai-writing-tells.md` for rationale and
+`content/ai-writing-tells.ts` for the current machine-readable rules. Never
+promise that copy is "undetectable" or optimize it toward one detector score.
+
+## Written GEO/AEO and citability gate
+
+Apply the semantic and evidence rules below to every written asset. Apply the
+technical rules only when the asset has an indexable public URL. Public
+non-indexable work still needs public-safe claims; private writing still needs
+clear entities and source fidelity. Spoken-only assets skip this section unless
+a written derivative is published.
+
+"Fully GEO/AEO optimized" means maximum practical extractability,
+verifiability, information value, and technical eligibility for the asset. It
+does not mean guaranteed retrieval, selection, quotation, citation, ranking,
+recommendation, referral, or sale. 2026 preprints provide descriptive,
+observational, and experimental evidence for testing modular, self-contained,
+evidence-bearing passages in some AI-search systems; they do not establish a
+universal ranking formula.
+
+### Content requirements
+
+1. **Resolve the real question early when the format calls for it.** Utility,
+   product, comparison, and question-led content should put the answer, offer,
+   or decision aid in the first sentence or paragraph. A narrative or editorial
+   opening may begin with a relevant scene, but should reveal the subject and
+   value before the setup becomes a detour.
+2. **Name the entity and scope.** Use the full brand, product, category, market,
+   date, and population where ambiguity would weaken an extracted passage.
+3. **Build self-contained evidence blocks.** A useful paragraph should preserve
+   its subject, claim, support, and material limit when quoted alone. Replace
+   vague dependencies such as "this," "it," and "as above" when context would
+   be lost.
+4. **Keep evidence next to the claim.** Name and link the exact source, date
+   material facts, distinguish fact from inference, and state relevant limits.
+   Prefer current primary or first-party sources.
+5. **Add information gain.** Contribute original data, first-hand experience,
+   a useful comparison, a reproducible demonstration, a disclosure-safe method,
+   or a defensible point of view. Do not paraphrase category consensus and call
+   it insight.
+6. **Structure around reader intent.** Use descriptive headings, definitions,
+   lists, steps, tables, or FAQs only when they make the answer easier to use.
+   Q&A formatting by itself is not a citability strategy.
+7. **Make provenance visible.** For public editorial work, show a verified
+   author or publisher, publication or review date, sources, and relevant
+   disclosures. Treat these as trust and verification practices, not universal
+   ranking factors.
+8. **Link contextually.** Connect the page to the most relevant evidence,
+   capability, comparison, trust, or next-step page. Do not chase arbitrary
+   internal-link counts.
+9. **Keep surfaces consistent.** Visible copy, metadata, structured data,
+   feeds, sales material, and machine-readable surfaces must describe the same
+   facts and scope.
+10. **Measure distinct events.** Record the engine or surface, prompt or topic,
+    market, date, and observed output. Keep brand mention, source selection,
+    answer absorption, visible citation, referral, recommendation, conversion,
+    and revenue separate.
+
+### Emotional marketing plus extractability
+
+Emotion can earn attention; a self-contained evidence block can support
+understanding and verification. A useful page-level sequence is:
+
+1. a concise hook that names a real tension or desired change;
+2. an early plain-language answer or value claim;
+3. self-contained blocks for mechanism, evidence, comparison, and limits;
+4. one primary CTA intent that follows from the supported conclusion.
+
+The hero may carry more emotion. Proof sections should carry more diagnostic
+facts. Both must use the same destination belief and claim scope. Do not repeat
+this four-part shape mechanically across every section.
+
+### Engine-specific boundaries
+
+These controls were verified against official documentation on 2026-07-17.
+Recheck them before an engine-specific recommendation; crawler names and product
+behavior can change.
