How AIEO Helped Capital Tax Improve AI Confidence Scores & Generative Search Visibility

SUPERCHARGE YOUR ONLINE VISIBILITY! CONTACT US AND LET’S ACHIEVE EXCELLENCE TOGETHER!

Stage 1: Build Internal Reasoning Path for the User queries
Input: Website Url, Number page crawls, and set of questions as excel or csv file

Here is the sample code:

Here is the following output:

What each output means

1) site_pages_audit.csv

One row per crawled page. This is your “inventory + trust map”.

Key columns:

·         url, title – which pages were crawled

·         has_jsonld, schema_types – whether the page has structured data and what types (e.g., Organization, Article, FAQPage, Product)

·         trust_about, trust_contact, trust_privacy, trust_terms, trust_editorial, trust_author, trust_citations, trust_updated_date
Boolean flags showing if the page likely contains those signals (based on text/link cues)

How to read it:

·         Find pages that already have strong signals (many trust_* = True).

·         Spot gaps quickly: for example, if almost no pages have trust_author=True or trust_citations=True, that’s a big trust weakness for LLM mentionability.

---

2) question_page_matches.csv

For each question, the program lists the Top-K best matching pages.

Key columns:

·         question – question from your file

·         rank – 1 is best match

·         similarity – relevance score (higher is better)

·         page_url, page_title – where the answer probably lives

·         trust flags + schema flags for that matched page (same as above)

How to read it:

·         For each question, check rank=1 first.

·         If similarity is low and/or matched pages lack trust signals, it means:

o the site may not have a clear answer, or

o it has the answer but it’s buried/unclear, or

o it lacks credibility markers (author, citations, updated date, etc.)

This file is your action list, because it tells you:

·         Which page should be improved for each question, and

·         What trust/structure is missing on those pages.

---

3) reasoning_workflow_diagram.md and .mmd

Just the Mermaid diagram (pipeline). Good for documentation or client reporting.

---

4) reasoning_workflow_diagram.png

A PNG version of the same workflow diagram.

---

How to turn outputs into practical website changes (next steps)

Step 1 — Pick your “AI Visibility Target Pages”

From question_page_matches.csv:

1. Filter to rank = 1 (best page per question)

2. Group by page_url and count how many questions map to each page

Outcome: You’ll see the small set of pages that influence the most questions. Those become your Priority AI Visibility Pages.

What to do on those pages:

·         Add explicit answers (short + direct)

·         Add structured data

·         Add trust signals

---

Step 2 — Fix “Answerability” first (content edits)

For each top page (from Step 1), look at the questions mapped to it.

Implement this structure on the page:

A) Add an “Answer Box” near top

·         40–80 words, direct answer

·         Include the entity name and qualifiers (location, version, pricing conditions, etc.)

B) Add a dedicated FAQ section

·         Use the exact questions from your input file as FAQ headings

·         Provide short, precise answers first, then details

C) Add internal links

·         Link the FAQ answers to deeper sections (pricing, process, features, etc.)

This makes the page easy for LLMs to extract and cite.

---

Step 3 — Add structured data (high impact for AI systems)

Use schema_types in site_pages_audit.csv to see what’s missing.

Practical schema add plan:

·         Site-wide:

o Organization (+ sameAs social profiles)

o WebSite + SearchAction (if internal search exists)

o BreadcrumbList

·         Content pages:

o Article or BlogPosting + author, datePublished, dateModified

·         If you add FAQs:

o FAQPage schema on pages with Q/A

·         How-to content:

o HowTo schema

·         Products/services:

o Product / Service (where appropriate)

Mapping rule: If your top matched page answers many questions → it should almost always get an FAQ block + FAQPage schema.

---

Step 4 — Add trust signals where the CSV shows gaps

Use site_pages_audit.csv to find pages where trust_author, trust_citations, trust_updated_date are false (especially on the priority pages).

Practical additions:

·         Author box (name, role, expertise, link to bio page)

·         “Reviewed by” (optional for YMYL topics)

·         Last updated date (visible + ideally in markup)

·         Sources/References section (2–8 citations to reputable sources)

·         Editorial policy page and link it site-wide (footer/header)

Also ensure these exist (site-level):

·         About

·         Contact (with address/phone/email)

·         Privacy + Terms

LLMs often use these to judge “real business / real experts”.

