Search for a command to run...
I work as a Tech Arch Senior Manager at Accenture and am a mother to two wonderful kids who test my patience and inspire me to be curious. I love cooking, reading, and painting.
No comments yet. Be the first to comment.
A Tale of Protocol Quirks, JSON Wrestling, and the Joy of Finding the Bug
In 2018, OpenAI pledged its "primary fiduciary duty is to humanity." Seven years later, the company launches erotic AI companions while families testify in courtrooms about children who never came home. We're conducting the largest unregulated behavi...
Did you know that 68% of organizations now identify data silos as their top operational concern? A 7% increase from 2023, according to a 2024 DATAVERSITY® Trends in Data Management Survey. In today's hyper-personalized business landscape, enterprises...
The Reality of AI Hallucinations: Recent Examples Last week, the AI coding assistant tool Cursor encountered a crisis when its front-line AI support bot named “Sam” confidently provided a user with a completely fabricated policy. The user raised a su...
This week, OpenAI introduced shopping capabilities within ChatGPT's search feature, coinciding with Mastercard unveiling Agent Pay, a new payment approach that enables AI agents to securely complete purchases. These developments signal a clear intention: generative AI platforms are aiming to keep users engaged within their ecosystems for all search needs, creating closed-loop environments for information discovery, evaluation, and action.
This observation led me to reflect on my search habits. For more than two decades, Google has been synonymous with internet searching and much more, but unconsciously, my search behavior has shifted dramatically. I now find myself turning to ChatGPT, Perplexity, or Claude for specific information searches, while relegating traditional search engines to simpler location-based queries like finding my nearest shopping center's trading hours.
The question arises: what changed, and is it just me? How have these AI tools managed to challenge decades of trust earned by traditional search engines, despite their known tendency to occasionally hallucinate? If you missed my blog on what AI hallucinations are, read here.
The traditional search model anchored in keyword-based SEO and fragmented user journeys is being rapidly disrupted by the rise of LLM-powered platforms like ChatGPT, Claude, and Perplexity. This shift is not merely behavioral, it’s architectural. Generative AI compresses the cognitive effort required for information discovery, offering synthesized, context-rich results in a single query, reshaping how users and enterprises interact with information.
This blog explores the deep-rooted technical implications of this transformation, from moving beyond keyword taxonomies toward knowledge graphs, to preparing internal data systems for AI-native interactions. It lays out a practical roadmap for organizations to realign both internal and external data architectures to remain discoverable, authoritative, and actionable in LLM-driven environments. The call to action is clear: technical leaders must move from SEO playbooks to knowledge-centric data strategy.
Now the fundamental question is, “why this paradigm shift is gaining traction so rapidly?” The answer lies not just in the technology itself but in its alignment with core human psychology. Traditional information foraging involves multiple cognitive steps: formulating search queries, scanning results pages, clicking through to various websites, evaluating information quality, and mentally aggregating insights from multiple sources.
Generative AI abstracts and consolidates these steps into a singular interaction, streamlining the user’s cognitive load and delivering synthesized results with zero clicks. In our era of increasing information overload, this efficiency in delivering direct results has rapidly won user acceptance. The reduced cognitive load leads users to accept occasional inaccuracies for convenience.
With the introduction of multimodality and citations to chatbots, enterprises particularly must rethink their content strategy to remain discoverable and relevant in this emerging search paradigm.
To understand the implications for enterprise data architecture, we need to examine the fundamental differences between how traditional search engines and LLMs process information:
SEO and Keywords: Traditional search engine optimization revolves around keywords, specific terms and phrases that users might input into search boxes. The data architecture supporting this approach involves:
Metadata optimization for search engine crawlers
Keyword density and placement throughout content
Link structures that signal authority
Content organized primarily for crawlability
This model worked well when search was a distinct activity separated from content consumption and action.
LLMs and Knowledge Graphs: In contrast, large language models operate on entities and relationships, similar to knowledge graphs. They seek to understand:
The entities (people, organizations, concepts, products) mentioned in content
How these entities relate to one another
The semantic context surrounding these entities
The factual assertions being made about these relationships
For organizations, this represents a fundamental shift in how information systems should be structured and exposed.
