I had the pleasure of returning to the 2nd annual Tech SEO Connect conference this week in Raleigh, North Carolina. This time was even more exciting for me because I got to DJ the conference afterparty.
As usual, I took rapid-fire notes of everything that was said by the brilliant speakers at the conference, and felt inspired to share the key takeaways and my thoughts below. These conference talks included some of the best insights I’ve heard all year – and I spend a lot of my life at SEO and AI search conferences.
Thanks to the generosity of the organizers, the conference also livestreams and publishes all the live video recordings of each talk in case you need additional context on anything I summarized below.
Note: I used AI to help with summarization here (after manually typing 10,000 words of notes at the conference), supported with transcripts from the speakers’ talks. The speakers have been listed alphabetically by last name.
Beyond GSC: Using First-Party Logs to Win with AI Crawlers – Rachel Anderson
• Advised large domains to be careful relying on GSC’s Crawl Stats, as they can be inaccurate for sites with a large number of URLs, and pivot to using first-party log files.
• Stressed the importance of collaborating with the Site Reliability Engineering (SRE) team, framing them as the crawl experts who maintain the advanced log tools needed for tracking all bots.
• Shared a success story where using an SRE-built dashboard identified a third-party security tool that was mistakenly blocking Googlebot (hitting 406s) and all AI bots, leading to the cancellation of that contract.
• Noted that AI crawlers generally do not read JavaScript, and that the conventional index is “out the window” when discussing AI crawlers, as they have their own knowledge base.
• Concluded that there is no evidence that the proposed LLMs.txt protocol is being reliably crawled by major AI agents and advised against wasting time on it.
From Guesswork to Growth: Building an SEO Strategy That Compounds – Brie Anderson
• Advocated for the strategy to Always Be Testing and not to rely on anecdotal evidence or external advice when making strategic decisions.
• Introduced the BEAST cycle (Benchmark, Explore, Analyze, Strategize, Test & Track) as a rigorous framework for continuous growth, aligning with the scientific method.
• Emphasized the necessity of defining KPIs, ensuring all parties agree on the metrics (e.g., leads vs. rankings), and confirming that GA4 tracking is correctly set up.
• Showed that analyzing data should center on finding trends and outliers, which are best found by visualizing GSC and GA4 data.
• Demonstrated a successful test using SEO Jobs data where optimizing the remote page, based on identified user trends, led to a 6.5% increase in organic applications from that page.
AI Internal Linking for E-Commerce Websites – Serge Bezborodov
• Discussed optimizing internal linking at scale using vector embeddings and cosine similarity to find semantically similar pages.
• Noted that the sheer number of permutations (millions for 10K pages) makes calculating internal linking difficult, especially for low-content thin pages.
• Advised solving the scale issue by splitting the linking logic by category level, which both speeds up the process and increases relevancy.
• The core concept is to pass link equity from donor pages (strong, high crawl budget) to acceptor pages (weak, low crawl budget) to improve indexation.
We Analyzed 250,000,000 AI Search Responses: Here’s What We Found – Josh Blyskal
• Presented empirical data showing that traditional SEO metrics (backlinks, authority) only predict between 4% and 7% of AEO citation behavior.
• Found that optimizing content structure for machine retrieval is key; for example, using natural language URLs (5–7 words) drove 11.4% more citations.
• Stated that answer engines do not have a summarization feature in the RAG pipeline, meaning answers must be condensed into semantic chunks to fit the snippet or context window.
• Highlighted a strong recency bias in AI search, noting that 50% of top-cited content is less than 13 weeks old.
• Observed that blogs and opinion content (hyper-opinionated content) now lead in citation type, as AI is “lazy” and seeks established frameworks of analysis.
• Confirmed that User Generated Content (UGC) is a critical information class, with Reddit being the #1 cited source and YouTube being the second most cited source across aggregated AI models.
