RESEARCH DOSSIER
Digital Insurance Sales Funnel Bias and Consumer Abandonment
Abstract. Digital insurance distribution measures customer progress through a sequence of exposure, product comparison, quote generation, disclosure, application, eligibility assessment, payment, and policy activation. The observed abandonment rate is generated jointly by customer decisions, distributor architecture, insurer actions, competing channels, and the measurement system. This monograph develops the statistical interpretation of attrition using coherent exposure denominators, competing outcomes, missing-data sensitivity analysis, experimental estimands, and consumer-protection evidence. It examines the distinction between unwanted exclusion and informed refusal, the selection of policyholders from applicant populations, and the consequences of optimizing a commercial conversion metric without evaluating coverage or welfare. Evidence is drawn from selected insurance field experiments, supervisory publications, legal instruments, and relevant cross-sector research. Worked examples are synthetic and establish mathematical relationships rather than market prevalence. International comparisons cover selected EU, UK, US, Australian, and international supervisory materials. The principal conclusion is that valid digital distribution evaluation requires a journey-level denominator, independently reconciled coverage outcomes, differentiated exit states, and direct evidence of customer understanding and access. Public materials identify plausible mechanisms and selected causal effects; they do not establish a universal abandonment rate or a causal diagnosis for an individual insurer.
Keywords: Digital insurance distribution; sales funnel; consumer abandonment; attrition; competing risks; selection bias; missing data; insurance access; conversion measurement; causal inference; consumer protection; market conduct; accessibility; product governance.
Introduction
Insurance acquisition is a process of establishing a contractual relationship that transfers defined financial consequences of uncertain events. The consumer must identify a need, locate a product, establish eligibility, evaluate price and protection, supply information, and obtain a valid contract. A digital distributor places these activities within interfaces that also collect data and allocate consumer attention. A sales funnel represents a statistical abstraction of that sequence, and its empirical validity depends on the relation between the abstraction and the customer journey.
The subject of this inquiry is stage-specific attrition and the systematic distortion of its measurement, including unequal opportunity to progress among customers. The analytical object includes consumers who never reach a quote, receive an adverse eligibility outcome, defer a purchase, or use a different channel. It also includes the differences between event recording and actual behavior. It is narrower than a general theory of persuasion, lead procurement, or underwriting classification. The study treats insurer acceptance as a separately observed event which can change the distribution of consumers in later stages.
The growth of digital distribution does not imply that customers travel along a single strictly digital path. The European Insurance and Occupational Pensions Authority (EIOPA, 2024) describes uneven digital readiness and the importance of physical and hybrid distribution in its digitalization survey summary. A consumer who compares policies online before arranging a telephone purchase might be a digital abandonment in the platform report and a successful insurance acquisition in the insurer's contract system. Both observations can be accurate relative to their distinct definitions.
The method is a targeted documentary synthesis rather than a registered systematic review or a proprietary data analysis. Searches addressed insurance distribution and enrollment, behavioral friction, website accessibility, missingness, experimental selection, and the institutional treatment of consumer outcomes. The source set includes full official HTML texts and summaries, accessible journal text, publisher or repository abstracts, and selected legal excerpts. The distinction is maintained when drawing conclusions. Legal examples identify selected documented requirements and their dates, not exhaustive current national transpositions or legal advice. No insurer contracts, underlying event-level records, or consumer interviews were supplied. There is no pooled estimate of dropout or universal ordering of mechanisms. The contribution is the integration of statistical and insurance-specific interpretation, with explicit worked examples and operational tests; no claim of novel theory is advanced.
Funnel Definitions and Exposure Denominators
From interface events to coverage states
A single customer need can produce several sessions, quotes, applications, and even policy contracts. A platform session is a technical unit. A visitor identifier is an imperfect proxy for a person. A quote represents a product configuration and price at a particular time. An application seeks acceptance; a binder or issued policy establishes a contractual state under product-specific conditions. Coverage may depend on payment and effective date. Each unit needs its own denominator. A report that counts page views at entry and people at exit divides unlike populations.
Define a cohort of N distinct eligible journey identifiers beginning at a recorded entry event during a stated recruitment period. Distinguish eligibility to enter the measurement cohort from underwriting eligibility for the requested product. Define a fixed follow-up window T relative to entry, such as 30 days when appropriate to the product. Define reaching stage j by a precisely observable event, with ordering rules for repeat visits. In a strictly nested journey, let N_j be the count of cohort members reaching stage j within T. For successive nested stages, conditional transition p_j = N_(j+1)/N_j. Overall progression to stage K is the product of all p_j, equal to N_K/N_0. An unconditional shortfall at stage j is (N_j - N_(j+1))/N_0. The corresponding conditional shortfall divides by N_j. Both are useful; neither specifies why a person failed to progress.
A journey that permits skipping, reversals, simultaneous quotes, or offline referrals is not strictly nested. A person can enter the application stage after speaking to an adviser rather than recording an online quote. Forcing this event into a sequential model creates false denominators. Establish a state graph with legal transitions, an entry state, and time stamps. Use a unique journey identifier for the need or product family, and retain offer identifiers for simultaneous quotes. Classify re-entry and referral rather than deleting them. A multi-state representation may include recurrent visits and absorbing states; where termination is defined by T, an unfinished state at time T is recorded as unresolved.
The outcome event must be matched to product obligations. A motor insurance purchase could become effective only at the stipulated start time; a life insurance submission can remain under consideration, and a payment authorization may precede actual issuance. A technical success event in a payment processor is insufficient evidence of an insurance contract. Define separate milestones for application submission, accepted offer, binder where relevant, policy issuance, activation, and subsequent early cancellation. Specify which of these establishes the chosen coverage outcome and reconcile it against administration records.