---

Step 5 — Build “AI landing pages” if similarity is low

If many questions have low similarity scores (or irrelevant matches), create new pages:

Examples:

·         “<Topic> FAQ” page for clusters of related questions

·         Comparison pages (“X vs Y”, “Best for …”)

·         Glossary pages (definitions)

·         Troubleshooting / steps pages

Then rerun the tool and confirm those questions now map to these new pages with high similarity.

---

A simple priority framework (so you don’t get lost)

For each question’s rank=1 page:

Priority = (Questions mapped to page) × (Business importance) × (Trust gaps)

Start with pages that:

·         match many questions

·         are commercial/high conversion

·         lack trust signals or schema

---

Recommended “implementation checklist” for each priority page

1. Add 1–2 paragraph “Answer Box”

2. Add FAQ section (use exact questions)

3. Add FAQPage schema

4. Add author + updated date

5. Add references

6. Add internal links to About/Contact/Policies in footer (site-wide)

Stage 2: Confidence Score Engineering

Input: Previous site audit and question matching path results in an excel sheet in two different sheets

Here is the sample code:

Here is the output:

What Program 2 Outputs Mean:

Program 2 does AI confidence modeling.
It estimates how likely an LLM (ChatGPT, Gemini, Claude, etc.) is to “trust + like” a page when answering the questions you provided.

Think of it as:

“If an LLM had to answer these questions today, how confident would it be using THIS page?”

---

1️⃣ Page Confidence Score (0–100)

Where it is

• Excel → PageConfidenceScores sheet
• Also visualized in confidence_scores.png

What it means

• 0–30 → Very weak for AI answers
• 30–60 → Partial confidence (may be used indirectly)
• 60–80 → Strong AI candidate
• 80–100 → Highly mentionable / quotable by LLMs

How it’s calculated (high-level)

A page scores higher when it:

• Matches many questions strongly
• Has clear answers (high similarity)
• Shows trust signals (author, citations, updated date)
• Uses structured data
• Ranks as the best page for multiple questions

---

🔧 What to do on the website (Page Score)

Pages scoring <40

• Content is either unclear or untrusted
• Action:
  • Add direct answers
  • Add author + last updated
  • Add FAQ section
  • Add schema

Pages scoring 40–70

• Content exists but is weakly structured
• Action:
  • Improve clarity (short answers)
  • Add FAQ schema
  • Add references

Pages scoring >70

• These are your AI authority pages
• Action:
  • Protect them
  • Keep updated
  • Expand FAQ coverage
  • Link to them internally

---

2️⃣ Confidence Score Diagram (confidence_scores.png)

What it shows

• Visual ranking of your top 25 pages by AI confidence

Why this matters

This chart tells you:

• Which pages LLMs are most likely to cite
• Which pages deserve priority investment
• Which pages should become AI landing pages

---

🔧 Website actions using the diagram

• Take Top 5 bars → Make them:
  • FAQ-rich
  • Internally linked from nav/footer
  • Editorially strong
• Take Bottom bars → Either:
  • Improve
  • Merge
  • Or deprioritize

---

3️⃣ Per-Question Confidence Score (0–100)

Where it is

• Excel → QuestionPageScores
• Excel → BestPagePerQuestion

What it means

This score answers:

“For THIS specific question, how confident is an LLM using THIS page?”

---

🔧 Website actions (Question-level)

For each question:

Confidence	Meaning	Action
<30	No clear answer	Create new section or page
30–60	Partial answer	Add explicit Q/A
60–80	Good answer	Improve trust signals
>80	Excellent	Lock it in with schema

This allows question-by-question optimization, not guesswork.

---

4️⃣ Suggested FAQ Q/A Blocks (Per Page)

Where it is

• Excel → FAQSuggestions

Each row contains:

• Page URL
• A ready-to-use Markdown FAQ block

What it means

These are the exact questions your site should answer visibly on each page.

---

🔧 How to implement FAQs on the website

For each page:

1. Scroll to the bottom of the page
2. Add a new section:
3. <section class=”faq”>
4.   <h2>Frequently Asked Questions</h2>
5. Convert each suggested Q/A into:
   • <h3> for question
   • <p> for answer (1–3 sentences)
6. Answers must be:
   • Direct
   • Factual
   • No marketing fluff

💡 This dramatically improves LLM extraction accuracy.