From Deck to Dev: Getting SEO Recommendations Shipped at Enterprise Scale – Bryan Casey
• Described the building of IBM’s mutually reinforcing “Inbound Program,” a compounding growth system (SEO, video, podcasts, newsletter) inspired by the Disney synergy map.
• Secured large paid media budgets by promising to deliver the same advertising KPIs but with 10x better efficiency, framing the budget as a “war chest” to raid.
• Enforced a policy of only doing fully contained programs and avoiding reliance on ad-hoc shared team resources, which previously had a 100% failure rate.
• Showed that grouping evergreen content and tutorials into “Hubs” made them durable references, which increased returning user traffic by 50%.
• Demonstrated that investments in search became 40% more valuable due to the complementary investments in the newsletter.
• Used internal communication (blogs) to control the narrative, such as framing the deletion of 80% of IBM.com content as fixing the “number one client experience problem in the company”.
Using GA4 to Surface Technical SEO Issues in Real-Time – Dana DiTomaso
• Advised SEOs to stop looking for problems and start using visitor experiences intel (GTM/GA4) so that issues “pop up and come” to them in real-time.
• Provided a specific GA4 tip: add a custom event parameter with a value of 1 (a count metric) to non-Key Events, allowing their frequency to be tracked in Looker Studio.
• Showed how to use GTM to track technical issues like 404 errors by checking the DOM element body class for error 404.
• Demonstrated tracking URLs with encoding (e.g., %20) and uppercase letters using page path triggers and regular expressions.
• Stressed the cardinal rule that UTMs should never be added to internal links, as this breaks attribution in GA4 by forcing the session to attribute to the last non-direct source.
Beyond Googlebot: Evidence-Based Retrieval Experiments with AI Crawlers – Jori Ford
• Stressed that SEOs must adopt a scientific approach, choosing to assume nothing and test everything.
• Her micro-tests showed that structured data (Schema) is the API for the logic and enables complex logical reasoning for Ragbots, whereas unstructured text makes the bot “dumber”.
• Found that RAG crawler activity is governed by appetite and session throttling, meaning bots are lazy and will only visit the best content once; they don’t want to go more than about three pages deep.
• Demonstrated that context bias is real; if the bot retrieves unstructured (flat) content first, it will use that information and ignore subsequently structured content in the same session.
• The actionable mandate is to not release pages that do not have structured data, as structure must be present from the first touch point to be effective.
Should We Do This Yet? Clear Advice On 5 High-Impact Technical SEO Decisions – Alex Halliday
• Advised against implementing emerging AI strategies like creating Markdown alternative pages or adopting the LLMs.txt protocol, as community tests show no evidence that this improves retrieval and no major crawler reliably consumes LLMs.txt.
• Warned that if a Markdown-only page accidentally becomes indexed, it is essentially cloaking/deceptive practice.
• Stressed that the problem LLMs.txt tries to solve does not exist, as RAG pipelines already convert well-structured HTML into markdown or equivalent formats efficiently.
• Recommended focusing on foundational elements: strong, clean HTML and using the main priority Schema types to create clarity around page entities.
From Deck to Dev: Getting SEO Recommendations Shipped at Enterprise Scale – Ross Hudgens
-
Provided strategic guidance on company selection, noting that it is the #1 skill marketers must master, and advised looking for companies with high change velocity (ability to implement changes quickly)
- Recommended asking prospective clients about their recent SEO implementations to assess their willingness to execute new strategies
- Stressed the importance of building business cases using competitor data (like traffic value from Ahrefs or Moz) and creating outcome models to predict the ROI of changes
- Advised that due to the impact of ChatGPT, all traffic projections must be increasingly conservative, suggesting that traffic numbers be cut in half in 2026
- Recommended providing executive summaries and TLDRs for recommendations, allowing leaders to focus immediately on high-priority actions.
Trashcat’s Guide to Black Boxes: Technical SEO Tactics for LLMs – Jamie Indigo
• Affirmed that AI ranking is probabilistic and not deterministic, meaning two users can get two different answers.
