Mapping Buyer Intent Graphs Into LLM Bidders

An intent graph encodes your buyer journey as nodes (states), edges (signal-bearing actions), and weighted transitions an LLM bidder reasons over — replacing keyword lists and audience segments as the structure paid media optimizes against. Mapping it correctly is the 2028 prerequisite for hosting an agentic buyer that actually performs.

Key takeaways

• Intent graphs encode buyer-journey nodes, edges, and transition probabilities — the same shape DeepMind uses for ETA prediction.
• DeepMind's Graph Neural Network work on Google Maps cut ETA errors by up to 50% in major cities, proving graph-structured reasoning beats flat feature sets.
• Sequoia's "reasoning era" thesis says inference-time reasoning unlocks long-horizon agents — graphs are the substrate that reasoning runs on.
• LLM bidders use the graph to pick creative, channel, and cadence in one decision instead of three.
• First-party data populates the graph; clean-room joins extend its reach by 2032.

What an intent graph is

A directed graph where nodes are buyer states ("researching", "comparing", "ready"), edges are signal-bearing actions ("downloaded benchmark", "viewed pricing"), and weights are transition probabilities learned from first-party data. The LLM bidder reads the current state of a user and decides what to show them, on what channel, at what bid.

DeepMind's research line on Graph Neural Networks is the clearest precedent. Their team showed GNNs "effectively encode combinatorial and relational input due to their permutation-invariance and sparsity awareness" (DeepMind: Combinatorial Optimization and Reasoning with GNNs), and applied the same primitives to real-time Google Maps ETAs with up to a 50% accuracy improvement in major cities (DeepMind: Traffic Prediction with GNNs). Paid media is the same kind of problem — a routing decision over a sparse relational state.

Why reasoning over a graph beats segment math

Sequoia's reasoning-era thesis is that inference-time reasoning — "thinking slow" — is what is unlocking the new cohort of agentic applications (Sequoia: Generative AI's Act o1). A reasoning bidder needs a structure to reason over; flat audience segments do not provide one. a16z's 2026 thesis adds the corollary: interfaces shift from chat to action, design shifts from human-first to agent-readable (a16z: Big Ideas 2026, Part 1). Agent-readable means graph-shaped.

And the bidder needs the right slice of that graph in its context window at the right time — Anthropic's context-engineering work is the operating discipline (Anthropic: Effective Context Engineering for AI Agents).

The four-step mapping (DeepMind-aligned)

1. Define the states. Five to nine nodes covering unaware → researching → comparing → ready → bought → retained.
2. Catalog the signals. Every action that moves a user between states — site events, ad interactions, CRM stages — labelled as edge types.
3. Compute transitions. Run historical first-party data through a transition-probability calculation; weight edges by recency and volume.
4. Publish the graph. Stream it to the warehouse view the bidder reads, refreshed at the cadence the platform's context window can consume.

How the bidder uses it

Given a user in state "comparing", the bidder knows the top three transition probabilities — to "ready" (via a case study), to "lost" (via no follow-up), to "researching again" (via price shock). It picks the creative, channel and bid that maximize the "ready" transition while respecting policy. This is the agentic-buyer pattern operating end-to-end; see the broader stack in our autonomous revenue OS blueprint.

2026 → 2040 evolution

In 2026, intent graphs are an internal artifact powering manual campaign decisions and feeding the post-pixel attribution stack. By 2028, they feed LLM bidders directly through the server-side tracking pipe. By 2032, clean-room joins extend the graph to off-site signal. By 2040, the autonomous revenue OS maintains and evolves the graph as a living document.

FAQ

Do I need a data scientist?

Helpful but not mandatory. The first version can be hand-built from CRM stages and site events. DeepMind-style GNN training comes later.

How does this interact with audiences in ad platforms?

Audiences become projections of the graph, not the source of truth. Each platform gets a derived segment list — but the graph in your warehouse is canonical.

Can this be done without first-party data?

No — first-party data is the substrate. Clean rooms extend reach; they do not replace the substrate.

Want the graph mapped for your account? Start with a free 48-hour audit or explore our paid advertising service.