What is the Signal Graph

The Signal Graph is a search index of raw signals, built for subsecond AI traversal. It indexes petabytes of structured signals in a shape an AI can reason over while it works — discovering, exploring, and retrieving in split seconds without flooding the model's context. One person with an LLM + the Signal Graph can perform the work of entire teams of data engineers, scientists, and analysts in minutes.

Reasoning time, not training time. Rather than work from what it absorbed in training, the model reaches into the index as it runs — pulling actual signals and the relationships measured between them, not recollection. The index recomputes daily, with roughly a third of its signals refreshing daily.

Raw signals, not pre-computed results

The Signal Graph indexes raw signals — the same signals data providers traditionally source, process, and aggregate into pre-computed results. A provider sells you the finished fields and segments, the same pre-baked answers your competitors buy; the graph serves the raw signals underneath, so you compose against your own intent. The closer you get to that raw, granular material, the more edge there is to find — big tech, hedge funds, insurers, governments have known it for decades and exploited it with resources no one else could match. The Signal Graph is our answer: the same raw material, open to anyone with an LLM.

That's how you find alpha — the defensible edge that comes from blending your subject-matter expertise, your first-party data, and raw signals into outputs that move your business. The craft of doing that is signal engineering, the subject of Learn.

The moat is the graph, not the sourcing

Anyone can buy the same raw signals we buy. The hard (and novel) part is structuring those signals so an AI can reason across all of them at once: without a schema, and without pre-aggregation into someone's idea of what matters.

Every other data product was built for a human consumer. A data warehouse assumes an analyst with a query in mind; an enrichment API, a single CRM field filled row by row. AI-ready vendors wrapping traditional data models in MCP servers work for basic retrieval, mimicking today's human behavior. But AI reasoning over petabytes of external data opens a whole new world of possibilities.

That's what lets a marketer chase an idea nobody anticipated and get back a 1.2M-row audience in a single session, with no data team in the loop.

Built for AI traversal

The graph runs on a novel database architecture built at Watt specifically for AI traversal of signals at enormous scale. Instead of a traditional entity-centric, schema-driven model, it uses a technique we call boolean-incidence: signals are represented inversely — by their relationships through shared membership, rather than as rows under a fixed schema.

That design makes the graph:

• Context-friendly — schema-less, self-discoverable access keeps the context footprint minimal.
• Composable — signals combine at reasoning time, not stitched together ahead of time.
• Fast — unbounded discovery, exploration, and retrieval in split seconds.

Signals attach to entities — the people and businesses they describe — and each signal is a field that changes over time.

Go deeper