Retrievable Beats Rankable: Build for Retrieval-Based AI

AI answers depend on retrieval. If systems can’t reliably extract your facts, definitions, comparisons, and proof, you will not appear—even if your pages are “good.” This session shows how to make content retrievable: clean entity statements, chunking, intent-aligned headings, answer blocks, evidence patterns, and freshness signals. You’ll also learn a lightweight evaluation method—prompt sets and change control—so teams can iterate with discipline instead of guessing. This session will help ensure attendees will walk away with practical information that can be implemented right away including: A retrieval-readiness checklist for content and templates; Page patterns that increase reuse in AI answers (definitions, decision criteria, comparisons, proof blocks); A repeatable test-and-learn loop (prompt library,logging, versioning, evaluation).

Transcript

1. #AthensSEO26 Judith Lewis Founder, Decabbit Consultancy Retrievable Beats Rankable: Build for Retrieval-Based AI decabbit@gmail.com
2. #AthensSEO26
3. #AthensSEO26 Expert trainer, speaker, fractional CMO, MC and compère Run DeCabbit Consultancy specialising in global digital marketing consultancy Non-Executive Director (NED) of the Academy of Chocolate Technical Lead of the WBA think tank in Switzerland working on AI in viticulture Contact: decabbit@gmail.com Who I Am & What I Do
4. #AthensSEO26 I have CHOCOLATE for questions relevant to the session so remember to ask!
5. #AthensSEO26
6. #AthensSEO26
7. #AthensSEO26 The company running a version of an LLM (like Claude) releases a bot
8. #AthensSEO26 The LLM ingests content including webpages
9. #AthensSEO26 The LLM chunks and tokenises the content
10. #AthensSEO26 Content is no longer recognisable
11. #AthensSEO26 Tokens placed within vectors
12. #AthensSEO26 Entities and their relationships built
13. #AthensSEO26 LLMs do not store facts the way we assume they do
14. #AthensSEO26 Factual recall in LLMs is additive rather than driven by one single “fact neuron”
15. #AthensSEO26 It treats facts as triples like (subject, relation, attribute)
16. #AthensSEO26 The right answer appears because these mechanisms reinforce one another
17. #AthensSEO26 LLMs combine several partial clues until one answer stands out
18. #AthensSEO26
19. #AthensSEO26
20. #AthensSEO26
21. #AthensSEO26 What does the machine REALLY want?
22. #AthensSEO26 Building Brand Trust Through AI Readiness Generally on Site https://www.aleydasolis.com/en/ai-search/ai-search-winning-brands-characteristics/
23. #AthensSEO26 Aleyda’s 10 points specifically for AI content ingestion for positive results https://www.aleydasolis.com/en/ai-search/ai-search-optimization-checklist/
24. #AthensSEO26 So just SEO?
25. #AthensSEO26 Rank well on Google, surface well in AI
26. #AthensSEO26 This is why content structure is important
27. #AthensSEO26 Single page, single topic
28. #AthensSEO26 It helps to have answers to popular questions already within content, structured to make it easy to understand and deliver in a result
29. #AthensSEO26 FAQs are really for PAA (and people)
30. #AthensSEO26 Be the answer
31. #AthensSEO26 On every platform
32. #AthensSEO26
33. #AthensSEO26 Optimise content for crawlability Are you using a CDN and if so, has it auto-blocked AI bots or are they allowed? (This is a known issue) Avoid Client Side Rendered JavaScript – AI cannot see it. Check that none of the pages are noindexing via meta robots any of your valuable content which should be surfaced. Avoid using a “nosnippet” rule which will prevent the content from being used as a direct input for AI Overviews and AI Mode (covered recently by MS at a conference – interesting news) Optimise internal linking so make sure if there is a relevant associated page or article on your site, you link to it in other articles.
34. #AthensSEO26 llms.txt
35. #AthensSEO26 llms.txt
36. #AthensSEO26 llms.txt Google search say they don’t use it but… agents are different… Agentic llms.txt is different Tells agents *how to do specific things* on your website Points to specific tools and APIs Points to pages for completing tasks like making a booking, filling out a form, or *purchasing a product* There are many use cases for llms.txt Wix websites have agentic llms.txt – check Crystal Carters posts on this for more information!
