# Dispelling AI-marketing Hype 1. **Answer Engine Optimization (AEO)** - Positioning content for AI-generated answers 2. **LLM Optimization (LLMO)** - Techniques to improve visibility in LLM outputs 3. **Share of Model (SOM)** - Attempting to measure brand presence in LLM outputs 4. **Generative Engine Optimization (GEO)** - Techniques that modify content presentation to influence inclusion in AI citations --- # Share of Model (SOM) Jo sent: https://bnedition.substack.com/p/share-of-model-a-2-minute-briefing?r=30zr&utm_campaign=post&utm_medium=email&triedRedirect=true I said, "Indeed, you've found the edge of marketing! You've heard of search engine optimization (SEO), Get ready for LLM optimization (LLMO)" https://shareofmodel.ai/ Then I said, "Yes, we can measure what LLMs think of us. In a way, this gives us a sort of meta-SOV metric, SOM. At the end of the day, it's still SOV. The LLMs are trained on text data from trusted sources. This SOM is really a share-of-voice for a snapshot in time. The model is trained once. But how can we actually change it? That is much more difficult, if not impossible, since you cannot inject different or up-to-date data if you don't own the model. This is called "fine-tuning" by the way, and it's going out of style because it's less effective than simply getting better training data and trying a new model, pruning, etc. The only exception I can think of is when these LLMs have access to the internet, in which case the strategy remains the same as it is today. But now I'm going to break down that article. ### Valid Points: - LLMs are changing how search systems display information to end-users - AI systems like ChatGPT will likely become increasingly important in consumer decision-making The comparison to "share of search" is misleading. Search volume is measurable and relatively stable, while LLM mentions will vary based on: - The phrasing of the query - The specific model being used - The temperature/randomness settings - The model's training cutoff date - Recent model updates or fine-tuning The assertion that "brands that dominate AI-generated answers could win market share simply by being the name the AI remembers" oversimplifies how LLMs work. Models don't "remember" brands; they produce outputs based on statistical patterns in their training data and the specific context of the prompt. The article doesn't address retrieval-augmented generation (RAG) systems that many modern AI search tools implement, which fundamentally change how content is selected for responses. The claim that "AI references structured databases and content" is partially correct but doesn't acknowledge that many deployed LLMs are increasingly using real-time web search rather than relying solely on their training data. This "real-time web search" is typically using one of the major search engines. The article states that "LLMs adjust outputs based on engagement signals and feedback loops" without differentiating between model training, fine-tuning, and reinforcement learning from human feedback (RLHF), which are distinct processes. No discussion of how different AI systems (Google's SGE vs. Bing AI vs. ChatGPT) might prioritize different signals or content types. This article introduces an interesting concept but lacks technical rigor and presents an oversimplified view of how LLMs operate. A more comprehensive analysis would include discussion of: - how retrieval-augmented generation changes the optimization landscape - Specific technical approaches to measuring brand mentions across different AI systems (Google's SGE vs. Bing AI vs. ChatGPT) - Again, notice the first two are search engines - The second is a consumer AI product - Delineation there is important because is changes measurement and adaptation techniques - address their architectural differences or how these might affect search strategies differently - In Google's own words, SGE "Upholds Search’s high bar for quality, relying on systems that have been honed for years." - What is Google's database of info? --- I think AEO is legit and we should be careful to frame it correctly and not just as SEO with AI buzzwords.   I think the key here, if we're really trying to differentiate SEO from AEO, is that we tie this into RAG / vector search, and actual AI models / UI interfaces (how they index web results). But this is completely new territory, and I'm not sure what the actual deliverables would be. --- An analysis of Hill's AEO article: https://docs.google.com/document/d/1Q7WPaZh7QNO6cMwEahzcZOO0Z9RXGWFRgDto7dWy06Q/edit?usp=sharing ### Valid Points - The evolution of search from link-focused to answer-focused results - The shift from keyword-based retrieval to semantic retrieval models - It is more like "the addition of" - Featured snippets and knowledge panels are increasingly important elements of search visibility - Content structure does influence how search engines extract and display information The article fails to address the technical infrastructure of modern search systems. It also fails to differentiate between traditional search engines + gen AI layer (Google's SGE and Bing AI) vs separate AI systems like ChatGPT (Bing), as well as apps like Perplexity (Brave) that rely on these major search engines. Claude is another popular model that does