Compare coverage before comparing percentages
Compare DMARC Reports by Identity, Coverage and Counts
Several DMARC files do not automatically make a complete picture of email activity. They may cover different receivers, domains, periods, or repeated deliveries of the same report. A careful comparison starts by defining that coverage, then distinguishing repeated report content from potentially overlapping observations. The resulting totals describe reported message instances in the accepted files. They are not a deduplicated ledger of unique emails, sent mail, delivered mail, or inbox placement.
Open the DMARC Report Analyzer, choose Files, and select your authorized report files under Report files. Paste XML accepts one supplied XML report as text. Load example fills input only; Analyze reports explicitly starts processing. Keep the source files privately because the analyzer does not retain a project or comparison history after the working session is cleared.
Define the comparison population before loading a batch
State the question precisely: comparing the same receiver across two periods, comparing several receivers for one domain, or reviewing several policy domains are different exercises. For each report, retain the declared reporting organization and contact identity, report ID, policy domain, and begin and end times. These fields establish the declared scope, not independent proof that the named organization generated the file. Use the report metadata rather than relying on a downloaded filename, email subject, or folder name. A file renamed after download still describes its original internal window. A neatly organized folder can still contain reports for the wrong domain or a repeated attachment.
Keep policy domain, author domain, and authentication domains distinct. A report belongs to its published policy context, while individual rows can name author domains and raw SPF or DKIM identities. A source-IP grouping spanning several contexts does not make those domains interchangeable. For a domain-specific question, inspect the relevant reports and rows rather than assigning all activity from a shared address to the first brand you recognize. If a comparison includes parent domains and subdomains, preserve their actual declarations. Do not infer that a reported policy for one domain applied identically to every related domain or that current organizational ownership establishes the historical configuration.
Choose comparable coverage deliberately, including which receiving organizations and UTC windows are represented. The Google report guide, Microsoft report guide, and Yahoo report guide provide separate collection contexts. A larger receiver set may increase the observed total without indicating a change in sending behavior. A missing receiver may reduce it without indicating successful remediation. Note the chosen population outside the tool before drawing a before-and-after conclusion, and keep a list of expected but unavailable reports so their absence does not become an invisible assumption in the final comparison.
Separate exact duplicate reports from identity conflicts
The analyzer checks a report key comprising the normalized reporter identity, report ID, policy domain, and beginning and ending timestamps. That is a compound identity, not just the report ID by itself. Two different receivers can use similar-looking identifiers without describing one report. Likewise, an identical ID attached to another domain or window is not automatically the same keyed report. Reporter identity includes the declared organization and contact context; it is not inferred from the file extension. Keep the displayed metadata available when explaining which reports were retained. The key is a local comparison rule for the supplied declarations, not an external verification of reporter identity.
When the same key also has exactly equal meaningful parsed content, the repeated report is skipped and the duplicate count is reported explicitly. This is report-level deduplication. Different wrappers or filenames do not create new observations, and comparing meaningful content is not the same as comparing the compressed file bytes. A copy of one report inside a ZIP and another copy as XML can therefore represent the same report. Do not subtract its row counts manually after the analyzer has already skipped it. Record both the accepted-report scope and the duplicate count when preparing an external comparison note so another reviewer understands why selected files and retained reports differ.
The same key with different meaningful content is a conflict, not a harmless repeat. Changed counts, row results, or other meaningful declarations cannot be resolved by keeping whichever file was selected first or appears newer in your downloads folder. A conflict aborts the batch atomically; there is no partial successful result for the remaining files. Preserve the competing originals privately and resolve their provenance through the authorized report-handling process outside the analyzer. Do not edit a report ID merely to bypass the conflict check. That would disguise the disagreement and make a later total look more certain without supplying evidence about which version, if either, should be used.
Treat overlapping windows as a warning about interpretation
Accepted reports can cover overlapping or touching periods, and the analyzer warns about those boundaries. A warning is not an instruction to delete one report automatically. For example, an invented window from 00:00 to 23:59 and another spanning noon into the next day share part of their stated coverage. Their aggregate rows do not identify which underlying messages, if any, appear in both. Touching endpoints deserve review too: do not silently invent a convention that proves they are disjoint. Preserve the declared timestamps, receiver and domain contexts and explain the overlap instead of claiming the tool measured a uniquely counted timeline.
An IP match cannot solve this problem. The same address may contribute many different messages, and different addresses may relate to one underlying message passing through different systems. Matching an address, count, and author domain across reports still does not create a message identifier. The analyzer therefore does not deduplicate messages by IP, discard distinct reports because their rows look similar, or estimate the overlap fraction from the duration of the intersecting windows. Subtracting half of a count because half of a period overlaps would assume a distribution that the report never supplied. Keep the sum labeled as reported instances whenever you retain both reports.
Missing coverage requires equally careful language. If a report expected for Tuesday is unavailable, Tuesday is not measured as zero activity. A report whose declared window spans several days is not a series of daily observations that can be split evenly. It cannot reveal whether all instances occurred on the first day or were spread across the interval. Keep gaps and irregular windows visible in external notes, and compare like-for-like subsets when possible. Where coverage cannot be matched, describe a difference between supplied report sets rather than a clean trend. Avoid annualizing, daily averaging, or inventing an improvement rate from a partial collection of attachments.
Start with identity and coverage
Inspect a bounded report batch
Review duplicate counts and overlap warnings before interpreting totals or comparing source rows.
