Signals and intent topics
Everything a model reads is a named signal in one registry. The same registry feeds search, the workspace columns and the models, so a filter and a model always mean the same thing by the same name.
Events
Section titled “Events”An event is one person’s presence against one topic on one day. Search and navigation data are known to carry information about later behavior (Goel et al., 2010; Moe, 2003); see research note 1. Repeat visits on the same day count once, and raw activity that does not resolve to a person is dropped. As of July 31, 2026 there were 62.6B resolved events.
The feed is read daily and runs about a day behind real time.
The signal registry
Section titled “The signal registry”Each signal in the registry is named and defined, and can be queried by name.
| Kind | What it holds |
|---|---|
| Intent topics | Subjects people research, such as a vehicle category or a type of trip. 10,312 topics as of August 3, 2026. |
| Demographic fields | Person and household facts from the consumer graph. |
| Model scores | The output of each Gravity model, as a signal other tools can read. |
Signals retired in a release keep their IDs reserved and are not shown to clients, so a retired name is never reused for something else.
Intent topics
Section titled “Intent topics”Intent topics are organized in a tree of 26 categories. A topic is read with two qualities:
- Recency. How recently the person engaged with it. Recent days count more than older ones.
- Strength. How much engagement there was, relative to the population.
Search can filter on both, so “researched this in the last few weeks, heavily” and “has an ongoing interest” are different filters. See Search and sizing.
Habits and buying windows
Section titled “Habits and buying windows”Two kinds of behavior are read differently.
- Habits read repeated engagement over weeks. They are relative: a habit score says how a person compares with everyone else.
- Buying windows read an episode that started recently, is still warm, and expires. They are absolute: a person is either in the window or not, and the window closes on its own.
The split is consistent with habit research, which finds past behavior predictive in stable settings (Ouellette and Wood, 1998), and research on timing, which finds response falls quickly after a moment of interest (Sahni et al., 2019). See research note 2.
Coverage
Section titled “Coverage”A behavioral model can only score people whose browsing is visible to it. Coverage therefore differs by model and is printed on each model card. A person a model cannot see gets no score, never a low one.