Insurance Lead Acquisition Glossary
Definitions of the terms used in insurance lead scoring, real-time lead auctions, and acquisition economics. Written to be useful whether or not you ever buy anything from us.
What is auc (area under the curve)?
AUC measures how well a model separates converters from non-converters across all possible thresholds. 0.5 means no better than chance; 1.0 means perfect separation.
Concretely, AUC is the probability that a randomly chosen converter receives a higher score than a randomly chosen non-converter. It describes ranking quality independent of where you set a cutoff, which makes it useful for comparing models.
Context determines what counts as good. In domains with strong signal, 0.85 might be unremarkable. In lead conversion prediction, where outcomes depend heavily on factors absent from the data — whether an agent reached the person, what was said, what competitors offered — values in the 0.6 to 0.7 range can represent genuinely useful models.
AUC alone is not sufficient. A model can have respectable AUC and still be poorly calibrated, meaning its probability estimates are systematically too high or too low. For bidding, where the number is used directly to set a price, calibration matters as much as ranking.
What is conversion propensity?
Conversion propensity is a model's estimate of how likely a specific lead is to result in a completed sale or enrollment, usually expressed as a probability or a score derived from one.
Conversion propensity is a prediction about commercial outcome, not about the person. It answers whether this lead is likely to become a customer, which depends on intent, timing, eligibility, contactability, and how well the product fits.
It should not be confused with risk scoring. An insurance risk model predicts claims or loss; a conversion model predicts purchase. A prospect can be an excellent underwriting risk and extremely unlikely to buy, and the two models will disagree about that person for entirely valid reasons.
Propensity scores are only useful if they are calibrated and if performance is measured against actual outcomes. A score that ranks leads correctly but is systematically overconfident will still mislead a bidding system that treats the number as a true probability.
What is cost per acquisition (cpa)?
Cost per acquisition is total acquisition spend divided by the number of resulting sales or enrollments. It captures what a customer actually costs, as opposed to what a lead costs.
Lead price and CPA move independently, which is why optimising lead price alone is a trap. Buying leads at half the price doubles CPA if they convert at a quarter of the rate. The only figure that matters commercially is cost per acquired customer.
CPA should be calculated on total spend, including leads that never converted, leads that were never reached, and leads rejected after purchase. Excluding them produces a flattering number that does not correspond to money actually spent.
Pre-bid scoring attacks CPA by changing the mix rather than the unit price. Declining leads unlikely to convert and paying more for leads likely to convert can raise average lead price while lowering CPA, which is the intended outcome and often surprises buyers who are managing to lead price.
What is decile lift?
Decile lift compares how often leads in the best-scoring tenth of a population convert versus the worst-scoring tenth. A lift of 10x means the top decile converts ten times as often as the bottom.
To calculate it, score a population, sort by score, divide into ten equal groups, then compare conversion rates between groups. The comparison is usually top decile against bottom decile, though top decile against the overall average is also used and produces a smaller number. Which comparison is being quoted should always be stated, because the two are not interchangeable.
Lift is the metric that maps most directly to bidding decisions. If the top band converts far more often than the bottom, there is a concrete argument for paying more for one and less or nothing for the other. That makes it more immediately actionable than aggregate accuracy measures.
Lift is also easy to inflate. A model evaluated on the same data it was trained on will report dramatic lift that disappears in production. Figures worth trusting come from out-of-time validation against outcomes the model did not see during training.
What is identity resolution?
Identity resolution is the process of matching a lead record to a known individual in a reference data set, so that attributes which were not submitted with the lead can be attached to it.
A lead form typically captures very little: a name, a phone number, maybe an age range. Identity resolution matches that fragment against a large consumer reference set to determine, with some confidence, which real person it corresponds to. Once matched, additional attributes become available to the scoring model.
Match confidence matters more than match rate. A high match rate achieved by accepting weak matches introduces errors that are worse than no match, because the model then scores against the wrong person's attributes. Serious implementations grade matches and treat low-confidence results as unmatched rather than guessing.
Identity resolution is distinct from data enrichment in emphasis. Enrichment describes appending attributes; identity resolution describes establishing which individual those attributes should come from. In practice the second is the hard part.
It is also distinct from identity verification, which asks whether a person is who they claim to be for fraud or compliance purposes. Resolution asks a narrower question: which record in the reference set does this correspond to.
What is lead fraud?
Lead fraud covers records that misrepresent what they are: fabricated contact details, falsified consent, or inquiries presented as new when they are not.
The common forms are distinct and worth separating. A fabricated lead contains invented or stolen contact information. A falsified-consent lead describes a real person who never agreed to be contacted. A recycled lead is a genuine older inquiry resold as fresh.
These require different defences. Fabrication is detectable through data-consistency and validation checks. Recycling is detectable by tracking inquiry timestamps and matching against records already seen. Consent problems are generally not detectable from the record at all, and require documentation from the vendor.
Fraud detection is distinct from quality scoring, though the two are often bundled. A lead can be entirely genuine, fully consented, and still very unlikely to convert. Treating a low conversion score as evidence of fraud conflates two different problems and leads to unfair vendor disputes.
What is ping-post?
Ping-post is a two-step lead transaction: the publisher first sends a minimal record (the ping) so buyers can bid without seeing the full lead, then sends the complete record (the post) to whichever buyer wins.
The ping contains only enough information to evaluate the lead — commonly a phone number, ZIP code, and product type. Buyers respond with a bid or a decline. The winning buyer then receives the post, containing the full contact record, and is charged.
The structure exists to protect both sides. Publishers avoid giving away complete lead data to buyers who will not pay for it. Buyers avoid paying for leads they would not have wanted had they seen them.