Calculate from weighted counts and preserve unknown outcomes
Use each record’s count as its weight. If one invented report contains 90 passing instances out of 100 and another contains 100 out of 1,000, their combined passing total is 190 out of 1,100, about 17.3 percent. Averaging the two report percentages would yield 50 percent and badly misrepresent the supplied population. Counting passing rows instead of instances creates a similar error because row sizes can differ widely. Calculate only for a clearly named set of accepted reports, and identify any narrowed report or source selection beside the number. A large report and a small report are not equal votes in a rate calculation.
The derived DMARC outcome comes from reported aligned SPF and DKIM, not arbitrary raw authentication passes. Either aligned pass yields pass, both aligned failures yield fail, and remaining combinations are unknown. Keep unknown instances in view and explain the denominator used for any percentage rather than quietly dropping them to improve a headline. With no counted instances, the rate is unavailable, not zero percent. A zero-count row can still be inspected as a declaration but adds no instances. The alignment guide explains why raw authentication entries and aligned results cannot be substituted for each other during comparison.
Compare disposition separately from alignment. Reported none, pass, quarantine, reject, and other values are handling declarations, not alternate calculations of DMARC success. Preserve raw override reasons and the policy options reported in each file, including legacy pct or newer np, testing, and discovery_method where present. A change in the handling distribution might accompany a policy or receiver-context change rather than a change in underlying aligned outcomes. Conversely, an improved aligned distribution does not establish inbox placement. Do not describe a count difference as recovered deliveries or prevented attacks without separate evidence that actually measures those claims.
Move from a batch difference to its contributing rows
Start with the accepted report list and warnings, then use the report filter to inspect a specific context. Review source-IP summaries and open row details to see which counted groups contribute to a difference. Keep author domains and report boundaries attached during this process; a pooled IP total can conceal one domain improving while another worsens. Use the plain raw XML viewer for the selected report to inspect source wording when a declaration needs clarification. The viewer does not resolve a disputed report or fetch missing evidence. It shows the supplied XML as text, not a live remote document or a validated account of the original mail transactions.
When a source becomes prominent, use the failures and source IPs guide to compare its context with an approved sender inventory outside the analyzer. A newly visible source can reflect a newly collected receiver, a changed window, or an approved service change, not necessarily new unauthorized activity. The tool does no ASN, PTR, reputation, geolocation, or DNS lookup and cannot name an IP operator or attacker. Preserve a concrete question for the domain’s administrator: which rows changed, how much reported volume they represent, which inventory entry might explain them, and which coverage differences remain unresolved.
For a specific delivered message, the Email Header Analyzer addresses a separate single-message reading task. It does not automatically correlate that header with aggregate rows or establish a unique-message count for this comparison. Keep external corroboration separate from the batch statistics. When writing a conclusion outside the tool, state the accepted reports, relevant UTC windows, domains, duplicate count, overlap warnings, and the narrow observed change. That makes the comparison reproducible from the originals without pretending that the analyzer supplies export, Copy, PDF, storage, history, saved projects, automated redaction, or an ongoing monitoring service.
Respect batch limits and close the local comparison cleanly
A batch accepts up to 10 XML, GZIP, or ZIP files, each no larger than 4 MiB, with 16 MiB total source and 16 MiB total expanded XML. Each expanded XML is limited to 4 MiB. The additional ceilings are 5,000 rows per report, 10,000 rows per batch, depth 24, 150,000 XML nodes, and 4,096 UTF-16 units per leaf text value; all apply together. UTF-8 is required, with an optional UTF-8 BOM. ZIP permits exactly one XML plus optional directories. Consult the file-format guide for archive rejection rules and supported namespaces rather than modifying evidence to bypass a limit.
The local outer worker has a 12-second deadline and publishes a batch only after complete success. A corrupt archive, rejected XML, limit violation, or identity conflict cannot leave a partial successful comparison. DTD and entity declarations and external processing instructions are rejected, with no external entity calls. Parsing is bounded structural interpretation, not full XSD certification. The analyzer connects no API or mailbox, uploads no reports, and performs no domain, SPF, DKIM, or DMARC revalidation. These boundaries are especially important when a comparison seems surprising: a successful parse cannot certify that the source files are authentic, complete, or sufficient for the operational conclusion you hoped to draw.
Clear removes the file input, pasted XML, statistics, report filter, source and row details, raw viewer, and all derived results while stopping active work. Hidden-page cleanup, pagehide, pageshow, and freeze do the same; returning to the page starts without the prior comparison. This is not secure memory erasure or protection from extensions, a compromised device, or a hostile parent interface. Retain authorized originals and external notes under your own rules. Return to the DMARC Report Analyzer overview for another bounded analysis, or choose a separate utility from Tools without putting report contents in the destination URL.
Questions about Compare DMARC Reports
Does the analyzer remove duplicate emails across reports?
No. It skips reports with the same compound identity and exactly equal meaningful parsed content, and counts those skips. Overlapping reports remain a warning because aggregate rows cannot establish which individual messages overlap. Batch totals are reported instances, not unique emails.
What happens when the same report identity has different counts?
Different meaningful content under the same normalized reporter identity, report ID, policy domain and window is a conflict. The whole batch fails rather than choosing a version or publishing partial results. Resolve the competing originals outside the analyzer without changing identifiers to hide the disagreement.
Can a missing report be included as zero percent failure?
No. Missing coverage is unavailable evidence, not a zero-valued observation. Compare matched contexts where possible, calculate from instance counts rather than averaging report rates, and describe unmatched receiver or time coverage explicitly. A zero denominator also produces an unavailable percentage, not a measured zero.
Continue with another DMARC-report guide
Read an aggregate report in the DMARC Report Analyzer & XML Viewer or choose a related task below. Explore more practical utilities in all free tools.
Keep every total attached to its scope
Compare the reports you actually have
Analyze supplied files locally and keep report-level duplication distinct from unknown message overlap.