МАТЧАСТЬ / NEXT BEST PUBLICATION / DOCUMENT 35 / 02.09.2026

Next Best Publication

Как превратить Search Proof, AI Visibility, Entity Graph, editorial gaps и результаты предыдущих публикаций в конкретный следующий шаг: не абстрактное «пишите больше контента», а приоритизированную рекомендацию вида «сейчас вам нужен кейс по теме X, потому что конкуренты регулярно появляются в 18 целевых prompts, у вас отсутствует подтверждённый материал по этой задаче, а AI-системы опираются на источники Y и Z». Документ определяет gap taxonomy, recommendation engine, scoring, evidence, briefing, human review, feedback loop и коммерческую роль модуля.

Gap → Actionаналитика должна заканчиваться конкретным действием

Not always contentиногда лучший шаг — исправить факт, профиль или внешний источник

Evidence firstкаждая рекомендация объясняет, откуда она взялась

Repeat engineглавный мост от первого отчёта ко второй покупке

1. Главное решение

Next Best Publication должен быть не генератором идей, а decision engine. Он получает реальные наблюдения о компании, конкурентах, поиске, AI-ответах, источниках и существующем контенте, определяет наиболее ценный информационный пробел и предлагает один из нескольких типов действий. Публикация — важный, но не единственный тип результата.

OBSERVE Search + AI + Reader + Entity Graph ↓ DETECT GAP ↓ CLASSIFY GAP ↓ CHOOSE ACTION TYPE ↓ PRIORITIZE ↓ SHOW EVIDENCE ↓ CREATE BRIEF / TASK ↓ HUMAN REVIEW ↓ EXECUTE ↓ MEASURE AGAIN

2. Почему это ключевой repeat engine

После первой публикации клиент обычно не знает, что делать дальше. Если «Матчасть» заканчивает продукт PDF-отчётом, следующий заказ снова нужно продавать вручную. Если отчёт заканчивается доказанным gap и готовым briefing, повторная покупка становится продолжением рабочего процесса.

Основная продуктовая петля: Publish → Measure → Find Gap → Recommend → Publish/Fix → Measure.

3. Рыночное направление уже видно

Semrush в 2026 году прямо связывает AI Visibility с gap analysis: показывает prompts, topics и sources, где конкуренты появляются, а бренд отсутствует, и предлагает закрывать эти пробелы через контент и внешнюю видимость. Их AI Search Optimizer анализирует существующие страницы и выдаёт приоритетные рекомендации по структуре, clarity и entity signals.

Profound пошёл ещё дальше: раздел Projects каждую неделю получает от Aim agent набор конкретных возможностей по visibility, citations и prompt coverage, превращает каждую возможность в Project, а затем позволяет запускать Agents для выполнения работы. FactCheck позволяет из inaccurate claim сразу перейти к исправлению Knowledge Base, outreach к источнику или созданию нового контента.

Это подтверждает, что рынок движется от dashboard → prioritized work queue. «Матчасти» нужно построить этот слой вокруг собственного publishing workflow, а не просто повторить чужой AI dashboard.

4. Чем «Матчасть» может быть сильнее

Semrush / Profound: observe external AI/search ecosystem → recommend optimization/content Mathchast additionally knows: verified company facts published Mathchast assets editorial formats publication quality client experts case relationships source provenance Search Proof reader behavior commercial workflow → recommendation can become real publication order immediately.

5. Next Best Publication — название продукта, не ограничение логики

Внутренний engine лучше назвать Next Best Action, а клиентский коммерческий модуль — Next Best Publication, потому что часть рекомендаций вообще не должна вести к новой статье.

NEXT BEST ACTION TYPES: PUBLISH_CASE PUBLISH_EXPLAINER PUBLISH_RESEARCH PUBLISH_EXPERT_OPINION PUBLISH_INTERVIEW UPDATE_EXISTING FIX_COMPANY_PROFILE ADD_EXPERT VERIFY_CLAIM UPDATE_CLIENT_SITE EARN_EXTERNAL_MEDIA CORRECT_EXTERNAL_SOURCE WAIT_AND_MEASURE NO_ACTION

6. Самый важный anti-pattern

Каждый gap → «купите ещё одну статью». Такой engine быстро станет рекламной машиной, которой клиент перестанет доверять.

7. Пример, где статья НЕ нужна

AI говорит: "Acme имеет бесплатный тариф" Verified fact: бесплатного тарифа нет Source: устаревшая pricing page клиента Correct action: UPDATE_CLIENT_SITE Not: "купите статью о тарифах".

8. Пример, где нужна внешняя PR-работа

Competitors appear because AI repeatedly cites: Industry Media X Client has: strong own content strong Mathchast case but no independent coverage Action: EARN_EXTERNAL_MEDIA Not: third Mathchast article on same topic.

9. Пример, где нужна новая публикация

Prompt cluster: "AI agent security" Client: Mention Rate 2% Competitors: 42% Source pattern: technical cases security research expert commentary Mathchast: no client case/research on security Verified expertise: CTO exists Action: PUBLISH_CASE or PUBLISH_EXPERT_OPINION Priority: HIGH.

10. Gap taxonomy

Gap	Что отсутствует	Типичный action
TOPIC_GAP	Компания не представлена в важной теме	Article / case / research
PROMPT_GAP	Конкуренты появляются в конкретных buyer prompts	Targeted content/source work
CITATION_GAP	Бренд упоминается, но его источники не цитируются	Source/content optimization
SOURCE_GAP	AI опирается на источники, где клиента нет	Earned media / research / new source
FORMAT_GAP	Нет нужного типа доказательства	Case/research/expert content
ENTITY_GAP	Не хватает подтверждённой компании/эксперта/связи	Verification/profile
FACT_GAP	AI не знает или путает объективный факт	Fix source / verify claim
NARRATIVE_GAP	Бренд описывается не в нужном контексте	Evidence/content/PR
SEARCH_GAP	Есть demand, но нет подходящей indexable page	Publish/update
READER_GAP	Статья видна, но не ведёт к следующему действию	Update UX/content

11. Topic Gap

Input: topic map prompt mentions competitor presence existing content graph Question: "Есть ли у компании содержательный публичный asset по теме, где покупатель реально её ищет?"

12. Prompt Gap

Semrush Competitor Research в 2026 году прямо показывает prompts, где конкуренты получают mentions/citations, а бренд отсутствует. Это хороший raw signal, но одной отсутствующей строки недостаточно для рекомендации.