Synthetic transition example
Suppose two groups each have 1,000 cohort entrants. A common quote stage produces 800 and 600 quotes, application submission produces 600 and 450, offer acceptance produces 540 and 405, and effective coverage is recorded for 486 and 365 respectively. Quote progression is 80% for the first group and 60% for the second. Subsequent submission rates are both 75%; offer acceptance conditional on submission is 90% for both. Coverage after an accepted offer is 90% and approximately 90.12%. Overall coverage is 48.6% versus 36.5%, a 12.1 percentage point difference. Counting only submitted applications yields 81.0% versus approximately 81.1%, obscuring the earlier gap. The second group loses 400 people before quotation compared with 200 in the first group. The excess early shortfall is therefore 200 people per 1,000 entrants, while the net coverage difference is 121 per 1,000. These quantities have distinct interpretations. Early absence of a quote does not identify a barrier, and eliminating that difference would not necessarily preserve all subsequent conditional transition probabilities.
The example is deliberately constructed with nearly equal later transitions. It demonstrates why auditing only underwriting decisions can be insufficient to characterize full distribution access. A later-stage parity measure is a conditional estimate for the selected population. A group difference in starting a journey can also arise before any first-party event exists, for example through different reach of distribution arrangements. These people require a population or target-market sampling frame outside first-party telemetry. The cohort defined above answers a narrower question about entrants.
Competing outcomes and time
Each departure must be distinguished from alternatives. At the chosen horizon an entrant could have effective coverage through the observed distributor, a verified purchase through another channel, a recorded informed decline, an insurer rejection, a technical blockage still unresolved, an ongoing application, or an unknown outcome. Choose mutually exclusive adjudication categories, for instance by the final verified status and a specified priority rule. Report counts and a separate history of intermediate states. An insurer rejection is not customer abandonment merely because a browser stopped recording.
Time-to-event analysis is useful when application duration varies. A cause-specific hazard describes the instantaneous occurrence of a given event among trajectories still at risk. Its relation to cumulative event probability depends on other event hazards. Treat competing terminal outcomes as such rather than censoring them as independent loss. If termination categories are incomplete, report a distinct unknown state and quantify the effect of plausible reclassification. Administrative censoring at a common follow-up cutoff is distinct from abandoning a task. A quote with a 30-day validity period and a life underwriting request requiring medical evidence may require different observation horizons.
A useful release of metrics includes the number of eligible entrants, counts reaching each milestone, cohort time, definitions, group inclusion rules, missing records, and terminal adjudications. Event-level behavioral metrics such as backtracking counts and time on task may identify where to investigate, but they do not establish the meaning of a departure. A long interval can indicate difficulty or appropriate deliberation. A short interval can indicate efficient completion or insufficient attention to material terms.
Consent, Tracking and Missingness
Measurement design and consent
Analytics measurement is itself a selected process. Cookie choices, browser restrictions, ad blocking, device changes, denied tracking authorization, asynchronous transmission, and vendor boundaries can prevent recording or linking of events. An insurance transaction system has its own completeness problems including duplicate quotes, manual corrections, and delayed posting. A consumer's decision about optional tracking cannot be interpreted as a decision about the insurance product. Reporting should separate the fraction consenting to analytics from the fraction establishing coverage. The General Data Protection Regulation (European Parliament and Council, 2016) establishes principles including data minimization and purpose limitation. It follows for research design that additional tracking should have a justified purpose, appropriate legal basis, and proportionate safeguards; a desire for more complete conversion data alone establishes none of these.
The recommended event dictionary specifies event name, semantic meaning, source system, customer or journey key and its issuance rule, offer identifier, event and receipt time, channel, product version, experiment assignment, configuration version, source reliability, and any lawful segmentation variables. Do not store unnecessary application answers in generic analytics payloads. Sensitive health and financial information can remain in protected transaction systems, with authorized minimal outcome aggregates used for statistical analysis. A pseudonymous identifier still requires appropriate protection where it remains personal data. Maintain a data-flow inventory of each party's role and separate retention schedules for operational and research records.
Consent rates must not be fabricated for consumers whose preference is unknown. An independent lawful server-side aggregate may make a denominator available, but instrumentation itself needs compliance assessment; server-side collection does not automatically escape privacy obligations. Aggregate counts, privacy-protective sampling, trusted linkage with defined retention, and survey measurement can reduce reliance on individual-level tracking. A dashboard should indicate which denominator it is actually capable of observing.
Missingness and identification
Let R equal one if the outcome is ascertainable in a chosen data source, and Y equal one for verified coverage. The complete-case conversion estimate is P(Y=1 | R=1). The target is P(Y=1) for the defined cohort. The equality of these two quantities requires assumptions about observation. In a simple setting with no covariates, equality holds when R and Y are independent. With covariates X, conditional independence can justify standardization or weighting if the model is correct and positive observation probabilities exist. Rubin's missing-data analysis (1976) supplies the classical distinction between missing-at-random conditions and the conditions required for likelihood or Bayesian ignorability. Neither missing at random nor positivity can be confirmed simply by fitting a model to observed outcomes.
If the outcome-dependent probability of observation remains after conditioning, observational data alone generally do not identify the missing outcomes. Missingness may be informative when a frustrated consumer blocks tracking, changes device, or transfers to an adviser. One should distinguish unobserved variables from observed outcomes truncated by eligibility: an offered product's satisfaction is undefined for consumers who never receive or use it. Imputing a satisfaction score for an impossible exposure would change the estimand.
Partial identification example
Suppose 1,000 legitimate journey entrants are known from a lawful independent cohort frame. Outcomes are ascertained for 700, with 280 verified purchases and 420 verified nonpurchases. The observed-case rate is 40%. With no assumption about the other 300, the full-cohort coverage rate is bounded by 28% and 58%. If the unobserved group has a coverage probability q between zero and one, the rate is 0.28 + 0.30q. Taking q=0.2 gives 34%, and q=0.6 gives 46%. The numbers are hypothetical and show why a seemingly precise dashboard estimate may be a poor basis for a consumer exclusion conclusion.
