DeepMind Insights: How LLMs Actually Work for Distribution

DeepMind Insights: How LLMs Actually Work for Distribution

Pierre-Alexandre Kamienny

Pierre-Alexandre Kamienny

Friday, Nov 15, 2024

During my time at Google DeepMind working on Gemini's information retrieval systems, I gained unique insight into how LLMs actually make recommendations. These insights are now powering Kinro's approach to financial services distribution.

The Retrieval-Augmented Generation Architecture

When you ask ChatGPT or Gemini about financial services, the process is more complex than most people realize:

Stage 1: Query Understanding

The model doesn't just parse your words—it infers:

  • True intent - What you're really trying to accomplish
  • Risk tolerance - Implicit signals about decision factors
  • Context - Life stage, financial situation, urgency
  • Constraints - Budget, timeline, requirements

Example: "Best home insurance for young professionals in San Francisco"

The model understands:

  • Urban location = earthquake risk
  • Young professional = price-sensitive, tech-savvy
  • "Best" = optimization across multiple factors
  • Implied constraints: digital-first, flexible, modern

Stage 2: Information Retrieval

This is where most financial services companies lose:

The model searches 20+ sources but most content fails because:

  • Gated information - Can't access content behind forms
  • Poor structure - Can't parse complex layouts
  • Outdated data - De-prioritizes stale information
  • Marketing speak - Prefers factual, comparative content

From my work on symbolic regression (283 citations, NeurIPS), I learned that models favor structured, mathematical representations of information. Financial services that provide this format win disproportionately.

Stage 3: Credibility Assessment

The model evaluates each source across multiple dimensions:

Authority Signals:

  • Domain reputation and age
  • SSL/security indicators
  • Regulatory mentions
  • Industry associations

Content Quality:

  • Factual accuracy (verified against multiple sources)
  • Completeness (answers related questions)
  • Recency (timestamp and update patterns)
  • Consistency (no contradictions)

Trust Markers:

  • Customer reviews (sentiment, not just volume)
  • Regulatory compliance statements
  • Clear terms and conditions
  • Transparent pricing

Stage 4: Synthesis and Ranking

This is the black box most companies struggle with. From my multi-agent research (469 citations, NeurIPS), I understand this as a consensus-building process:

Different "evaluation agents" assess:

  1. Best fit agent - Match to user's specific needs
  2. Risk mitigation agent - Safety of recommendation
  3. Value optimization agent - Price/feature balance
  4. Practical feasibility agent - Can user actually get this?

Only sources that satisfy all agents make the final recommendation.

Why Financial Services is Unique

High Stakes = Conservative Recommendations

LLMs are more cautious with financial services than other categories because:

  • Regulatory consequences - Wrong advice has serious implications
  • User trust - One bad recommendation breaks trust forever
  • Verification difficulty - Hard to fact-check financial claims

This means traditional marketing approaches actively hurt your chances:

  • Hyperbole = red flag
  • Vague claims = filtered out
  • Hidden information = disqualified
  • Aggressive positioning = downranked

Comparative Analysis is Mandatory

Unlike product recommendations where "best" is subjective, financial services recommendations MUST be defensible through comparison.

The model asks:

  • Why this option vs named alternatives?
  • What trade-offs is the user making?
  • Which user types prefer which options?
  • What are the hidden costs or catches?

If your content doesn't provide this comparative context, the model must infer it—usually to your disadvantage.

The Gemini Optimization Principles

From my work improving Gemini for financial services applications, I identified three core principles:

1. Verifiability

Every claim must be:

  • Checkable against other sources
  • Timestamped and current
  • Specific rather than general
  • Attributed when appropriate

2. Comprehensiveness

Content must address:

  • The main question
  • Obvious follow-up questions
  • Edge cases and exceptions
  • Alternatives and trade-offs

3. Accessibility

Information must be:

  • Publicly available (no gating)
  • Parseable (structured data)
  • Maintained (regular updates)
  • Discoverable (proper markup)

Practical Application: The Kinro Framework

At Kinro, we've translated these DeepMind insights into an operational framework:

Phase 1: Information Architecture

Audit:

  • Can an LLM access your key information?
  • Is it structured for machine parsing?
  • Does it answer comparative questions?
  • Is it current and maintained?

Optimize:

  • Expose critical data publicly
  • Add structured data markup
  • Create comparison-friendly formats
  • Implement regular update cycles

Phase 2: Credibility Building

Audit:

  • What trust signals are present?
  • Are regulatory details visible?
  • Do customer reviews exist?
  • Is pricing transparent?

Optimize:

  • Highlight compliance and licensing
  • Surface customer sentiment
  • Make terms and conditions clear
  • Show pricing with context

Phase 3: Competitive Positioning

Audit:

  • How do you compare to named competitors?
  • What makes you different?
  • Which use cases favor you?
  • What trade-offs exist?

Optimize:

  • Create explicit comparison content
  • Document differentiation clearly
  • Map products to use cases
  • Be honest about limitations

The Technical Edge

Most financial services companies treat AI distribution as a marketing problem. It's actually a technical problem that requires:

  • Deep understanding of LLM architecture
  • Information retrieval expertise
  • Structured data implementation
  • Continuous measurement and optimization

This is why Kinro's founding team combines:

  • Google DeepMind AI research (me)
  • Fintech founding and scaling experience
  • Deep financial services domain knowledge

What This Means for You

If you're optimizing for ChatGPT using traditional marketing approaches, you're fighting against how these systems actually work.

Success requires:

  1. Technical understanding - How do LLMs really function?
  2. Structural optimization - Make your information LLM-friendly
  3. Credibility building - Earn algorithmic trust
  4. Continuous measurement - Track actual impact

The financial services companies that win in AI distribution will be those that understand the technology at a fundamental level—not those applying old marketing playbooks to new platforms.

At Kinro, we're building the first true AI distribution intelligence platform for financial services, grounded in world-class AI research and real-world fintech experience.

The future of distribution is technical. Are you ready?