Prompt gap becomes meaningful if: buyer intent relevant + repeats across runs/time + competitor pattern persistent + topic strategically relevant.

13. Citation Gap

Brand mentioned: YES Owned/Mathchast citation: NO Competitor: cited repeatedly Possible causes: brand known from generic sources owned assets weak wrong format retrieval prefers third-party evidence.

14. Citation gap does not automatically mean «optimize page»

Иногда лучше создать independent evidence или получить external coverage, чем endlessly переписывать owned page.

15. Source Gap

Semrush описывает AI citation gap как ситуацию, где AI цитирует competitor pages, но не бренд. Profound Citation Pages/Watched Pages позволяют анализировать конкретные URLs и использовать их для content/outreach workflows.

SOURCE GAP: For target prompts: sources repeatedly cited when competitors win Classify: owned earned media reference community research competitor Mathchast Then: which source class is missing from client's footprint?

16. Format Gap

Topic: "enterprise migration" Client has: landing page blog explainer AI answers cite: implementation cases migration guides customer proof Gap: CASE Action: collect real migration case.

17. Evidence-type matrix

Buyer question	Сильный content format
«Как это работает?»	Explainer / Guide
«Кто реально делал?»	Case
«Есть ли данные?»	Research / Dataset
«Кто подтверждает?»	Expert / Interview / external source
«Что изменилось?»	News / Analysis
«Можно ли доверять факту?»	Verified profile / source correction

18. Entity Gap

Company has: publication But no: verified expert official domain product entity client relation in case Action: complete/verify entity graph Before: write more content.

19. Fact Gap

Profound FactCheck показывает современный action pattern: AI claim сравнивается с Knowledge Base, определяется inaccurate claim и конкретные citation URLs, после чего можно запустить action — outreach, update own content или creation.

«Матчасть» может сделать тот же цикл более строгим, используя verified Claims + Sources из документов 19–20.

20. Narrative Gap

Desired: "enterprise-ready" Observed: "small-business tool" Why? sources old positioning competitor comparison product history Action may be: case research owned update PR not pure opinion article.

21. Search Gap

Search Console: many impressions around topic Current page: weak CTR / wrong intent or no relevant page AI: same topic also weak → high-confidence cross-channel opportunity.

22. Reader Gap

Article: search traffic good Reader: high exit low second-click no company profile click Action: improve page/related links/CTA Not: publish duplicate article.

23. Gap sources

INPUT SOURCES ENTITY GRAPH claims experts products relations CONTENT GRAPH formats topics existing assets versions SEARCH PROOF queries impressions clicks index state AI VISIBILITY mentions recommendations citations competitors sources narratives accuracy READER reads second-click outbound actions EDITORIAL topic gaps quality format diversity COMMERCIAL client goals campaign capacity.

24. Recommendation is evidence graph

Каждая рекомендация должна хранить machine-readable reasons, а не только LLM-generated paragraph.

recommendation_id action_type: PUBLISH_CASE target_topic: AI_AGENT_SECURITY signals: PROMPT_GAP COMPETITOR_GAP FORMAT_GAP EXPERT_AVAILABLE evidence: 18 prompts 3 competitors 7 citation sources 0 existing cases CTO verified score: 82/100 internal status: PROPOSED

25. Клиенту не показывать «82/100» как истину

Internal priority score полезен для сортировки. Client-facing view лучше объясняет reasons: «Высокий приоритет — пробел устойчивый, тема коммерчески важная, конкуренты представлены, у вас нет case evidence».

26. Recommendation priority

CRITICAL HIGH MEDIUM LOW WATCH No: fake precision 83.7194.

27. Внутренний scoring нужен

Priority score = Gap Strength × Business Relevance × Evidence Quality × Actionability × Expected Information Gain × Freshness − Redundancy − Risk − Cost/Capacity.

28. Gap Strength

Signals: brand absent competitors present persistent over time multiple platforms large SoV difference repeated source pattern.

29. Business Relevance

Critical buyer topic high-intent client product priority revenue/product importance target market Not: random topic with high AI activity.

30. Evidence Quality

HIGH: locked cohort +30 persistent multiple platforms clear sources LOW: single run one obscure prompt ambiguous entity match.

31. Actionability

Can client produce: real case? expert? data? proof? If no evidence available: publishing priority drops.

32. Expected Information Gain

Очень полезный критерий: создаст ли действие новый публичный факт/evidence, которого сейчас действительно нет?

HIGH: real case metrics original research verified expert evidence new methodology LOW: generic opinion rewrite same landing page AI summary of existing article.

33. Redundancy penalty

Existing: 3 strong articles 2 cases research Recommendation: 4th generic explainer → strong penalty.

34. Risk penalty

regulated claim weak evidence legal risk commerciality site reputation abuse topic outside editorial scope → lower / manual review / reject.

35. Cost / Capacity

Research: high effort high info gain Case: medium Update profile: low Recommendation engine: can prefer low-cost fix when it solves same gap.

36. Simple MVP scoring model

0–5 each: Gap Strength 25% Business Relevance 25% Evidence Quality 15% Information Gain 15% Actionability 10% Freshness 5% Client Goal Fit 5% Subtract: Redundancy 0–20 Risk 0–30 Use: ranking only.

Weights are project hypotheses for pilot calibration.

37. Why scoring must remain explainable

Если recommendation появляется только потому, что embedding similarity = 0.83, редактор и клиент не смогут её проверить.

38. Rule engine first, ML later

MVP: deterministic/rule-based candidate generation + LLM explanation. Не строить black-box recommender до появления outcome dataset.

RULES: detect gaps score filter LLM: summarize evidence suggest angle draft brief HUMAN: approve strategy.

39. Почему не LLM-first

Prompt: "Что писать Acme дальше?" LLM alone: generic 20 ideas Mathchast engine: knows actual missing topic, competitors, sources, existing articles, facts and reader behavior.

40. Candidate generation

For each company: 1. inspect critical topics 2. inspect prompt gaps 3. inspect citation/source gaps 4. inspect content graph 5. inspect entity/fact gaps 6. generate candidate actions 7. deduplicate 8. score 9. policy filter 10. editorial review.

41. Topic universe

Company: product/service topics customer problems verified expertise buyer questions Search queries AI prompt topics editorial taxonomy.

