Customer Journey Analytics · Stitching
How many people visited us?
Three people visit the shop website in one week, from four devices, and one of them buys. CJA does not count people: it counts the ID left on each event after stitching. Change the method and see how many people you get, and what conversion rate.
Example assumptions
Example people, devices, times and events. The split of events on shared devices is simplified.
Customer Journey Analytics · data view
What CJA counts
People and conversion against what really happened: 3 people and 1 purchase.
Web dataset · event by event
Which ID each event ends up with
One row per device and one column per day. Each event shows the person ID it carries and the ID stitching leaves on it.
Architecture
Where it fits in the platform
Comparison between the identity graph and Customer Journey Analytics stitching, top to bottom in five stages. Data: web events from the Web SDK with the ECID, plus the CRMID on login, and a CRM in batch with CRMID and email. Data Lake: events and records land in their datasets, with the primary identity flagged as primary=true. Identity: the Data Lake passes identities to Identity Service, which links them in a graph with its Linking Rules. Separately, the event datasets reach the CJA connection, which sets one person ID per dataset. Resolution, to activate: the graph reaches Profile, which merges in real time according to its merge policy, and the merged profile goes out to AJO and Destinations to act. Audiences are evaluated on that merged profile under its merge policy, which uses the graph to gather related identities. Identity settings cannot be switched off without resetting the sandbox. Resolution, to measure, by two routes. Field-based stitching, from CJA Select, takes the persistent ID and the person ID from the same dataset. Graph-based stitching, from CJA Prime, takes the persistent ID, looks it up in the graph and gets back one namespace. It ignores timestamps and inherits graph quality. Use: both deliver one person ID per row to Workspace, for People and attribution, and replay restates history inside the lookback window. The shared key: the top-priority unique namespace in the graph should be the connection person ID. CJA audiences published to Profile send that person ID, and if no profile has that identity, a new one is created.
The recommendation
Choose stitching by the question, and do not compare people across data views that stitch differently
No method gives the real number. Each one is wrong in its own way: too high, by counting devices, or too low, by merging people who share one.
- 01
Identifiers on events
The more often an event carries the CRMID, the less stitching has to guess. Login is the best measurement tool.
- 02
A window that fits behaviour
If people take days to identify themselves after arriving, a 24-hour window leaves almost everything out. Choose the window from how long people really take to log in.
- 03
Watch shared devices
With graph-based, a device with two people in the graph goes to one of them. Review the Identity Graph Linking Rules and namespace priority before using it to count people.
- 04
Normalise the ID
Stitching is case-sensitive: c-1042 and C-1042 are two people. Normalise the ID before it reaches Experience Platform.