Skip to content
Fresh

Google AI Mode Architecture

Google AI Mode processes queries through a 4-stage pipeline: Prepare, Retrieve, Signal, Serve. The Signal stage applies 7 distinct ranking signals. This documentation is based on reverse-engineering the system's behavior through controlled queries, response analysis, and parameter observation.


The 4-Stage Pipeline


Stage 1: Prepare

The Prepare stage transforms the raw user query into a structured search intent that the retrieval system can act on.

Query Understanding

Google AI Mode parses the query into:

  • Intent classification (informational, navigational, transactional, local)
  • Entity extraction (people, places, products, concepts)
  • Temporal signals (is this about current events, historical facts, or timeless information?)
  • Complexity assessment (single-hop factual vs. multi-step reasoning)

Synonym Expansion

The system expands the query with synonyms and related terms. This is not simple keyword matching -- it uses embedding-based similarity to find semantically equivalent phrases. A query for "best laptop for coding" also retrieves results for "developer notebook," "programming computer," and related terms.

Autocomplete Signals

Google's autocomplete data feeds into query understanding. Popular completions for the query prefix signal common user intents and help disambiguate vague queries.

Schema Processing

If the query targets structured information (recipes, products, events, FAQs), the Prepare stage identifies the expected schema type and flags it for structured extraction in later stages.


Stage 2: Retrieve

The Retrieve stage fetches and chunks web content for the ranking pipeline.

Chunking at 500 Tokens

Retrieved documents are split into ~500-token chunks. This is the unit of ranking -- Google AI Mode does not rank entire pages. It ranks passages.

Chunk Boundaries Matter

If your key answer spans a heading boundary or falls at the end of a long paragraph, it may be split across two chunks -- weakening its signal. Structure content so that each section under a heading delivers a complete, self-contained answer within ~500 tokens (roughly 350-400 words).

Heading Context Injection

When a chunk is extracted from the middle of a page, the system injects the parent heading hierarchy as context. A chunk from under "H2: Installation > H3: macOS" gets that heading chain prepended so the ranking model knows what section the content belongs to.

Implication: Your heading hierarchy is not just for human readers. It is metadata that travels with every chunk into the ranking pipeline. Descriptive, keyword-rich headings directly improve chunk relevance scoring.

Layout Parser

Google's layout parser identifies and separates:

  • Main content vs. sidebar/navigation
  • Article body vs. comments
  • Primary content vs. ads and promotional blocks

Content in the main body gets retrieved. Sidebar content and navigation are typically excluded. This means your key information must live in the primary content area, not in widgets or sidebars.


Stage 3: Signal

The Signal stage applies 7 ranking signals to each chunk. These signals are combined into a final relevance score.

Signal 1: Base Score

The foundational relevance score derived from traditional ranking factors -- query-document match, domain authority, and page-level quality signals. This is the starting point before the AI-specific signals are applied.

Signal 2: Gecko Embeddings

Gecko is Google's lightweight embedding model. It converts both the query and each content chunk into dense vectors and computes semantic similarity. This captures meaning-based relevance that keyword matching misses.

What Gecko rewards:

  • Content that uses natural language aligned with how people ask questions
  • Semantically complete answers (all key concepts present)
  • Clear, direct statements over verbose explanations

Signal 3: Jetstream (Natural Language Inference)

Jetstream is Google's NLI model. It classifies the relationship between the query and each chunk as:

  • Entailment (the chunk supports/answers the query)
  • Contradiction (the chunk contradicts the query premise)
  • Neutral (no clear relationship)

Jetstream Negation Handling

Jetstream specifically handles negation queries well. If a user asks "Does X cause cancer?" and your content says "X does not cause cancer," Jetstream correctly classifies this as entailment (your content answers the question), not contradiction. This is a major improvement over keyword-based systems that would penalize the word "cancer" appearing near a negation.

Signal 4: BM25 (Lexical Matching)

Classic BM25 term-frequency scoring. Despite all the neural ranking signals, lexical matching still matters. Exact keyword matches in your content boost this signal.

Practical takeaway: Do not abandon keyword optimization for AEO. BM25 is one of seven signals, and it rewards exact-match terminology.

Signal 5: Engagement Tiers

Google AI Mode uses a three-tier engagement system:

TierSignal SourceWeightWhat It Measures
Tier 1: HighClick-through + dwell time > 60sHighestUsers chose this result AND found it valuable
Tier 2: MediumClick-through + dwell time 15-60sModerateUsers clicked but left relatively quickly
Tier 3: LowImpressions without clicks, bounces < 5sLowestUsers saw it but did not engage

Engagement is Measured at the Chunk Level

The engagement signal is not just page-level. Google tracks which sections users scroll to and how long they spend in specific content areas. A page with high overall engagement but poor engagement in one section will see that section's chunks scored lower.

Signal 6: Freshness

Freshness scoring varies by query type:

Query TypeFreshness WeightExample
Breaking newsVery high"earthquake today"
Trending topicsHigh"new iPhone review"
SeasonalModerate"tax filing deadline 2026"
Evergreen informationalLow"how does photosynthesis work"
HistoricalNone"when was the Eiffel Tower built"

The system uses datePublished and dateModified schema markup, HTTP Last-Modified headers, and content-level temporal signals to determine freshness.

Signal 7: Boost / Bury Rules

Hard rules that override the combined signal score:

Boost triggers:

  • Official source for the query topic (e.g., IRS.gov for tax questions)
  • Government/institutional authority for regulated topics (YMYL)
  • Verified authorship with E-E-A-T signals

Bury triggers:

  • Known misinformation domains
  • Thin content / doorway pages
  • Excessive ad density
  • Content that triggers safety classifiers

Stage 4: Serve

The Serve stage takes the ranked chunks and generates the AI Mode response.

Search Type Mapping

The system maps the query to a response format:

Query TypeResponse Format
Factual / single answerDirect answer with 1-2 citations
ComparisonStructured comparison with multiple citations
How-toStep-by-step with citations per step
Opinion / subjectiveMultiple perspectives with balanced citations
LocalMap + local results + AI summary
ShoppingProduct cards + AI comparison

LLM Configuration Selection

Different query types trigger different LLM configurations:

  • Temperature varies by query type (lower for factual, higher for creative)
  • Max tokens varies by complexity
  • Citation density varies by topic sensitivity (higher for YMYL)

Safety Gates

Before the response is served, it passes through:

  1. Factuality check against known knowledge bases
  2. Harmful content filter
  3. YMYL sensitivity screening (elevated standards for health, finance, legal)
  4. Attribution verification (cited sources actually support the claims)

Optimization Implications

Content Structure

  • Write self-contained sections of ~350-400 words under descriptive H2/H3 headings
  • Place the direct answer to the likely query in the first sentence of each section
  • Use heading text that matches how people phrase questions

Signal Optimization

SignalHow to Optimize
Base ScoreStandard SEO: authority, relevance, quality
GeckoUse natural language that matches query phrasing
JetstreamMake clear, direct statements. Handle negation explicitly.
BM25Include exact-match keywords naturally
EngagementDeliver value fast. Reduce time-to-answer.
FreshnessUpdate content regularly. Use dateModified schema.
Boost/BuryBuild E-E-A-T signals. Avoid thin content.

Chunk-Level Thinking

Stop thinking about page-level optimization. Start thinking about chunk-level optimization. Each 500-token section of your content is independently ranked. A page can have some chunks that rank well and others that don't. Optimize every section as if it were a standalone answer.

The Shift

Google AI Mode does not rank pages. It ranks passages. Your optimization unit is no longer the URL -- it is the section.