42. Do not generate arbitrary topics outside company reality

«AI говорит про blockchain, значит SaaS-клиенту нужен blockchain article» — плохая рекомендация, если topic не связан с продуктом/expertise.

43. Topic relevance guard

Require: entity relation OR client goal OR verified expertise OR strong buyer relation Otherwise: reject as opportunistic.

44. Site-reputation abuse guard

Doc 12 становится обязательным policy filter: платный client demand не создаёт право публиковать нерелевантную тему ради host ranking signals.

45. Context Bridge interaction

New client topic: editorially relevant but far from current Mathchast graph Engine: CONTEXT_BRIDGE_REQUIRED Generate: 3–5 genuine supporting topics for editorial review Not: SEO camouflage.

46. Context Bridge as cluster recommendation

Primary: client case Support: editorial explainer industry research expert overview Measure: topic cluster not fake isolated effect.

47. Recommendation formats

Gap pattern	Recommended format
Competitors win on implementation/use-case prompts	Case
Category understanding weak	Explainer
Sources cite data/reports	Research
Expertise prompts weak	Expert opinion/interview
AI factual errors	Profile/source correction
Strong existing content, weak third-party evidence	External earned media

48. Case recommendation rule

Recommend CASE if: use-case/buyer prompt gap + company has real client/project evidence + no recent equivalent case + topic editorially relevant.

49. Research recommendation rule

Recommend RESEARCH if: AI/search questions need data + competitor/source landscape dominated by datasets + company/Mathchast can collect original data + information gain high.

50. Expert recommendation rule

Recommend EXPERT if: topic strongly connected to verified expert + AI answers use practitioner commentary + no current expert asset.

51. Explainer recommendation rule

Recommend EXPLAINER if: foundational buyer question + site/entity lacks clear definition + current content fragmented + no stronger case/data need.

52. Update-existing rule

If relevant strong page exists: do not create duplicate Check: freshness coverage source gaps structure facts target prompts → UPDATE_EXISTING.

53. Semrush lesson: optimize existing pages

Semrush Content Toolkit in 2026 explicitly analyzes drafts/pages against factors correlated with AI citations and issues recommendations for structure, clarity and entity signals. This reinforces a crucial rule for «Матчасти»: new publication is not always superior to improving a relevant existing asset.

54. Profound Pages lesson

Profound Pages now combines citation share, bot visits, page health, readability, freshness, structure, information density and machine readability, then generates optimization recommendations for the page. It also benchmarks pages against a wider network.

Наш difference: page optimization should sit behind reader/editorial quality and verified evidence, not become AEO score gaming.

55. Update vs New classifier

Existing page fit: HIGH freshness: LOW citation: LOW → UPDATE Existing fit: LOW / wrong intent → NEW ASSET.

56. Cannibalization / duplication guard

Before new content: search semantic neighbors same topic same intent same entity same format If overlap high: recommend update/merge.

57. No keyword-cannibalization superstition

Дублирование оцениваем по reader job/topic/intent и canonical content architecture, а не по мифическому правилу «два URL не могут содержать одно ключевое слово».

58. Source opportunity classification

Cited source is: OWNED → update own source MATHCHAST → improve/reuse existing asset EARNED_MEDIA → PR/outreach RESEARCH → produce/or earn inclusion in data COMMUNITY → maybe participate authentically COMPETITOR → create independent evidence not copy competitor.

59. Third-party opportunity

Semrush onboarding material отдельно выделяет third-party brand mention opportunities: места, где AI цитирует внешние источники и где бренд может быть недопредставлен.

Это нужно встроить как action family EARN_EXTERNAL_MEDIA, чтобы продукт не замыкал всю стратегию на mathchast.com.

60. Community source gap

Если AI часто цитирует Reddit/форумы, нельзя советовать клиенту создавать скрытые рекламные аккаунты и массово постить отзывы. Action может быть authentic community participation, support quality or no action.

61. Directory/reference gap

Wrong company details on authoritative directory/reference Action: correct/claim profile Not: new article.

62. FactCheck action tree

AI factual error ↓ source identified? YES: ├─ owned source outdated │ → UPDATE_CLIENT_SITE ├─ Mathchast wrong │ → CORRECT_MATHCHAST ├─ external source wrong │ → OUTREACH/CORRECTION └─ source accurate but AI wrong → strengthen corroboration / monitor NO: → build stronger verified source footprint.

63. «Ничего не делать» тоже valid action

Если gap слабый, sample нестабилен или business relevance низкая, system should recommend WAIT_AND_MEASURE.

64. Why this builds trust

Клиент видит, что engine не оптимизирует собственную выручку в каждом случае.

65. Recommendation eligibility gate

Before HIGH priority: data quality sufficient? gap persistent? business relevant? entity resolved? content inventory checked? policy allowed? action evidence available? capacity reasonable?

66. Persistence threshold

MVP: High priority gap should appear in more than one monitoring snapshot or be strong across multiple prompts/platforms. Exact threshold calibrated after pilots.

67. Cross-platform confirmation

ChatGPT gap only: MEDIUM ChatGPT + Gemini + Perplexity: stronger Unless: client says ChatGPT is only strategic platform.

68. Client goal weighting

Goal: enterprise security Security gap: boost Generic SMB pricing gap: lower But: goal cannot override editorial relevance/policy.

69. Commercial intent

"best tool for..." "X vs Y" "software for..." "how to choose..." higher buyer intent "history of..." "definition..." may be strategic but lower commercial.

70. Demand signal

На MVP demand может приходить из Search Console/search tools/client data. Не использовать synthetic «AI prompt volume» как будто это реальный traffic volume без licensed methodology.

71. Search + AI convergence

Один из сильнейших recommendation signals.

Search: topic receives impressions AI: client absent, competitors present Content: no asset Entity: expert available → HIGH priority publication.

72. Search strong, AI strong

Both already strong + content exists → no new publication unless freshness/fact issue.

73. Search weak, AI strong

Could mean: AI source ecosystem different Action: analyze citations before writing search article.

74. Search strong, AI weak

Candidate: content/source format gap AI citation gap entity clarity → inspect cited sources.

75. Reader signal in recommendation

Existing article: strong impressions weak meaningful read weak second-click Before new article: UPDATE_EXISTING or UX fix.

