Quick answer
A unified data model is a shared schema in which data from every platform — web, app, mini-program, server — uses one definition of a person and one definition of each event, so counts, funnels and retention can be compared in a single view.
What is a unified data model?
A unified data model is a shared schema in which every platform — web, app, mini-program, server — reports data using one definition of a person and one definition of each event. The key is not moving bytes into one place; it is agreeing on those two definitions. Get them right and data from all four sources can be compared in a single chart: one active-user count, one funnel, one retention curve. Get them wrong and a shared database just stores four incompatible reports side by side.
That distinction matters because most "we unified our data" projects stop at storage. The Segment Spec makes the two definitions concrete: an identify call ties a user to their actions through a unique user ID, and a track call records an action with a standard name and properties (Segment Spec: Identify). A unified data model is what you get when every platform in your stack follows that contract — not when every platform merely ships its rows to the same destination.
What has to be unified
Three things need one definition each. Miss any of them and cross-platform numbers stop being comparable:
| Layer | Before unification (typical) | After unification | What breaks if you skip it |
|---|---|---|---|
| Identity (who) | Web cookie ID, app device ID, mini-program openid, server customer ID — four ID spaces | One person ID; platform IDs mapped to it at login | The same person counts as up to four users; active users inflated |
| Events (what) | purchase, OrderPaid, pay_success — three names for one action; value in dollars vs cents | One event name per action, one property dictionary with fixed types and units | Event counts split across aliases; revenue off by 100× |
| Time & metric definitions (how counted) | Client local time vs server UTC; "active" = any ping vs a real event; bots mixed in | One timezone, one active-user rule, one deduplication rule, bots separated | Daily actives shift across midnight; denominators disagree between reports |
The identity layer is the hard one, because it requires a rule for merging IDs, not just a mapping table — see identity resolution for how anonymous and identified records get stitched, and identified vs anonymous users for what can and cannot be merged. The event layer is old data-warehouse wisdom: Kimball's conformed dimensions are "defined once … reused across fact tables" precisely so separate fact tables stay comparable. A cross-platform event schema is the same idea applied to behavioral data.
Microsoft's Customer Insights documentation describes the mechanics as four steps: map source columns to common types, deduplicate rows within a source, match records across sources, then build the unified profile (data unification overview). Whatever tool you use, those four steps are the work.
Unified data model vs data warehouse vs CDP
These three get conflated constantly. They sit at different layers:
| What it is | What it guarantees | What it does not do | |
|---|---|---|---|
| Unified data warehouse | Storage + compute for all sources | Data is in one place, queryable together | That rows mean the same thing across sources |
| Unified data model | Shared definitions of people, events, metrics | Counts and funnels are comparable across platforms | Store data or move it anywhere by itself |
| Customer data platform | Identity resolution + audience activation | Profiles can be synced to ad, email and CRM tools | Guarantee your analytics definitions are consistent |
Two practical rules follow. If your problem is "our reports disagree", you need the model, and a bigger warehouse will not fix it — that is a single source of truth problem at the definition layer, not the storage layer. If your problem is "marketing can't target buyers across channels", you need a CDP on top of a model you already have. Buying the CDP first, without agreed event definitions, activates audiences built on inconsistent data.
A worked example: one day across three platforms
Sample data. A subscription product tracks one day of activity on web, iOS and a WeChat mini-program, plus purchase confirmations sent from the server. Every buyer logged in on every platform they used, so ID merging is exact — no probabilistic guessing.
Step 1: count active users before and after unification. Per-platform reports count 1,000 (web) + 600 (iOS) + 400 (mini-program) = 2,000 daily active "users". But the platforms share people. The overlap breakdown, where each person is counted once:
| Segment | People |
|---|---|
| Web only | 600 |
| iOS only | 300 |
| Mini-program only | 200 |
| Web + iOS | 200 |
| Web + mini-program | 120 |
| iOS + mini-program | 30 |
| All three | 10 |
| Total distinct identified people | 1,460 |
Check against the platform reports: web = 600 + 200 + 120 + 10 = 930 identified, plus 70 anonymous-only visitors = 1,000; iOS = 300 + 200 + 30 + 10 = 540 identified, plus 60 anonymous = 600; mini-program = 200 + 120 + 30 + 10 = 360 identified, plus 40 anonymous = 400. The 170 anonymous visitors never logged in, so they stay separate — merging what cannot be matched would invent people.
Distinct daily actives are the 1,460 identified people plus the 170 anonymous visitors, each counted once: 1,630. Summing platforms overcounts by 2,000 − 1,630 = 370, a 22.7% inflation (370 ÷ 1,630; factor 2,000 ÷ 1,630 ≈ 1.23). This is the thesis in arithmetic: without one definition of a person, "daily active users" is not one number.
Step 2: reconcile purchase events. The platforms report 300 (web) + 200 (iOS) + 150 (mini-program) = 650 purchase events. Of these, 60 are client-side events the server also sent as confirmations — the same order, two records. After deduplicating on order ID: 650 − 60 = 590 purchases. Without a deduplication rule in the model, revenue reports differ by platform depending on which side you count.