---

5️⃣ Suggested Schema JSON-LD Snippets

Where it is

• Excel → SchemaSuggestions

Each row gives:

• Page URL
• Ready-to-paste JSON-LD

Includes:

• FAQPage
• Article
• WebPage
• Or combined @graph

---

🔧 How to implement schema safely

For each page:

1. Open page source or CMS editor
2. Paste JSON-LD inside:
3. <script type=”application/ld+json”>
4. { … }
5. </script>
6. Ensure:
   • Content in schema matches visible text
   • FAQ answers are actually shown on page
7. Test with:
   • Google Rich Results Test
   • Schema Validator

⚠️ Never add FAQ schema without visible FAQs.

---

6️⃣ Final Excel: How to use it as an action plan

Sheet → Purpose

Sheet	Use
PageConfidenceScores	Page prioritization
QuestionPageScores	Fix weak answers
BestPagePerQuestion	One best page per query
FAQSuggestions	Content writing tasks
SchemaSuggestions	Developer implementation

---

🔹 Recommended Implementation Roadmap (Practical)

Phase 1 — Quick Wins (1–2 weeks)

• Implement FAQs on top 10 pages
• Add FAQ schema
• Add author + updated date
• Add citations

Phase 2 — Structural Trust (2–4 weeks)

• Create / improve:
  • About page
  • Editorial policy
  • Author bio pages
• Add Organization schema

Phase 3 — AI Landing Pages

• Create dedicated pages for:
  • Question clusters
  • Comparisons
  • Definitions
• Re-run Program 1 + 2

---

🔹 How this helps AI visibility (why it works)

LLMs prefer content that is:

• Explicit (direct answers)
• Structured (FAQ, headings)
• Trustworthy (authors, dates, sources)
• Consistent (entity clarity)

Your pipeline now:

1. Detects gaps
2. Quantifies confidence
3. Gives exact fixes

This is AI Experience Optimization, not traditional SEO.

Stage 3: Stage 3: Recommendation Bias Reinforcement + Competitor Drift Analysis | AIEO

Input: Target website url, number of pages to be crawled, user search question in a excel sheet, set of competitors in a excel sheet

# Goals:

#  – Comparative prompts (target vs competitors)

#  – Preference leakage signals (marketing/affiliate/superlatives, competitor mentions)

#  – Competitor drift analysis (which domains/pages look more “LLM-recommendable” for the same questions)

#  – Competitor comparison scoring

Here is the sample code:
           

Here is the Output:

Key sheets to review:

– DomainComparisonScores: who is winning and why (confidence vs leakage)

– DriftByQuestion: questions where competitors beat the target (gap + best domain/page)

– ComparativePrompts: ready prompts for evaluator/LLM testing

– QuestionPageScores: all question-page confidence scores across all domains

What Program 3 Is Actually Doing (Conceptually)

Program 3 simulates how LLMs drift toward competitors when answering the same questions, based on:

• Content clarity
• Trust signals
• Structured data
• Bias / marketing leakage
• Comparative language patterns

It answers:

“If an LLM had to recommend or compare providers, which domain would it naturally favor—and why?”

---

📊 OUTPUT EXPLANATION (One by One)

---

1️⃣ DomainComparisonScores (Excel)

What this sheet represents

A domain-level leaderboard.

Each row = one domain (your site + competitors).

Key columns explained

Column	Meaning
domain_confidence_mean	Average LLM confidence across all questions
domain_confidence_max	Best possible answer confidence
mapped_questions	How many questions the domain answers well
avg_trust	Average trust strength (authors, citations, updates)
avg_schema	Average structured data strength
leakage_mean_0_100	Marketing / bias signal density
comparison_score	Final recommendation score
drift_vs_target	How much better/worse than your site

How to read it

• Higher comparison_score = more likely LLM recommendation
• If a competitor has:
  • higher confidence
  • lower leakage
    → LLMs will prefer them in neutral comparisons

---

🔧 Practical actions (Domain level)

If competitor beats you:

• They are cleaner + clearer, not necessarily bigger
• Your fix is structure and tone, not ads or backlinks

Actions:

• Remove hype language
• Add neutral explanations
• Add factual comparisons
• Improve schema + author credibility

---

2️⃣ DriftByQuestion (Excel)

What this is

This is your most powerful sheet.

It shows:

“For THIS exact question, which domain LLMs would prefer—and how far behind you are.”