For two groups with missing fractions m_A and m_B and observed purchase counts y_A and y_B relative to their full denominators, the difference is bounded below by y_A - (y_B + m_B) and above by y_A + m_A - y_B, with y terms expressed as proportions. If the interval contains zero, the direction of disparity is not identified without additional evidence. These bounds concern known cohort membership; they cannot account for people missing from the entry frame itself. Report both limitations.
Weighting by inverse estimated observation probability can help when factors associated with observation and coverage have been measured with sufficient overlap. Inspect weight distributions and effective sample sizes; compare weighted estimates with administrative totals. Trimming extreme weights trades variance against bias and changes interpretation. Multiple imputation can propagate sampling uncertainty conditional on an assumed data model; sensitivity analysis must separately address violations of the missingness assumptions. A successful predictive fit is not proof that missingness is ignorable.
Identity resolution and reconciliation
A quote repeated after an email reminder can be a continuation of the same need or a new offer. Deduplication rules should distinguish unique customer, unique journey, quote, application, contract, and renewal. Test exact agreement of identifier mappings and use uncertain probabilistic linkage only with documented error estimates. False matches inflate cross-channel completion; missed matches inflate unknown departures. Validate against appropriately sampled, authorized records with known linkage, and quantify both precision and recall. Do not infer identity from high-risk device-fingerprinting techniques without legal and ethical assessment.
Reconciliation is strongest when a verified contract system provides a coverage event independent of the interface's success message. The observed purchase count can be reconciled to contracts by issue date, product and intermediary, accounting for reversals, multiple policies per need, pending records, and clock differences. The reconciliation does not establish the outcomes of customers who never acquired a policy elsewhere. A study that needs market-wide take-up must obtain a valid additional source or label the outcome distributor-specific.
Checkout Frictions and Customer Abandonment
The mechanisms behind departure
A consumer can rationally discontinue a purchase after receiving a premium that exceeds willingness or ability to pay, discovering an exclusion incompatible with the intended protection, identifying a superior competing offer, or deciding that the risk can be retained. Departure can also arise because required information is difficult to locate, an identity check is inaccessible, an error disables continuation, or a needed adviser is unavailable. Observable behavior alone cannot distinguish these mechanisms.
Operational frictions consume time, effort, attention, or resources. Examples include redundant fields, repeated requests for information, errors without specific recovery instructions, login requirements before quote visibility, unclear document formats, and payment interruptions. Useful deliberative friction can include a clear review of meaningful exclusions, an opportunity to compare total cost, a confirmation of an optional add-on, or a cooling period. The purpose and effect matter. A design evaluation should measure informed decisions and access to protection alongside speed.
Evidence from enrollment experiments supports the relevance of administrative and informational costs. In California, Domurat, Menashe, and Yin (2021) randomized information mailings among approximately 87,000 households; deadline reminders increased enrollment by 1.3 percentage points, or 16% relative to the baseline described in the published abstract. The intervention concerns marketplace participation and mailed outreach, not an online checkout page. Ericson et al. (2023), in the January administrative barriers working paper, found that simplified enrollment in the Massachusetts marketplace increased participation more than reminder letters. The working-paper version and later revisions must be kept separate; this study uses the qualitative ordering rather than mixing numerical results.
Comparison has its own cost. Kling et al. (2012) studied Medicare drug plan comparison: personalized cost information increased plan switching from 17% to 28%, with roughly $100 lower predicted annual consumer cost. The outcome is switching among plans, not acquisition among uninsured people, and the cost measure is predicted rather than necessarily realized expenditure. The studies together establish that selected interventions can affect insurance decisions and that the relevant effect depends on the decision context. They provide no pooled commercial funnel effect.
Price changes, disclosure, and timing
A quote can change as information improves. A consumer entering a tentative vehicle description may be shown an indicative price which is revised after risk factors are validated. The statistician must link the original and revised offers with a product and configuration identifier. A price increase after information submission is not automatically deceptive: the initial quote may have been clearly conditional. Conversely, a misleading initial low price can create wasted search effort and impair comparison even if the consumer exits before purchase. Measure the price change, the time it becomes visible, and what the earlier display represented.
Premium finance creates a distinct choice about spreading payment over time. In its historical 2015 premium finance review, the Financial Conduct Authority examined 43 websites before payment details and reported limitations in information on payment costs and arrangements. It supports the analytical importance of the disclosure point but supplies no population abandonment estimate. Audit should compare annual premium, instalment total, APR where applicable, fees, exclusions, and payment eligibility. The consumer can abandon because an annual amount is unaffordable or because the financing arrangement is undesirable. Combining the two mechanisms under payment failure can misdirect remediation.
Policy comparability requires attention to deductibles, limits, coverage definitions, exclusions, provider networks where relevant, contract duration, and claims arrangements. A cheap product might remain unsuitable. A high completion rate after a visually prominent low premium is not evidence of equivalent protection. The evaluation should test comprehension of consequential terms before concluding that simplification improves outcomes.
Accessibility and recoverability
Digital access can fail without an intentional discriminatory rule. Authentication requiring an inaccessible cognitive exercise or blocking pasting can obstruct progress. W3C's accessible authentication guidance explains how autofill and support for copying codes can help users complete authentication. Test the whole journey rather than only screen contrast: keyboard navigation, focus handling, readable error feedback, assistive technology, alternative verification paths, save and resume, mobile layout, upload recovery, and access to human assistance. Record failure type, recovery option, and ultimate outcome; avoid collecting disability status unnecessarily.
A product journey may reasonably require a document or identity check. The research question is whether the requirement is lawful and relevant, whether an equally effective less burdensome alternative exists, and which customers cannot complete it. An accessible alternative that requires a week-long appointment rather than a minute-long automatic process may still produce unequal practical access. Test elapsed time and resolution rates, not just whether the alternative is nominally available.