76. Reader strong, discovery weak

When people arrive: they engage But: search/AI visibility low → distribution/source/discovery action may have high value.

77. Recommendation evidence card

HIGH PRIORITY Publish a security case Why: • Client absent in 16/20 critical prompts • 3 competitors appear repeatedly • Cases are cited in 9/20 answers • No security case in your content graph • Verified CTO available • Search impressions for security cluster rising Expected information gain: HIGH [See evidence] [Create brief]

78. Avoid «Expected uplift +23%»

До накопления intervention outcome dataset нельзя прогнозировать точный uplift от recommendation.

79. Future outcome model

After hundreds of interventions: features → action → measured change Could estimate: historical success probability or typical observed range But: never before data exists.

80. Recommendation provenance

evidence_ids: AI observations Search query clusters content nodes entity claims source URLs reader metrics measurement reports.

81. Explainability

Любой analyst/editor должен открыть recommendation и восстановить, почему она появилась.

82. Candidate lifecycle

DETECTED → SCORED → REVIEW_REQUIRED → RECOMMENDED → ACCEPTED → BRIEF_CREATED → IN_PRODUCTION → PUBLISHED/FIXED → MEASURING → EVALUATED alternatives: DISMISSED DEFERRED INVALIDATED SUPERSEDED.

83. Recommendation expiry

Gap observed: Sep 1 Competitor disappears: Sep 20 Recommendation: re-evaluate Do not: keep selling stale opportunity forever.

84. TTL

High-volatility AI recommendations should automatically revalidate after 7–30 days depending data frequency.

85. Recommendation version

recommendation v1: Publish case new evidence: existing external case discovered v2: Earn external amplification / update profile relation History preserved.

86. Human review

Первые версии Next Best Publication обязательно проходят editor/strategist review.

87. Human reviewer checks

Is gap real? Does company have expertise? Does format fit? Is it redundant? Is evidence sufficient? Is topic allowed? Could a cheaper fix solve it? Would reader care?

88. Editor can override

Recommendation: PUBLISH_EXPLAINER Editor: OVERRIDE → CASE Reason: buyer needs proof, not definition Override becomes training/eval data.

89. Override reason codes

WRONG_TOPIC WRONG_FORMAT DUPLICATE INSUFFICIENT_EVIDENCE BETTER_EXISTING_UPDATE POLICY_RISK CLIENT_NOT_QUALIFIED LOW_READER_VALUE OUTDATED_GAP OTHER.

90. Feedback loop

Recommendation ↓ Editor decision ↓ Client decision ↓ Execution ↓ Before/After ↓ Observed outcome ↓ Recommendation evaluation.

91. What gets learned later

Which gap types: lead to accepted actions? produce useful content? get cited? improve targeted prompts? generate reader value? lead to repeat purchase?

92. Recommendation quality metrics

Metric	Что показывает
Editor acceptance rate	Candidate quality
Client acceptance rate	Business relevance
Execution rate	Action feasibility
Time to action	Workflow friction
Measured positive movement	Outcome association
Repeat purchase	Commercial value

93. Do not optimize only for client acceptance

Самая «продаваемая» рекомендация может быть editorially useless. Quality KPI must include editorial acceptance and information gain.

94. Recommendation false-positive rate

Dismissed as: not relevant duplicate bad evidence noise Track: by rule/model version.

95. Recommendation precision first

На старте лучше 3 сильных предложения, чем 40 «возможностей».

96. Client interface

NEXT ACTIONS 1. HIGH Security case [Why] [Create] 2. MEDIUM Correct outdated pricing source [Why] [Open task] 3. WATCH Enterprise integration topic Recheck in 14 days.

97. One primary CTA

Dashboard after +30 should show one strongest action, with «See all opportunities» secondary.

98. Why not listicle dashboard

Задача модуля — помочь принять решение, а не переложить сортировку 30 gaps обратно на клиента.

99. Recommendation detail

WHAT: Publish case TOPIC: AI agent security WHY NOW: persistent gap + rising demand EVIDENCE: prompts competitors sources search content inventory WHY THIS FORMAT: AI answers cite implementation proof WHAT WE NEED: real client/project metrics CTO comment security evidence EXPECTED: close information gap not guaranteed AI uplift [Create brief]

100. One-click brief

Profound уже предлагает Create Content Brief node, который использует citation analysis, answer patterns, target prompts, platform considerations and internal linking. «Матчасти» нужен похожий workflow, но grounded in our editorial formats and verified entity data.

Recommendation → Create Brief Auto-populate: reader question target prompt cluster gap evidence competitors source landscape entities required proof format angle suggested headings internal links measurement plan.

101. Brief ≠ article

Автоматический результат engine — в первую очередь research-backed brief. Статья создаётся следующим этапом через client/editor workflow.

102. Why this matters

Если engine сразу пишет 1 500 слов, он перескакивает через самый важный этап: получение реальных данных, кейса, эксперта или источников.

103. Brief skeleton

1. Reader job 2. Target topic 3. Why this is a gap 4. Target prompts 5. Existing client coverage 6. Competitor/source pattern 7. New information target 8. Required evidence 9. Recommended format 10. Questions for client/expert 11. Source plan 12. Internal links/entities 13. Distribution idea 14. Measurement plan.

104. New Information Target обязателен

Связь с Doc 28: briefing должен отвечать «какую новую полезную информацию эта публикация добавит в web?».

105. Weak brief

"Напишите статью: Почему AI важен бизнесу" No: new information specific gap evidence target reader.

106. Strong brief

"Case: как B2B SaaS внедрил agent security controls Why: client absent from 16/20 security prompts; competitors cited via implementation cases Need: architecture before/after incident reduction limitations CTO quote client confirmation Target: security consideration prompts."

107. Evidence request generated from gap

For CASE: client problem baseline implementation result measurement limitations confirmation For RESEARCH: sample methodology raw fields period sources.

108. Client feasibility check

Before checkout: Do you have: real case? YES/NO measurable result? YES/NO expert? YES/NO permission? YES/NO If no: choose another format/action.

109. This prevents low-quality paid content

Engine cannot recommend a «case» solely because case format performs well if client has no real case.

110. Commercial SKU mapping

Recommendation: PUBLISH_CASE Client can choose: Publish Edit + Publish Create + Publish Publish + Visibility Same strategic brief, different service level.

111. Don't auto-select highest-priced SKU

Service level depends on how much production help client needs, not on engine revenue optimization.