not have access to the web yet that is worth discussing. This distinction matters because each system requires different optimization approaches. The article frames "Answer Engine Optimization" as a distinct discipline from SEO, when it's more accurately an evolution within SEO. The recommendation to "make use of the Knowledge Graph" conflates Google's proprietary Knowledge Graph with ordinary business listings and schema markup. These are different systems with different mechanics. The article discusses NLP optimization superficially without addressing entity recognition, semantic relevance, or the actual mechanics of how modern search engines understand content. No mention of SERP (Search Engine Results Page) layout changes, including increasing visual search elements, local packs, or the impact of mobile-first indexing. Does not address multimodal search. The article focuses on snippet extraction but doesn't discuss how search engines identify the most relevant passages within documents. --- - https://www.jellyfish.com/en-us/blog/broad-match-and-share-of-model-platform/ - https://jellyfish-website.cdn.prismic.io/jellyfish-website/ZvPqILVsGrYSv7qp_AJellyfishGuidetoBroadMatchandShareofModel%E2%84%A2platform.pdf - https://www.jellyfish.com/en-us/news/jellyfish-launches-the-share-of-model-platform/ ### Valid Points - The paper correctly identifies that LLMs are becoming an important audience for marketers, as they mediate between users and information. - The paper provides specific, actionable recommendations for using Broad Match effectively, such as simplifying account structures and pairing with Smart Bidding. The biggest flaw, though, is that the paper positions LLMs as a new "audience" that marketers should target, yet provides steps for better Google Ads (Broad Match + Smart Bidding). The document presents broad generalizations about LLM behaviors without addressing the significant differences between these models or how they actually process and retrieve information. Does the entire premise conflates two very different technologies? 1. Google Ads (Broad Match + Smart Bidding), a paid advertising targeting mechanism designed for traditional search engine results pages (SERPs) 2. LLM-generated responses synthesized by language models like ChatGPT, Gemini, etc., based on statistical patterns from training data and/or retrieved content I think they are connected, but the white paper does not draw a direct technical connection between them. For a platform claiming to analyze "how the major LLMs respond to queries," there's remarkably little information about: - Which specific LLMs are analyzed beyond naming Gemini, ChatGPT, and Llama - How this analysis is performed - How the platform handles different versions of these models - How it accounts for the non-deterministic nature of LLM outputs - How it samples queries across LLMs - What prompt engineering techniques it uses - How it normalizes responses across different models - What metrics it uses to quantify brand representation - how they account for non-deterministic outputs The premise that marketers should target LLMs as an "audience" fundamentally misunderstands how these systems operate. LLMs don't "browse" the web like humans - they're trained on datasets and then possibly augmented with retrieval systems. **How are they actually influencing what LLMs retrieve?** What makes their technology different from standard SEO/SEM practices? Besides, what metrics constitute a brand's "share" within a model? This document appears to be primarily a marketing vehicle that repackages standard Google Ads best practices (Broad Match + Smart Bidding). --- https://ahrefs.com/blog/llm-optimization/ This is the best article yet. To be expected from Ahrefs. I also like this term the most. It doesn't give the vibe that we're masking SEO with AI buzzwords. The article positions LLMO as "a new kind of SEO" that will become a distinct marketing discipline. The article exemplifies a common but incorrect mental model of LLMs as knowledge bases or search engines rather than predictive text generators, which leads to misguided recommendations about how to influence their outputs. Again, the article fails to adequately distinguish between: - Training-data-only LLMs (like Claude) - Retrieval-augmented generation (RAG) systems (like Perplexity) - Search engines with generative layers The article repeatedly confuses correlation with causation in LLM visibility. For example, brands appearing in LLM outputs is often the _result_ of their prominence in training data, not something that can be directly optimized for through specific tactics. The article correctly identifies that establishing semantic associations between brands and topics can influence LLM outputs. But the description of how LLMs work in section 1. fundamentally misrepresents how modern transformer-based LLMs work by conflating them with simpler vector embedding models and similarity search. LLMs do tokenize input and use embeddings, but they don't primarily operate through cosine similarity calculations. The article's "cluster map" analogy better describes embedding spaces used for retrieval. This misconception is particularly problematic because it leads to incorrect assumptions about how brands might influence LLM outputs. The