As organizations adapt to this new reality, technical leaders must fundamentally rethink how data is structured, both internally and externally. The challenge extends beyond SEO tweaks. It requires a comprehensive data architecture transformation.
For over a year, a client company I consult for has invested in Microsoft 365 Co-pilot licenses for all their employees. Apart from the occasional handy meeting notes that Co-pilot does, they haven’t really explored its full potential. With the introduction of AI agents, mainly within document repositories like SharePoint, there are untapped capabilities to improve their business processes and streamline their governance, but here comes the hard question: are their data assets restructured to be LLM consumable? Or is their SharePoint structure LLM friendly?
Simply investing in these LLM tools is not going to magically improve your organization’s productivity, and to see that ROI, organizations must change their data patterns and behavior. Below are critical changes that must be addressed:
Traditional Enterprise Data Model:
Tables → Relationships → Applications → Siloed Knowledge
LLM-Ready Enterprise Knowledge Architecture:
Entities → Attributes → Relationships → Unified Knowledge Graph
This transformation requires
Converting relational database schemas into semantic entity models
Explicitly modeling relationships between business entities (customers, products, services)
Preserving context in data storage rather than reconstructing it at query time
Maintaining provenance and authority indicators throughout the knowledge graph
Enterprise document repositories need restructuring:
Implement semantic chunking that preserves contextual units rather than arbitrary page breaks
Extract and explicitly model entities mentioned in documents
Create vector embeddings of document chunks for similarity-based retrieval
Maintain bidirectional links between structured data and unstructured content
Internal search capabilities must evolve:
Design data access layers that support natural language query translation
Implement hybrid retrieval mechanisms (keyword, semantic, vector-based)
Create context-aware query expansion capabilities
Build explanation mechanisms that justify why information was retrieved
Stepping back to our original news about the introduction of shopping capabilities and payment agents, another clear indication that the agentic AI workflows are here to stay, providing business avenues to users. For organizations with public APIs or publicly exposed data to external LLMs and search engines, is your data optimized for this?
Below are a few strategies that can guide you on this transformation path:
While we understand LLMs may have difficulty processing JSON-LD injected via JavaScript, if we look at today’s search engines like Google, which have come a long way to understand entities, their relationships, and context. The expectation is that the structured data understanding is also part of the training data incorporated to build the knowledge graphs that LLMs use.
This calls for leveraging Schema.org, especially FAQ schema, How-To Schema, Article Schema, and other relevant schemas to define content structure in plain HTML and guide the LLM towards the intended context.
Organizations need:
Deep entity markup that extends beyond basic product and organization schemas
Explicit relationship modeling between entities
Machine-readable assertions about capabilities, limitations, and specifications
Temporal metadata indicating freshness and update frequency
LLMs increasingly call APIs directly or recommend them to users:
Design deterministic, self-describing API endpoints optimized for LLM understanding
Implement context-preserving pagination that maintains semantic coherence
Include metadata about data freshness, confidence, and limitations
Support multiple representation formats (JSON, JSON-LD, GraphQL)
Here's a sample LLM-optimized API response, one I often use as a reference on how to make systems more machine-readable. This example is annotated to show the structural and metadata attributes that enhance semantic clarity, trust, and retrieval efficiency for generative models:
{
"data": {
"productComparison": [
{
"productId": "EDP-2025", // Unique identifier for entity resolution
"name": "Enterprise Data Platform",
"keyDifferentiators": [ // Explicit value props help LLMs distinguish offerings
"Real-time processing capability",
"Integrated knowledge graph"
],
"limitations": [ // Transparency on constraints improves trust
"Requires minimum 16GB RAM"
]
}
]
},
"metadata": {
"lastUpdated": "2025-04-28T14:22:17Z", // Temporal context for freshness
"confidenceScore": 0.97, // Helps LLMs prioritize results
"dataSource": "Official product specifications", // Provenance for reliability
"citation": "https://example.com/products/edp-2025/specifications" // Link for verification
}
}
As the data principles to enable semantic understanding and entity recognition, structure formats, context, disambiguation, accuracy, and credibility apply to both internal and external data sources, it is recommended that organizations use this excuse to bridge their internal and external data approaches to define a consistent governance. A unified strategic framework for data preparation is crucial when addressing this shift to generative AI and LLM-driven search.
