Identify the moving part
A news page may revise prose, a dashboard may load new numbers daily, a database may re-rank results and online documentation may replace text without displaying a version. Establish whether the content, underlying data, presentation or all three can change. That determines the evidence you need.
Keep the source's displayed data date separate from access time. “Data to August” and “accessed in September” are different claims. If no update date appears, do not invent one. Your screenshot timestamp documents access, not publication.
Treat filters and parameters as part of the source
Dashboards and databases construct the visible result from selections. Record query, period, geography, unit, category, sort and filters. Copy a parameterised URL if it retains the state and test it in a fresh browser. Session links may expire or disclose tokens and should never be published.
Capture defaults that the interface applies silently. A later change to the default range can alter the result. For paginated tables, state whether you exported all records or inspected one screen. Record the platform's definition for calculated rates, seasonal adjustments or suppression symbols.
Preserve a lawful research state
Use official CSV, PDF or report exports where terms and data protection allow. Store export date, filename and checksum. PDF preserves presentation; CSV better preserves values, so both may have a role. A screenshot supplements the record but loses hidden values and machine-readable metadata.
A web archive may capture static prose but fail on interactive applications. Test the result. For important data, prefer a versioned download or API response with documented request parameters. Do not place personal, restricted or licensed data in a public package merely to make the citation convenient.
Split work between the reference and research log
The reference list can contain organisation or creator, page title, displayed date, URL and access date in the required style. The method or data note records filters and export details. Long parameter lists may sit in a reproducibility log if the main citation points to it unambiguously.
A generic form is: “Organisation (date/state), Title [dynamic database], URL, query …, data state …, accessed [date, time, zone].” Use only fields you verified. A version DOI or fixed dataset should take precedence over a volatile dashboard link.
Reconstruct the result without memory
Open the source in a private session and reproduce the selection from the record. Compare result count, units, columns and data date. Preserve a changed result as a new access event rather than overwriting the earlier state. Decide which one supports the sentence and whether the manuscript needs revision.
Inspect the retained file for completeness and rights. Remove tokens from addresses and document properties. For archived HTML, use the original–archive–access date schema. Where change is the research object, cite and compare both states.
Do not confuse a stable snapshot with a true claim
A snapshot shows what was observed, not that the source was accurate. For central statistics, preserve the definition and underlying publication and seek an independent source. If the platform later posts a correction, record it and update public work under the relevant process.
Review dynamic citations shortly before submission. Test addresses, filters and exports, and log failures. When preservation is impossible, narrow the claim to what your dated evidence genuinely supports. Explain interface limitations instead of presenting an unrecoverable number as permanent.
Read calculated dashboard indicators critically
Record numerator, denominator, unit, population and revision policy rather than the displayed value alone. Check notes on seasonal adjustment, estimation, disclosure suppression and rounding. A percentage can move because records were added, earlier data revised or a definition changed. Preserve the displayed definition with the export. If an underlying table is available, cite it separately and reconcile your calculation with the interface, documenting filters and missing values rather than choosing whichever figure is more convenient.