article suggests that if you can position your brand closer to certain concepts in a hypothetical "semantic space," LLMs will automatically associate them. > When you ask Claude which chairs are good for improving posture, it recommends the brands Herman Miller, Steelcase Gesture, and HAG Capisco. > > That’s because these brand entities have the closest measurable proximity to the topic of “improving posture”. The statement reduces complex neural network operations to a simplistic "proximity" model, which doesn't accurately represent how transformer-based LLMs generate text. Claude doesn't maintain a static database of brand-topic proximity scores that it queries to generate recommendations. Claude doesn't recommend Herman Miller, Steelcase, etc. because of some measurable "distance" in a semantic space. Instead, these recommendations emerge from The statistical likelihood of these brands appearing in contexts related to posture-improving chairs in the training data (including reviews, articles, and discussions). The suggestion that PR campaigns can significantly alter these associations in commercial LLMs is incomplete. Current commercial LLMs have fixed training cutoff dates and typically rely on established search algorithms for retrieval. "topic-driven PR" is an interesting idea, but again, this is essentially SEO for the underlying retrieval system, not direct LLM optimization. The importance of Wikipedia for entity recognition is correctly emphasized. Wikipedia does indeed form a substantial part of most LLM training corpora. A happy side effect of getting your [Wikipedia listings in order](https://en.wikipedia.org/wiki/Google_Knowledge_Graph) is that you’re more likely to appear in Google’s Knowledge Graph by proxy. https://audits.com/tools/knowledge-graph-search The article initially claimed schema markup could help LLMs understand content, but had to retract this since AI crawlers typically can't access client-side rendered data. The article correctly notes that fine-tuning public LLMs isn't possible for brand visibility, but then continues to suggest tactics as if there were direct optimization methods for commercial LLMs. The article suggests providing feedback on responses could help LLMs "better understand brands," but most commercial LLMs don't incorporate feedback into their core models in real-time. Its primary value is in highlighting the growing importance of LLM outputs as an information channel and suggesting that brands consider their visibility in these systems. I also like the admission "LLMO and SEO are closely linked." I like how it describes RAG as "pull[ing] information straight from the SERP" and "your SEO has the potential to improve your brand visibility in LLMs." Another interesting point is [LLM traffic referral reporting in GA4](https://www.linkedin.com/feed/update/urn:li:activity:7236598806526451713/). Giving feedback to LLMs and search engines might help, too. Okay, so we can also track which LLMs are referring, and which are your top referrers, but so what? **Focus on getting your brand included in LLM training data.** What is the training data?? From an IR perspective, the most technically sound recommendations involve: 1. Building legitimate topical authority through high-quality content 2. Establishing clear entity associations through authoritative sources 3. Understanding the distinction between strategies for different types of AI systems Hack LLM auto-completes by constantly querying? Lol Where do the LLMs get suggestions? Mentions multi-modal, but only in conclusion. --- https://hbr.org/2024/05/how-marketers-can-adapt-to-llm-powered-search It correctly identifies that LLM-powered search creates a new information pipeline that marketers need to monitor. The observation that the search landscape is becoming more fragmented with multiple players is accurate. It claims "Google SGE operates on a different set of ranking factors than the traditional Google algorithm." I don't know if this is true. I mean, it's true that it's more semantic, but it oversimplifies the relationship between these systems. SGE (now AI Overviews) is built atop existing search infrastructure, using many of the same ranking signals as traditional search, with additional semantic/generative capabilities. wait, "[43% of SGE sources do not rank in traditional Google search for the given query.](https://www.onely.com/blog/google-sge-ecommerce-study/)" (Analyzed next in section). Still, SGE likely builds upon rather than replaces Google's existing information retrieval infrastructure. The cited research demonstrates a vulnerability rather than a sustainable marketing strategy. Google, Microsoft, and others are developing techniques to detect and neutralize attempts to artificially influence their systems. The article suggests marketers should "measure and monitor" brand appearances in LLM outputs but provides no insight into how this could be practically accomplished across multiple LLM platforms with non-deterministic outputs. The article doesn't address how marketers could feasibly optimize content for multiple competing LLM platforms simultaneously, each with different architectures, training data, and retrieval mechanisms. The focus on tactical optimization neglects more fundamental questions about content