Exploring SEO and AI Search: The Questions Keeping Me Awake

I can’t help but feel restless as I ponder the evolving landscapes of SEO and AI search. Treating ChatGPT like Google seems like a recipe for failure in today’s world of RAG, reranking, and probabilistic systems.

As someone engulfed in SEO for years, I’ve tried to relate each new technology to the tools I know well.

Remember the buzz around “mobile SEO” when mobile search surged or when “voice search optimization” became the new must-know with voice assistants?

In my journey, I once thought I had Google all figured out. That belief shattered after examining how ChatGPT selects citations, analyzing Perplexity’s ranking process, and digging into Google’s AI Overview criteria.

I’m not claiming that SEO is obsolete or that we’ve encountered a total paradigm shift. I want to share the lingering questions that suggest we might need to fundamentally alter our methods of understanding.

These questions have emerged from months of intense analysis of AI search systems, documented observations of ChatGPT’s behavior, and reverse-engineering Perplexity’s ranking factors.

The Questions That Won’t Let Me Sleep

The questions reflecting on AI’s complexities have dismantled much of what I once confidently believed about search optimization.

When Math Doesn’t Add Up

While I grasp PageRank and link equity, encountering Reciprocal Rank Fusion in ChatGPT’s code led to moments of realization where I comprehended my gaps:

• Why does RRF prefer consistency over singular excellence in query results? Is securing the #4 spot across multiple queries superior to achieving #1 once?
• How do vector embeddings alter semantic distance from conventional keyword matching? Are we striving for semantic intent or mere words?
• Why does temperature=0.7 cause unpredictable rankings? Are repeated tests now mandatory?
• How do cross-encoder rerankers approach query-document pairs versus PageRank? Is now the time to shift towards real-time relevance?

These questions echo traditional SEO concepts but seem rooted in entirely different mathematical frameworks when juxtaposed with LLMs. Or are they?

When Scale Feels Unbreachable

While Google indexes trillions, ChatGPT retrieves a measly 38-65 results. This stark 99.999% reduction leads to pressing inquiries that linger:

• Why does ChatGPT retrieve so few results compared to Google’s billions? Is this a short-term anomaly or a foundational shift?
• How do token limits imbuing rigid confines differ from traditional search’s freedom? When did search results shrink in their dimensionality?
• Does the k=60 constant in RRF conceal a ceiling on visibility? Has position 61 supplanted the secondary page?

Are these mere modern-day constraints? Or do they signal a novel information retrieval ideology?

The Questions that Continue to Haunt Me

Here are 101 questions that persist, gnawing at what I believed I knew about SEO in the AI era:

1. Is OpenAI employing CTR for citation rankings?
2. Does AI perceive our page layout as Google does or focus just on text?
3. Should our writing gear towards shorter paragraphs for AI to digest content adeptly?
4. Can interaction metrics like scroll depth or mouse movement influence AI ranking signals?
5. What is the effect of low bounce rates on our citation potential?
6. Could session data like reading order prompt AI model rerankings?
7. How might a nascent brand integrate into offline training data to earn visibility?
8. What strategies optimize a web/product page for probabilistic systems?
9. Why do citations transform inexplicably?
10. Is running multiple tests necessary to gauge variance?
11. How can Google’s “blue links” aid in acquiring specific answers to long-form questions?
12. Do LLMs mirror the same reranking algorithms?
13. Does web_search act as a binary switch or a probabilistic trigger?
14. Should our focus pivot to accolades or citations?
15. Is reranking deterministic or stochastic?
16. Do Google and LLMs utilize identical embedding models, and if so, what’s the corpus variance?
17. Which pages garner maximal requests by LLMs and maximum visits by users?
18. Should we monitor drift post-model updates?
19. Why does EEAT manipulate seamlessly in LLMs contrary to traditional Google search?
20. Who among us amplified traffic tenfold post-Google algorithm revelation?
21. Why does the answer structure morph even within a mere day’s interval?
22. Could post-click engagement amplify our odds of inclusion?
23. Is session memory gearing citation bias towards preliminary sources?
24. Why inherently are LLMs more prone to bias than Google?
25. Does offering a downloadable dataset escalate citation potential?
26. Why does content in Turkish retain anachronistic data despite contemporary queries?
27. Are vector embeddings capturing semantic difference distinctly from keyword associations?
28. Should we master LLMs’ “temperature” value henceforth?
29. How can a modest website emerge in ChatGPT or Perplexity answers?
30. What events unfold if our entire site optimizes solely for LLM targeting?
31. Might AI agents evaluate images alongside pages at an instant, or simply focus on surrounding text?
32. How could we ascertain AI tools leveraging our content?
33. Could AI models quote a lone sentence from our blog posts?
34. How do we ensure AI comprehends our business purpose?
35. What differentiates pages showing in Perplexity or ChatGPT but absent from Google?
36. Does AI preference newer content over steadfast, older references?
37. Once retrieved, how might AI rerank content?
38. Could LLMs retain our brand voice enveloping their outputs?
39. Is there a mechanism enabling AI summaries with direct links to our pages?
40. Can we monitor when our content is quoted without linked acknowledgment?
41. Can we identify prompts or themes fostering additional citations?
42. What shifts when monthly client SEO reports rebrand as “AI Visibility AEO/GEO Reports”?
43. Is there a facility to estimate brand mentions within AI results akin to search volume metrics?
44. Could Cloudflare logs reveal AI bot exposure to our domain?
45. Do schema changes measurably affect AI mentions?
46. Do AI agents recall our brand after initial interactions?
47. How might a local business with a mapping result amplify visibility within LLMs?
48. Might Google AI Overviews and ChatGPT responses share similar signals?
49. Can AI establish trust metrics for our domain over temporal spans?
50. Why should query fanouts prioritize visibility for several concurrent queries, and why do AI models fabricate synthetic responses?
51. How frequently do AI systems recalibrate their understanding of our site? Are there search algorithm updates involved?
52. For LLMs, does freshness remain sitewide or page-specific?
53. Could form submissions or downloads signal content quality?
54. Do internal links facilitate more rapid bot navigation across sites?
55. How does semantic affinity between content and queries affect ranking?
56. Can nearly identical pages vie within the same embedding cluster?
57. Do internal links enhance a page’s ranking signals for AI evaluation?
58. How are high-confidence passages distinguished during reranking?
59. Does content freshness eclipse trust when conflict arises?
60. How many rerank layers precede AI’s citation selections?
61. Might extensively cited paragraphs bolster sitewide trust scores?
62. Could model updates reset pre-existing reranking preferences while retaining partial memory?
63. Why are traditional queries often more definitive sans AI hallucinations?
64. Who in the system dictates final citation preferences?
65. Can human feedback loops reshape LLM source rankings?
66. When might an AI initiate midpoint searches amid answers, and why are multiple continuous AI searches within a single chat window observed?
67. Does being once-cited predispose future citation allocation? Can LLM ranks sustain visibility likened to Google’s top 10?
68. Do frequent citations autonomously elevate a domain’s retrieval priority?
69. Are user clicks on linked sources embedded in feedback signals?
70. Do Google and LLMs employ identical deduplication protocols?
71. Might citation velocity be traced akin to SEO link velocity?
72. Will LLMs someday curate a lasting “citation graph” paralleling Google’s link constructs?
73. Do LLMs correlate brands entwined in related subjects or question clusters?
74. How long elapses before repeated interactions etch into durable brand memory within LLMs?
75. Why doesn’t Google reveal 404s while LLM responses do?
76. Why fabricate citations while Google directs only to accessible URLs?
77. Do LLM retraining phases present reset opportunities post-visibility slump?
78. How should we construct recovery roadmaps against AI model misinformation?
79. Why might some LLMs cite while others disregard?
80. Are ChatGPT and Perplexity leveraging identical web data repositories?
81. Do OpenAI and Anthropic gauge trust and freshness identically?
82. Do source-specific limits apply to maximum AI citations per response?
83. How shall we verify citation following content evolution?
84. What’s the simplest route to trace prompt-level visibility over extended periods?
85. How can we persuade LLMs to regard our assertions as factual?
86. Does a topic-aligned video linked to the page fortify cross-format grounding?
87. Could identical questions lead to divergent brand suggestions for differing users?
88. Might LLMs register previous brand engagements?
89. Can previous click histories skew subsequent LLM endorsements?
90. How do retrieval and reasoning converge on citation attributions?
91. Why does ChatGPT retrieve 38-65 outputs while Google spans billions?
92. How do cross-encoder rerankers diverge from PageRank in query-document evaluations?
93. How does a backlink-void site surpass authorities within LLM result sets?
94. Why impose token barriers absent in conventional search?
95. Why does LLM temperature determination yield erratic rankings?
96. Does OpenAI allocate a dedicated crawl budget to web properties?
97. Do Knowledge Graph recognition and LLM token embedding methods diverge?
98. How is crawl-index-serve distinct from retrieve-rerank-generate dynamics?
99. Do temperature settings in LLMs generate inconsistent rankings?
100. Why is tokenization integral?
101. How does a knowledge cutoff induce unintentional blind spots versus real-time crawling dynamics?

When Trust Turns Probabilistic

I grapple with how Google reliably links to tangible URLs while AI systems, astoundingly, can fabricate information:

• Why might LLMs fabricate citations while Google anchors existing URLs?
• How do hallucination rates of 3-27% stand against Google’s 404 incidence?
• Why do similar queries yield conflicting “facts” in AI over search indices?
• How does obsolete data prevail in Turkish content despite contemporary inquiries?

Are we orienting ourselves around systems liable to mislead users? How does one manage that eventuality?

Where We Stand

I’m not suggesting AI search optimization/AEO/GEO is utterly unlike SEO. Yet, I confront 100+ unanswered questions challenging my foundational SEO acumen at this moment.

Perhaps solutions await folks with more advanced insights. For now, I remain entwined in seeking answers but know these queries will persist, with brand new ones arising on the horizon.

The mechanisms generating these queries aren’t vanishing. We must engage, scrutinize, and potentially innovate approaches to fathom and leverage them.

The victors in this novel expanse won’t inevitably own the totality of wisdom. But they will bravely ask, probe, and identify workable solutions amid ambiguity.

This article originally appeared on metehan.ai (as 100+ Questions That Show AEO/GEO Is Different Than SEO) and has been republished with permission.

Inspired by this post on Search Engine Land.