112. Recommendation → Order

ACCEPT → create draft → attach recommendation_id → choose service → payment/credit → production → publication → measurement plan → evaluate.

113. Recommendation attribution

publication: origin_recommendation_id Later: Before/After outcome linked back to recommendation.

114. This creates proprietary training data

Через время «Матчасть» будет знать не только «что AI цитирует», но и какие рекомендованные действия реально были выполнены и что после этого наблюдалось.

115. This is a stronger moat than AI writing

Generic LLM: can write text Mathchast dataset: gap → recommendation → execution → external outcome → repeat Harder to copy.

116. Search opportunity integration

Search Console: query cluster high impressions client no dedicated asset AI: competitors also win → priority up Search only: maybe update existing AI only: source/citation analysis first.

117. Search CTR opportunity

Existing page: high impressions low CTR Action: title/snippet/intent review Not: new article by default.

118. Search position opportunity

Average position is supporting signal only. Recommendation cannot be «write article because position 8.2» without intent/content/source analysis.

119. Reader opportunity

Research article: high saves/shares strong external citations Related topic: not covered → editorial expansion candidate.

120. Editorial recommendation engine

Этот же backend можно использовать не только для клиентов, но и для собственной редакции:

topic gaps search demand AI source gaps reader behavior entity gaps research opportunities → editorial backlog.

121. Commercial/editorial queues stay separate

Одна и та же opportunity может быть редакционно интересна, но клиент не должен автоматически получать право купить её как partner material.

122. Editorial route

Opportunity: important general topic Editor decides: Mathchast should cover independently → editorial assignment not commercial SKU.

123. Commercial route

Opportunity: specific client case with evidence and reader value → paid/contributed workflow with correct disclosure.

124. Free contributed route

Exceptional strong material selected by editor → free editorial/contributed not fake discount.

125. Contextual commercial conflict

Engine must know whether opportunity is better served by independent editorial research than client-sponsored content.

126. Example

Gap: market-wide pricing data missing Client wants: "research proving we're cheapest" Correct: Mathchast independent research or reject biased framing.

127. Topic ownership

topic owner/editor sees: client gaps editorial gaps source map content inventory Can merge: recommendations into broader research plan.

128. Small Business Pulse interaction

Field Snapshot data reveals: common SMB payment problem Reader/search/AI signals: high Action: Mathchast editorial research Not: sell every participant article.

129. Recommendation to small business participant

Possible: complete profile confirm source add current channel participate next survey Not automatically: buy publication.

130. Recommendation states by evidence

State	Meaning
RECOMMENDED	Enough evidence, actionable
WATCH	Interesting but insufficient/premature
NEEDS_DATA	Potential gap but client evidence missing
EDITORIAL_REVIEW	Topic/policy judgement required
NO_ACTION	No valuable intervention currently

131. «Needs data» UX

We see a possible case opportunity. To confirm, answer: • Do you have a real implementation? • Can the client relationship be disclosed? • Is there a measurable outcome? [Answer 3 questions]

132. Recommendation can trigger research, not sales

Это повышает precision before showing paid CTA.

133. Recommendation freshness

generated_at evidence_cutoff valid_until last_revalidated_at If data stale: recompute before checkout.

134. Data-source freshness gate

AI data stale 45d → no HIGH recommendation Search data delayed → note Entity fact expired → verify first.

135. Recommendation audit

why generated rules fired input snapshots score version LLM explanation version human override client decision.

136. Scoring versioning

NBA_SCORE_V1 effective Sep 2026 later V2: different weights Historical recommendations: retain V1.

137. Re-score old candidates?

Можно показывать current priority based on V2, но original recommendation remains auditable.

138. Recommendation confidence

HIGH EVIDENCE MEDIUM LOW Derived from: coverage persistence multi-source agreement entity certainty Not: opaque 93% probability.

139. Explain evidence quality

HIGH: +30 persistent 3 platforms 18 prompts clear source pattern LOW: single +7 snapshot one platform 4 prompts.

140. Recommendation status after client dismisses

Dismiss: not relevant no evidence not priority already planned too expensive wrong format other Use: product learning.

141. Snooze

Defer: 14d 30d 60d custom Revalidate at wake-up.

142. Client can mark «already doing this elsewhere»

Это important confounder and avoids duplicate work.

143. External execution tracking

Recommendation: EARN_EXTERNAL_MEDIA Client later adds URL: Media X article → mark executed externally → add watched URL → measurement continues.

144. Mathchast still provides value even without selling publication

Мониторинг и recommendation layer становятся standalone intelligence product.

145. Pricing implication

Basic Publish: simple next action hints Visibility package: full ranked opportunities Monitoring subscription: continuous Next Best Actions Agency: multi-client opportunity queue.

146. Free version

Можно показывать один teaser gap after basic audit, но evidence/detail/continuous monitoring paid. Не делать fake personalized «AI recommends 97 fixes» lead magnet.

147. Paid report ending

YOUR NEXT BEST ACTION Publish: Implementation case Topic: Security Why: 3 evidence bullets [Create brief] Alternative: Fix outdated pricing source.

148. Sales motion

Old: "Хотите ещё статью?" New: "У вас сохраняется конкретный security gap. Вот 18 prompts, конкуренты и sources. Мы можем закрыть его кейсом."

149. More credible upsell

Следующая покупка привязана к диагностике клиента, а не к календарю sales manager.

150. But no guaranteed closure

Правильная формулировка: «создать сильный источник по gap и измерить изменение». Неправильная: «закрыть gap = гарантированно попасть в ChatGPT».

151. Recommendation performance

accepted executed published first cited target mention delta recommendation delta search reader repeat Store outcome.

152. Success label for recommendation

EXECUTED MEASURED_POSITIVE MEASURED_MIXED MEASURED_FLAT MEASURED_NEGATIVE INSUFFICIENT_DATA No: "AI optimization succeeded" from single metric.

153. Learning dataset

gap_features action_type format topic company type evidence strength cost execution quality outcomes +7/+30/+60 → future recommender.

154. When ML becomes justified

После сотен/тысяч executed recommendations с measured outcomes, не после 20 клиентов.

155. Future ranking model

Could estimate: probability client accepts probability action feasible historical target-metric movement cost-adjusted information gain Still: policy/editorial constraints deterministic.