The FCA's 2025 digital acquisition journey review is a small consumer-credit study, not an insurance sample. Its examples of useful friction, support for vulnerable customers, and multiple success indicators are informative design analogies. Transfer to insurance requires product-specific validation and appropriate regulatory scope. The 2022 Competition and Markets Authority online choice architecture review provides broader multi-sector context for design-related consumer harm.
Abandonment and manipulation
Recorded continuation can itself represent undesirable influence. Luguri and Strahilevitz (2021) report dark pattern experiments in a non-insurance service setting in which manipulative designs increased acceptance. Their evidence warns against equating completion with exercised preference; it does not quantify mis-selling in insurance. A distributor may simultaneously cause unwanted departure among people who need coverage and unwanted continuation among people who do not.
Evaluation of a suspicious stage compares product selection, understanding, active consent where required, cancellations, complaints, and subsequent reinstatement. Customer interviews can distinguish inability to progress from considered refusal, although response bias must be measured. Avoid experimentation that knowingly exposes consumers to unlawful, deceptive or harmful designs. The legitimate scientific use of manipulative-interface literature is detection, prevention, and repair.
Selection, Fairness and Access Outcomes
Selection before underwriting
Digital distribution acts as an upstream filter of the applicant pool. Insurance risk classification, eligibility, price, and capital constraints determine available coverage under applicable law. Customer attention, language, device reliability, affordability, and interface requirements determine who reaches those decisions. A consumer can be excluded by practical barriers before the insurer has sufficient information to make an underwriting decision.
Let G be a legally and ethically permissible group indicator, Q denote receipt of a quote, A submission, and Y coverage. For a strictly nested process, P(Y=1 | G) is the product of sequential conditional probabilities. A group gap may arise through any factor, and each stage conditions on prior selection. In an observed cohort, differences in these probabilities can reflect variation in product need, offered risk, geography, channel, time, or design. Merely labeling a conditional rate as fair or unfair is not a causal or legal adjudication.
Suppose a platform allocates leads on predicted purchase probability. Past successful customers supply the training labels; failures arising from inaccessible design may be treated as evidence of low consumer demand. If the system then sends fewer accessible offers to consumers who share the associated device or location variables, future data contain fewer opportunities to observe their true demand. This is a potential feedback mechanism, not an empirical claim about any particular firm. Correcting such a loop requires exposure data, versioned decision logs, suitable exploration constrained by consumer protection, and outcome monitoring outside the purchase-only sample.
Conditioning and actuarial consequences
An analysis solely among buyers cannot establish the risk composition of all interested households. In a synthetic example, a target population contains 500 higher-risk and 500 lower-risk consumers, with expected claim costs of $600 and $200 respectively over a common exposure period. The full-population average expected cost is $400. If 30% of the higher-risk group and 60% of the lower-risk group obtain coverage, the insured population contains 150 and 300, with average expected claims of (150 times 600 + 300 times 200)/450, or $333.33. If the take-up rates reverse, the selected average becomes $466.67. These are expected-claim calculations assuming identical period exposures and no further risk heterogeneity. Premiums, expenses, reserves, and risk adjustment are absent. They demonstrate that selective attrition can change observed loss costs in either direction, not that those rates describe actual markets.
The commercial implication is that observed portfolio profitability can improve through the removal of friction for appropriate risks or through undesirable exclusion of consumers. The welfare implication is different. Lost protection should be measured using coverage needs and product terms rather than a change in the insurer's average modeled loss. Where risk classification is authorized and genuinely necessary, audit whether digital filtering corresponds to those criteria rather than an unrelated proxy or technical limitation.
Selection can also distort the evaluation of user understanding. Suppose only purchasers take a comprehension survey. An intervention causes additional consumers to buy; the treated and control survey samples no longer represent the same population. Even randomized assignment does not eliminate outcome selection. Lee's sample selection study (2009) establishes the importance of this problem in another domain. In insurance, use outcomes observed for the original assignment cohort when possible; state assumptions for any estimand restricted to those who would purchase under either condition. Do not interpret a simple comparison of purchasers across conditions as an unbiased treatment effect.
Fairness metrics and consumer outcomes
Report absolute group differences in transition probabilities, rates per 1,000 entrants, relative differences where denominators are meaningful, and uncertainty intervals. Standardize for clearly defined pre-exposure covariates when the question concerns comparable populations; keep crude estimates as descriptive measures of actual distribution. Adjustment variables must be selected according to the research question: controlling for an intermediate stage blocks part of the pathway under study and can introduce selection bias. A model that uses risk variables may be relevant for a particular underwriting question while failing to describe accessibility.
For standardization, define a reference distribution of X, estimate stage or coverage probability for each group within X, and average over that reference. Positivity requires sufficient observations in the comparable cells; when some combinations never occur, identify the restricted comparison population instead of extrapolating silently. Differences in missingness, failed linkage, and outcome definitions must be shown before comparing groups. Small-cell noise, multiple testing, reidentification risk, and legal restrictions on sensitive data all matter. Report suppressed cells and uncertainty rather than assigning zero.
Fair treatment includes customers who do not purchase. The IAIS (2025) announcement of its fair treatment application paper links the work to Insurance Core Principle 19. For the present inquiry, one implication is to evaluate access and treatment across the journey rather than using purchased policies as the sole consumer population. The paper's full content and jurisdiction-specific implementation would be needed to assess compliance in a specific market.
A welfare evaluation cannot be reduced to a single presumed utility weight. Record access to appropriate protection, affordability, understandable information, consumer choice, and the prevalence of verified unwanted outcomes. On the insurer side, include processing cost, fraud prevention, underwriting integrity, and risk pool sustainability. Trade-offs should be explicit. An increase in coverage achieved by obscuring material terms is analytically different from increased coverage after removing a redundant field; the two may generate identical conversion statistics.
