AI tools make mistakes: what you can learn from hallucinations

Hallucinations are an inherent part of AI answers, so they are worth your focus.
Discover how to find and fix them and how to prevent them as much as possible in this 20 minute talk given at Search 'n Stuff in Antalya 2026.

Transcript

1. AI tools make mistakes What you can learn from hallucinations Search ‘n Stuff Antalya
2. Hi 👋 Pieter Serraris Strategy Director at Connect with me Pieter Serraris
3. Hi 👋 Pieter Serraris Strategy Director at Connect with me Pieter Serraris
4. Why should I care about hallucinations?
5. Hallucinations are an inherent part of how AI models work
6. “ AI is far from a high-precision technology. If AI is too precise, it is considered flawed because it is memorising instead of learning United Nations University Source
7. They are not exceptional 23% Brand-related questions: at least 1 factual error
8. Not every model is the same Prompt: “Find the original article, publication and URL for this news item.”
9. Not every language is the same 15-35% more frequently in non-English languages than in English.
10. Not every language is the same 15-35% more frequently in non-English languages than in English. Training imbalance
11. Not every language is the same 15-35% more frequently in non-English languages than in English. Training imbalance Token cost
12. Where do hallucinations originate?
13. Hallucinations = LLM filling gaps?
14. Sometimes, but there is more to it.
15. Google Research: almost no more “gaps” +95% saturation Source: https://research.google/blog/empty-shelves-or-lost-keys-recall-is-the-bottleneck-for-parametric-factuality/
16. Hallucinations can develop anywhere in the pipeline
17. Pre training Training data Encoding Parametric memory
18. Cut off date = gaps Training data Encoding Parametric memory Knowledge cut off: - GPT 6 (Astra): April 2026 - Claude Fable: January 2026
19. From vague prompts … Training data Encoding Parametric memory
20. To constraint loss Training data Encoding Parametric memory
21. Model “recalls” facts Training data Encoding Parametric memory Parametric recall
22. Failure to “recall” facts Training data Encoding Parametric memory Parametric recall Models fail to recall 26–34% of the facts -> “Reversal curse”
23. The “Reversal Curse” A -> B
24. The “Reversal Curse” A -> B B -> A ?
25. Issues during retrieval Training data Encoding Parametric memory Parametric recall Retrieval • retrieval staleness • attribute transfer ◦ Entity ◦ Locale
26. Locale attribute transfer
27. Locale attribute transfer
28. Locale attribute transfer Products NOT SOLD in the US.
29. Chunking the data Training data Encoding Parametric memory Chunk 1 Parametric recall Retrieval Chunk 2 Chunk 3 Chunk 4 Chunk 5
30. Chunking process Training data Encoding Parametric memory Chunk 1 Parametric recall Retrieval Chunk 2 Chunk 3 Chunk 4 Chunk 5 • Chunking process • Lost in the middle
31. Document/chunk reranking Training data Encoding Parametric memory Chunk 1 Parametric recall Retrieval Chunk 2 Chunk 3 Chunk 4 Chunk 5
32. Memory vs retrieval Chunk > memory Chunk > retrieval “Authority over recency” Retrieval overrules memory (Petrovic: “Primary bias”)
33. Memory vs retrieval Chunk > memory Chunk > retrieval “Authority over recency” Retrieval overrules memory (Petrovic: “Primary bias”) Depends on the confidence in the initial response (via measuring token probabilities)
34. Memory vs retrieval Prompt date: 24/8/2026
35. Memory vs retrieval Prompt date: 24/8/2026
36. Memory vs retrieval Prompt date: 24/8/2026
37. … Next steps? 1: discover & fix 2: prevent
38. STEP 1 6 ways to find & fix fabrications
39. 1. Citation verification “Open” every cited URL
40. 1. Citation verification ISSUE FOUND HOW TO FIX Dead links (404) cited - Old link (Wayback machine): fix or redirect - Fabrication: create content (and redirect) Claim not found on URL Fix or contradict the incorrect claim in the first 100 words
41. 2. Cross-check multiple LLMs “What are the main products and services of [BRAND X]?” Answer Answer = Answer ≠ Answer = Answer ≠