+
+| System | Safe rule |
+|---|---|
+| ChatGPT Search | `OAI-SearchBot` is the Search crawler; `GPTBot` concerns potential model training; `ChatGPT-User` performs user-initiated fetches. Permission can affect eligibility, not inclusion. ChatGPT referrals can be identified with `utm_source=chatgpt.com`. |
+| Claude web search | `Claude-SearchBot`, `ClaudeBot`, and `Claude-User` have different purposes. Blocking Search or User access may reduce visibility. Do not claim a first-party citation or referral analytics method unless Anthropic documents one. |
+| Perplexity | Treat `PerplexityBot` and `Perplexity-User` separately. Current official descriptions of robots behavior are not fully consistent, so name the agent, action, source, and date rather than saying Perplexity simply respects or ignores `robots.txt`. |
+| Google AI Overviews and AI Mode | Google Search controls and Googlebot govern Search eligibility. `Google-Extended` is separate and does not control inclusion or ranking in Google Search. Google requires no special AI schema, AI text file, or special page structure; `llms.txt` is ignored by Google Search. |
+| Bing and Copilot | Bing Webmaster Tools AI Performance and Microsoft Clarity can distinguish some citations, queries, and referrals. Citation count is not proof of prominence, answer contribution, accuracy, or conversion. IndexNow is a Bing-supported freshness mechanism, not a cross-engine ranking tactic. |
+| Grok | xAI documents web-search results and citations for its API. Its `citations` field can include every URL encountered during tool search, while inline citations are answer-linked and not guaranteed; do not count every encountered URL as visible support. Public site-owner controls for Grok crawling, indexing, and training are not established in the reviewed official material. Mark them unknown rather than inferring them. |
+
+Never import one engine's crawler, metadata, structured-data, or measurement
+rule into another engine without first-party support.
+
+### Indexable-public technical handoff
+
+Copy alone cannot complete GEO/AEO. Confirm or hand off checks for:
+
+- one stable, accessible canonical URL and intentional indexability;
+- important claims and evidence available in robust, accessible page content;
+- intentional, engine-specific crawler policy;
+- HTTPS, stable delivery, useful status behavior, and crawlable links;
+- sitemap discovery and truthful freshness dates where applicable;
+- descriptive metadata aligned with visible copy;
+- contextual internal links and useful image alternatives;
+- valid structured data that matches visible facts and a supported use case;
+- consistent author, publisher, organization, and product identities.
+
+Prefer important evidence in normal accessible HTML for retrieval robustness,
+but do not claim that every engine ignores client-rendered content, PDFs,
+images, or other formats. Structured data can clarify meaning or qualify a page
+for documented features; it does not guarantee AI selection or citation.
+Google documents no special AI schema. OpenAI's documented structured metadata
+use is currently specific to Shopping, not a general ChatGPT ranking rule.
+Never invent questions, authors, dates, ratings, products, or organization
+facts for markup. Do not require `llms.txt`; Google ignores it and a citation
+effect is unproved elsewhere.
+
+### GEO/AEO anti-patterns
+
+Reject:
+
+- keyword or entity stuffing;
+- long copy written only to satisfy a word count;
+- thin definition pages and minor-query variants;
+- hidden answer text or machine-only claims;
+- fake statistics, false precision, or source laundering;
+- FAQ or Q&A blocks added only as a citability tactic;
+- duplicated passages with swapped keywords;
+- crawler permission presented as retrieval, selection, or citation;
+- structured data, authorship, dates, direct answers, or HTML rendering
+  presented as universal ranking factors;
+- citation count treated as answer contribution, accuracy, endorsement,
+  prominence, referral, or conversion;
+- `llms.txt`, IndexNow, or an engine-specific crawler rule presented as a
+  universal requirement;
+- observational movement presented as causal lift;
+- generic AI phrasing that lowers specificity and information value.
+
+## Operational workflow
+
+1. Route the asset.
+2. Confirm audience, decision stage, offer, action, product state, evidence,
+   channel constraints, and disclosure limits.
+3. Write the current belief and one supportable destination belief.
+4. Select the minimum argument moves and attach proof to each premise.
+5. Draft for the channel: one idea per block and one primary action placed at
+   format-appropriate decision points.
+6. Apply the written GEO/AEO layer to every written output. Add technical
+   publication checks only for indexable public work.
+7. Cut setup, repetition, padding, and secondary arguments. Preserve evidence
+   and qualifications; use format ranges only according to their label.
+8. Apply the competitor-swap, read-aloud, defense, and density tests.
+9. Verify the CTA, then compare every final claim with its source and ledger
+   status.
+
+For CiteSurge public changes, pass every changed marketing file or directory
+explicitly to `npm run dev:ai-writing-tells:scan --`, then run:
+
+```powershell
+npm run dev:public:claim-safety:smoke
+```
+
+No-argument AI scan covers only its default gated page set. Pass every changed
+marketing path explicitly. Inspect rendered copy and all changed metadata,
+schema, feed, sitemap, or AI-readable surfaces.
+
+Published Insights also require a reviewer independent from the draft owner.
+That reviewer checks rendered article, sources, claim scope, methodology
+boundary, metadata, schema, feed, and AI-readable summary. Automated checks do
+not replace this review; publication stays blocked when it is unavailable.
+
+## Applied examples
+
+CiteSurge examples below show structure, not standing publication approval.
+Recheck claim ledger, source freshness, offer, and destination route before use.