Step 3: the funnel that only exists after unification. Of the 600 web-only people, all 600 viewed an item and 210 purchased: a web-only funnel reads 210 ÷ 600 = 35%. But many buyers browsed on web and paid in the mini-program. Across all platforms, 1,040 distinct people viewed an item (including all 590 buyers), so the true view-to-purchase rate is 590 ÷ 1,040 ≈ 56.7%. The web-only funnel understates conversion by 21.7 percentage points — and no amount of web-side optimization changes that, because the missing conversions happened on other platforms. That is why omnichannel analytics starts with the model, not with the charts.
Common mistakes
- Treating unification as an ingestion project. Rule: if your project plan ends at "all sources land in the warehouse", it is data integration, not a unified data model. The plan must also name the person ID, the event dictionary and the metric definitions.
- Merging IDs greedily. Merging on shared device or household attributes fuses family members into one person. Rule: merge on logged-in identity or an explicit alias event (the links between them form an identity graph); leave unmatched anonymous visitors separate, as the 170 anonymous visitors in the example above stay separate.
- Unifying names but not units. Two platforms both send
value, one in dollars and one in cents. The schema validates; the revenue is 100× wrong for one platform. Rule: the property dictionary must fix type and unit and currency, not just the field name. - Unifying once, then drifting. A new app version ships
checkout_v2while web keepspurchase. Rule: new events should enter a review queue before they count in reports — otherwise the model decays every release.
How to implement a unified data model
- Write the event dictionary first. One name per action, properties with fixed types and units, and the platform list each event applies to (web / iOS / Android / mini-program / server).
- Define the person. Decide the canonical ID (usually your login user ID), how platform IDs map to it, and what happens to anonymous visitors — keep them counted, just not merged.
- Fix time and metric rules. One timezone for day boundaries, one definition of "active", one deduplication key for server-confirmed events, bots excluded from human counts.
- Instrument all platforms against the same plan, including server-side events for anything involving money, so the client is never the only witness of a purchase.
- Validate arrivals, not just intentions. Check per-event arrival counts and required-property completeness per platform; route unknown events to a review inbox before they enter official definitions.
- Reconcile before trusting. Recompute one day the way the example above does: platform sums vs distinct people, raw events vs deduplicated events. If the two views cannot be explained by your merge and dedup rules, the model has a hole.
How TapCub unifies data across platforms
Take the example above into TapCub: three platform reports that sum to 2,000 actives, 650 purchase events that should be 590, and a web-only funnel that says 35%. TapCub is built around the two definitions this article is about: web, app, mini-program and server events land in one project and share one identity model and one event table.
- Register every platform in one project. Website domains, app IDs and mini-program IDs are registered as sources in All-platform analytics › Data › Data sources; the website snippet and the SDK wizard for the other platforms are on the same page. Send purchases from your server with the order number as
insert_id; retries with the sameinsert_idare deduplicated, and the server stays a witness for every order, which is the dedup rule Step 2 needed. - Define the person once. With identified mode on (Pro plan or above), call
tapcub.loginwith your user ID on every platform. All-platform analytics › Users › Identity map has a difference card that compares visitor, device and person counts for the selected dates, which is the 2,000-versus-1,630 reconciliation done for you, including how much of the gap is one person on several devices. Visitors who never log in stay anonymous; nothing is merged on a guess. - Define each event once, and watch for drift. In All-platform analytics › Data › Tracking plan, each planned event has its platforms (
web,ios,android,harmony,mp,server), properties with type and required flag, an owner and a status; Live verification (24 h) shows hits and required-property completeness. When an app release starts sendingcheckout_v2, it shows up as Unplanned in All-platform analytics › Data › Event inbox. It is still stored, so the inbox is where you spot the drift and decide to fix the code or block the name, not a filter that hides it from reports. - Read the numbers side by side. All-platform analytics › Overview › Platforms side by side lists users, new users, sessions, events, duration and errors per platform, with server events as their own row and a deduplicated total for active users. For the funnel, build view → purchase once in All-platform analytics › Behavior › Behavior analysis with person identity, so a buyer who browsed on web and paid in the mini-program counts as one conversion: the 56.7%, not the 35%.
The boundary: TapCub is not a CDP or a data warehouse. It does not sync audiences to ad platforms, and raw rows leave through an async export in All-platform analytics › Data › Data export. What it does is the core of a unified data model: one definition of a person and one of each event, so all four sources fit in one chart.
See also: What is data unification? Steps and a before/after example.
Sources
- Twilio Segment Docs, Spec: Identify — tying a user to their actions via a unique user ID. Checked 2026-11-15.
- Microsoft Learn, Data unification overview — Customer Insights — map, deduplicate, match, unify. Checked 2026-11-15.
- The Kimball Group, Conformed Dimensions — define once, reuse across fact tables. Checked 2026-11-15.
- Mixpanel, What is a unified data model? — industry definition reference (2019-11-12). Checked 2026-11-15.
Was this explanation clear?
We use your answers to fix and expand this page.
Couldn’t send. Please try again in a minute.
Thanks for the feedback. We’ll use it to improve this page.You’ve already sent feedback on this page. Thanks.

Product Analyst
Focuses on event analysis, conversion funnels, and retention, using clear metric definitions and structured analysis to explain product performance.