Technical leaders should implement
Consistent information classification frameworks across all data assets
Automated validation pipelines that ensure data quality and consistency
Real-time synchronization mechanisms for time-sensitive information
Comprehensive attribution systems that maintain data provenance
As LLMs become increasingly multi-modal:
Ensure consistent entity identification across text, images, and structured data
Implement cross-modal entity resolution systems
Create unified knowledge repositories that maintain relationships regardless of data format
Design media annotation systems that make visual and audio content LLM-comprehensible
The transformation to LLM-ready enterprise architecture requires a holistic approach that recognizes knowledge as a unified asset while acknowledging different access and security requirements. Strategic imperatives for enterprise-wide data transformation include:
The core of an LLM-ready architecture is a consolidated knowledge graph that serves as the single source of truth:
Conduct comprehensive semantic mapping
Audit existing data schemas across all systems
Identify entities, attributes, and relationships across previously siloed domains
Create standardized entity definitions with unique identifiers
Build a centralized knowledge graph
Implement entity resolution systems to connect disparate data sources
Define explicit relationship models with context preservation
Establish clear provenance tracking for all knowledge assertions
Create bidirectional links between structured data and unstructured content
Transform content repositories
Deploy intelligent semantic chunking that preserves contextual meaning
Generate vector embeddings for similarity-based retrieval
Extract entities from unstructured content and link to the knowledge graph
Implement automated tagging and classification systems
With a unified knowledge foundation in place, create specialized access mechanisms for different contexts:
Internal knowledge interfaces
Develop natural language query capabilities for enterprise data
Implement role-based access controls with appropriate security
Create context-aware middleware for translating user intent
Build explanation mechanisms that provide reasoning transparency
External knowledge exposure
Enhance Schema.org implementations with deep entity and relationship modeling
Design LLM-optimized API endpoints with self-describing capabilities
Implement structured data markup that preserves semantic relationships
Create machine-readable factual assertions with confidence indicators
Bridging mechanisms
Deploy automated validation pipelines ensuring consistency across interfaces
Implement real-time synchronization for time-sensitive information
Create governance frameworks controlling information flow between layers
Build feedback systems to detect discrepancies between internal and external representations
An effective LLM-ready architecture must evolve based on interaction patterns:
Monitoring and measurement
Track how LLMs represent your enterprise information
Analyze query patterns to identify knowledge gaps
Measure accuracy and completeness of LLM responses
Identify frequent hallucination triggers
Feedback integration
Create automated processes to detect and correct misrepresentations
Update knowledge assertions based on observed inaccuracies
Enhance knowledge graph connections for commonly confused entities
Refine content structure based on comprehension patterns
Knowledge enhancement
Deploy machine learning to identify potential new relationships
Automate discovery of implicit connections between entities
Enhance context representation for frequently accessed information
Continuously refine semantic structures based on usage analytics
The shift from traditional search to LLM-mediated information discovery represents both a challenge and opportunity for organizations. By understanding the fundamental differences in how LLMs process and present information, we can design systems that remain discoverable and authoritative in this new paradigm.
The organizations that will thrive will be those that recognize this isn't simply about tweaking SEO strategies—it requires a fundamental rethinking of how information is structured, exposed, and maintained. Technical leaders who embrace knowledge-centric design principles will position their organizations for success in the rapidly emerging LLM-first search ecosystem.
Heads-up! Watch this space, as I will be exploring the potential of Salesforce Data Cloud in this context and how it can help in unifying data from different sources.
https://www.iloveseo.net/a-guide-to-semantics-or-how-to-be-visible-both-in-search-and-llms/
https://www.csiro.au/en/news/all/articles/2023/october/generative-thinking-ai
https://document360.com/blog/technical-writing-ai-guidelines/
https://www.bain.com/insights/goodbye-clicks-hello-ai-zero-click-search-redefines-marketing/
Enjoy my blog? For more such awesome blog articles - follow, subscribe, and let's connect.
Disclaimer: The views expressed in this blog are my own; AI was used solely for editing and spell-checking purposes, and image generation is done through a beta AI tool and Canva.