+
+### Hero
+
+Weak:
+
+> Transform your brand's future with better AI visibility today.
+
+Stronger:
+
+> See the evidence behind AI answers. CiteSurge records observed answers with
+> citation URLs, prompt, engine, and market, then turns verified gaps into a
+> prioritized plan. Review the evidence.
+
+The stronger version moves from uncertainty to control through a concrete
+mechanism and action. It does not ask hype words to supply emotion.
+
+### Short sales email
+
+Fictional evidence for structure only. Replace every observation before use.
+
+> Hi Sam,
+>
+> I checked how answer engines describe Acme. Acme appeared in two buyer
+> questions, but neither answer cited an Acme page.
+>
+> I put the evidence and three possible fixes into a short audit. Want me to
+> send it?
+
+The sequence is situation, concrete stake, available agency, then one
+low-friction question. No category lecture or inflated claim is needed.
+
+### CTA and microcopy
+
+Weak button: `Unlock your potential`
+
+Stronger button: `Review your gaps`
+
+Supporting line: `See the observed answers before deciding what to fix.`
+
+The button is concise because its action is clear. The nearby line carries
+context and reduces uncertainty.
+
+## Release checklist
+
+- [ ] Asset route, audience, destination belief, and one primary CTA intent are
+  explicit.
+- [ ] Stakes come from a real situation and leave the reader with agency.
+- [ ] Narrative, fear, loss, guilt, scarcity, curiosity, and social proof obey
+  their conditional and ethical controls.
+- [ ] Each premise is verified, qualified, or removed.
+- [ ] Mechanism is concrete, disclosure-safe, and passes the education-leakage
+  test.
+- [ ] Every block advances, proves, resolves, or directs; setup and recaps are
+  cut.
+- [ ] Length and structure fit the format; no `HOUSE` or `TEST` range is treated
+  as a universal performance law.
+- [ ] CTA names a real next action and adds no new claim.
+- [ ] Competitor-swap, read-aloud, defense, and density tests pass.
+- [ ] No fabricated detail, fake urgency, prompt leakage, or detector promise.
+- [ ] Written utility, product, comparison, and question-led copy answers early;
+  all written copy names entities and keeps key evidence blocks self-contained.
+- [ ] Material written claims have exact, current, preferably primary sources.
+- [ ] Material inconsistencies across sources are reconciled or disclosed.
+- [ ] Structure, authorship, freshness, disclosures, and links serve reader
+  intent.
+- [ ] Public visible and machine-readable surfaces agree; indexable work has a
+  technical handoff.
+- [ ] No stuffing, thin variants, machine-only text, or citation guarantee.
+- [ ] Engine-specific crawler and measurement rules were reverified and were not
+  generalized across engines.
+- [ ] CiteSurge public claims map to non-expired `LIVE` rows.
+- [ ] CiteSurge methodology and customer-safety boundaries hold.
+- [ ] Brand mention, source selection, answer absorption, visible citation,
+  referral, recommendation, conversion, and revenue stay distinct.
+- [ ] GEO/AEO remains CiteSurge's sales frame; every SEO occurrence is reviewed.
+- [ ] Required smokes and independent editorial review pass where applicable.
+- [ ] Author can defend every sentence.
+
+## Experiment and learning standard
+
+Use local testing to adapt a `HOUSE`, `C`, `EVIDENCE-INFORMED`, or `TEST` rule.
+Do not convert a weak experiment into doctrine.
+
+1. Change one variable family and pre-register the audience, decision stage,
+   primary outcome, minimum useful effect, and guardrails. If several elements
+   change, label the test a full-page or full-message variant.
+2. Randomize the correct unit and control repeat exposure, sales-rep effects,
+   account overlap, and channel contamination where relevant.
+3. Calculate sample size from the current baseline, minimum useful effect,
+   power target, clustering, and multiple comparisons. Never copy a source
+   study's sample size as the new requirement.
+4. Run through the relevant business cycle. Do not stop when a result first
+   becomes favorable unless a sequential design and stopping rule were declared.
+5. Keep click-through, conversion, reply, meeting, qualified lead, close,
+   citation, referral, and revenue metrics separate.
+6. Track quality and harm guardrails: accessibility, comprehension, negative
+   replies, spam reports, unsubscribes, cost, downstream fit, and trust signals.
+7. Segment only when pre-declared and sufficiently powered.
+8. Record positive, null, mixed, harmful, and inconclusive outcomes.
+9. With low traffic, prefer comprehension tests, usability sessions, message
+   interviews, or sequential learning batches over underpowered causal claims.