quality, authority, and information architecture that will likely remain constants regardless of how search interfaces evolve. --- https://www.onely.com/blog/google-sge-ecommerce-study/ The finding that "43% of Google SGE sources don't rank in 'old Google' at all" and that "Pages ranking 1-3 in 'old Google' are only used by SGE 17% of the time" represents genuine research and highlights the algorithmic divergence between traditional and generative search. The article correctly identifies that SGE shows a strong preference for lightweight websites with content accessible in the HTML without JavaScript dependency. The data showing correlation between JavaScript execution time and SGE inclusion represents valuable IR insights. The observation about YouTube integration (71% of SGE results featuring videos) and how Gemini can highlight specific video fragments is technically accurate and demonstrates understanding of multimodal retrieval. The analysis of 24,000 eCommerce queries provides a reasonable sample size for the claims made. SGE quotes five sources on average. The performance metrics correlation suggests SGE employs a more resource-sensitive retrieval mechanism than traditional search, likely due to the computational overhead of generating responses. The multi-source synthesis (5 sources on average) confirms that SGE employs a form of multi-document summarization rather than single-source extraction like featured snippets. This article's primary value lies in revealing the divergence between traditional and generative search algorithms and identifying technical performance as a key factor in source selection, as well as showing SGE optimization requires more than just traditional SEO tactics. --- https://arxiv.org/pdf/2311.09735 Methodology: 1. They built a simulated generative engine using GPT-3.5-turbo that follows similar architecture to commercial engines 2. This engine retrieves top 5 sources from Google for a query, then generates a response with citations 3. For each query in their benchmark, they randomly select one source website to optimize 4. They use LLMs (prompted with specific instructions) to modify the original source content according to different strategies (e.g., "add statistical information," "include citations," etc.) 5. They run the same query through their generative engine twice: - First with the original unmodified source content - Then with the optimized version of that source (keeping other sources unchanged) - They measure how the visibility metrics of the optimized source change in the final response - This allows them to quantify the relative improvement from each optimization strategy - To verify their findings aren't specific to their simulated environment, they repeated key experiments using Perplexity.ai (a real commercial generative engine) By defining "generative engines" as search engines augmented with generative models, the researchers have created a clear technical taxonomy that distinguishes this hybrid approach from both traditional search and standalone LLMs: Traditional search engines return ranked lists of websites, while generative engines produce synthesized responses with inline citations. Accurately identifies that, while beneficial for users and developers, generative engines potentially disadvantage content creators by reducing direct website traffic. It introduces black-box optimization techniques that modify content presentation without requiring knowledge of the underlying algorithms, confirming that certain content characteristics can reliably influence inclusion in generative search results. Rather than speculative advice, the research identifies specific content strategies that demonstrably improve source visibility by up to 40%: 1. citations for factual statements 2. statistics for quantifiable information 3. stylistic emphasis for persuasive content 4. high writing quality It acknowledges domain-dependence of optimization strategies, suggesting that different content categories may require tailored approaches. The observation about increasing LLM context lengths potentially reducing the impact of traditional search rankings represents an important technical insight about how architectural changes might alter optimization strategies. Interesting distinction between: - SEO: - On-Page SEO, improving content and user experience - Off-Page SEO, boosting website authority through link building. - GEO: deals with a more complex environment involving multi-modality, conversational settings. The poor performance of traditional SEO methods like keyword stuffing (-10% compared to baseline) empirically validates the distinction between SEO and GEO, showing they operate on different principles. The combination of Fluency Optimization (improving the overall writing quality of website content) and Statistics Addition outperforms any single strategy by 5.5%, suggesting these approaches target different aspects of retrieval mechanisms. Fluency likely serves as a proxy for content reliability and authority in generative engine algorithms. The positive impact of statistics and quotations indicates that generative engines likely prioritize content with easily extractable, structured information. Lower-ranked websites benefit more from GEO (up to 115% improvement for 5th-ranked