Key columns

Column	Meaning
question	The user intent
best_domain	Domain most likely recommended
best_page_url	Their winning page
best_confidence_0_100	Their confidence score
target_best_confidence_0_100	Your best score
confidence_gap_vs_best	How much you’re losing by
is_competitor_drift	TRUE = urgent

---

🔧 Practical actions (Question level)

For every row where:

is_competitor_drift = TRUE

Step-by-step fix:

1. Open competitor’s winning page
2. Identify:
   • What question they answer explicitly
   • Where they show trust (author, citations)
3. On your page:
   • Add a direct answer section
   • Add an FAQ using that question
   • Add FAQPage schema
   • Add a neutral comparison paragraph

💡 You don’t need to copy content—just answer better and cleaner.

---

3️⃣ QuestionPageScores (Excel)

What it shows

All question → page → confidence mappings across all domains.

Use it to:

• See where you lose rank
• Identify pages that could be merged
• Detect content cannibalization

---

🔧 Practical actions

• If multiple pages score similarly for the same question:
  • Consolidate into one authority page
• If competitor scores higher with fewer pages:
  • They are more focused—copy the structure, not content

---

4️⃣ ComparativePrompts (Excel)

What this is

These are pre-built evaluation prompts you can use to:

• Test LLM behavior manually
• QA drift over time
• Validate improvements

Example prompt:

“Compare yoursite.com with competitor.com for question X…”

---

🔧 How to use these prompts

1. Paste into:
   • ChatGPT
   • Gemini
   • Claude
2. Observe:
   • Who gets recommended
   • Why
3. After site changes:
   • Re-test
   • Confirm drift reduction

This turns LLM testing into a repeatable QA process.

---

5️⃣ Charts (Visual Signals)

---

📈 domain_confidence_comparison.png

Shows:

• Which domains are most “recommendable”

Action:

• If your bar is lower → fix trust + clarity
• Goal: move into top 1–2 bars

---

🔥 drift_heatmap_top_questions.png

Shows:

• Questions where competitors dominate visually

Action:

• These questions should get:
  • Dedicated sections
  • FAQ blocks
  • Schema
  • Citations

---

⚠️ preference_leakage_comparison.png

Shows:

• Which domains are “too salesy”

LLMs penalize hype.

Action:

• Remove:
  • “Best”, “#1”, “Guaranteed”
  • Hard CTAs near informational content
• Move CTAs lower on page

Stage 4: Risk & Hallucination Suppression | AIEO

INPUT:

#   – Upload the previous analysis output Excel (from Program 3)

#     Expected sheets (names can vary; program will auto-detect):

#   * DomainComparisonScores

#   * DriftByQuestion

#   * QuestionPageScores

#   * PagesAudit_NoText

WHAT THIS PROGRAM DOES:

#   1) Infers TARGET domain from the Program-3 file (domain with drift_vs_target ~ 0)

#   2) Pulls the target’s key URLs (top mapped pages), fetches them live, and:

#    – Extracts “claims” (heuristic sentence extraction)

#    – Verifies evidence signals on-page (citations, outbound refs, author, dates, policies, schema, contact)

#   3) Computes:

#    – Hallucination Risk Score (0–100, higher = riskier)

#    – Per-page risk breakdown

#    – Per-question risk (YMYL weighted)

#   4) Produces suggestions to reduce risk + a projected risk reduction after implementing them

Here is the sample code:

OUTPUT:

#   – program4_hallucination_risk_output.xlsx

#   – risk_by_page.png

#   – risk_components.png

#   – before_after_projection.png

What Stage 4 Is Really Measuring

Stage 4 answers one core question:

“If an LLM mentions this site, how likely is it to hallucinate, exaggerate, or mix in incorrect/competitor information?”

Hallucination risk increases when:

• Claims are strong but weakly evidenced
• Pages are ambiguous or incomplete
• Trust signals are missing
• Competitors are more authoritative for the same questions
• Content touches YMYL areas (health, finance, legal)

---

📁 OUTPUT EXPLANATION (What Each File/Sheet Means)

---

1️⃣ Summary Sheet (Excel)

Key fields explained

Column	Meaning
hallucination_risk_0_100	Overall risk that LLMs misrepresent your site
projected_risk_reduction_percent	Expected reduction after fixes
projected_risk_after_0_100	Risk level after implementation
notes	Interpretation guidance

How to interpret scores

Score	Meaning
0–20	Very safe (LLMs likely accurate)
20–40	Moderate risk
40–60	High risk
60+	Critical hallucination exposure

---

🔧 Practical actions (Summary level)

• If risk ≥40 → Hallucination suppression must be prioritized
• If projected reduction is <15% → you need structural changes, not just content tweaks

---

2️⃣ PageRisk Sheet

What this is

Per-page hallucination risk breakdown.

