Product analytics
Analytics that shows its work.
Ask a question in plain English and Tangent hands back the answer and the exact query it ran — loaded into the builder, ready to change one filter and run again. Nine SDKs, one event contract, and a warehouse you can audit line by line.
Free plan, no card · your raw events never reach a model
Who it is for
Six people, one event stream.
Most teams end up with a product analytics tool, a replay tool, a flagging tool and a spreadsheet nobody trusts. Tangent is one stream feeding all of it, so two people asking the same question get the same number.
Product managers
Find out which release, cohort or surface moved the metric — without briefing anyone or waiting a sprint for a dashboard.
“Did the new onboarding lift week-two retention?”
Growth & marketing
Follow a campaign from first touch through activation to revenue, with acquisition data resolved at the moment of arrival.
“Which channel brings users who are still here in 30 days?”
Engineers
One documented event contract, nine SDKs, and flags and experiments in the same place you measure them. No bespoke tracking layer to maintain.
“Is the rollout degrading checkout for anyone?”
Support & CX
Open the session behind a ticket. See what the customer actually did, with inputs masked, instead of asking them to describe it.
“What happened just before they gave up?”
Founders & leadership
Scheduled reports and alerts that arrive on their own, so the weekly number is not a task somebody has to remember.
“What changed this week, and why?”
Data teams
A warehouse you can query directly and export from. Tangent is a layer over your events, not a place they disappear into.
“Can I reconcile this against our own tables?”
Ask in English
Type the question. Read the SQL.
Write what you would have asked a data analyst. Tangent turns it into a query, runs it on your own warehouse, and puts the generated SQL on screen. An answer you cannot check is not an answer — so every one arrives with its working attached.
> Which countries drove the drop in checkout completion last week?
SELECT country,
uniqExact(user_id) AS users,
countIf(event_type = 'checkout_completed') / users AS rate
FROM events
WHERE tenant_id = {tenant}
AND event_time >= now() - INTERVAL 7 DAY
GROUP BY country
ORDER BY rate ASC- 01It refuses rather than guesses. If a generated query names an event or property your workspace has never seen, Tangent declines to run it and tells you which one — instead of returning a confident, wrong number.
- 02The query is yours. Every answer loads straight into the visual builder. Change a filter, widen the window, break it down differently — no need to start the sentence again.
- 03Your events stay put. What leaves your workspace is the shape of your data — event and property names — plus your question. The query runs here, on your own rows.
- 04Or bring your own model. Use ours, your provider account, or a model inside your own network. Same feature, three places it can run.
Root-cause analysis · no model involved
Why did this move?
One click on any metric scans your breakdown dimensions — country, device, platform, release, plan, campaign, referrer — and ranks the segments carrying the change. It runs entirely on your warehouse, so it is included on every plan and nothing leaves.
| Segment | Share of change | q-value |
|---|---|---|
| Safari 17 · iOS | −41.2% | 0.003 |
| Checkout v4.2 rollout | −22.8% | 0.011 |
| Paid search · brand | +6.1% | 0.28 |
| Enterprise plan | — | too small to assess |
It tells you what it cannot know
- →Scanning many segments invents false positives. Test two hundred at a five-percent threshold and about ten look meaningful by luck alone. Tangent corrects every p-value across everything it tested and shows you the corrected figure, next to how many would have fooled you.
- →Thin segments are not tested. Below the minimum size a segment is listed as too small to assess, with the counts that made it so. No statistic is computed from too little data.
- →"Nothing shifted" is a finding. When no dimension's mix changed, the headline says the metric moved everywhere at once — look for a release, an outage, a campaign, a tracking change. Not for a segment that isn't there.
- →Correlation is labelled as correlation. Observational data can rank suspects; it cannot convict one. Every finding is marked exploratory — on the panel, on each chip, and in the caveats beneath. A screenshot cannot lose the warning.
The product
Follow one user the whole way through.
Every analysis below runs on the same event stream, so a cohort you define at acquisition is still the same cohort when you look at revenue. The stages are how people move through your product — not how the software happens to be organised.
Arrive
Where they came from, resolved on arrival rather than reconstructed later.
- Acquisition & attribution
- Campaigns and referrers
- Geography
Activate
Whether they reach the point where your product becomes useful.
- Funnels
- Paths
- Session replay
- Heatmaps
Return
Who comes back, who quietly stops, and what separated them.
- Retention
- Segments & cohorts
- Accounts & groups
- Experiments
Pay
Revenue attached to the behaviour that produced it.
- Revenue analysis
- Feature flags
- Alerts
- Scheduled reports
Define a segment once and every chart accepts it. Ask a question in English and it can use it too. Nothing here needs a data team to model it first.