37. #AthensSEO26 https://www.geeksforgeeks.org/blogs/what-is-technical-seo/
38. #AthensSEO26 Teaching the AI: Topic Breadth and Depth AI search breaks complex queries into multiple related searches This behaviour rewards sites with detailed content pieces (but note examples are about topics likely to have general search volume). For example, a detailed article about clonal selection relating to the plantings of pinot noir a winery chose to plant in its vineyard. Or an article about the history of Gyros and how it came to be one of the most popular foods. Stick to the same topic though so depth on topic, not breadth. Don’t cover all different vines and their parentage in a single article unless it is your pillar article lining out to detailed ones. Or an article about the history of Gyros start with the assumed point of invention
39. #AthensSEO26 Teaching: Topic Breadth and Depth Create pillar (hub) pages, main broad topics, and then cluster pages around specific facets: Pillar (broad/general) pages should summarize each topic element, covering key sections briefly, with links to deeper cluster pages. Cluster (indepth) pages should target specific facets, that should be covered in-depth, ensuring unique purpose and query intent for each page. Cross-link between cluster pages where relevant and back to the hub, as a central resource, establishing semantic relationships across content, and helping AI understand the full context and span connections between topics.
40. #AthensSEO26
41. #AthensSEO26 Optimise for “Chunk” Storage & Retrieval CHUNKS ARE WHAT ARE DONE TO TEXT, not what we do to them! AI search engines don’t index or retrieve whole pages They break content into chunks as part of their ingestion process. You should optimise each section/paragraph like a standalone snippet (for PAA, and other SEO reasons). Don’t rely on AI using the whole page for context, each segment should be independently understandable but still work within the whole… for PEOPLE and SEO not just AI One idea per section: keep each passage tightly focused on a single concept. Use structured, accessible, and well-formatted HTML with clear subheadings (H2/H3) for every subtopic.
42. #AthensSEO26 Be the Answer from Fan-out & RAG AI search engines put together multiple results from its analysis from different sources into a single (usually) coherent response. This means your content must be easy to extract and logically structured to fit into a multi-source answer. Summarise complex ideas clearly, then expand. Start answers with a direct, concise sentence. Favor plain, factual, non-promotional tone. Use Structured Data to help search engines and by extension AI better understand your page (LLMs don’t use schema themselves – it’s more complicated than that). Use natural language Q&A format.
43. #AthensSEO26 Be Citation Worthy LLMs will cite content when it’s perceived as factually accurate, up- to-date, well-structured, and authoritative. Your content must meet higher trust and clarity criteria to be included in a result with a link within an LLM (or be outside certain training data which does not get a link). Use specific, up-to-date, verifiable claims, fact-based statements, not vague generalities. Include source citations (link to studies, stats, or experts) when you include facts and figures (or specify these are from your data/info). Show authorship and/or credentials for EEAT signals such as “certified tour guide for 10 years in Athens” or “viticulturalist graduated from Geisenheim 2018 but working on Alpha Estate since 2015”. Use author, organization, and other structured data to really dominate search results outside LLMs.
44. #AthensSEO26 Content Authoritativeness Signals Authority increases the likelihood that your content will be included in AI-generated answers. Use clear authority signals such as expert bylines, structured data (dev help), external citations (links), and mentions on reputable sites Optimise your brand presence across web platforms, including social channels, in a consistent way, linking back to your main site, engaging with your community, answering reviews, etc. Publish original research, reports, or data studies, conduct surveys, compile unique datasets, or run industry studies. Promote them to journalists and bloggers who create content roundups. Secure coverage in industry and expert publications, contribute quotes or guest content to respected newsletters and blogs in your industry.
45. #AthensSEO26 So it’s just SEO
46. #AthensSEO26 Ensure Multi-Modal Elements On-Page AI search systems are increasingly retrieving and synthesizing multimodal content, including images, charts, tables, videos, to better answer user queries. This then creates opportunities to provide more useful, scannable and engaging answers for users. Ensure images and videos have well optimised names. Images should use descriptive alt text that includes topic context. Add captions to images and videos with explanation right below or beside the visual.