A/B Testing and Causal Attribution
Define the intervention and estimand
A randomized trial should assign an ethically permissible design change, such as an accessible replacement for a burdensome form or clearer cost presentation. Specify the unit, cohort, intervention versions, assignment time, primary outcome, follow-up, interference risks, and handling of delayed or missing outcomes in advance. Assign at the customer or journey level when cross-session consistency is needed. Randomizing pages independently can expose the same consumer to multiple variants and obscure effects. Where brokers or agents share resources, cluster assignment may be justified; inference must reflect the assignment unit.
For a binary coverage outcome, let Y(1) and Y(0) be potential outcomes under the new and current versions. The intention-to-treat contrast is the expectation of Y(1) minus Y(0) in the assigned eligible population. Under valid random assignment, consistent treatment definition, suitable interference assumptions, and complete outcome ascertainment, the difference between observed assignment-group means identifies this contrast. The estimand includes clients who do not encounter every downstream change after assignment. A conditional analysis limited to visitors who reach the changed page can be valid if assignment occurred at that point and defines the population, but conditioning on a post-assignment event after earlier assignment generally changes the question.
Distinguish effect on coverage through the distributor from effect on any coverage, and define the latter only if verified across relevant channels. The latter may remain unidentified. A purchase displacement from agent to web is commercially relevant but need not increase consumer protection. Trial alternatives must meet legal and accessibility obligations; harmful or noncompliant controls should not be maintained merely to estimate an effect.
Worked randomized example
Consider 2,000 assigned entrants per arm, with 480 effective policies for the control and 540 for a redesigned accessible form after equal follow-up. Coverage is 24% and 27%, an intention-to-treat difference of 3 percentage points and a relative difference of 12.5%. Under independent individual assignment and complete observation, the usual unpooled large-sample standard error is the square root of 0.24 times 0.76 divided by 2,000 plus 0.27 times 0.73 divided by 2,000, approximately 0.013775. A 95% Wald interval is approximately 0.3 to 5.7 percentage points. It is illustrative; clustering or adaptive assignment would require a different variance estimator. The result does not determine whether additional policies are suitable, persist, or protect households who would otherwise remain uninsured.
Suppose the control has 1,800 observable outcomes and the redesigned arm 1,600. If the same observed purchases were retained but the remaining outcomes were unknown, the respective full-cohort ranges become 24% to 34% and 27% to 47%. The effect range becomes -7 to +23 percentage points. The missingness can defeat a positive effect conclusion even when assignment was valid. This separate hypothetical variation uses no assumptions about the unknown cases and shows why tracking and randomized design must be evaluated jointly.
Sample planning should specify a minimum meaningful effect and expected baseline, adjusted for clustering, loss to follow-up and multiple outcomes. For a conventional two-sided 5% test with 80% power, equal arms and rates 24% and 27%, the normal approximation based on pooled variance and alternative variance gives roughly 3,315 participants per arm before any adjustment; actual planning should use the chosen exact design and variance estimator. Do not interpret statistical significance as practical sufficiency. Evaluate whether any improvement justifies implementation costs while preserving consumer outcomes.
Causal attribution in observational data
Observational funnel analysis is often useful for locating failures and prioritizing investigations. It cannot by itself assign causal responsibility for abandonment. A release of a revised form may coincide with a promotion, pricing changes, changed risk appetite, market events, seasonality, or new traffic sources. A before-and-after comparison conflates these changes. Preserve version identifiers, calendar timestamps, cohort definitions, marketing exposure, and policy rules.
A directed causal graph helps clarify which variables precede design exposure, which are mediators, and which determine measurement. For example, device type and channel may affect exposure to a mobile-only interface change and baseline probability of completion; quote price may be a mediator if the design affects chosen coverage options. Conditioning on it can remove part of the effect. If price is fixed before design assignment, it may be a baseline variable. The temporal ordering is essential.
Difference-in-differences with an unaffected comparison channel requires credible parallel trends in the absence of the change, no relevant spillover, coherent time windows, and a stable target population or explicit adjustment for compositional change. An event-study graph can reveal departures from pre-intervention trends but cannot establish the counterfactual in the post-period. Regression discontinuity at an eligibility threshold needs a continuity argument and evidence against manipulation around the threshold. Such a study identifies a local effect under its assumptions. Matching and weighting require sufficient measured confounders and overlap. Each method answers a specific identification question rather than automatically upgrading a descriptive gap into a causal estimate.
Experiments with competing outcomes
An intervention can affect several terminal states. The trial should report separately verified purchase, genuine refusal, rejection, other channel, and unresolved outcome. If an accessible form shifts cases from unknown to observed decline, the apparent digital abandonment rate might rise while measurement improves. Cause-specific treatment contrasts on each terminal cumulative probability at a common time are often easier to interpret than changes in hazard ratios. Ensure the terminal probabilities sum to one within each arm, accounting for follow-up.
Trials of reminder timing can alter decision timing without changing long-run coverage. Report pre-specified horizons such as deadline and later persistence when meaningful. If an intervention affects renewal or subsequent claim behavior, the short-run acquisition effect is insufficient for a lifetime-value or welfare claim. Post-acquisition outcomes might require longer follow-up and different denominators, and effects conditional on acquisition face the selection problem described earlier.
An experiment governance record should contain the scientific hypothesis, consumer-protection clearance, sample plan, assignment integrity test, stopping rules, incident escalation, outcome data access, and decision. Sample-ratio mismatches, uneven logging, incompatible concurrent experiments, deviations from planned assignment and interference deserve investigation before interpreting a result. Sequential monitoring must use an appropriate design; repeated ordinary significance tests inflate false positive risk.