+10. Promote a rule only when the observed effect matches the intended outcome
+    and survives the declared guardrails.
+
+## Evidence register
+
+This register records the durable sources used to turn the supplied research
+into operating rules. It deliberately excludes the reports' non-portable
+session citation tokens.
+
+### Persuasion, emotion, mechanism, and proof
+
+| Evidence | Grade | Operational use and limit |
+|---|---:|---|
+| [Verbal persuasion in marketing meta-analysis](https://doi.org/10.1007/s11747-025-01095-4) | `B` | Across 308 studies, argument quality, mental imagery, and positivity were positively associated with persuasion while information load was negative. Outcomes and moderators vary; this is not a conversion formula. |
+| [Delayed narrative-persuasion meta-analysis](https://doi.org/10.1093/joc/jqaa005) and [registered narrative-correction report](https://doi.org/10.1186/s41235-020-00266-x) | `B` | Narrative effects are small and conditional; equating information and comprehension can remove the advantage. Use narrative to illustrate, not replace, proof. |
+| [Anecdotal versus statistical evidence meta-analysis](https://doi.org/10.1016/j.obhdp.2020.01.006) | `B` | Statistics generally persuade more; anecdotes can help under high emotion or self-relevance. Match evidence to context without treating one story as representative. |
+| [Mechanism-based claim experiments](https://doi.org/10.1093/jcr/ucae066) | `B` | A coherent causal chain can increase fluency and perceived efficacy. The experiments do not establish actual product efficacy. |
+| [Online review meta-analysis](https://doi.org/10.1016/j.dim.2023.100058), [eWOM predictor meta-analysis](https://doi.org/10.1007/s10796-019-09924-y), and [source-credibility meta-analysis](https://doi.org/10.1016/j.jretconser.2019.01.005) | `B` | Valence, argument quality, usefulness, message trust, expertise, and relevance matter. Purchase intention is not always observed purchase behavior. |
+| [Multi-source inconsistency research](https://doi.org/10.1016/j.intmar.2021.01.003) | `B` | In an automobile experiment and sales analysis, conflicting sources reduced intent and early sales. Generalize cautiously beyond high-consideration durable goods. |
+| [Fear-appeal meta-analysis](https://doi.org/10.1037/a0039729) and [efficacy follow-up](https://doi.org/10.1093/joc/jqab049) | `B` | Average effects were positive, mainly outside commercial marketing; credible response efficacy is important. This does not license exaggeration or coercion. |
+| [Guilt meta-analysis](https://doi.org/10.3389/fpsyg.2023.1201631), [scarcity-backfire experiments](https://doi.org/10.3389/fpsyg.2024.1225011), and recent [loss-framing attention](https://doi.org/10.1093/cercor/bhaf006) and [reactance](https://doi.org/10.1016/j.jbusres.2025.115684) studies | `B` | Guilt effects are small, scarcity can backfire, and loss framing can trade attention for reactance or lower intent. Use only truthfully and test the target outcome. |
+
+### Copy length, structure, and CTA design
+
+| Evidence | Grade | Operational use and limit |
+|---|---:|---|
+| [Sentence-length study with screen-reader users](https://doi.org/10.1007/978-3-030-49282-3_18) | `B` | A small study of 21 users favored 16-20-word sentences for comprehension. It supports an accessibility editing trigger, not a universal marketing optimum. |
+| [WCAG 2.2 guidance on headings and labels](https://www.w3.org/WAI/WCAG22/Understanding/headings-and-labels) | `A` | Headings and labels should describe topic or purpose and need not be long. Descriptiveness outranks minimal word count. |
+| Gong analyses of [executive cold email length](https://www.gong.io/blog/do-execs-really-reply-to-cold-email-here-s-what-the-data-says), [cold-email structure](https://www.gong.io/files/gong-guide-how-to-master-cold-email-get-the-data-backed-guide-based-on-85-million-emails.pdf), [cold CTA type](https://www.gong.io/blog/this-surprising-cold-email-cta-will-help-you-book-a-lot-more-meetings), and [later-stage sales email](https://www.gong.io/blog/how-to-write-a-sales-email) | `C` | Large observational sales datasets support short first-touch messages, interest CTAs when cold, and more specific asks later. Vendor population, selection, and outcome limits remain. |
+| LinkedIn guidance for [recruiter InMail](https://www.linkedin.com/business/talent/blog/product-tips/tips-for-writing-inmails-from-linkedin-recruiters), [Sponsored Messaging](https://business.linkedin.com/advertise/ads/sponsored-messaging/tips), and [LinkedIn ad formats](https://www.linkedin.com/business/marketing/blog/linkedin-ads/a-b2b-marketers-guide-to-every-linkedin-ad-type) | `C` | Shorter message bodies performed better in the named LinkedIn products and datasets. Do not generalize their character ranges to all direct messages. |
