Relationship & Affinity turns a customer base into a network you can read and act on.
A sub-component of the abstraction layer built on a simple premise: a customer is not an atom, they are a node in a relationship network — and the value is in the edges. On the member side it computes a real relationship graph whose nodes are identity-resolved members and whose edges are governed, multi-type links: converted referrals, ambassador-to-customer ties, and likely households. It reads that graph's structure — communities, hubs and isolated members — so plays can be chosen by structural role. On the product side it derives affinity from real baskets: which products are bought together and which categories a member gravitates to. Reach, households, recommendations and referral plays then read from those real connections rather than guesses.
Treats customers as nodes in a network, not isolated atoms — a multi-type member graph of real referrals, ambassador ties and likely households, plus product co-purchase and category affinity.
Customers are nodes, not atoms — and the value is in the edges.
Most retail marketing prices every customer as an isolated atom, throwing away the most valuable thing about a base: the connections between the people in it. This layer makes those edges real — a computed, governed graph of resolved members and the relationships between them — so retention can map the network, read its structure, and work it by role.
- Hub (high referral degree)
- Member
- Isolated (churn-risk)
- Converted referral
- Likely household
- Co-purchase affinity
The node is the customer; the edge is the relationship
Every node is an identity-resolved loyalty member — a real, consented person, not a device or a duplicate. What makes the graph valuable is the edges between them, and each edge is a specific, governed kind of relationship rather than a vague similarity.
- Converted referral — the backbone edge, drawn only when a referral actually signed up. It is the one link that feeds hub influence, so it stays a clean record of real word-of-mouth.
- Ambassador-to-customer — every customer an ambassador brought in becomes a durable link, turning a fan base into a browsable network.
- Likely household — members inferred to share an address, from a privacy-preserving fingerprint that never stores the raw identifier, kept only for plausible small groups.
- Product affinity — co-purchase pairs (bought X, also bought Y) and per-member category leaning, derived from real baskets, connect the people-graph to what they actually buy. The statistical similarity itself — product/product and member/member lookalikes — is modeled in detail on the sister Affinity & Similarity layer.
What the edges let you compute
Once relationships are real and materialized, structure falls out of them — and structure is what you can act on. A bag of isolated atoms has none of this.
- Communities — clusters of connected members (households, friend groups, local clusters) that a single message can seed rather than paying to reach one by one.
- Hubs — members with high referral degree: natural influencers who earned the position by referring members who converted, not by follower counts.
- Isolated members — nodes with no converted-referral edge and no social anchor, which turn out to be disproportionately a churn risk.
- Next-best-product — recommendations grounded in a member's real co-purchase and category patterns instead of generic store-wide popularity.
Why this matters for AI decisions
An Agent reasoning over atoms can only ever compute similarity to an average customer who does not exist. Reasoning over the graph, it can pick a play from a member's real structural role — and explain why.
- A recommendation, a referral ask or a win-back is grounded in a role — a hub, a household, a co-purchase pattern — not an ad-hoc guess made on the spot.
- Isolated members get connected instead of blasted; hubs get recruited; communities get seeded — three different plays the atom view cannot tell apart.
- Every edge is stamped declared / inferred / imported with a High / Medium / Low confidence band, so a decision can weigh how sure the relationship behind it is — and inferred links are operator-only, dismissable in one click.
The full case for treating customers as a network — mapping the member graph, reading its communities and hubs, and compounding loyalty through governed referral and advocacy.
This is how the layer improves AI decisions.
Without relationship and affinity modeling, a base is priced as a bag of atoms: referral and word-of-mouth are invisible and unmanaged, hubs and isolated members are treated identically, and recommendations fall back to popularity. Connections between members and between products get recomputed ad hoc, so reach, households, influence and recommendations all rest on guesses instead of real ties.
A referral is attributed to the exact referrer, and a hub's fan base is a browsable network rather than a headline number that resets each report.
Plays can be chosen by structural role — recruit hubs, seed communities, connect isolated members — instead of blasting an average customer who does not exist.
Household-level reach can be reasoned about because the household relationship is in the model, inferred from a privacy-preserving fingerprint and confidence-banded.
A next-best-product pick is backed by a real co-purchase pattern, not generic popularity.
The value is operational, not just technical.
The Agent can reason over who is connected to whom, where a member sits in the network, and what is bought with what — so a recommendation, a referral ask or a win-back is grounded in a real structural role (a hub, a household, a co-purchase pattern) rather than an ad-hoc similarity computed on the spot for an isolated individual.
Models the member base as a graph — nodes are identity-resolved members, edges are real, converted relationships — instead of pricing every customer as an isolated atom.
Carries three governed link types on one graph: converted referral (the backbone), ambassador-to-customer, and likely household — each stamped with provenance and a confidence band.
Reads structure, not just dots: communities, high-degree hubs (natural influencers) and isolated members (a churn-risk signal) each become a targetable role.
Derives product affinity from real baskets — co-purchase pairs and per-member category affinity — for next-best-product from real patterns, not generic popularity.
Keeps the relationships materialized in the model, so reach, household and hub logic are read, not re-derived ad hoc on every call.
What this layer is responsible for
Reads orders and converted-referral records from the raw store; an edge is drawn only when a referral actually converts, never for an invite merely sent or expired.
Assembles a multi-type member graph — referral, ambassador-to-customer and likely-household edges — with each link marked declared / inferred / imported and a High / Medium / Low confidence band.
Detects communities with Label Propagation (preserving hub-and-spoke sub-structure) and ranks hubs by real referral degree; isolated nodes fall out as a churn-risk set.
Materializes a daily, deterministic snapshot per program (rebuilt on demand when stale) so the same data always yields the same structure.
Computes product co-purchase and per-member category affinity from itemized order lines, and exposes all of it to the ontology and recommendation surfaces as governed model outputs.