Product-Line and International Comparative Studies
Health and social insurance
Health insurance marketplaces involve enrollment windows, eligibility determinations, subsidies, household information, plan benefits and sometimes the continuation of public coverage. The field experiments by Domurat et al. (2021) and Ericson et al. (2023) show that non-price interventions can change take-up in two specific US settings. The experiments are not commercial digital checkout studies. In a subsidized setting, abandonment may leave an eligible household without protection because of administrative burden; the interpretation in an optional supplementary market can differ.
A comparable study of health enrollment should record eligibility determination, subsidy amount where relevant, plan selection, application completion, effective coverage, and continuity. A renewal failure needs a separate cohort from a first purchase. An eligibility redetermination that lawfully terminates participation is distinct from a paperwork failure. Health information requires particularly careful data governance. Survey linkage may be justified only with appropriate safeguards and a clear research purpose.
Medicare plan comparison evidence addresses an enrolled population considering alternative protection, illustrating how information costs can alter switching. The same outcome cannot stand in for coverage among uninsured households. Different hypotheses are needed for initial participation, switching, renewal, and cancellation. Cross-product meta-analysis would require comparable outcomes and intervention definitions; the present evidence is deliberately not pooled.
Motor and property insurance
Motor and property policies can involve instant or near-instant quotes for relatively standardized configurations, but policy terms, declarations, location and risk characteristics differ. Aggregator traffic can include price comparison across many insurers; an individual distributor's unsuccessful quote may correspond to a successful purchase elsewhere. A person-level or need-level denominator matters particularly when one consumer requests several quotes for the same exposure.
Price drift after additional risk information, availability constraints, payment options, and document requirements should be distinct events. A risk-address verification failure requires diagnosis: a genuinely unsupported risk differs from a data validation error. Compare product versions, declared coverage and effective dates, not just website URLs. For renewals, the counterfactual includes retaining existing coverage; in mandatory insurance settings the consequences of noncoverage can extend beyond the household, but those legal conditions differ across jurisdictions.
Historical UK premium finance evidence demonstrates why payment disclosure deserves explicit treatment in personal general insurance. It does not show which stage has the largest attrition in modern motor or property markets. Any local empirical comparison needs dated quotes, comparable offers, channel-resolved outcomes, and information on consumers unable to complete financing.
Life insurance and hybrid advice
Life insurance can involve needs assessment, medical disclosures, additional evidence, underwriting decisions, and considerable time between initial contact and effective coverage. The importance of physical and hybrid distribution identified in EIOPA's 2024 survey summary makes short session windows particularly inappropriate here. A consumer may reasonably seek advice about coverage amount, beneficiaries and policy features. Record referral, advice delivery, application, pending evidence, acceptance, premium, and policy commencement separately.
The distinction between insurer decision and consumer decision is especially important when the application remains pending. When different groups face distinct document collection paths, compare elapsed times and unresolved outcomes as well as formal underwriting disposition. Medical data, where relevant, must remain subject to lawful handling. A research design based on a fixed 24-hour digital conversion rate would systematically misclassify many prolonged but valid journeys.
Add-on and embedded distribution
An embedded offering can appear within a separate principal purchase. The observed purchase funnel denominator may count retail shoppers who never intended to buy insurance. It is important to distinguish exposure to an optional product, expressed interest, separate informed consent, and effective insurance. A high attachment rate is not evidence of appropriate coverage if consumers misunderstand the add-on.
Australia offers a useful institutional counterexample to the assumption that reducing all friction is desirable. ASIC (2021), in its deferred sales model announcement, describes a four-day pause for covered add-on insurance sales and notes exemptions. The precise applicable legal requirements and exemption categories must be verified for an individual product and date. Here the example demonstrates that delaying a transaction can serve a consumer-protection purpose. A metric that treats purchases not completed within the principal transaction as loss would misrepresent the intended timing of such a regime.
Comparative institutional scope
In the European Union, the Insurance Distribution Directive requires attention to customer interests, fair communications, and customer demands and needs. Article 20 requires the distributor to establish demands and needs on the basis of customer information and to provide understandable objective product information before contract conclusion. National transposition and scope determine the operative obligations. EIOPA's 2026a third IDD report summary notes that greater mystery-shopping journey thoroughness had limited association with product alignment. This is supervisory observational evidence, not a randomized session-length experiment. It reinforces the need for tests of meaningful consumer outcomes rather than treating either speed or length as a sufficient proxy.
In the United Kingdom, the FCA Handbook PROD 4.3 distribution rules addresses the distribution strategy and value of non-investment insurance products, including the influence of remuneration and arrangements. The specific rule application must be assessed within the handbook's scope. A regulator or distributor can use funnel information to investigate whether the distribution strategy reaches the identified target market and whether barriers compromise value. The rules do not supply a universal conversion target.
In the United States, insurance supervision is predominantly state-based. The NAIC's 2026 artificial intelligence overview records adoption of the 2023 model bulletin and discusses insurer obligations and governance in AI-enabled activities. A model bulletin and a state's operative implementation are distinct. The present work makes no claim about adoption in every state. A proposed investigation of algorithmic filtering should identify which state rules and insurance products apply before evaluating discrimination or documentation requirements.
EIOPA's 2026b digital platforms commentary highlights questions about AI-mediated distribution roles and potential steering. Platform, aggregator, intermediary and insurer functions should be mapped in actual commercial arrangements. A front-end platform might determine ranking, another entity collect application data, and an insurer make a final offer. A missing transition may occur at any boundary. Regulatory conclusions need responsible-party allocation and evidence; aggregated conversion data cannot identify the obligated entity.
International supervisory guidance, domestic regulation, and private interface standards are different types of evidence. A comparative table of enrollment rates across countries must define whether the denominator is a target population, digital traffic, quote seekers, applicants or offers, and whether an outcome denotes issued or effective coverage. Differences in product mix, compulsory coverage, public subsidies and distribution arrangements can dominate the apparent gap. No global abandonment ranking follows from the records reviewed here.
