LeafprintTechnical guide
Deterministic vs. probabilistic identity resolution
These are the two main ways of deciding that separate signals belong to the same person. Understanding them makes it much easier to judge what a vendor is really selling.
Deterministic matching
Deterministic matching links records only when they share an exact identifier, such as the same email address or a login. If a visitor signs in with an address that is also in your customer list, those two records belong together. There is little guesswork, which is the appeal.
The limit is reach. It only works where a shared identifier exists, and that usually means people who have already logged in, signed up, or clicked a tracked link. Identifiers can also go stale. People share devices and email accounts, and a clean rule can still give a wrong answer.
Probabilistic matching
Probabilistic matching weighs indirect signals such as device type, browser, IP address, location, and behavior, then estimates how likely it is that they belong to one person. It can make a connection when no exact identifier exists, so it reaches further.
The cost is uncertainty. A system tuned to match more people will also make more mistakes, and a system tuned for caution will match fewer. Every probabilistic result is a judgment, so treat it as one.
How they compare
| Deterministic | Probabilistic | |
|---|---|---|
| Basis | Exact shared identifier | Statistical likelihood across signals |
| Certainty | Generally higher | Varies by model and threshold |
| Reach | Narrower | Wider |
| Good for | Connecting known customers | Understanding unknown traffic |
| Main risk | Stale or shared identifiers | False matches |
Most products blend both
In practice, few tools rely on one method alone. A vendor may use exact matches where it has them and fall back on modeled connections elsewhere. That is reasonable, but it makes a single accuracy figure hard to interpret. Results from the certain part and the uncertain part are often reported together, and you may not be able to tell them apart.
A simple example
Say a shopper logs in to a loyalty account on your site with the same email that is on your customer list. Connecting those records is deterministic, because an exact identifier matches. Now say a different visitor arrives on a phone with no login. A system may notice that the device, network, and location resemble a profile it has seen before, then estimate that this is probably the same person. That is probabilistic. The second connection may well be right, but nobody has confirmed it, so it deserves less trust than the first.
Questions to ask a vendor
- What share of your matches are exact, and what share are modeled?
- Can I see or filter by confidence level?
- How do you measure accuracy, and on what kind of traffic?
- How often does the underlying data get refreshed?
- Where does your identity data come from?
You may not get a complete answer to all of these, but how a vendor responds tells you a lot. For the wider list, see how to evaluate a vendor.
How to test it yourself
The most useful test is simple. Take a sample of matched profiles and compare them against people you know: existing customers, loyalty members, or staff who agree to help. Note how many fields are right, how many are partly right, and how many are wrong. Run the same check over a few different traffic sources, because quality often differs between them.
Reading a confidence score
If a tool shows confidence levels, use them. Reserve lower-confidence matches for broad pattern reading, and look harder at higher-confidence ones when you research a specific account.
Why it matters for how you use the data
The right level of certainty depends on what you plan to do. Using matches to judge which channel brings local shoppers tolerates some errors, because you are looking at patterns. Acting on a single record calls for much more care, since a wrong match can mean contacting the wrong person about a sensitive category.
In cannabis, that risk is especially real. Someone who receives a message about a dispensary they never visited may be confused or upset, and that person may be in a job or household where the message causes problems. Use matched data for insight first. Build direct communication on consent that people gave you themselves, as explained in our SMS consent article.
The takeaway
Neither method is simply better. Deterministic matching gives you confidence with less coverage, while probabilistic matching gives you coverage with less confidence. Know which you are getting, test it on your own traffic, and match your use of the data to how sure you really are. For the basics, read what identity resolution is or see how Leafprint works.
See it on your own traffic
We will walk through the portal, the pixel, and how consent works, then answer questions about your stores and systems.