• Advised SEOs to stop relying on synthetic AI rank tracking, which she called the “Lighthouse of AI,” and instead focus resources on first-party log files for real user data.
• Recommended partnering with Site Reliability Engineers (SREs) for log access and expertise, as they are the true crawl experts.
• Noted that most LLMs cannot render JavaScript during retrieval, and that core performance metrics (like CWV and TTFB) still matter for real-time RAG.
• Stressed the importance of the meta description as a marketing vehicle to tell crawlers what the page is about, aligning with the idea that AI prioritizes freshness.
Everything You MFs Should Know About Query Fan Out – Michael King
• Predicted that Retrieval Augmented Generation (RAG) is the future, explaining that RAG systems take a prompt and extrapolate a variety of synthetic queries (Query Fan Out or QFO) to pull content chunks.
• Shared a critical data point that 95% of queries in QFO have no Monthly Search Volume (MSV), indicating these are unique, machine-generated longtail queries not found in traditional tools.
• Noted that AI systems route queries based on expectations of content format (index, API, images, videos); winning depends on whether your content maps to the expected format.
• Showed that success is probabilistic, recommending that SEOs must maximize inputs by ranking for multiple synthetic subqueries.
• Advised optimizing content using semantic triples (subject, predicate, object)—the idea that structured data isn’t just Schema—to provide clear and specific data points for the language model.
• Provided open-source tool equivalents, such as Qforia (for query generation) and n8n (for custom workflows), to close the gap in commercial SEO software that hasn’t adapted to these new mechanics.
Video SEO – The Shortcut to AIO Visibility – Cindy Krum
• Argued that video is the shortcut to effective branding and AI visibility.
• Noted that YouTube is one of the top most cited sources across AI platforms, and Google Discover now surfaces social feeds and short-form video, indicating a shift in user consumption towards dopamine-focused, quick hits.
• Stressed that basic SEO is table stakes; success relies on optimizing for engagement signals (likes, comments, re-watching), which is how algorithms determine what content to push.
• Advised using a strong “hook” in the first three seconds to maximize engagement, noting that Instagram tracks statistics on viewership past this point.
• Showed that tools like Opus Clips use AI to automatically convert a long-form video URL into multiple optimized short-form, vertical videos with transcripts, allowing for rapid, multi-platform distribution.
• Cautioned that relying on one platform carries risk (citing recent YouTube channel cancellations); content should be distributed across YouTube, Instagram, Tik Tok, LinkedIn, and Facebook to maximize investment and mitigate platform risk.
Beyond Blue Links: Freshness, Control and Visibility – Krishna Madhavan
• Stated that visibility belongs to content that AI can trust, understand, and ground, defining grounding as the anchor LLMs rely on for verifiable, trusted sources before answer generation.
• Emphasized that freshness is an imperative because AI systems are often asking for fresh content in their query expansion, and stale information is likely not selected by grounding models.
• Advised balancing crawl efficiency with freshness by using the IndexNow open protocol to immediately notify crawlers of new, updated, or deleted content.
• Advocated for using precision control via the data-no-snippet HTML attribute to exclude specific parts of a page from AI summaries/snippets while keeping the page eligible for ranking.
Stacking Signals: How International SEO Comes Together (And Falls Apart) – Max Prin
• Stated that despite hreflang being introduced 14 years ago, its implementation is still difficult, and the goals of International SEO remain focused on sending the right URLs to the right users.
• Stressed that domain strategy (ccTLD vs. gTLD) must be dictated by the “reality of the SERP”; for example, Condé Nast was compelled to launch a .de ccTLD because 78% of the German SERP was already .de.
• Confirmed that using the signals noindex and canonical together is acceptable in the same language/market to prevent syndicated content from competing with the original, a strategy Condé Nast uses heavily.
• Noted that hreflang tags have no impact on Google Discover traffic, which is a growing problem as Discover traffic rises.