Reproducibility and Regulatory Implications
A reproducible empirical specification
The empirical research package for a real distribution investigation should initially fix the research question, the institution and regulated roles, jurisdictions, product classes, customer population, calendar period, inclusion logic, channel boundaries, and proposed intervention or observed exposure. Record software versions, schema version and data snapshot fingerprints; separate evidence derived from administrative records, client analytics, inspections and surveys. Prepare an analysis plan before examining outcome differences where practicable.
A minimum event dictionary should distinguish a journey's entry, quote request and result, display of key terms, application submission, eligibility decision, payment attempt and confirmed contract activation. Document which events are client-generated, server-generated or supplied by another organization. An event acknowledgment is evidence of receipt by a system, not necessarily evidence of consumer understanding. A test fixture should include duplicate events, events received out of order, stops with no later record, pending applications, multi-offer journeys, and a referral ending in an offline contract.
Create cohort-level counts using reproducible joins and a documented rule for assignment of a person to a journey. Reconcile transitions to the event log and terminal states to authorized administrative records. Each row of calculated counts should retain the product, channel, period, unit and follow-up window. Retain explicit missing and unknown status. Report operational failures separately from consumer refusals. The data can support descriptive patterns without causality; assign the interpretation level to each statistic.
For accessibility, conduct standardized task testing across supported devices and assistive technologies, with independently selected test scenarios and human observation. Avoid recruitment solely from converted customers. For surveys, retain sampling probabilities where known, response dispositions, question wording, translated versions and estimates of response bias. Qualitative accounts can explain mechanisms and generate hypotheses, yet cannot determine their prevalence in a population without an appropriate sampling design.
Audit and management information
A useful monitoring register contains rates at every meaningful transition, denominators, known error types, elapsed time, human-assistance requests and resolution, offer changes, rejection reasons, pending outcomes, alternate-channel outcomes where verifiable, early cancellation, complaints, results of comprehension tests, and confidence or sensitivity intervals where appropriate. Data owners should explain changes to definitions and show parallel series during instrument migration. A drop in recorded visits after a privacy-interface update may be a measurement break rather than a consumer behavior change.
An insurer should distinguish target-market coverage and commercial lead efficiency. A distributor should distinguish product selection and conversion. Technical providers should support event integrity and accessible recovery paths. Cross-party agreements should clarify evidence ownership, authorized exchanges, incident review and escalation, without transferring regulatory responsibility by assumption. Where privacy prevents detailed linkage, agree on compliant aggregate reconciliation and report the residual uncertainty.
Monitoring triggers should be tied to the potential severity of consumer harm, the reliability of the measure, baseline variation and sample size. A sharp increase in failed address verification across many customers warrants immediate investigation even before causal estimates are available. A modest gap in a small group with broad uncertainty calls for stronger measurement rather than a categorical conclusion. For repeated monitoring, account for false alarms and changes in population composition. The aim is timely correction without overinterpreting noise.
Regulatory inference and safeguards
Market conduct analysis examines both access and quality of treatment. Conversion data can signal an issue under the EU customer-interest framework or UK product governance arrangements, but establish neither a breach nor its absence in isolation. An observed differential rate can motivate investigation into inaccessible pathways, proxy targeting, disclosure timing, pricing, or lawful eligibility. Specific legal rules, product details, and causal evidence determine the final assessment.
A robust investigation creates a trace from the customer's experience to the operational cause, responsible process owner and remediation. For example, a repeated document-upload failure should be reproduced on the relevant devices, linked to the configuration version, tested with a suitable fix, and followed through to verified coverage or informed refusal. A mere reduction in the error event after removal of logging would be an invalid success criterion. Reconcile actual task completion and ensure the error can still be detected.
Experiments should prioritize removing unnecessary burdens and improving comprehensibility, and should stop if material harm arises. Design oversight must pay attention to vulnerable and underserved consumers while protecting personal data. Independent red-team or mystery-shopping studies can reveal hidden pathways and inconsistencies, but they require realistic scenarios and limits on interpretation. External validation is appropriate if a consequential claim relies on an unobserved mechanism. Scientific transparency includes recording what cannot be determined.
Limits of the evidence and research priorities
The evidence corpus combines local causal experiments, regulatory observations, methodological studies, and design guidance. It establishes the importance of plausible mechanisms and demonstrates how to measure them, but it supplies no common cross-market event data, insurer-level causal decomposition, representative worldwide benchmark, or direct calculation of aggregate welfare loss. Abstract-based findings are restricted to the claims explicitly visible in the abstracts. The legal discussion is comparative and selective. A document-level legal audit would require current local implementing rules, product classifications and actual distributor agreements.
A future multi-insurer study should use harmonized event semantics, channel-resolved outcome verification, a common cohort definition within product classes, and transparent selection of institutions. A separate experiment could test a specific identifiable nonessential form burden with randomized assignment and verified outcomes. A third study could combine independently sampled consumer interviews with telemetry to estimate reasons for abandonment and their measurement error. These questions are empirical extensions of established methods. They do not require a new general theory of insurance choice.
Conclusion
Digital insurance abandonment has several distinct empirical meanings. A session may end, a customer may defer a purchase, a platform may fail to transmit an event, a distributor may redirect the customer, an insurer may reject an application, or a consumer may obtain an appropriate contract elsewhere. Scientific reporting requires distinct outcomes and consistent journey denominators. Where coverage is the target, an independently reconciled effective contract event is preferable to a checkout success message.
Stage-specific differences in progression can be substantial even when formal underwriting or final payment rates are similar among those who reach them. Selection changes the observed portfolio and can obscure unmet protection needs. The mathematical examples demonstrate why later-stage parity does not imply equal full-journey access and why missingness can prevent identification of even the direction of a coverage disparity. A causal interpretation requires suitable evidence, explicit assumptions and a defined counterfactual.