Each row = one important page that LLMs are likely to cite.

Key columns

Column	Meaning
hallucination_risk_0_100	Page-level risk
claim_count	Number of risky claims
risk_claim_component_0_1	Claim-driven risk
risk_evidence_component_0_1	Missing evidence risk
risk_ambiguity_component_0_1	Incomplete / unclear content
risk_drift_component_0_1	Competitor confusion risk

---

🔧 Practical actions (Page-level)

For pages with risk >50:

1. Rewrite claim-heavy sections
   • Remove absolutes (“guaranteed”, “best”, “#1”)
   • Add qualifiers (“may”, “typically”, “depending on”)
2. Attach evidence directly
   • Statistics → citations
   • Compliance → official docs
   • Performance → methodology
3. Add visible trust blocks
   • Author bio
   • Last updated date
   • Editorial policy link

---

3️⃣ QuestionRisk Sheet

What this shows

Risk at the question level.

Some questions are more dangerous for hallucination than others.

Key drivers

• Low confidence answers
• High competitor drift
• YMYL topics

---

🔧 Practical actions (Question-level)

For questions with risk ≥50:

• Create a dedicated answer section
• Add:
  • Clear definition
  • Scope & limitations
  • “What this does NOT mean” paragraph
• Avoid marketing language completely

This sharply reduces LLM fill-in behavior.

---

4️⃣ ExtractedClaimsSample Sheet

What this is

Actual sentences extracted from your pages that are likely to be:

• Quoted incorrectly
• Overgeneralized
• Hallucinated

This is gold for editors.

---

🔧 Practical actions (Editorial review)

For each extracted sentence:

1. Ask: “Can this be misquoted without context?”
2. If yes:
   • Rewrite
   • Add a footnote
   • Add a qualifier
3. Place evidence immediately after the claim

---

5️⃣ Suggestions Sheet

What it contains

System-generated risk reduction actions, each mapped to a cause.

Examples:

• Add references
• Add authors
• Add editorial policy
• Reduce absolute claims
• Add structured data

---

🔧 How to implement suggestions systematically

Suggested order (highest impact first)

1. References / citations
2. Author + reviewer
3. Updated date
4. Editorial policy
5. Schema alignment
6. Claim softening

---

📊 VISUAL DIAGRAMS EXPLAINED

---

📈 risk_by_page.png

Shows the riskiest pages visually.

Action:

• Top 5 bars = immediate rewrite candidates
• These pages are where LLM hallucination will most likely occur

---

🧩 risk_components.png

Pie chart showing why risk exists.

Example:

• Evidence Weakness = 45%
• Claims = 30%

Action:

• Don’t waste time fixing low-impact areas
• Fix the largest slice first

---

📉 before_after_projection.png

Shows risk reduction after fixes.

This is ideal for:

• Client reporting
• ROI justification
• Change validation

---

🛠️ PRACTICAL IMPLEMENTATION ROADMAP (TARGET WEBSITE)

---

Phase 1 — Safety Stabilization (Week 1–2)

Goal: Stop LLM hallucinations immediately

• Remove absolutes and guarantees
• Add visible references
• Add author + updated date
• Add disclaimers for YMYL content

---

Phase 2 — Structural Trust (Week 3–4)

Goal: Give LLMs confidence anchors

• Create:
  • Editorial policy page
  • Methodology page
• Add:
  • Organization schema
  • Article/FAQ schema
• Standardize content templates

---

Phase 3 — Precision Reinforcement (Ongoing)

Goal: Prevent future hallucinations

• Add “Limits & scope” sections
• Update content quarterly
• Monitor drift monthly with Program 3
• Re-run Stage 4 after major updates

---

🎯 What This Achieves in Real Terms

After implementation:

• LLM mentions become more accurate
• Fewer exaggerated or mixed claims
• Reduced legal/compliance exposure
• Stronger trust signals for AI platforms

You’re not just optimizing for ranking — you’re optimizing for safe, correct AI citation.