Skip to main content
Exploring the New
Architecture of Search
Dawn Anderson – Bertey
SMX Advanced Boston 2026
RAG, Context
Graphs &
Agents
One might assume a vision of a future of agentic search...
Where agents autonomously collaborate to...
REDUCE THE
VOLUME ON ALL
HOME DEVICES
SIMULTANEOUSLY
RESEARCH SERVICES
ONLINE
SHORTLIST BASED
ON REVIEWS AND
RATINGS
BOOK
APPOINTMENTS
CHECK DIARIES
Could be Sundar Pichai
on 'The Future of
Search'... in April 2026
But that was Sir Tim Berners-
Lee who wrote of this 'Agentic
Search Future'... in 2001
25 years later...
Agentic web is
getting there
as the next
stage of AI
search...
Who is Dawn
Anderson?
• SEO practitioner for almost 20 years
• International SEO conference
speaker since 2017
• Boutique agency owner - Bertey
• Enterprise SEO consultant
• Information retrieval & AI search
world interloper
• Pomeranian lover and avid runner
;P
• Commissioned to write a book on AI
SEO currently
The current scenario is an architecturally hybrid blend of classic, agentic & AI
Search
The misconception
• LLMs don't search
• They generate (and guess)
Since LLMs
are simple
'next word'
guessing
machines
Currently LLM & AI Search
answers can feel
intelligent until they're
confidently wrong...
And an LLM alone is like a very shaky and wobbly jelly
LLMs
alone...
• Stale data
• Hallucinations
• Bias risk
• Guessing
• Overly generic
• Mistrust
• Reputation loss
• Legal claims
Reliable Agentic
& AI Search
MUST have a
'rock solid'
foundation
The 'Disambiguating Difference' is required
Hello... Retrieval Augmented Generation (RAG). Bridging
gaps in knowledge and context for LLMs
RAG (originally introduced
by Facebook Research
Team) is evolving quickly...
since 2020
Evolution of RAG
The RAG Paradigm
Naive RAG
• Keyword matching
• Lexical
• Lacks context
• Cheap and
cheerful
• Poor on intent
Advanced RAG
• Vector database
• Includes pre-
retrieval
• Includes post-
retrieval
• Semantic context
• Still stuck in time
• Greater context
but not recent
Modular RAG
• Hybrid model
• Cost control
(modular model
switching)
• Can combine
lexical RAG and
semantic RAG
• Reranking is
optional
Graph RAG
• Traverses
knowledge graphs
• Maps entity
relationships
• Context / intent
Agentic RAG
• Reasoning
• Iterative
• Multi-turn
• Learns when to
stop / when
enough info has
been received to
decide
From simple 'Naive RAG'… Keywordesque
Simple
RAG
Flow
To more 'Modular RAG'...
RAG
uses adaptations of
classic information
retrieval approaches
Meaning... To some
VERY large extent... It's still
just search engines and SEO
With Nuances...
Pre-Retrieval RAG Techniques
TECHNIQUE PURPOSE HOW IT WORKS
Query Expansion Improves recall by capturing synonyms.
Uses an LLM to generate multiple versions or related
terms for the user’s query to ensure no relevant
documents are missed.
Query Transformation Clarifies intent for better matching.
Re-writes a complex or poorly phrased user query into a
more "search-friendly" format that aligns with the vector
database.
Hypothetical Document Embeddings (HyDE)
Bridges the gap between query and
document.
The LLM generates a "fake" ideal answer to the query.
This fake answer is then used to search the database,
as it shares more semantic similarity with real answers
than a short question does.
Multi-Query Retrieval Reduces sensitivity to phrasing.
Generates several variations of a query from different
perspectives and retrieves documents for all of them,
aggregating the results.
Step-Back Prompting Handles complex reasoning.
The LLM generates a broader, more "high-level" query
(a step back) to retrieve context about underlying
concepts before addressing the specific question.
Query Routing Directs the query to the right source.
A classifier determines which specific database, tool, or
index (e.g., "Technical Docs" vs. "HR Policy") should be
searched based on the query.
Data Cleaning & Enrichment Removes "noise" from the source.
Involves stripping HTML tags, resolving acronyms, and
adding metadata (like headers or summaries) to chunks
before they are ever indexed.
Query Fan Out is a Form of Pre-Retrieval Technique
Post-Retrieval RAG Techniques
Technique Purpose How it Works
Reranking Optimizes relevance order.