The selected field studies establish that administrative and comparison costs can influence particular insurance enrollment and switching decisions. Supervisory evidence and cross-sector research identify additional mechanisms requiring investigation. Their institutional settings differ, preventing direct transfer of effect sizes to digital commercial sales. Consumer welfare depends on whether protection is appropriate, understandable, accessible and freely chosen; completion alone cannot establish these attributes.
The research therefore supports an operational conclusion: measure conversion, coverage and consumer treatment as separate outcomes; establish exposure and transition denominators; distinguish competing and unobserved states; validate linkage and accessibility; and use ethically permissible causal tests to evaluate remediations. Under these conditions, funnel analysis can contribute to distribution efficiency, risk management, market conduct oversight and evidence-based improvement of insurance access.
References
EIOPA (2024). EIOPA report takes the pulse of digitalisation in the European insurance market. April 30. https://www.eiopa.europa.eu/eiopa-report-takes-pulse-digitalisation-european-insurance-market-2024-04-30_en
EIOPA (2026a). EIOPA publishes third Report on the application of the Insurance Distribution Directive. March 30. https://www.eiopa.europa.eu/eiopa-publishes-third-report-application-insurance-distribution-directive-2026-03-30_en
EIOPA (2026b). AI, digital platforms and the future of insurance distribution. Eurofi Magazine, September. https://www.eiopa.europa.eu/publications/ai-digital-platforms-and-future-insurance-distribution_en
European Parliament and Council (2016). Directive (EU) 2016/97 on insurance distribution. OJ L 26, 19–59. Consolidated text, October 9, 2024. https://eur-lex.europa.eu/legal-content/EN/TXT/?qid=1773178225329&uri=CELEX%3A02016L0097-20241009
European Parliament and Council (2016). Regulation (EU) 2016/679 (General Data Protection Regulation). OJ L 119, 1–88. https://eur-lex.europa.eu/eli/reg/2016/679/oj/eng/
Financial Conduct Authority (2015). TR15/5: Provision of premium finance to retail general insurance customers. May 10; marked historical August 2025. https://www.fca.org.uk/publications/thematic-reviews/tr15-5-provision-premium-finance-retail-general-insurance-customers
Financial Conduct Authority (2025). Digital design in customers' online journeys: good practice and areas for improvement. July 31; updated December 3. https://www.fca.org.uk/publications/good-and-poor-practice/digital-design-customers-online-journeys
Financial Conduct Authority (2026). FCA Handbook, PROD 4.3: Distribution of insurance products. Online rulebook, consulted November 19, 2026. https://handbook.fca.org.uk/handbook/prod4/prod4s3
International Association of Insurance Supervisors (2025). IAIS publishes Application Paper on fair treatment of a wide range of consumers. July 7. https://www.iais.org/2025/07/iais-publishes-application-paper-on-fair-treatment-of-a-wide-range-of-consumers/
National Association of Insurance Commissioners (2026). Insurance Topics: Artificial Intelligence. Updated April 3. https://content.naic.org/insurance-topics/artificial-intelligence
Australian Securities and Investments Commission (2021). ASIC releases guidance and customer information requirements to implement the new add-on insurance deferred sales model. Media release 21-189MR. https://www.asic.gov.au/about-asic/news-centre/find-a-media-release/2021-releases/21-189mr-asic-releases-guidance-and-customer-information-requirements-to-implement-the-new-add-on-insurance-deferred-sales-model
World Wide Web Consortium (2024). Understanding Success Criterion 3.3.8: Accessible Authentication (Minimum). WCAG 2.2 supporting documentation, online version. https://www.w3.org/WAI/WCAG22/Understanding/accessible-authentication-minimum.html
Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3), 581–592. doi:10.1093/biomet/63.3.581. https://academic.oup.com/biomet/article-abstract/63/3/581/270932
Lee, D. S. (2009). Training, wages, and sample selection: Estimating sharp bounds on treatment effects. Review of Economic Studies, 76(3), 1071–1102. doi:10.1111/j.1467-937X.2009.00536.x. https://academic.oup.com/restud/article-abstract/76/3/1071/1590707?login=false
Domurat, R., Menashe, I., & Yin, W. (2021). The role of behavioral frictions in health insurance marketplace enrollment and risk: Evidence from a field experiment. American Economic Review, 111(5), 1549–1574. doi:10.1257/aer.20190823. Published abstract indexed by RePEc. https://ideas.repec.org/a/aea/aecrev/v111y2021i5p1549-74.html
Ericson, K. M., Layton, T. J., McIntyre, A., & Sacarny, A. (2023). Reducing administrative barriers increases take-up of subsidized health insurance coverage: Evidence from a field experiment. NBER Working Paper 30885. doi:10.3386/w30885. January working-paper version. https://www.nber.org/papers/w30885
Kling, J. R., Mullainathan, S., Shafir, E., Vermeulen, L. C., & Wrobel, M. V. (2012). Comparison friction: Experimental evidence from Medicare drug plans. Quarterly Journal of Economics, 127(1), 199–235. Publisher abstract. https://academic.oup.com/qje/article-abstract/127/1/199/1833354?login=false
Luguri, J., & Strahilevitz, L. J. (2021). Shining a light on dark patterns. Journal of Legal Analysis, 13(1), 43–109. doi:10.1093/jla/laaa006. https://doi.org/10.1093/jla/laaa006
Competition and Markets Authority (2022). Evidence review of online choice architecture and consumer and competition harm. April 5. https://www.gov.uk/government/publications/online-choice-architecture-how-digital-design-can-harm-competition-and-consumers/evidence-review-of-online-choice-architecture-and-consumer-and-competition-harm