Metehan YESILYURT - From Rankings to Citations: How LLMs Actually Choose Sources - Salon du Search Marketing Paris 30/01/2026

Bonjour France, Paris!

Transcript

1. From Rankings to Citations How LLMs Actually Choose Sources Metehan Yesilyurt metehan.ai
2. The Big Shift in Search Traditional Search You search → You read → You decide AI-Powered Search AI searches → AI reads → AI decides → AI cites The question: How do you get cited? Metehan Yesilyurt - FePSeM Paris 2026
3. Les LLMs sont une boîte noire. Même quand ça marche. Metehan Yesilyurt - FePSeM Paris 2026
4. Credit: iPullRank
5. The Stakes ChatGPT 900M+ weekly users 1B+ queries // STILL FAR BEHIND OF THE GOOGLE SEARCH, HYPE Google AI Overviews 30 to 50%+ of searches Rapidly expanding Perplexity Growing AI-native search Metehan Yesilyurt - FePSeM Paris 2026
6. Not Every Prompt Triggers Search Knowledge-Only Training data only, no web search "What is photosynthesis?" Search-Triggered Retrieves fresh web data "Best smartphones 2026" Hybrid Combines both sources "Compare React vs Vue in 2026" Search Triggers: Time-sensitive, specific facts, current events Metehan Yesilyurt - FePSeM Paris 2026
7. Metehan Yesilyurt - FePSeM Paris 2026
8. The Dependency Revelation ChatGPT Relies on Google Data?!!! OpenAI uses SCRAPERS TO FETCH & RERANK Google search results Confirmed by The Information investigation (Aug 2025) SerpAPI listed OpenAI as customer (removed May 2024) Also serves Meta, Apple, Perplexity Google attempted to block SerpAPI crawler The Irony: ChatGPT competes with Google while depending on Google's index Metehan Yesilyurt - FePSeM Paris 2026
9. RESONEO Discovery ChatGPT Uses Google Shopping Exclusive Finding (2025) • ChatGPT shopping results scraped from Google Shopping • Uses SCRAPER service (similar to SerpAPI) • Product IDs match Google Merchant Center exactly • Named Entity Recognition uses Google Shopping Graph Implication: Optimize Google Merchant Center = Better ChatGPT Shopping visibility Source: RESONEO / think.resoneo.com Metehan Yesilyurt - FePSeM Paris 2026
12. Three Types of Query Fanouts ALSO THANKS TO OLIVIER DE SEGONZAC 1 Search Fan-Out Text queries • 1-4 per conversation IN AVG • Up to 20 in thinking mode • Traditional web search 2 Shopping Fan-Out Product queries • 1-3 per conversation IN AVG • Shorter, product-focused • Google Shopping (SearchApi.io) 3 *VISUAL Fan-Out (GOOGLE) Visual content • 5-10 per conversation • Bing or proprietary index • High volume queries Not all searches use fanouts! Simple factual queries = single search Metehan Yesilyurt - FePSeM Paris 2026
13. Single Query vs Multiple Fanouts Single Query (No Fanout) • "GDP of France" → Direct answer • "iPhone 17 price" → One search • "Capital of Japan" → Simple lookup Simple, factual lookups Multiple Fanouts (Complex) "Best project management tools" Generates 3-5 variations: 1. "project management 2025" 2. "team collaboration tools" 3. "Asana vs Monday comparison" 4. "PM software reviews" Then aggregates with RRF Metehan Yesilyurt - FePSeM Paris 2026
14. The Retrieval Window ChatGPT Retrieval Pattern 38-65 sources per search (varies by query complexity) Rank 1-20 High probability Rank 21-40 HIGH & MODERATE probability (RERANKING) Rank 41-65 Low probability (DEEP RESEARCH CAN CHANGE THIS) Rank 66+ PROBABLY Not retrieved = Zero chance The Math: If you're not in the retrieval window, RRF score = 0 Metehan Yesilyurt - FePSeM Paris 2026
15. The ACTUAL Retrieval Window Metehan Yesilyurt - FePSeM Paris 2026
16. A SMALL TOOL FOR POTENTIAL AI’S TRAINING DATA FOR DOMAINS Metehan Yesilyurt - FePSeM Paris 2026 lınk -> https://webgraph.metehan.aı
17. FOR TRAINING DATA Metehan Yesilyurt - FePSeM Paris 2026 lınk -> https://webgraph.metehan.aı
18. Grounding Explained Grounding = Citations Required Without Grounding ✗ Model's training data only ✗ No sources cited ✗ Can hallucinate ✗ "Based on my knowledge..." With Grounding ✓ Must reference retrieved URLs ✓ Citations mandatory ✓ Facts tied to sources ✓ "According to [Source]..." Why It Matters: Grounded responses = Traffic opportunity Metehan Yesilyurt - FePSeM Paris 2026
19. The Complete Pipeline 1 Prompt Classification Search needed? 2 Query CLASSIFICATION Single or fanouts? 3 Data Source SerpAPI / SearchApi / Bing 4 Retrieval 38-200 sources per query* 5 Scoring/INTENT CLASSIFICATION RRF or HYBRID 6 Filtering Top X RESULTS 7 LLM Synthesis Grounded answer 8 Citations clickable sources * If DEEP RESEARCH IS IN PROGRESS, IT MAY FETCH UP TO 200 SOURCES Metehan Yesilyurt - FePSeM Paris 2026
20. ChatGPT's RRF System (CONSIDER QUERY FANOUTS) Reciprocal Rank Fusion (RRF) Score = 1 / (60 + rank) • Combines multiple fanout rankings mathematically • k=60 discovered via reverse engineering • Each query variation contributes to total score Key Insight: Multiple moderate rankings beat one #1 ranking Metehan Yesilyurt - FePSeM Paris 2026
