What is a signal
A signal is the atomic unit of behavioral data — a single, raw, composable fact about an entity at a specific moment.
A signal is a single fact about an entity (e.g. a person) at a specific moment: something they did, something they own, something that happened to them, or something they were exposed to. A few examples from the Signal Graph:
- Searched buying a Peloton bike in the last 90 days
- Renewed a commercial driver's license in 2025
- Household with at least one child under the age of 5
- Lives within 25 miles of a Tesla service center
- Company hired its first VP of Sales
- Recently relocated from a single-family home to a multi-unit
- Researched bariatric surgery in the last 30 days
- Consistently reads The Information
None of those is a profile, a segment, a score, or an audience — those are downstream artifacts built out of signals. A signal is the atomic unit: the smallest piece of truth at a moment in time about an entity that's useful on its own and composable with others.
What a signal isn't
Signals are easy to confuse with three things that aren't signals: fields, traits, and segments. Each is a layer of pre-aggregation — a human deciding, in advance, what mattered.
| Term | Example | The limitation |
|---|---|---|
| Field | income | One bracket per person, fixed at write time. |
| Trait | income = "$100K–150K" | A human drew the brackets; everyone gets rounded into one. |
| Segment | Affluent households | A pre-built audience — target it, but you can't take it apart or recombine it. |
| Signal | a person's means, read many ways — household income, net worth, home value, investable assets — each its own signal, on its own cadence | None — there's no single "income." Compose the reads that fit the goal: a lender weighs income and credit; a luxury brand weighs net worth and discretionary spend. |
The difference is structural, not semantic. Fields, traits, and segments are pre-aggregated — the output of a one-way compression. Somebody decided which behaviors counted, which signals collapsed into which categories, and which dimensions made it into the schema. Whatever didn't fit the chosen shape is gone, and the downstream consumer never sees it.
A signal is raw. It's the thing the schema would have been built from, if anyone had known in advance which questions you'd ask.
That pre-determination is what makes pre-aggregated data hard for an AI to work with. It's both limiting — the model can only ask the questions a human already anticipated — and counter-productive — with the underlying detail aggregated away, there's no context left to reason over, so the work collapses from reasoning into retrieval.
What makes a signal useful
- Granularity — how specific is the signal? Owns a car is coarse: billions of people qualify. Owns a 2019 Brinkley Model G is fine-grained: a few thousand people qualify, and the audience is half-defined the moment the signal fires. Fine-grained signals do more work per signal — and covering the full range of what people actually do takes a lot of them.
- Freshness — when did the signal fire, and how recently was it refreshed? Renewed a commercial driver's license in the last five years is a different signal than in the last 30 days. Behavioral data decays fast; a signal you can't refresh gets less true every day.
- Lineage — where did the signal come from, who observed it, and what was the chain of custody between the original event and the version in the graph? Most vendors hide this. Every signal in our graph has a lineage we can name, and most have passed through fewer hands than the equivalent signal from a legacy aggregator — which means less compression and less staleness.
- Composability — can the signal be combined with others at reasoning time, without being pre-joined into a fixed schema first? A signal that lives in its own silo, reachable only through one vendor's API in their pre-chosen schema, isn't composable. A signal that lives in the same graph as 160,000 others, in a structure built for an agent to traverse, is.
Why one signal is never enough
A single signal is noisy. One that's right 30% of the time can't carry a decision on its own — too many false positives. But stack several of those with the right boolean composition and the noise cancels: each signal you add narrows the cone of possible interpretations, until what's left at the intersection is genuinely predictive of what you're after.
What separates Signal Engineers: they ask what can I find?, not what data can I get? The first composes an answer out of many partial signals; the second shops for one source that already holds it. Signal Engineers find alpha; the rest of the market licenses a commodity.
The result is a composition: a boolean expression of signals, often with weights and exclusions. It describes something no single signal could express, and no vendor's catalog would have sold. The graph indexes the raw signals; the work is composing them.