156. Never optimize for revenue alone

Objective function «expected order value» превратит recommendation system в sales spam.

157. Multi-objective future model

maximize: reader value client relevance information gain evidence quality measured outcome probability subject to: policy editorial quality risk capacity revenue: commercial constraint, not sole objective.

158. Agency Workspace future

20 clients Queue: 3 critical 9 high 14 medium Filters: client topic action type evidence campaign due → weekly agency planning.

159. Agency bulk brief

P2. Не массовая генерация статей, а bulk opportunity management with individual evidence.

160. API future

GET recommendations GET evidence POST accept/defer POST external execution URL No API: force publish override editorial policy.

161. Technical architecture

Data sources ↓ Feature builders ↓ Gap detectors ↓ Candidate actions ↓ Rule scoring ↓ Policy filters ↓ LLM explanation/brief ↓ Human review ↓ Client UI.

162. Feature builders

topic_visibility_features prompt_gap_features citation_gap_features source_gap_features content_inventory_features entity_gap_features search_features reader_features freshness_features.

163. Gap detector example

PROMPT_GAP rule: IF target_prompt = true AND client_presence < threshold AND competitor_presence > threshold AND observations >= minimum THEN candidate PROMPT_GAP.

164. Source Gap detector

IF competitor-winning answers share recurring citation domains AND client absent THEN candidate source opportunities Classify source type before action.

165. Format detector

Look at: cited pages content type answer language buyer job current inventory Infer: case/research/explainer/expert Human validates.

166. Content type detection caution

External URL format classification can be imperfect. Store classifier confidence and manually inspect top sources on launch.

167. Duplicate detector

Candidate topic → search internal graph → exact topic → semantic similarity → intent → format If strong existing asset: update candidate instead.

168. Entity eligibility

Case needs: company client/vendor relation evidence Expert article needs: verified person/topic relation Research needs: methodology/data capability.

169. Client capacity

Client has no spokesperson → don't recommend interview No case permission → don't recommend named case Can anonymize? editor decides.

170. Editorial capacity

High-effort research but newsroom overloaded Recommendation: DEFER or lower-effort valid action Do not sell impossible SLA.

171. Commercial inventory capacity

Connect to Doc 30 capacity units before checkout.

172. Recommendation UI should show effort

Effort: LOW / MEDIUM / HIGH Needs from you: 30 min case metrics expert interview source file This makes action realistic.

173. Estimated price

After strategy approved: service options Publish: 7 900 / current price Edit: ... Create: ... Never: price changes recommendation score.

174. Editorial-only recommendations

Some opportunity: assigned to Mathchast newsroom Client sees later: new editorial coverage if relevant Not: "you must sponsor it".

175. Recommendation integrity wall

Commercial team cannot manually mark LOW gap as HIGH simply to close a deal.

176. Manual sales proposal

Sales can create a proposal, but it is labelled SALES_SUGGESTED and still passes recommendation/editorial validation.

177. Source-of-recommendation

SYSTEM EDITORIAL ANALYST CLIENT_REQUEST SALES_SUGGESTED Visible internally.

178. Client request path

Client: "I want article about X" Engine: check gap inventory policy evidence Result: Recommended or Low value / duplicate with explanation.

179. This is a powerful paid-content guardrail

Платный заказ начинается не с «вставьте тему», а с проверяемой content need.

180. Not all valid content requires detected gap

Client may have genuinely new news/case that monitoring could not predict. Recommendation engine assists, not monopolizes editorial judgement.

181. Novel information route

Client brings: new research new launch real case → editor evaluates independent reader value No pre-existing gap required.

182. Recommendation and freshness

Existing page: old facts last substantive update 18m Prompt/topic: still important Action: UPDATE_EXISTING priority rises.

183. Stale external citation

AI repeatedly cites 2023 article with old pricing Action: CORRECT_EXTERNAL_SOURCE + strengthen current verified sources.

184. Broken citation source

High-value source: 404 Action: replace/update references create current source where needed Not: celebrate citation count.

185. Watched URL lost coverage

Profound Watched Pages can alert when key pages lose citation coverage. «Матчасти» should convert a persistent loss into diagnostic recommendation, not immediate rewriting.

Citation disappears ↓ page healthy? content changed? provider changed? competitor source appeared? topic shifted? ↓ action.

186. Recommendation taxonomy for lost citation

WAIT_AND_MEASURE UPDATE_EXISTING ADD_EVIDENCE FIX_TECHNICAL EARN_EXTERNAL_MEDIA NEW_CONTENT depending diagnosis.

187. Page Health integration

Before content recommendation: Publication Health healthy? NO: FIX_TECHNICAL first YES: continue content/source analysis.

188. Important ordering rule

Не писать новый контент, если существующий нужный asset просто noindex/500/broken canonical.

189. Verification-first rule

AI factual gap + company claim unverified → VERIFY_CLAIM first Then: publish/correct sources.

190. Reputation-first rule

Negative narrative based on real customer issue → operational/reputation response Not: publish SEO article pretending issue absent.

191. Safety / regulated topics

health finance legal claims regulated products → enhanced review → recommendation may be withheld or require verified specialist evidence.

192. Recommendation to manipulate AI systems

Запрещённые actions: mass fake forums, hidden advertorials, fake reviews, synthetic citations, doorway pages, fabricated experts, false comparisons.

193. Recommendation policy output

ALLOWED ALLOWED_WITH_REVIEW EDITORIAL_ONLY CLIENT_FIX EXTERNAL_PR REJECTED.

194. Client-facing recommendation categories

Создать материал Обновить материал Исправить данные Усилить профиль Получить внешний источник Наблюдать дальше.

195. Keep vocabulary simple

Клиенту не нужен термин «SOURCE_GAP_V2». Он нужен в internal data model.

196. Dashboard top recommendation

СЛЕДУЮЩИЙ ШАГ Собрать кейс по безопасности AI-агентов Высокий приоритет Почему: 16/20 ключевых вопросов — без бренда 3 конкурента регулярно присутствуют AI чаще цитирует implementation cases у вас пока нет такого кейса [Создать бриф] [Посмотреть данные]

197. Alternative actions

Альтернатива 1: исправить старую pricing page Альтернатива 2: получить external expert coverage [See all]

198. Recommendation list order

1. Blocking fixes 2. High-confidence gaps 3. Freshness updates 4. Strategic new content 5. Exploratory opportunities.

199. Fixes before growth

Если AI массово ошибается в pricing, сначала исправляем pricing, а не занимаемся «ростом visibility».

