Identity & Addressability
Part of Advertising identity and addressability
Comparing deterministic and probabilistic audience matching
Compare direct and inferred audience matches using eligibility, match coverage, validation quality and the cost of wrong links.
Deterministic matching joins records through a directly observed attribute or agreed key. Probabilistic matching infers a link from a pattern of signals. Compare them by the identity level needed, eligible coverage and consequences of a wrong or missed link. Neither label is an accuracy score, and a provider may use both methods in one audience.
Identify the link being made
An email address held by two eligible services may provide a direct match key. It does not prove that one person exclusively controls the account or that every connected media identifier remains current. A hashed email may still be linkable to someone.
A probabilistic method may use device or connection signals to estimate that observations belong together. Shared networks or devices can create mistaken associations; one person’s unlinked devices can be missed. Ask whether each proposed link represents a person, household, account, device or browser. A household-level inference does not validate person-level targeting.
A “deterministic audience” can contain inferred edges. A provider might match an email directly and then infer that another email or device belongs to the same person. Request a breakdown of direct, inferred and unmatched records at the level the provider can substantiate.
Deterministic vs Probabilistic Audience Matching: Key Differences
- Matching Method
- Deterministic
- Link Basis
- Directly observed attribute (e.g. hashed email)
- Accuracy
- High, when input is clean and consistent
- Coverage
- Lower – limited to records with matching identifiers
- Use Case Suitability
- Best for person-level targeting with verified data
- Error Risk
- Low false positives; high risk of missed links if identifiers change
- Matching Method
- Probabilistic
- Link Basis
- Pattern of signals (device, IP, behaviour)
- Accuracy
- Moderate – depends on signal quality and validation
- Coverage
- Higher – can link across devices and accounts
- Use Case Suitability
- Good for broader reach when direct data is sparse
- Error Risk
- Higher false positives; lower risk of missing valid links
Keep four measures separate
| Measure | Question it answers | What it cannot establish alone |
|---|---|---|
| Input eligibility | How many supplied records were permitted and technically usable? | Match accuracy or media reach |
| Match coverage | How many eligible records received a usable match, using which denominator? | Reach in the planned inventory |
| Link quality | How often are proposed links correct, and how many true links are missed in a suitable validation set? | Quality outside the tested population |
| Activated delivery | How many eligible opportunities and impressions were delivered? | Whether matching caused an outcome |
For link quality, ask for precision and recall with the identity unit stated. Precision is the share of proposed links that are correct in the validation set. Recall is the share of true links in that set that the method found. Measuring recall requires a credible way to identify true links beyond those the method proposed.
A stricter confidence threshold can reduce wrong links while missing more valid ones. Request the threshold, truth-set method, sample dates and relevance to the Australian audience. If a credible truth set is unavailable, mark link quality unverified; match rate cannot replace it.
Key Metrics for Evaluating Audience Matching Quality
- Input Eligibility
- Number of usable records from supplied data
- Match Coverage
- Share of eligible records that received a match
- Link Quality – Precision
- Proportion of proposed links that are correct
- Link Quality – Recall
- Proportion of true links found in validation set
- Activated Delivery
- Impressions delivered to matched audiences
Choose the error the campaign can tolerate
For a customer exclusion, a missed link may leave a customer exposed to a prospecting ad; a wrong link may exclude an actual prospect. For person-level frequency control, joining two people may suppress one person’s ads because the other saw them. Failing to join one person’s devices may allow extra exposures. Decide which error has the greater cost before accepting a threshold.
A proposed comparison can use the same permitted input file, Australian market, inventory and period where feasible. Request separate counts for direct, inferred and unmatched records, then delivered impressions for those groups where reporting permits. Some providers expose only aggregates.
Check permitted use and changes
Establish the source and permitted use of each input attribute. Ask how withdrawn choices, deleted records and changed identifiers affect the graph and activated destinations.
For covered tracking using data collected in an Apple app, App Tracking Transparency permission is required. Apple’s examples of tracking include sharing email lists with a data broker or a third-party advertising network for retargeting or finding similar users. Other environments have their own rules.
Prefer direct matching when its eligible attributes, identity level and reachable inventory suit the task. Consider inferred links when additional coverage matters and validation supports the campaign’s error tolerance. An advertised match rate without its denominator, validation method and method mix is an incomplete answer.
Pre-Activation Checklist for Audience Matching Providers
- Confirm source and permitted use of input attributes (e.g. email, device ID)Ensure compliance with OAIC guidance and App Tracking Transparency (Apple) requirements
- Verify whether withdrawn consents or deleted records affect the matching graphCheck how the provider handles opt-outs and data deletion under GDPR and Australian privacy laws
- Request breakdown of direct, inferred, and unmatched recordsRequired for transparency and error assessment in campaigns targeting Australians
- Ask for precision and recall metrics with stated identity unit and validation methodMust include truth-set methodology and sample dates relevant to the Australian market
- Avoid accepting match rates without denominator or validation detailsA match rate alone is misleading—must be contextualised with eligibility and quality measures
