Real-time inputs for recommendation engines

Your recommendation engine doesn't know what happened in this visit.

Signals computes features from every customer interaction and serves them to your model at inference time, in 6ms, so the next recommendation reflects this visit.

  • 6ms p50 read, in-region
  • <1s event to attribute current
  • Free tier, no card
Teams running on Snowplow
Experian
1Password
Strava
HelloFresh
AutoTrader
Supercell

Signals, about you

connecting
What Signals knows about your visit, right now

Signals is computing your first attributes from the events this page just sent.

Every highlighted part is an attribute Signals computed from the behavioral data your browser sent during this visit. Open the Product menu above and watch the sentence change.

Keyed on an anonymous session id. Signals drops these values a day after your last event.
Why recommendations miss

Recommendations lag behind what the customer just did.

Today

Features computed before the visit

Whether your features refresh nightly or hourly, what the customer does during the visit rarely reaches the model in time.

session_events_used = 0
With Signals

Features current to the event

Each feature updates <1s after the event that moved it, so inference reads the session in progress.

categories_viewed_10m = 3
Result

The next recommendation reflects this visit

A shopper who moved from running shoes to rain jackets sees jackets in the next carousel.

same session, new ranking
What you get

What fresh features are worth.

None of this needs a better model.

Relevant recommendations from the first visit

Most sessions are new or logged out, so most get generic picks. With features keyed on the session, the model has something real to score from the second page view.

share of sessions with personalized recommendations

The lift you measured offline, in production

One definition builds the training set and serves live requests. The lift your offline test promised is the lift you ship, without a quarter spent tracing the skew.

offline-to-online lift gap

More clicks when intent changes

A customer who has switched category twice in four minutes wants something different from one still on the same product. Recommend for the visit they are having and they are more likely to click.

recommendation click-through, session conversion

No streaming pipeline to own

No Kafka or Flink job to build, tune and keep on call. Your team spends its time on the model, not the plumbing behind it.

engineering time on pipeline upkeep
How it works

One definition. Live at inference, correct in training.

The streaming engine keeps each feature current for serving. The same definition builds a point-in-time correct training set from your warehouse history.

  1. 01Collect. Add a Snowplow SDK to your app and events start flowing. No stream to stand up first.
  2. 02Define. Say what the model should know about the customer. Signals calls each feature an attribute. Define them in the console or in code, versioned through CI/CD.
  3. 03Serve. An event updates its features in <1s. Your model server reads them at 6ms p50 on every request.
Signals → your app6ms p50
categories_viewed_10m3
items_viewed_session7
seconds_on_product_page47
basket_value184.00
sessions_last_7d4
Trigger intent_shiftdelivered to your model server
Live attributes showing price hesitation and interest in trail gear feed an OUTPOST recommendation strip that now leads with a price match and free returns note and puts the trail runner first.
Works with
  • Snowflake
  • Databricks
  • BigQuery
  • Kafka
  • Vertex AI
Questions

Before you sign up

Is Signals a recommendation engine?

No. If you need a recommendation engine, this is not it. Signals feeds the one you have, whether you built it, bought it or run it on rules.

Will it fit inside my inference budget?

6ms p50 and 10ms p95 for an in-region round trip, and one call returns a whole service of features, not one at a time. It is a number with a boundary on it. Whether it fits your budget is your call.

How do I train on these features?

The dataset builder computes each feature from only the events before the moment you are predicting from, using the same definitions the live path reads. Nothing that happened afterwards leaks into training.

Do I need an event stream already?

No. Collection is part of Signals. Add a Snowplow SDK to your app and events start flowing.

What does the free tier include?

The full product on your own traffic, up to 5 million events a month. No card, no sales call. You should see an attribute updating against your own traffic in the first session.

See what your model would know about this visit.

Add the SDK, define one attribute, then browse your own product and watch what your model would read change as you go.

The real engine · no card · no sales call
Or hand it to your coding agent
npx plugins add snowplow/skills
Then: Let's add Snowplow Signals to this app. Understand how recommendations are served first, then define an attribute for categories viewed this session and read it back before the model scores.