• Advised that geo-redirects should be used in specific cases where sister brands are outranking original content and a brand wants to ensure users are sent to the correct version of the page.
When Traffic Gets Weird: ML for Anomalies & Forecasts – Sam Torres
• Warned that LLMs are unsafe for analyzing structured data (tabular, numerical data) because they tend to hallucinate when dealing with numbers.
• Advocated for using Machine Learning (ML) models for numerical data analysis, noting their transparency and mathematical rigor make them superior for this task.
• Recommended using Google Colab—a free, cloud-based Jupyter notebook environment—for running ML models, which is secure for GSC/GA4 client data and easy to share.
• Identified three key ML models for anomaly detection and forecasting: Isolation Forest (fast, multi-dimensional), Local Outlier Factor (LOF) (good for A/B testing or finding outliers), and the Forecast Residual Approach (best for time series and seasonality).
Behind the Scenes of Building Scalable Data Products for SEOs – Baruch Toledano
• Stressed that to build trust in data, SEOs must follow the “garbage in, garbage out” principle and be aware of LLM hallucination when models are not used properly.
• Demonstrated using triangulation—modeling the relationship (absolute ratio) between GA search traffic and GSC clicks—to handle data uncertainty between different data sources.
• For high-volume, repeatable tasks (like classifying 100 million pages daily), he uses model distillation, where a strong LLM trains a smaller, cheaper “student” model to maintain high accuracy at a lower cost.
• When optimizing LLM prompts, he uses few-shot prompting (giving the LLM an example of the desired output) to ensure correct formatting and specific results, rather than relying on pure instruction.
• Advised using clickstream data as a “compass” to determine which AI models are being adopted, how often users return to AI search, and how users interact with different modalities.
From Schema Markup to Knowledge Graphs: Powering AI with Connected Data: Martha Van Berkel
• Focused on preparing for the Agentic Web by emphasizing the critical role of Structured Data and Knowledge Graphs
• Defined the de facto standard for AI readiness as storing Schema markup in a high-quality Knowledge Graph, noting this improves LLM grounding and can achieve up to 91% accuracy compared to 43% for GPT4
• Advised SEOs to define entities deeply using Schema properties to create semantic triples (subject, predicate, object), rather than just basic Schema types
• Cited an example where a company used robust location schema markup connected to their Knowledge Graph, which resulted in AI Overviews citing the correct location page instead of outdated information from other sites
• Affirmed that LLMs can and do use structured data as part of their RAG pipelines
Decision Intelligence for Enterprise SEO – Tyler Gargula
• Introduced Decision Intelligence as a discipline and presented the DECIDE Model framework to help enterprise SEOs move from poor decisions to informed decisions.
• Stressed the necessity of defining the primary goal and rigorously avoiding biases like selection bias (analyzing only successful data) and recency bias (overweighing recent data spikes).
• Showed that analyzing pages flagged as “crawled, not indexed” led to identifying content attributes—such as using unique images vs. placeholders and stock status—that correlated positively with indexation.
• Advocated for visualizing data using tools like the Lorenz curve (for prioritized sampling of high-traffic pages) and violin/box plots (to visualize metric spread) to identify outliers and opportunities.
JavaScript Rendering and the Search Bot Viewport – Giacomo Zecchini
• Confirmed that while Google can render, most AI crawlers are not capable of rendering JavaScript during content retrieval.
• Detailed Google’s non-standard viewport expansion technique, which resizes the viewport height once to match the full page height to fire lazy-loaded content.
• Advised that developers should use the Intersection Observer API for lazy loading, as automated bots do not scroll or click.
• Stressed that content in CSS pseudo-elements (::before/::after) is not part of the DOM tree and will not be indexed.
• Recommended using visibility:hidden instead of display:none to hide content, as display:none removes the element’s dimension and position, potentially lowering its perceived value in search models.
• Cautioned that pages using viewport height (vh) without a max-height risk having critical content pushed below the fold when Google expands the viewport, reducing its perceived value.