Preserve the denominator before comparing the trend
An authentication-adoption trend can change because domains changed configuration, because the sample changed or because the measurement method changed. Those explanations are not equivalent. A longitudinal comparison should retain the population, metric definition, parser and error handling for every period so an apparent improvement is not created by silently removing difficult cases or adding a different mix of domains.
This article proposes a reproducible method and does not present a study result. RFC 9989 defines current DMARC behavior, while RFC 7208 defines SPF publication and evaluation. Standards help specify what a measurement means; they do not supply the population or justify a causal interpretation of a trend. No domains were queried for this manuscript.
Define the measured property and unit
State whether the metric counts exact-name record presence, valid applicable policies, enforcement requests, observed aligned messages or another clearly defined property. A configured domain and a message are different units, and a domain can support multiple sending workflows. The denominator must correspond to the numerator. An observed-message pass rate cannot be substituted for the share of domains publishing a policy without changing the question.
Record the sample's origin, inclusion criteria and intended scope. Website domains, customer domains and actual SMTP identities can produce different results. If the study uses a convenience list, retain that limitation. Avoid saying the internet improved when the evidence concerns a selected customer cohort. This draft contains no population count or adoption percentage to generalize.
Track membership as a first-class dataset
Keep a versioned membership table with each unit's stable identifier, normalized name, entry date, exit date and reason for inclusion or exclusion. Preserve domain renames, ownership changes and mergers where they affect interpretation. Do not assume that the same text label represents the same organizational relationship throughout the study. A domain can remain in DNS while its business role changes.
Choose whether the analysis uses a fixed cohort, a repeated cross-section or both. A fixed cohort can support comparison of the same units but may become less representative over time. A repeated cross-section can reflect the current population but mixes membership change with configuration change. Report the chosen design and, where useful, show both views separately rather than hiding their different denominators.
Version the measurement implementation
Retain query code or method description, resolver configuration, timeouts, retries, normalization rules, parser version and standards references for each run. A parser upgrade can reclassify the same raw record even if DNS did not change. RFC 9989's current discovery rules, including the DNS Tree Walk, illustrate why a method based on a different discovery algorithm may not be directly comparable without analysis.
If the method changes, preserve the previous results and identify a break in the series. Where raw evidence permits, reprocess earlier observations with the new parser and report that as a separate comparable series. Do not quietly overwrite historical classifications and present the revised trend as the original measurement. Reprocessing cannot reconstruct DNS answers that were never retained.
Keep failures in the accounting
Distinguish name error, no requested data, timeout, server failure, invalid policy and unsupported interpretation. A failed query should not automatically become absence of deployment, and excluding all failures can inflate the apparent adoption rate. Retain attempted observations and explain the treatment of each error category. If a retry succeeds, preserve both the initial failure and the rule used for the final classification.
RFC 1035 defines DNS response and record structures. Preserve raw record boundaries before protocol parsing, especially when multiple TXT strings or records are present. A change in the parser's handling of split strings or duplicate candidates can create a false trend if the underlying evidence is lost. Record structure is part of reproducibility, not merely display formatting.
State every fraction explicitly
For each period, retain the number of units selected, attempted, successfully observed, classified, excluded and unresolved. Then define the numerator for the reported property. Adoption among all selected units differs from adoption among successfully classified units. If both measures are informative, report both with their labels and uncertainty rather than presenting whichever is more favorable.
Do not invent confidence intervals, statistical significance or representative coverage without a justified sampling model. Missing observations may not be random, and the domains that time out can differ systematically from those that respond. A transparent descriptive comparison can still be useful without unsupported inferential claims. This manuscript supplies no numerical estimates or fabricated uncertainty calculations.
Separate configuration change from message outcomes
A domain moving from no observed policy to an observed policy is a configuration observation. It does not demonstrate that every legitimate message now aligns or that fraud decreased. RFC 8601 describes trusted authentication-result evidence that could support message-level analysis, but that analysis needs its own population and retained receiver observations.
RFC 9990 defines aggregate reporting. If reports contribute to a longitudinal study, preserve reporter coverage, periods and count dimensions separately from DNS observations. Changes in reporting coverage can alter apparent authentication performance even without changed sender behavior. Do not join a DNS cohort trend and a report-count trend into one metric unless the method explains the compatible units and limitations.
A longitudinal reconciliation worksheet
Create a worksheet with period, sample version, membership changes, method version, source references, attempted count, error categories, exclusions, classified count and metric numerator. Add the rule used for each denominator and a comparison status indicating whether periods are directly comparable, reprocessed or separated by a method break. Retain raw evidence hashes and the error/exclusion log.
The review should ask whether the same question is being answered in each period. If a new receiver source, parser rule or domain class enters the study, document its effect before interpreting the direction of change. Have an independent analyst reproduce a small set of classifications and the aggregate accounting when original data becomes available. No such reproduction was performed here.
Fictional diagnosis: a better percentage after removing failures
A fictional team publishes a higher adoption percentage in its second period. The first denominator included every attempted domain, while the second excluded all timeouts and invalid responses. The apparent improvement may therefore reflect changed accounting rather than additional deployment. The proposed review reconstructs both periods using the same declared rule and separately reports the unresolved cases.
The team would also check whether membership or parser behavior changed. It would not claim a causal improvement from the original percentages until those differences were understood. This example supplies no actual counts, domains or study observations. It is a fictional illustration of denominator drift and not a measured finding about an authentication ecosystem.
Publish comparability limits with the result
A defensible trend report states the population, time windows, measurement versions, numerator, denominator and errors. It explains changes to the method and does not silently convert deployment observations into message compliance, delivery or prevented losses. The reader should be able to determine what changed in the measured systems and what changed only in the measurement process.
The catalogue requires a versioned dataset, query method and error/exclusion log that another analyst can reproduce. Those original artifacts remain open. Human specialist review, source applicability, production routes, accessibility and owner release also remain open. This draft provides a methodological design and fictional diagnosis, not an executed longitudinal study or a claim of improved adoption.