For scoring, ping-post is what makes latency the binding constraint. Any evaluation has to happen inside the ping window, on ping-stage data only, which is why pre-bid models are built differently from models that run after a lead is purchased.
What is post-acquisition scoring?
Post-acquisition scoring evaluates a lead after it has been purchased, using the full contact record to produce a more accurate estimate than was possible at the pre-bid stage.
Once a lead is bought, the buyer has the complete record — name, age, gender, full address — and can perform identity resolution against a consumer reference set. That additional information supports a substantially more accurate model than the sparse ping-stage data allows.
The purpose is not to decide whether to buy, since that decision has already been made. It is to decide what to do next: which leads to route to the most experienced agents, which to call first, which to put into a lower-cost follow-up track, and which to stop working.
Buyers who use both stages together get compounding benefit. Pre-bid scoring controls what enters the funnel and at what price; post-acquisition scoring controls how effort is allocated across what did enter.
What is pre-bid lead scoring?
Pre-bid lead scoring is the practice of predicting how likely an insurance lead is to convert before paying for it, using only the limited data available at the ping stage of a real-time auction.
In a real-time lead auction, a publisher sends a minimal record — typically a phone number and ZIP code — to prospective buyers before the full lead is sold. Pre-bid scoring runs on that partial record and returns a conversion propensity estimate fast enough to inform what the buyer bids, or whether they bid at all.
The defining constraint is latency. A scoring model that needs a second to respond is useless in an auction that resolves in a few hundred milliseconds, so pre-bid models are deliberately built to run on sparse inputs and answer in single-digit milliseconds. This is a genuine engineering trade-off: pre-bid models are less accurate than models that run after purchase, because they have far less to work with.
Pre-bid scoring is distinct from lead filtering. Filtering rejects leads that fail a rule, such as an out-of-area ZIP code. Scoring produces a graded estimate that supports paying different amounts for different leads rather than a binary accept or reject.
Note on terminology: in construction and procurement, 'pre-bid' refers to meetings and qualification steps before submitting a tender. That is an unrelated use of the word. In insurance lead acquisition, pre-bid refers specifically to the moment before a lead is purchased in an auction.
What is price-per-probability?
Price-per-Probability is a lead buying principle: pay for each lead in proportion to its predicted probability of converting, rather than paying one flat price for every lead from a given source.
Most lead buying is priced by category. A Medicare Advantage lead from a given vendor costs a fixed amount regardless of whether that specific lead is likely to enroll. Because conversion rates within a single source vary enormously, a flat price systematically overpays for weak leads and underpays for strong ones.
Price-per-Probability treats the predicted conversion probability as the thing being purchased. If one lead is several times more likely to convert than another, it is worth several times more, and the bid should reflect that. The practical effect is that budget moves toward leads that convert instead of being spread evenly across everything on offer.
This requires two things to work: a scoring model accurate enough that the probability estimates are meaningful, and a bidding mechanism capable of acting on them in real time. Without both, the principle is just a description of what buyers wish they could do.
Note on terminology: in prediction markets and sports betting, contract prices are often read as implied probabilities — a contract trading at $0.20 implies a 20% chance. That is a different concept that happens to use similar language. Price-per-Probability as used here refers to lead acquisition pricing, not to market-implied odds.
What is recycled lead?
A recycled lead is a genuine but older consumer inquiry that is resold as though it were new, without disclosing how long ago the original inquiry was made or how many buyers already received it.
Recycling is distinct from legitimate lead sharing. A shared lead is sold to several buyers under disclosed terms, and buyers price it accordingly. A recycled lead is sold as fresh when it is not, so the buyer prices it wrongly.
The practical harm is that a consumer who inquired months ago and has already spoken to several agents behaves nothing like someone who submitted a form minutes ago, yet both arrive priced the same. Contact and conversion rates on recycled inventory are usually far worse.
The most reliable defence is maintaining your own lead history rather than relying on vendor freshness claims. Normalising phone numbers and matching every incoming record against everything previously received reveals both duplicates within a vendor and the same record arriving from multiple vendors.
What is retrospective validation?
Retrospective validation tests a model against historical data where the real outcomes are already known, scoring past leads as if they were new and comparing predictions to what actually happened.
A retro study takes a partner's historical lead file along with the outcomes — who enrolled, who did not — and scores those leads using only information that would have been available at the time. The resulting predictions are compared against the known results to measure AUC, lift, and calibration.
Done properly it is a strong test, because the outcomes are real rather than simulated. Done carelessly it is misleading, with data leakage the most common failure: if any feature encodes information that only existed after the outcome, the model will appear far better than it is.
Retrospective validation is not the same as a live result. It demonstrates that a model would have ranked past leads well. It does not establish that acting on those scores would have produced the modelled benefit, since changing bidding behaviour changes which leads are won and how they are worked.
What is snp lead scoring (d-snp and c-snp)?
SNP lead scoring means predicting conversion for Medicare Special Needs Plans using models built specifically for those populations, rather than applying a general Medicare Advantage model to them.
Dual Eligible Special Needs Plans (D-SNP) serve people eligible for both Medicare and Medicaid. Chronic Condition Special Needs Plans (C-SNP) serve people with qualifying chronic conditions. Both have eligibility requirements, enrollment timing, and demographic concentrations that differ from general Medicare Advantage.
Because the populations differ, the signals that predict conversion differ too. A single blended Medicare model averages those differences away and will systematically misrank leads for the subpopulations, usually underestimating the ones that are least like the majority.
A scoring model cannot establish eligibility. D-SNP eligibility depends on actual Medicaid status and C-SNP eligibility on a verified qualifying condition, both of which are confirmed through the enrollment process. A model can only estimate the likelihood that someone will turn out to qualify and enroll.