47. #AthensSEO26
48. #AthensSEO26 You can win brand recognition, help people learn about your brand (and win at TikTok)
49. #AthensSEO26 So now we need some frameworks
50. #AthensSEO26
51. #AthensSEO26 And you don’t always need to have a framework you made (Note: This is created with a history of Hotel Chocolat analysis so it already has URLs and other information in memory like my requests for a priority checklist and other requests regarding this website)
52. #AthensSEO26 This enables us to quickly populate a complex audit
53. #AthensSEO26 And even help us make decisions!
54. #AthensSEO26 But we need to make sure we ask the LLM the right things
55. #AthensSEO26 Or what we get back will be worse than bad…
56. #AthensSEO26 We can use prompt engineering to improve the output
57. #AthensSEO26 How to create a framework Structure your prompts using a systematic approach, such as the C.R.I.S.P.Y. framework, to ensure the AI has all necessary information Context: Set the background and provide relevant information. Role: Define the persona or perspective the AI/LLM should adopt (e.g., "Act as an expert data scientist"). Instructions: Clearly state the action required. Specifics: Add constraints (e.g. "keep it under 200 words" "use bullet points"). Parameters: Specify tone, format, or style. Yielding: Describe the expected output or result.
58. #AthensSEO26 Lightweight evaluation method Create a prompt library as you work Ensure you are logging everything, including failures in responses to improve your prompting Once you have a lot of personas, scenarios, and data, versioning is essential (remember to clear memory and start a new chat) Then finally evaluation of the results by a human is critical Basically… spreadsheet everything with dates and times and be the words most boring person…
59. #AthensSEO26
60. #AthensSEO26 Other frameworks Chain of Thought (CoT): Instruct the model to "think step by step" to increase logical accuracy. Ask Before Answering: Instruct the AI to ask clarifying questions before providing a final answer to reduce ambiguity. Persona Assignment: Assigning a specific role instantly alters the depth, perspective, and tone of the response, yielding better, more tailored results. Define Output Format: Clearly state if you want a table, JSON, code block, or a summary, which reduces post- processing work.
61. #AthensSEO26 Ways to improve output Iterative Generation/Evaluation: Use a "Draft-Evaluate- Refine" system where the AI generates an initial response, self-evaluates against criteria, and improves it. Meta Control Framework (MCF): When stakes are high, break down the goal into smaller, manageable sub-tasks and create specific, testable components for each. Reality-Anchored Framework (RAF): For enterprise use, instruct the AI to use specific data sources, documents, or external references to prevent hallucinations.
62. #AthensSEO26 And finally… Define Metrics: Evaluate answers based on accuracy, relevance, coherence, and helpfulness. Human in the Loop (HILCS): For critical tasks, ensure a human reviews the AI's output to own the risk and make final decisions. Continuous Improvement: Standardise successful prompts and create a "system prompt" to avoid repeating instructions.
63. #AthensSEO26 AEO and GEO are SEO in a red dress
64. #AthensSEO26
65. #AthensSEO26 Remember… even ChatGPT uses search results
66. #AthensSEO26 https://www.nngroup.com/articles /ai-changing-search-behaviors/ Google is a hard habit to break – and now pushes to AI
67. #AthensSEO26 Which means optimising for Google
68. #AthensSEO26 Because the AIO pulls from search results
69. #AthensSEO26 As does ChatGPT
70. #AthensSEO26 How people search will continue to evolve
71. #AthensSEO26 Good SEO is all you really need.
72. #AthensSEO26
73. #AthensSEO26 Ask questions
74. #AthensSEO26 You Can Find Me Online Bluesky/X: @JudithLewis Threads/Insta: @Decabbit LinkedIn: https://uk.linkedin.com/in/judithlewis Web: https://www.decabbit.com Food Blog: https://MostlyAboutChocolate.com Slides: https://speakerdeck.com/decabbit https://www.slideshare.net/DeCabbit/