How it works
Five hops, and none of them lose your data.
Worth knowing before you commit a product's telemetry to it: this is the whole path an event takes, and what happens at each step when something fails.
- →Failures are visible, not silent. An event that genuinely cannot be processed is kept and surfaced with the reason, so "the numbers look low" has an answer.
- →Recovery does not depend on backups. Because the log is replayable, most storage incidents are resolved by re-reading it — losing nothing at all.
Install
One contract, nine ways in.
Every SDK speaks the same event protocol, so a property you send from Swift means exactly what it means from Python. Offline queueing, batching and retries are built in — a dropped network does not become a hole in your data.
Bring your own model
Three places the AI can run.
Some teams have an approved AI vendor. Some cannot let anything touch a third party at all. The feature is the same in every case — only the address changes.
Tangent-hosted
Works the moment you sign up. A pinned model version, so answers do not change under you without notice. We pay for the inference.
Your own key
Point Tangent at your provider account — OpenAI-compatible or Anthropic. Your vendor, your terms, your bill. We test the connection with a real question before saving it.
Your own endpoint
A model inside your own network — vLLM, Ollama, TGI, Azure OpenAI, anything OpenAI-compatible. Nothing about your schema leaves your infrastructure.
On every tier, your keys are encrypted at rest, never returned by any API once saved, and never written to a log. Root-cause analysis uses no model at all and is included everywhere.
Built to be checked
Boring where it counts.
Analytics is infrastructure. It should be dull, durable and legible — especially when somebody asks where a number came from.
- →Nothing is discarded quietly. Events queue on the client, buffer in the pipeline, and retry until they land. Anything genuinely unprocessable is kept and shown to you, not dropped.
- →Isolation is structural. Every read and write is scoped to one workspace and one project, and a request that arrives without scope is refused rather than widened.
- →Privacy by default. Replay masks inputs, IP addresses are anonymised, and consent state travels with the event rather than being applied afterwards.
- →Answerable. Single sign-on, two-factor, granular roles, and an audit trail covering every administrative change — including who changed which limit, and when.
Check it yourself
We would rather be verifiable than impressive.
Tangent is new, so we are not going to show you a wall of logos and ask you to infer quality from it. Here is what you can inspect before you speak to a single person.
Documentation
Written for the person doing the install.
Every SDK page carries a working snippet, the full method surface, and the same event contract stated field by field. The ingest and read APIs ship as OpenAPI specifications you can generate a client from.
Pricing
Start free. Pay when it is load-bearing.
Priced on tracked users and events, not on seats — inviting a colleague to look at a chart should never cost anything.
Starter
Free
For individuals and explorers getting started
10,000 monthly tracked users · 1M events / month
- 30-day data retention
- 25 AI questions a month
- Report builder, up to 10 saved reports
- Analytics: Overview dashboard, Event stream and 14 more
- Data: Tracking plan, Deep analysis and 4 more
- Experience: Sessions & replays, Heatmaps and 1 more
- Operations: Scheduled reports, Metric alerts
- Growth: Experiments / A-B testing
- AI: Why did this move? (root-cause analysis)
- Apps: Release Impact
Plus
$61 / month
For small teams finding product-market fit
100,000 monthly tracked users · 10M events / month
- Everything in Starter
- 90-day data retention
- 500 AI questions a month
- Unlimited saved reports
- SQL explorer: 1M rows, 30 s per query
- Nightly Parquet sync to your own bucket
- API rate limit raised to 3,000 requests a minute
- Roles & permissions
- Data export
Growth
Custom
For businesses looking to scale
Custom volume
- Everything in Plus
- 1-year data retention
- 5,000 AI questions a month
- SQL explorer: 20M rows, 120 s per query
- Hourly Parquet sync to your own bucket
- Live BI connection for Power BI, Tableau and Looker (3 keys)
- API rate limit raised to 6,000 requests a minute
- Audit trail
- Churn risk
- Bring your own model key (AI)
Enterprise
Custom
For larger enterprises with complex requirements
Custom volume
- Everything in Growth
- 2-year data retention
- Unlimited AI questions
- Live BI connection for Power BI, Tableau and Looker (10 keys)
- API rate limit raised to 12,000 requests a minute
- Enterprise single sign-on
- Data residency
- Bring your own model endpoint (AI)
Find out what moved, before someone asks.
Two ways in, depending on whether you would rather read or talk.
Start free
One script tag or one import, and your first chart in about five minutes. The free plan includes analytics, session replay, heatmaps and AI questions — no card, no call.
Talk to us
For migrations from another tool, your own model endpoint, single sign-on, data residency, or a security review. We will tell you honestly whether Tangent fits yet.