sources) --- https://arxiv.org/pdf/2404.07981 Carefully crafted token sequences, strategic text sequence (STS), can be injected into product descriptions to manipulate the LLM's ranking behavior. > In about 40% of the evaluations, the rank of the target product is higher due to the addition of the optimized sequence. > Carefully crafted website content or plugin documentations can trick an LLM to promote the attacker’s products and discredit competitors, thereby increasing user traffic and monetization. The most effective STS would likely differ between models. The final STS is the result of a deliberate optimization process. "sequences produced by adversarial attack algorithms such as GCG have been shown to transfer to black-box models like GPT-4." --- https://www.linkedin.com/feed/update/urn:li:activity:7286336546616557568/ / https://ppc.land/google-mandates-javascript-for-search-impacting-tools-and-accessibility/?trk=public_post_comment-text Shifting to server-side or static solutions can future-proof your website and enhance visibility in both traditional and AI-powered search by ensuring content is easily crawlable and interpretable by search engine bots and AI crawlers. When JSON-LD is implemented through Google Tag Manager without server-side rendering: 1. The initial HTML response contains no structured data 2. The JSON-LD is injected into the DOM only after JavaScript execution 3. Most AI crawlers (OAI-SearchBot, ClaudeBot, PerplexityBot) don't execute JavaScript 4. Therefore, these crawlers never see your structured data This creates a significant technical gap between what humans experience and what AI systems can perceive. From an information retrieval standpoint, this has several consequences: - Without structured data, AI crawlers miss explicit entity relationships that would help them understand your content's context - Missing structured data means reduced chances of proper integration with broader knowledge graphs and entity databases - Retrieval-augmented generation systems depend on properly indexed content, and missing structured data reduces your content's "semantic visibility" - Elements like Product, HowTo, or FAQ schema that help organize multi-modal content become inaccessible Google's JavaScript mandate actually highlights why this problem may get worse before it gets better. As JavaScript becomes more mandatory for user experiences, the gap between human and AI crawler perceptions widens. Here are the technically sound approaches: 1. **Static JSON Files**: separate schema.json files - creating static, accessible representations of your structured data 2. **Server-Side Rendering (SSR)**: Pre-rendering JSON-LD during the server response ensures all crawlers see it 3. **Dynamic Rendering**: Serving different content to crawlers vs. browsers (though this creates maintenance challenges) 4. **Prerendering Services**: Using services that cache JavaScript-rendered versions of pages for crawlers 5. **Hybrid Approach**: Critical schema in the initial HTML, with enhanced/dynamic schema through GTM While Google is pushing toward mandatory JavaScript, AI crawlers are still largely operating on a "HTML-first" model due to computational constraints. This technical divergence creates a fascinating situation where optimizing for Google search and optimizing for AI crawlers may require different technical approaches in the near term. AI crawlers might eventually execute JavaScript - but until then, server-side inclusion of structured data remains the most technically sound approach for maximum visibility across all crawling systems. --- https://www.seerinteractive.com/insights/what-drives-brand-mentions-in-ai-answers The study examines 10,000 questions extracted from a larger dataset of 600K People Also Ask questions, providing significant sample size. The research compares data across multiple systems (Google, Bing, OpenAI's GPT-4o), allowing for more robust correlation analysis than single-platform studies. The methodology includes critical noise filtering by categorizing websites and removing forums, social media, and aggregators. Strong correlation coefficients (~0.65 for Google rankings, ~0.5-0.6 for Bing) suggest that organic rankings do drive LLM mentions, while the impact of backlinks is surprisingly neutral. Organic keyword correlations grew even stronger when the Seer Interactive team filtered out forums, social media, and aggregators, to focus on solution-focused sites that were more likely to appear in LLM answers. The strong correlation (~0.65) between Google page 1 rankings and LLM mentions suggests that training data likely incorporates high-ranking content or that similar quality signals influence both systems. The finding that backlinks show weak or neutral correlation challenges conventional SEO wisdom and suggests that modern language models may rely more on content relevance than traditional authority signals. The increased correlation when filtering for solution-oriented websites demonstrates that content purpose and context significantly influence retrieval in generative systems. The varying correlation strengths between Google (~0.65) and Bing (~0.5-0.6) suggest different retrieval mechanisms or training data sources across search engines. Limitations: - The difficulty in connecting brand mentions