Data & Decisions
Identity Merging Changed Your Analytics Counts: Check the User Mapping
Check how anonymous and known identifiers map to reporting profiles before treating a count change as growth. A completed eight-event reconciliation and separate missed-link and incorrect-link cases preserve raw events, known-user boundaries and unresolved evidence.
When identity merging changes an analytics count, reconcile the same events under the old and new user mappings before interpreting the movement as growth. Check which anonymous identifiers link to known users, what evidence supports those links, and whether the dashboard applies a mapping that the raw export does not. Fewer profiles or more completed funnels can reflect changed attribution rather than changed behavior.
This LandOffer.ai guide includes a completed fictional identity reconciliation and two investigation records, supported by an executed local fixture. No analytics project, profile, login, Identify call, merge, reset or historical data correction was performed. All identifiers and authentication-link evidence below are synthetic; the fixture is a local grouping model, not an implementation of a provider's identity engine.

Name the identifier and the reporting view
A device identifier, an authenticated user identifier and an event identifier have different jobs. One account may use several devices, several people may share a device, and one person may generate many events. For each “users” count, record the identifier and mapping version that produced it.
Amplitude's unique-user documentation describes combining device IDs, user IDs and Amplitude IDs into profiles. It recommends stable user IDs for authenticated people and leaving anonymous users without a user ID. These documented rules are product-specific; they do not prove that your current implementation sends the intended identifiers.
The same documentation explains that merged-ID mapping is applied to dashboard queries, while its described raw Redshift data does not receive that mapping. A raw-ID count and a dashboard profile count may therefore differ without any event being added or removed. Verify the actual export and query path instead of assuming every warehouse applies the same transformation.
For fictional PatternRoom, an internal sample-library portal, analyst Sora compares a legacy report that groups raw identifier keys with a proposed reviewed mapping. Engineer Davi owns the link evidence and implementation investigation. Both reports use the same eight selected events and the same observation window.
The raw view prefers a supplied user ID; otherwise it groups by an anonymous device token. That is an authored warehouse-view rule, not a claim that Amplitude leaves the example's sign-ins unmerged. Likewise, the reviewed local profiles are not actual Amplitude IDs or a count of independently verified human beings.
Complete an event-to-profile reconciliation
The following selected traces all occur on one fictional day. Times are included to check ordered paths; no late-arrival, timezone or session-boundary change occurs in this example. The eight rows are deliberately chosen diagnostic inputs, not representative traffic for estimating real conversion.
| Event | Time | Action | Device / supplied user | Raw key | Reviewed mapping |
|---|---|---|---|---|---|
| E1 | 09:00 | Viewed | D1 / none | anon:D1 | user:U1 |
| E2 | 09:05 | Started | D1 / U1 | user:U1 | user:U1 |
| E4 | 09:30 | Viewed | D2 / none | anon:D2 | user:U1 |
| E3 | 10:00 | Completed | D2 / U1 | user:U1 | user:U1 |
| E5 | 11:00 | Viewed | D3 / none | anon:D3 | user:U2 |
| E6 | 11:05 | Completed | D3 / U2 | user:U2 | user:U2 |
| E7 | 12:00 | Viewed | D4 / none | anon:D4 | anon:D4 |
| E8 | 13:00 | Completed | D5 / U3 | user:U3 | user:U3 |
Sora's completed link register accepts D1→U1 and D2→U1 because the fictional authentication receipts associate those anonymous histories with U1. It accepts D3→U2 on equivalent authored evidence. D4 has no accepted link. The source events retain their original fields; the reviewed mapping is a separate versioned interpretation.
These receipts are assumptions supplied to the teaching case, not authenticated accounts examined by this guide. In real work, the permitted link evidence, its timing and the validity of historical attribution require an approved identity policy. Sharing a network, a device description or a similar display name is not enough to establish the link in this example.
The raw-key view has seven distinct keys. The reviewed view has four profiles: U1, U2, U3 and still-anonymous D4. Both retain the same eight unique event IDs and unchanged source-event bytes, as checked by the saved fixture hash. Reporting this as “three users disappeared” would confuse the raw counting keys with verified people and obscure the mapping change.
Recompute the ordered path without claiming growth
The diagnostic path is Viewed followed later by Completed within this selected day, grouped by the reporting identity. Started is retained as context but is not a required step. Changing that path definition while changing identity would introduce a second explanation for the result.
In the raw-key view, four anonymous keys have Viewed events, while three known-user keys have Completed events. No raw key connects both steps, so the ordered-path count is zero. After reviewed linking, U1 and U2 each have an earlier view and a later completion. D4 has only a view; U3 has only a completion.
| Reconciliation measure | Unstitched raw keys | Reviewed mapping | Meaning |
|---|---|---|---|
| Unique event IDs | 8 | 8 | No event growth, loss or deduplication change |
| Reporting keys or profiles | 7 | 4 | Anonymous histories consolidate under accepted links |
| Identities with Viewed | 4 | 3 | D1 and D2 views now belong to U1 |
| Identities with Completed | 3 | 3 | Known completers remain U1, U2 and U3 |
| Ordered Viewed→Completed paths | 0 | 2 | Previously separated steps now connect for U1 and U2 |
The reconciliation explains the reporting change; it measures no improvement in product performance. Neither the interface nor the selected event behavior changed. The count moved because the report connected observations that its earlier raw-key rule separated.