Uses a more powerful (but slower) cross-encoder model
to re-score the initial search results, moving the most
"semantically rich" documents to the top.
Context Compression Reduces noise and saves tokens.
Identifies and extracts only the specific sentences or
paragraphs within a document that are relevant to the
query, discarding the "fluff."
Prompt Compression Maximizes efficiency.
Uses specialized algorithms to remove redundant tokens
from the retrieved context without losing the core
meaning, helping fit more info into the context window.
Context Filtering Removes low-quality matches.
Sets a similarity threshold; any retrieved document that
doesn't meet a certain "confidence score" is discarded to
prevent irrelevant data from confusing the LLM.
Context Reordering Combats "Lost in the Middle."
Strategically places the most important documents at the
very beginning and very end of the prompt, as LLMs tend
to ignore information placed in the center of long contexts.
Summarization Distills large volumes of data.
If too many documents are retrieved, an LLM creates a
concise summary of all findings to provide a "bird's eye
view" instead of raw text chunks.
Reality - Hybrid retrieval – Combined strength via dense + sparse + rules
BUT... RAG is a process, NOT magic
RAG & LLMs
STILL have
MANY
limitations
Documents Vs Chunks (PRECISION VS
NOISE BALANCES)
'Lost in The
Middle' -
Positional
Bias
Challenge in
RAG & LLMs
Freshness &
Recency Bias
LLMs Don't Know 'Now' – RAG
is also sensitive to latency
Most AI search
failures are still
retrieval
failures
You've got to be in it to win
it... You can't generate what
you can't retrieve
Relies on fast
and relevant
indexing and
updated content
indexed
Bloated index could possibly lead to... Curse of dimensionality
Possibly calls for increased 'dynamic index pruning'
Stale Retrieval == Wrong Answers
And... Semantic
blur -
Embeddings Aren't
Magic - Similarity is
NOT relevance
… You can't CONFIDENTLY
generate without STRUCTURE
Structure as
signal can
Help
Architecture,
Headings, lists,
tables, strong
internal
structures &
focused intent
Structure Aligns with Graph RAG
Structure is the Input to Context Graphs
CLEAR ENTITIES EXPLICIT
ATTRIBUTES
DEFINED
RELATIONSHIPS
UNAMBIGUOUS
BOUNDARIES
Relationships, Context & Entities Mapped to Augment
Contextual Relevance – Same Query,
Different Need
• Could this mean?
• DIY guides
• Find a plumber locally
• Parts needed for a leaky faucet?
Context Graphs vs Knowledge Graphs
Knowledge Graphs + Context Graphs – The 'What' + The
'Why'
Why
Context
Graphs
Matter in
AI Search
Context is not a single document
Chunking fragments meaning
Retrieval returns partial views
Context graphs restore coherence
Context graphs enable agentic action through structured
understanding
Context graphs power AI commerce
Meaning emerges across entities, attributes and
relationships
Agentic RAG is the Glue...
Agents don't just
answer. They ACT,
collaborate on user
intent... Then reflect and
improve
Universal
Commerce
Protocol (UCP)
& Agentic
Commerce
Protocol
Enable Agent-Merchant
Interaction at Scale...
Discover, compare, complete
purchases...
Why UCP and ACP
Matter to SEO?
Visibility now includes agent accessibility, not just
rankings
SEO in An Agentic Commerce World
Optimizing for agentic-commerce readiness
Structured,
agent-
friendly
product
meta-data
Capability
meta-data
(eg
available
actions)
High
precision
feed,
timeliness
& accuracy
Protocol
compliance
(UCP, ACP,
MCP)
Structure
optimization is
no longer
optional
• Structured data is retrieval infrastructure
• Schema is NOT rich snippets
• Structured data feeds retrieval
• Enables graph construction
• Tabular data, connected data is structured data too
• Additional product schema is NO coincidence
• Don't forget semi-structured data
• Context and knowledge graph building with internal link
graphs aligned with intent and context
Key
Actions
THIS is the
new search
stack
In reality...
Hybrid Search is
Likely The Norm...
"It Depends" is
sooooo true ;P
AI Search via
Agentic RAG is
closer than we
think... BUT AI
doesn't fail loudly,
it fails fluently
The 'Disambiguating Difference' is required
The AI SEO
Playbook on
pre-order
Thank you!