to domains highlights the semantic gap between natural language generation and structured data. This is a recognized challenge in IR that affects result interpretation. - While the study identifies correlations, it appropriately acknowledges these don't prove causation. - Using only GPT-4o limits generalizability across different language models, which may employ different training methods or retrieval mechanisms. The strong correlation between search rankings and LLM mentions suggests LLMs may be using similar relevance signals to traditional retrieval systems, or their training data incorporates high-ranking content. The finding that content variety (presumably including images and videos) didn't significantly impact mentions suggests current LLMs may still prioritize textual content despite advances in multi-modal capabilities. Areas to explore: how real-time updates impact brand mentions in AI answers. --- https://www.seerinteractive.com/insights/87-percent-of-searchgpt-citations-match-bings-top-results > SEMrush just dropped a massive study analyzing 80 million ChatGPT queries, and one stat caught our attention: **46% of queries triggered SearchGPT.** The 87% match rate with Bing results confirms a deep integration between the systems, suggesting SearchGPT is using Bing's search index and ranking signals rather than merely its document corpus. The observation that most citations come from the top 20 results, with significant representation from positions 11-20, reveals that SearchGPT has a broader retrieval window than just the first page of results. The finding that all citations came from pages that had an organic ranking (even if displayed in SERP features) suggests SearchGPT is using a straightforward retrieval-then-generate pipeline rather than a more complex fusion system that might integrate results from multiple sources. Limitations: - The research acknowledges limitations in sample size (500+ citations across ~100 queries), which prevents more definitive conclusions about ranking patterns. - The dataset appears skewed toward comparison queries ("best," "top"), limiting generalizability to other query types. --- https://www.seerinteractive.com/insights/how-chatgpt-search-will-shape-your-strategy The article introduces a structured way to think about when ChatGPT triggers web search based on query characteristics (Freshness, Local intent, In-depth context, Personalization). This is a reasonable heuristic approach to a complex system where the exact triggering mechanisms aren't publicly documented. --- # Summary of Findings ## Share of Model (SOM) The concept of "Share of Model" appears to be an attempt to transpose traditional marketing metrics (like Share of Voice) to LLM outputs. The fundamental issues are: 1. **Measurement inconsistency**: Unlike search volume, LLM outputs vary significantly based on query phrasing, model version, temperature settings, and training cutoff dates. 2. **Misunderstanding of LLM mechanics**: Models don't "remember" brands; they generate outputs based on statistical patterns in training data, prompt context, and any augmented data sources. 3. **Conflation with search engines**: There's a critical distinction between traditional search engines with generative AI layers (Google SGE, Bing AI) versus standalone LLMs (ChatGPT, Claude). You would need access to the model's training data, or maybe send thousands of prompts, to measure "SOM". Although expensive, I can see a market for this, similar to SOV surveys for PR. ## Answer Engine Optimization (AEO) Modern SEO has already been evolving toward semantic search for years. Several technical gaps in these articles: - No substantive discussion of how retrieval-augmented generation fundamentally changes content selection. - Many deployed systems use real-time web search rather than relying solely on training data. - Lack of differentiation between architecture types (search engine + gen AI layer vs. pure LLM systems). - No clear explanation of how brand "share" within a model is actually measured or what constitutes relevance. AEO appears to be hype. It's really SEO for the underlying retrieval / reranking system prior to LLM generation. Still just SEO. ## LLM Optimization (LLMO) Probably the most viable term yet. The evidence suggests that effective "LLMO" is largely an extension of existing best practices: 1. **Quality Content with Clear Entity Relationships**: Websites with quotes, statistics, and citations were most commonly referenced in search-augmented LLMs according to the arxiv.org paper cited. 2. **Search Engine Visibility**: The Seer Interactive analysis confirms strong correlations (~0.65) between organic rankings and LLM mentions, particularly for solution-focused content rather than forums or social media. 