21. QUICK WIN ACTION STEPS - NOW! CONTENT FRESHNESS STILL WORKS, LLMS HAVE RECENCY BIAS, UPDATE YOUR POSTS/PAGES OLDER THAN 6 months chatgpt ıs usıng longer fanouts, run base prompts multıple tımes CHECK FOR LLM CONSISTENCY SIMILAR/SAME PROMPTS ON GOOGLE AI MODE, PERPLEXITY, MISTRAL SAVE FANOUTS, RUN ANALYSIS, IDENTIFY TOPIC GAPS Metehan Yesilyurt - FePSeM Paris 2026 CHECK REDDIT CITATIONS, LOOK FOR USERS QUESTIONS, IDENTIFY GAPS NICHE DOWN AS POSSIBLE, CREATE PROMPTS FOR DIFFERENT BUYER PERSONAS, THEY LIKELY PROMPTING; “I’m thinking of, fınd me, search for me x,y,z”
22. Metehan Yesilyurt - FePSeM Paris 2026 A QUICK RESEARCH ON CLAUDE, I ASKED BUYER PERSONAS FOR Fepsem
23. Metehan Yesilyurt - FePSeM Paris 2026
24. Google AI Mode's 7 Signals Discovery Engine Reveals: 1 Base Ranking Core algorithm score 2 Gecko Embedding similarity 3 Jetstream Context understanding 4 BM25 Keyword frequency 5 PCTR Engagement (3-tier) 6 Freshness Time-sensitive boost 7 Boost/Bury Manual adjustments Technical: 500-token chunks with heading hierarchy preserved Metehan Yesilyurt - FePSeM Paris 2026
25. Chunking: The Industry Mystery The Truth About Chunking No industry consensus on optimal chunk size. No published case studies. What We Know: • Google: 500 tokens (Discovery Engine) • Chunks preserve heading hierarchy • Different for each LLM • No proven "best" size What Actually Matters: • Good for UX regardless of SEO • Clear section breaks help humans • Logical content organization • Scannable structure PrACTICAL Approach: Structure in ~400-word sections with clear H2/H3 headings Metehan Yesilyurt - FePSeM Paris 2026
26. Most Cited Domains in ChatGPT Top Domains (2025 Data) // IT CHANGES DYNAMICALLY Pattern: Authority + Freshness + User Engagement = Citations Metehan Yesilyurt - FePSeM Paris 2026
27. OPENAI PARTNERSHIPS Metehan Yesilyurt - FePSeM Paris 2026
28. ChatGPT Shopping Shopping Research GPT-5 mini trained for shopping Researches products, asks questions Provides comprehensive buyer's guide Instant Checkout Buy without leaving ChatGPT Shopify, Etsy, Walmart Seamless in-chat purchase How It Works: 1. User: "Find quiet vacuum for small apartment" 2. ChatGPT: Researches, asks clarifying questions 3. User: Reviews buyer's guide → Clicks "Buy" → Completes in-chat Data Source: Google Shopping (via SearchApi.io) Metehan Yesilyurt - FePSeM Paris 2026
29. ChatGPT Ads (Coming 2026) Status: Internal testing, launching February 2026 Ad Format • Bottom of answers • Clearly labeled 'Sponsored' • Click to chat with brand • Conversational (revolutionary!) Pricing Model • ~$60 CPM (vs Meta $20-25) • Impression-based • $1M minimum commitment • Early access required Will NOT • Influence ChatGPT answers • Show to users under 18 • Appear in health/politics • Sell user conversation data New Revenue Channel: Projected $25B by 2029 Metehan Yesilyurt - FePSeM Paris 2026
30. The Math Behind Multiple Rankings Strategy A Rank #1 for ONE query Position #1 → RRF: 0.0164 Not ranking for related queries Total Score: 0.0164 Strategy B Rank #5-8 for FOUR queries Query 1, Position #5 → 0.0154 Query 2, Position #6 → 0.0152 Query 3, Position #7 → 0.0149 Query 4, Position #8 → 0.0147 Total Score: 0.0602 Strategy B wins by 367% Lesson: Coverage beats perfection Metehan Yesilyurt - FePSeM Paris 2026
31. Optimize Your Google Feeds New Priority: Google Merchant Center Since ChatGPT scrapes Google Shopping: Complete product data - All attributes filled High-quality images - Multiple angles Accurate pricing - Real-time updates Rich descriptions - Natural language Schema markup - Product structured data Double Win: Google Shopping + ChatGPT Shopping visibility Metehan Yesilyurt - FePSeM Paris 2026
32. Topic Clusters Strategy Why Clusters Win: • Each page ranks for different fanout variations • Hub gains authority from cluster links • Combined RRF across all pages • 5 pages at #10 > 1 page at #1 Hub Page "Complete Email Marketing Guide" Cluster Pages • Email Marketing Tools 2026 • Email Copywriting Tips • Automation Workflows • Metrics & Analytics • GDPR Compliance Metehan Yesilyurt - FePSeM Paris 2026
33. Optimization Checklist Content Answer in first 100 words Clear HEADING hierarchy (~400 words) Cover semantic variations Include data, stats, examples Technical Core Web Vitals optimized Google Merchant Center complete Schema markup implemented (useful for Google) Mobile-first responsive Authority Quality backlinks Topic cluster architecture Regular content updates Industry citations Metehan Yesilyurt - FePSeM Paris 2026
34. ANY QUESTIONS? metehan.ai @metehan777 on X linkedin.com/in/metehanyesilyurt