200. Recommendation report section

WHAT TO DO NEXT 1. Fix inaccurate pricing source 2. Publish security implementation case 3. Watch enterprise integrations topic Each: why effort evidence measurement target.

201. Measurement target generated automatically

Recommendation: Security case Measurement: 15 security prompts Recommendation Rate Mathchast Citation Rate Watched URL +7/+30/+60 This becomes: Doc 34 measurement plan.

202. Recommendation closes the loop technically

Doc 33: detect metrics Doc 34: measure change Doc 35: choose next intervention → repeat.

203. Recommendation quality before monetization

Первые 20–30 clients: recommendations можно генерировать internally/manual first, проверяя precision, прежде чем делать glossy automated feature.

204. Concierge MVP

System: candidate gaps Analyst: reviews Editor: chooses action Client: receives 1–3 recommendations Product logs: all decisions.

205. Why concierge first

быстрее понять реальные action patterns;
не автоматизировать неверную taxonomy;
собрать override dataset;
понять, какие recommendations клиенты действительно выполняют;
снизить риск self-serving upsells.

206. P0 rule engine

Rules: missing first case missing expert persistent prompt gap citation gap fact error technical blocker stale asset source gap Simple: high precision.

207. P1 engine

topic/source clustering format inference cross-channel scoring persistent gaps automatic briefing client goal weighting revalidation.

208. P2 engine

historical outcome model expected action value content cluster planning agency prioritization external source recommendations fact accuracy actions.

209. P3 engine

learned ranking cost-adjusted intervention policy historical probability bands cross-client anonymized priors advanced causal feedback.

210. Что не входит в MVP

Не строим	Почему
Black-box ML recommender	Нет outcome dataset
100 recommendations/client	Noise, not decision support
Exact predicted uplift	Нет empirical basis
Every gap → paid article	Destroy trust
Automatic article publication	Bypasses evidence/editorial workflow
Fake external source actions	Policy/reputation risk
SEO/AEO score gaming	Wrong optimization target

211. P0 data model

recommendation_candidates recommendations recommendation_evidence recommendation_actions recommendation_scores recommendation_reviews recommendation_feedback recommendation_executions recommendation_outcomes.

212. Recommendation record

id company_id action_type topic_id format priority evidence_state score_version generated_at valid_until status reviewer client_decision origin_report_id.

213. Evidence record

type source_object_id metric value window platform/topic description weight freshness.

214. Action execution

execution_type publication_id external_url profile_change_id claim_id started_at completed_at.

215. Outcome

measurement_plan_id +7 result +30 result +60 result search result reader result AI result overall state.

216. Recommendation engine events

MeasurementReportReady → RecomputeRecommendations PublicationPublished → invalidate duplicates → schedule measurement ClaimCorrected → resolve FACT_GAP ExternalMediaAdded → re-evaluate SOURCE_GAP.

217. Recompute strategy

Event-driven: major new evidence Periodic: weekly/monthly for active monitoring Not: re-score every second.

218. Cache expensive features

source clusters content embeddings topic overlap competitor patterns Reuse across: recommendations.

219. No Neo4j requirement

Doc 19 remains valid: PostgreSQL relations + vector candidate matching are enough for MVP recommendation graph.

220. LLM role

GOOD: summarize evidence classify external content propose angles draft brief explain recommendation NOT SOURCE OF TRUTH: priority itself company facts metric counts policy decision.

221. Recommendation evals

Gold cases: analyst/editor chosen action Evaluate: gap classification format choice duplicate detection reason correctness policy compliance brief usefulness.

222. Target eval metric

Top-3 recommendation precision editor acceptance wrong-action rate duplicate rate policy violation rate.

223. Client value metric

% reports where client: accepts or meaningfully acts on ≥1 recommendation.

224. Commercial retention metric

Report viewed → recommendation opened → brief created → order started → next publication within 60–90 days.

225. This can become north-star loop

Активная компания завершила measurement loop и совершила следующий осмысленный action.

226. Don't overcount free fixes

Meaningful next action: publication profile correction verified expert external source monitoring continuation Track: commercial and non-commercial separately.

227. Recommendation-to-order conversion

Useful commercial KPI, but not recommendation quality KPI alone.

228. Strategic moat

Generic media: sells next placement Generic AI tracker: shows next gap Mathchast: knows gap + owns publication workflow + verified facts + measures result + learns from intervention.

229. Why this can justify recurring subscription later

Monthly: new AI/search data new competitor moves new source gaps new facts new opportunities → continuous decision layer.

230. Monitoring subscription promise

«Каждый месяц вы получаете не просто графики, а приоритетный список того, что реально стоит исправить, опубликовать или усилить следующим.»

231. Avoid recommendation fatigue

Max: 1 primary 2–4 secondary watchlist Archive: resolved/stale.

232. Weekly digest

This week: 1 new high-priority gap 1 resolved 2 watching No: 20 unchanged recommendations.

233. Alert vs recommendation

ALERT: something changed now RECOMMENDATION: what to do Example: Alert: old price appears in AI Recommendation: update pricing page + source correction.

234. Project/workspace pattern

Profound Projects is a useful contemporary benchmark: weekly Aim agent suggestions become collaborative Projects with tasks, chat and reusable artifacts. «Матчасти» не нужно копировать generic project-management suite, но recommendation should become a persistent work object, not ephemeral notification.

235. Recommendation work object

Recommendation ├─ evidence ├─ discussion ├─ assigned owner ├─ task/brief ├─ status ├─ execution └─ outcome.

236. No Asana clone

Статусы и комментарии нужны только вокруг recommendation/publication workflow.

237. Client team assignment

Assign: Content Manager Expert Agency Owner Due date optional for accepted action.

238. Expert request

Brief needs CTO input → invite verified CTO → answer structured questions → feed publication draft.

239. Data request

Research recommendation: need survey sample → don't create order immediately → data feasibility task first.

240. Recommendation economics

Low-cost fix: profile update vs: 24.9k full article Engine: choose based on gap not revenue.

241. Why economically still good

Даже бесплатная correction reinforces trust and keeps client in monitoring; strong trust can increase long-term LTV more than forcing one unnecessary article.