42. Cross check matrix … ChatGPT Gemini Claude Perplexity Product 1 V V V V Product 2 V V X V Product 3 V V V V Product 4 X V X X Incorrect Product V X X X Delivery time - 72 h 24 h - partner exclusive exclusive “works with” 10 y 10 + 2 y 30 y 30 & 10 + 2 Parent / daughter company Warranty
43. 2. Cross-check multiple LLMs ISSUE FOUND HOW TO FIX “Row” error: ≠ mistakes Clarify this statement throughout your website “Row” error: = mistake Find and fix the source that created this mistake “Column” error Compare the sources of the mistaken model with sources of correct models
44. 3. Parametric vs grounded > Dan Petrovic “What are the main products and services of [BRAND X]?” Search on No search Answer ≠ Answer
45. 3. Parametric vs grounded ISSUE FOUND HOW TO FIX Mistake w/ Search on Fix the retrieved/grounded source(s). Mistake w/ Search off Try to influence training (long term), check training bot access.
46. 4. Locale cross checking “What are the main products and services of [BRAND X]?” Belgium Netherlands USA France Answer Answer Answer Answer
47. 4. Locale cross-check ISSUE FOUND HOW TO FIX One country mistake(s) Analyze localisation of content, other CMS?, … All country mistake(s) Check crawlability, JS-usage, … across the entire website infrastructure
48. 5. Data-driven analysis AI-Bot visits - Log files, Cloudflare Analytics, Clarity, … Referral Traffic - GA4: 404s in AI Assistant channel group
49. 5. Data-driven analysis ISSUE FOUND HOW TO FIX Training bots 403: - Check bot access 404: - Old URL (Wayback machine, …): redirect - Fabricated URL: content gap Search and User bots Referral clicks 404s +1 visit: higher urgency -> redirect
50. 5. Data-driven - log files 404 AI bots: +3K visits to 404
51. 5. Data-driven - log files In Wayback Machine (> training data) 404
52. 5. Data-driven - log files Fabricating the same directory in +4 TLD’s - .gt/sourcing .co.cr/sourcing .ee/sourcing .ba/sourcing
53. 5. Data-driven - log files “Sourcing” = Content gap
54. 5. Data-driven - GA4 … Session Source: AI Assistant
55. 6. Centralized logging WHAT? Keep track of all fabrications WHO? You / your colleagues / your client and their team WHERE? Centralized document (excel)
56. STEP 2 Prevent them when possible
57. By design, hallucinations cannot be ruled out completely
58. But you can prevent them with …
59. … good SEO - Structured data No client side JS Disambiguation & NAP consistency Localisation instead of translation Crawlability …
60. About Us page optimization Default owned source for brand information for AI tools
61. Get your Wikis in order
62. New industry suggestions Most promising so far: entitymap -> / Dixon Jones
63. Idea makes sense, results seem promising
64. The benefits of hallucination monitoring
65. Benefits Gives leverage with clients & stakeholders - Big fixes & technical tickets - Business decisions beyond SEO It moves the conversation from rankings & impressions to brand safety
66. Benefits Gives leverage with clients & stakeholders Helps me understand the “SERP” - You can’t do SEO without opening Google - You can’t optimize for AI-search without opening an AI model
67. Benefits Gives leverage with clients & stakeholders Helps me understand the “SERP” Great addition to AI visibility measuring - vs. prompt tracking (narrowing, polluting) - vs. log file monitoring (complex, modelling)
68. What can you do on Monday?
69. Monday: what we want to do? Connect with me Pieter Serraris
70. Monday: what we should do? • One research method • Analyze one prompt manually • Find & understand the hallucinations • Start the conversation Connect with me Pieter Serraris
71. Thank you! Connect with me Pieter Serraris Search ‘n Stuff Antalya
72. Quick afterthought. In case we have a minute left…
73. Why NOT “hallucinations” 1. Anthropomorphism 2. Misleading Scale AI is NOT human It’s not an exception, it’s the rule Suggestions: Fabrication OR system error
74. Sources … Worth reading: - … - Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality (https://arxiv.org/pdf/2602.14080) - https://seranking.com/blog/hallucinated-urls-chatgpt/ - Across languages? https://aclanthology.org/2025.emnlp-main.1481.pdf -