Preserve when each relationship became known as well as which historical events the report attributes through it. A report evaluated with the latest map can answer a different question from one restricted to identity evidence available at an earlier cutoff. PatternRoom deliberately compares both versions over the same selected events; it does not claim that the accepted links were known at every event's occurrence time. If a business decision requires an as-known-then view, keep that temporal mapping rule explicit rather than silently applying today's links to yesterday's report.
A real funnel additionally needs its conversion window, repeated-step rules, identity scope and exclusion settings checked. Save the query and mapping version with each screenshot so a changed link can be separated from a changed event filter or reporting window. A different event filter or observation cutoff could move the result independently of merging.
Investigate a missed link and an incorrect link separately
Completed missed-link record. “Davi omits D3→U2 from local mapping v2-missing. E5 remains anon:D3 and E6 remains user:U2. The model now has five profiles and only one ordered path, U1. Sora identifies the missing relationship by comparing E5's mapping with the accepted link register, not by assuming every anonymous view needs a completion.”
Completed incorrect-link record. “A deliberately faulty custom rule attaches D4→U3 using a shared-network hint that the approved register does not accept. E7's view is now attributed to U3, whose E8 completion occurs later. The model falls to three profiles and reports three ordered paths. Sora rejects this proposed link and keeps D4 anonymous pending appropriate evidence.”
The incorrect example does not merge two distinct known user IDs. U1, U2 and U3 remain separate in every model version. Amplitude explicitly documents that it cannot merge distinct user IDs; supplying a new known user ID for an existing person creates a different recognized user. Do not turn an offline grouping rule into a claim that the provider supports arbitrary known-user consolidation.
The two defects push counts in different directions. Missing a legitimate link can leave duplicate profiles and fragmented paths. Accepting an unsupported link can combine unrelated observations and invent a path. “Make the unique-user count lower” is therefore not a valid repair criterion.

Keep property changes and identity changes distinct
An Identify operation can also update properties such as a plan label. Amplitude's Identify API documentation states that property updates apply to future events rather than retroactively updating earlier events, and that Identify calls are not events themselves. Those statements should not be confused with dashboard query-time merged-ID mapping.
Changing a property called “customer type” is not a demonstrated repair for an anonymous-to-known identity link. Likewise, observing a different historical user count does not prove that the raw event payloads were rewritten. Inspect the specific mapping, property and query mechanisms involved.
Twilio Segment's Identify specification distinguishes anonymous identifiers from persistent user IDs and recommends stable database identifiers rather than changeable email addresses or usernames. Its call syntax and destination behavior belong to that system. This guide sends no Segment call and assumes no automatic cross-provider synchronization.
For PatternRoom, Davi records the source identifier fields, the link-register version, the affected event IDs and the metric query version separately. He does not substitute real email addresses or other personal identifiers into this sample. An internal investigation should use only the evidence and access authorized for that purpose.
Finish with a bounded correction decision
Sora's completed decision reads: “Accept the reviewed local map containing D1→U1, D2→U1 and D3→U2 for the saved diagnostic dataset. Reject D4→U3 because its evidence does not meet the case's policy. Preserve raw events and all mapping versions. Describe the seven-to-four profile change and zero-to-two ordered-path change as attribution reconciliation, not product growth.”
Davi's next-step record remains open: “Validate the deployed identifier fields and approved link timing through the authorized process. Compare the actual dashboard's merged-user view with the relevant export. Confirm shared-device and logout handling using the applicable SDK guidance before considering any real change. Obtain provider-specific repair options rather than assuming historical merges can simply be undone.”
The decision avoids promising an account reset, historical unmerge or immediate dashboard convergence. Correcting a local lookup table is reversible in this fixture; changing a production identity system can have different persistence, processing and recovery constraints. None of those product actions is performed here.
Preserve conflicting evidence rather than forcing every event onto a known user. D4 remains anonymous in the accepted case. Keeping D4 anonymous preserves the rejected-link decision in the map and prevents its view from being attributed to U3 in this accepted case.
Preserve the reproducible evidence and its limits
The saved Python fixture executes 24 assertions over the same eight synthetic events. It checks event preservation, raw and reviewed profile counts, ordered paths, the missed-link and incorrect-link branches, deterministic replay and unchanged raw-event bytes. It also checks that all three supplied known user IDs remain distinct.
The model does not reproduce an analytics provider's full identity graph, Amplitude event-ID semantics, authentication system, SDK lifecycle, cross-domain tracking or session rules. Its evidence is the stated local mapping and ordered-path calculation. No production data, consent configuration, account property, destination or dashboard was accessed or changed.
For more practical work guides, visit LandOffer.ai. Before describing a user-count jump, give the metric owner the unchanged event set, the old and proposed maps, the accepted and rejected link evidence, and a versioned reconciliation. Record the owner and missing evidence for unresolved links. Describe any accepted count revision as an attribution change, and assess growth separately with a consistent measurement definition.
Sources and Further Reading
- Amplitude: Track unique users — identifier relationships, merged dashboard mapping, raw Redshift distinction and known-user-ID limitation; freshly reviewed October 27, 2026.
- Amplitude: Identify API — property-update timing, identity considerations and Identify calls distinct from events.
- Twilio Segment: Identify specification — anonymous and persistent user identifiers, with product-specific call semantics.


