User Habits

The problem

Before this patent, audience quality in digital advertising was mostly volume accounting — how many times did the ad fire, how many clicks. That flattened all users to the same denominator and hid the fact that some audiences are scarce, expensive, and strategically valuable while others are cheap and abundant. The patent reframes audience worth as a qualitative index anchored to the ratio between impressions seen and opportunities realized, with physical location as a first-class axis.

The mechanism

The method tracks two distinct counters per user: impressions (how often the user is online and presentable) and opportunities (how often an actionable targeting condition is satisfied for that user). It computes a cost associated with those opportunities and combines the three signals into a single qualitative index — a per-user scalar expressing "what is this person worth to reach, given how rarely the conditions to reach them occur and how much competition exists when they do." Location data is appended to the behavior profile so the index can vary by where the user physically is, not just what they're browsing. The index is anonymity-preserving — it operates on profile tokens, not identities.

What it proved

Scarcity is the signal. The per-user index reframes marketplace value as availability × cost × fit rather than raw reach, which is the conceptual ancestor of modern bid-shading and predicted-LTV bidding — but applied to humans, places, and moments rather than impressions in the abstract.

If it were built today

The per-user index idea migrated into every yield-management system in the decade since — programmatic ad exchanges, retail-media-network floor pricing, dynamic surge in ride-hail, dynamic room pricing in hotels all use a variant. What's newly possible in 2026 is computing the index on-device with Core ML, ONNX Runtime, or Apple ANE so the raw impression stream never leaves the user; emitting only a privacy-budget-constrained differentially-private scalar to the server; and letting an LLM reason over the index in context — "this user is in a scarce, high-intent micro-moment, bid accordingly" — instead of hard-coded bidding logic. Federated learning lets the index-computation model itself improve without the raw behavior ever pooling. The clever part in 2026 isn't the index — it's proving to a counterparty that your index is calibrated (honestly trained, no adversarial drift) without exposing the training set. Cryptographic attestation of model weights + on-device evaluation + server-side bid is the end state.