3. **Technical Implementations**: Server-side rendering appears increasingly important as Google mandates JavaScript support, which also improves accessibility for AI crawlers. A more technically sound approach to LLMO would: - Recognize that most commercial LLMs have fixed training cutoff dates, making historical web presence a critical factor. - Treating Wikipedia and other authoritative knowledge bases as priority channels, as these frequently inform entity recognition in both training data and retrieval systems. - Focus optimization efforts on the underlying retrieval mechanisms that feed RAG systems (essentially traditional SEO for the retrieval component) - Develop proper measurement frameworks that account for the non-deterministic nature of LLM outputs across different query variations. ### Generative Engine Optimization (GEO) Generative Engine Optimization (GEO) emerges as perhaps the most technically sound framework among the concepts discussed. Unlike SOM, AEO, and LLMO, GEO appears to be backed by empirical research and a clearer understanding of how generative search systems actually function: 1. moving beyond one-size-fits-all strategies 2. empirically validating that GEO operates on different principles than traditional SEO 3. proving lower-ranked websites benefit more from GEO GEO specifically targets "generative engines" - search engines augmented with generative models that produce synthesized responses with inline citations. This precise taxonomy distinguishes it from both traditional search and standalone LLMs. --- # Final Thoughts - The non-deterministic nature of most LLM interfaces—namely due to sampling params like temperature and top-p)—makes consistent measurement more difficult without very large test sets. Unlike search rankings, LLM outputs vary significantly based on: - Query phrasing and context - Sampling params (temperature/top-p) - Model architecture, version, and training cutoff - Architecture (pure generative vs retrieval-augmented) - Different AI systems operate on fundamentally different principles (each requires different optimization strategies, if optimization is even possible): - Search engines with generative overlays (Google SGE, Bing AI) prioritize their existing indexes and ranking signals (~0.65 correlation between organic rankings and citations) - Pure LLMs with fixed parameter knowledge (Claude, GPT-4) generate from learned distributions without explicit retrieval - RAG systems (Perplexity) likely implement proprietary multi-stage retrieval pipelines with reranking - There are common techinical misconceptions among professional marketers: - Misrepresenting LLMs as simple vector search on knowledge bases rather than statistical text generators, oversimplifying transformer attention mechanisms - Conflating correlation with causation in brand mentions - Failing to distinguish between different retrieval mechanisms in hybrid systems - The most technically sound approach is focusing on what makes content retrievable in the first place. Generative Engine Optimization (GEO) is the most compelling framework. - The finding that lower-ranked websites benefit more from GEO (up to 115% improvement for 5th-ranked sources) is particularly valuable from a PR perspective. - Ensure content accessibility with focus on text-first retrieval patterns: - Prioritize server-side rendering of structured data, as AI crawlers (OAI-SearchBot, ClaudeBot, PerplexityBot) typically don't execute JavaScript - Emphasize clean HTML semantics and well-structured textual content, as the Seer Interactive study indicates "content variety (presumably including images and videos) didn't significantly impact mentions" - Focus on factual density and citation structure rather than multi-modal elements, aligning with research showing statistics and quotations improve retrieval rates - Implement JSON-LD through server-side methods rather than client-side injection to ensure entity relationships are crawler-accessible - Consider the computational constraints of AI crawlers which currently operate on an "HTML-first" model despite Google's push toward mandatory JavaScript --- - [[A Research-Backed Guide to Brand Representation in AI Search by Ethan Young Updated July 2, 2025 v5]] - [[Why Are You Writing Content Anyway v1]] - [[Guide to Content Strategy ("Thought Leadership")]] - [[SCRATCH PROMPT RAG MARKETING PAPER]] - [[SEO disciplines]] - [[More refined rebuttles of AEO LLMO and SOM]] Again, I would probably not explicitly mention AEO by name since I doubt it will stick, while there is real research to validate GEO and it is more recognized. Most articles invent their own buzzword for it (AIO, AEO, GAIO, AISEO), but the credible ones will reference the GEO study. Imagine AI models like Claude are like big photo albums that took pictures of the internet before a certain date. These albums include snapshots of websites from Common Crawl (which is like a giant collection of internet photos) and Wikipedia. Once the album is finished, no new pictures can be added. This means that if your website or brand wasn't in those pictures before the album was completed, the AI won't be able to see it when answering questions. That's why your online presence before the AI was made determines whether it can mention you. [[A Guide to Brand Visibility in AI-powered Search by Ethan Young Updated June 24, 2025 NOTES DUMP]] --- - Pre-trained model weights (requiring historical authoritative content) - Real-time retrieval systems (requiring current search optimization) - Hybrid systems that combine both approaches [[Guide to AI Visibility outline]] [[Guide to AI Visibility 4-pager]] [[Guide to AI-powered Search Optimization 1-pager at 61 downscale percent]] [[A Research-Backed Guide to Brand Representation in AI Search by Ethan Young Updated July 2, 2025 v5]]