242. Sales dashboard

Qualified opportunities: client gap priority action evidence commercial eligibility last contact Sales sees: only reviewed recommendations.

243. Sales cannot see confidential editorial notes unnecessarily

Role boundaries from docs 14/29 remain.

244. Recommendation and discounts

Priority does not create automatic discount. Price is determined by SKU/pricing version.

245. Recommendation and editorial pick

Recommendation never promises homepage/editorial promotion.

246. Recommendation and AI guarantee

Recommendation never says «закроет gap». It says «addresses the identified information gap; result will be measured».

247. Public methodology

/methodology/recommendations Explain: data inputs gap types priority human review limitations no guaranteed uplift commercial/editorial separation.

248. Why public methodology matters

Поскольку recommendation directly leads to potential purchase, клиент должен понимать, что система не просто придумывает upsell.

249. Conflict-of-interest disclosure

Стоит публично сказать: «Матчасть может рекомендовать собственные publishing services, но engine также может рекомендовать update, external media, correction or no action.»

250. Recommendation disclaimer

Recommendations are based on: observed data + current content graph. They do not guarantee: search ranking AI mention citation traffic lead revenue.

251. Launch workflow

First 10 monitored clients: manual analyst recommendations 20–50: rule candidates + human review 50–100: automatic client suggestions after review 100+: learned prioritization experiments.

252. Pilot questions

Did recommendation surprise client? Was evidence credible? Could client act? Was format correct? Was cheaper fix possible? Did editor agree? Did client execute? What changed after +30?

253. Pilot success criteria

≥70% editor acceptance ≥40% client meaningful-action rate low duplicate/noise rate no policy violations at least several measured repeat loops.

Targets are initial product hypotheses.

254. Recommendation quality before automation

Если editor acceptance ниже 50%, автоматизацию нужно улучшать, а не показывать больше recommendations.

255. Example full journey

ACME Baseline: Security Recommendation Rate 5% Gap: competitors 35% case sources dominate Recommendation: Security implementation case Client: accepts Brief: real implementation metrics CTO limitations Publication: Sep 4 +30: Recommendation Rate 17% Mathchast URL cited in 4 prompts Next: external earned-media source gap Action: PR/outreach not duplicate article.

256. Example fact journey

AI says: free plan exists FactCheck: wrong Source: client pricing page stale Recommendation: update client site Client fixes: Sep 5 +30: incorrect claim drops from 8 to 2 observations Next: monitor no article needed.

257. Example no-action journey

One prompt: brand missing Other 39: healthy No persistent pattern Recommendation: WATCH 14 days No spend.

258. Example update-existing journey

Prompt gap: migration guide Existing: strong 2024 guide but outdated Recommendation: UPDATE_EXISTING add current implementation data sources expert quote Not: new duplicate URL.

259. Example research journey

Source gap: AI answers rely on thin surveys Mathchast editorial sees: original dataset opportunity Recommendation: independent research 50 agencies Result: research asset participant entities citations new topic hub.

260. This connects commercial and editorial moat

Some of the best gaps will become own editorial research, strengthening the whole domain rather than only one client.

261. MVP product screen

NEXT BEST ACTION [HIGH] Create case AI agent security Why: ▸ 16/20 target prompts absent ▸ competitors appear in 72% of runs ▸ 9 answers cite case-type evidence ▸ no equivalent client case exists Needs: real implementation metric expert client confirmation Measure: +7/+30/+60 Recommendation + Citation [Create brief] [Defer] [Not relevant]

262. MVP engine scope

P0: 7–10 gap rules content inventory topic map AI prompt gaps competitor gaps citation/source gaps fact/profile gaps technical blocker rule update-vs-new simple explainable priority human review one-click brief feedback states link to measurement plan.

263. P1

cross-channel scoring format inference persistent gap detection external-media actions source classification goal weighting client feasibility forms weekly recompute recommendation digest agency queue basics.

264. P2

historical outcome priors Fact Accuracy actions content cluster planning advanced duplicate detection cost/capacity optimization external execution tracking expected outcome ranges API.

265. P3

learned ranking multi-objective recommendation model portfolio optimization causal feedback cross-client benchmark priors automated experiment design.

266. Что НЕ строить

Не строим	Почему
Generic AI idea generator	Не является moat
Automatic 100-topic content plan	Noise/scaled-content risk
Black-box 0–100 opportunity score	Псевдоточность
Guaranteed uplift predictor	Нет данных и causal basis
Every recommendation = Mathchast sale	Conflict of interest
Auto-publish recommendation	Bypasses editorial/evidence
Fake community/PR actions	Reputation/policy risk

267. Decision

Утвердить Next Best Publication как главный repeat engine «Матчасти», но реализовать backend как Next Best Action. Система использует Search Proof, AI Visibility, Entity Graph, existing content, reader behavior, client goals и editorial policy; обнаруживает Topic, Prompt, Citation, Source, Format, Entity, Fact, Narrative, Search и Reader gaps; затем выбирает действие из publish/update/fix/verify/external media/wait/no action. На MVP candidate generation и scoring детерминированы и объяснимы, LLM используется для summarization и brief creation, а human editor/analyst подтверждает recommendation. Каждая recommendation показывает evidence, effort, required proof и заранее связанный measurement target. Новая публикация рекомендуется только после проверки существующего контента, technical health, verified expertise и более дешёвых alternatives. Recommendation engine не оптимизирует только revenue и имеет публичную conflict-of-interest methodology. После исполнения action результат измеряется через Mathchast_34 и возвращается в recommendation dataset. Это создаёт основную долгосрочную петлю: Verify → Publish → Measure → Diagnose → Act → Measure again.

268. Что этот документ разблокирует

Mathchast_35 Next Best Publication → Mathchast_36 reputation/reviews/media portfolio → Mathchast_37 agency workspace → Mathchast_38 sales/GTM first 100 → Mathchast_40 technical architecture → Mathchast_43 MVP scope & roadmap

Источники исследования

Gap taxonomy, action taxonomy, priority formula, weights, evidence thresholds, recommendation lifecycle, client UX, conflict-of-interest guardrails, P0/P1/P2/P3 scope and pilot acceptance targets are project decisions of «Матчасть». Competitor products are used as evidence that the market is moving from visibility dashboards toward actionable opportunity workflows; «Матчасть» should differentiate through verified entities, its own publication workflow, explicit evidence and closed-loop outcome measurement.