+| [Computational-advertising text-style field study](https://doi.org/10.1016/j.fmre.2021.11.004) | `B` | Subjective, personalized text produced higher click-through while objective, brand-consistent text produced higher conversion in the studied system. Treat this as a channel-specific test hypothesis, not a universal funnel rule. |
+
+### AI-search visibility and citability
+
+| Evidence | Grade | Operational use and limit |
+|---|---:|---|
+| [OpenAI crawler documentation](https://developers.openai.com/api/docs/bots) and [publisher FAQ](https://help.openai.com/en/articles/12627856-publishers-and-developers-faq) | `A` | Distinguishes Search, training, and user-fetch agents and documents ChatGPT referral tagging. Permission affects access or eligibility, not guaranteed inclusion. |
+| [OpenAI Shopping with ChatGPT Search](https://help.openai.com/en/articles/11128490-shopping-with-chatgpt-search) | `A` | Documents structured product metadata in the Shopping context only. It is not evidence for a general ChatGPT ranking or citation factor. |
+| [Anthropic crawler guidance](https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler) | `A` | Distinguishes Claude's crawler purposes. Do not infer analytics, ranking factors, or cross-engine rules. |
+| [Perplexity crawler documentation](https://docs.perplexity.ai/docs/resources/perplexity-crawlers) and [robots.txt help](https://www.perplexity.ai/help-center/en/articles/10354969-how-does-perplexity-follow-robots-txt) | `A` | Official pages verify the named agents and published controls. Their descriptions of robots behavior are not fully consistent, so the unresolved behavioral inference remains `D`; state the agent, action, source, and date. |
+| [Google AI-features guidance](https://developers.google.com/search/docs/appearance/ai-features) and [AI optimization guidance](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) | `A` | Normal Google Search eligibility and controls apply; no special AI schema, text file, or page structure is required. `Google-Extended` is separate from Search inclusion and ranking. |
+| [Bing AI Performance](https://blogs.bing.com/webmaster/February-2026/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview), [Microsoft Clarity AI citations](https://learn.microsoft.com/en-us/clarity/ai-visibility/ai-citations), and [IndexNow documentation](https://www.indexnow.org/documentation) | `A` | Supports engine-specific observation of citations, queries, referrals, and freshness submission. None alone proves answer contribution, prominence, accuracy, or business effect. |
+| [xAI web-search documentation](https://docs.x.ai/developers/tools/web-search) and [xAI citation documentation](https://docs.x.ai/developers/tools/citations) | `A` | Documents API search, encountered URLs, and inline citations. Site-owner crawl, index, and training controls remain `D` because they are unknown in the reviewed official material. |
+| [Selection and answer-absorption preprint](https://arxiv.org/abs/2604.25707) | `B` | Descriptive analysis with controlled prompts supports measuring source selection separately from answer absorption; it does not establish a structural intervention effect. |
+| [Multi-engine content-structure preprint](https://arxiv.org/abs/2603.29979) | `B` | Reports experimental structural effects and supports testing self-contained, evidence-bearing passages. It remains a preprint and not a universal ranking rule. |
+| [Google AI Overview sourcing preprint](https://arxiv.org/abs/2605.14021) | `B` | Longitudinal observational evidence supports separating visible citation from claim fidelity. It is Google-specific and non-causal. |
+
+## Evidence maintenance
+
+- Recheck official engine documentation every 30 days when this file is used
+  for engine-specific technical advice, and immediately after a crawler,
+  product, analytics, or policy change.
+- Check cited preprints quarterly for revisions, peer review, or retraction.
+- Recheck product claims, prices, routes, offers, and legal constraints at the
+  time of use; do not rely on this file as their source of truth.
+- When adding a rule, record the channel, audience, population, measured
+  outcome, date, evidence grade, and material limit.
+- Preserve contradictory and null evidence. Do not silently promote a vendor
+  association or local test into a universal rule.
+- Use durable direct URLs or DOIs. Do not copy temporary research-session
+  citation markers into this file.
+
+## Change log
+
+| Date | Change |
+|---|---|
+| 2026-07-17 | Integrated three commissioned research reports; replaced session-bound research artifacts with durable sources; added portable LLM activation, evidence grades, conditional persuasion controls, format-specific length and CTA guidance, engine-specific GEO/AEO boundaries, and a testing standard. |