Media research methods and limitations

Here, a media record represents one source web address (URL), with its saved title, short description and other available metadata. RSRC adds a relevance decision, themes and connections to related coverage. It is a research record about a page, rather than a copy of that page.

GDELT describes its Article List (GAL) as a standardized and minimized basic set of metadata for every article GDELT monitors. Read GDELT’s description and field definitions (6 December 2021). Metadata describes a page; it is not the page’s full contents.

Our collection includes online news coverage, commentary and blog posts, with pages from news organizations, broadcaster websites and blogs. A page may contain text, video or audio. We have not classified every record by media format; a television or radio website does not establish that the item is a broadcast report.

The main analysis uses saved titles and descriptions. We did not watch or listen to every report, collect broadcast recordings, or analyze audio/video transcripts or the full text of every page. Selective deeper reading is documented separately. This collection uses GAL metadata, not all of GDELT’s other news, television or visual datasets.

Official complaints, academic studies and media reports show different parts of the problem. This collection makes one of those sources easier to examine: reporting about romance scams.

The inspiration is Herrera and Hastings’s 2024 paper, The Trajectory of Romance Scams in the U.S.. That study compared web-search interest, news metadata, research publications and government reports. Its news searches used several synonyms and removed duplicate results and keyword false positives.

This project develops the media-research question further. It uses GDELT’s multilingual article metadata, records contextual inclusion decisions, codes reported themes, and distinguishes repeated wording from stronger links between reports. It has a different source, scope and procedure, so its totals should not be spliced into the paper’s trend series.

1. 1Collect
2. 2Screen
3. 3Review
4. 4Code
5. 5Check

1. Collect available GDELT files. Files are requested for defined UTC observation windows. Checksums preserve their identity. An unavailable slot means no usable file was retrieved for that timestamp. It does not establish why it was unavailable, whether GDELT produced a file, or whether any news was published.
2. Screen titles and descriptions. A saved multilingual vocabulary finds candidates. A separate terminology supplement and fixed samples of nonmatching records help expose gaps. A keyword match alone does not establish relevance.
3. Review the available context. Records are included, excluded or left unresolved. Romance involvement must be supported; investment fraud and money laundering alone are insufficient. Inclusion means relevant coverage, not a verified crime.
4. Assess themes and relationships. Each included record receives the applicable codebook assessment. Allegations, warnings, attempted actions and completed reported developments stay distinct. Countries have explicit roles. Direct connections retain their reasoning; articles are not merged into cases through chains of links.
5. Validate and preserve. Source bindings, complete assessments, database totals and repeat imports are checked. Corrections create new versions and preserve earlier decisions. Code and same-system review checks are not independent proof of accuracy.

Bulk processing takes place locally. Codex assists with contextual review, coding and quality checks. This is model-assisted research without an outside review team; there is no independent inter-rater reliability estimate. The website serves curated results and does not generate new analysis when you open a page.

The story register groups related coverage: possible case families, initiatives and media works. Each membership keeps its source-record revision and a reason. A group is a research hypothesis, not an independently verified case.

The expanded review spans 1 January to 4 October 2026. Local programs screened all 6,484 included records for recurring coverage. We examined 324 saved connections, screened 431 additional title groups, and read the saved source passages for 265 selected groups. Named-subject searches found further reports, including coverage in different languages.

The result contains 334 story hypotheses covering 1,734 source URLs, with 23 direct follow-up connections across the collection. The 83 saved possible connections keep their uncertainty. Similar headlines about different victims, losses or proceedings were separated. General advice and vague similarities were not turned into cases.

This pass reused saved titles, descriptions and evidence passages. It made no new publisher-page downloads or paid API calls. It is a systematic, bounded screen across the collected period, not exhaustive discovery of every recurrence or full-text verification of every article. A name can connect several proceedings or experiences; groups can overlap and must not be added as incident totals.

Timelines sort URLs by first observed metadata date. They do not reconstruct incident dates or assume that a later record is a new development. Arrest, charge, requested sentence, imposed sentence and other stages remain distinct. A follow-up needs a direct explanation of what the reporting adds.

Several URLs can repeat one story. Source hosts are not necessarily independent publishers, and repetition does not measure audience size or public awareness. A record without a story assignment is not automatically a new case. All included records remain searchable with their existing theme assessments.

The first public transfer omitted the original five-week story register and three saved follow-up links. These were restored before the wider review. Transfers now check for missing stories, members and outdated revision bindings. Reuse is allowed only when the saved source and interpretation remain unchanged.

Across the 40 collected windows, the screening programs read 129,882,128 metadata rows, including 129,882,128 valid record objects. These describe all topics in the retrieved files, before relevance screening.

Each collection window is counted once, even if its processing was rerun. A URL can occur in many metadata rows, so the first figure is not a count of unique articles. We have not measured distinct URLs or duplicates across the entire all-topic corpus. The smaller candidate and outcome counts are deduplicated by normalized URL across the reviewed collection.

Screening volume documents the scale of the work. It does not measure global coverage, search recall or how often scams occur. Most rows were screened programmatically; only selected candidates and audit samples received contextual review.

A second set of search terms checks for wording the primary search might miss. We also review a fixed sample of records that did not match: normally 56 observations per full week, spread across days and four language groups. These checks can reveal gaps, but they cannot tell us what percentage of all relevant reporting we found.

An observation is not an incident date

“Observed” means the record appeared in the collected GDELT metadata. An old court story may reappear in a later window. Only explicit dates support a reported-event date; missing years are not guessed.

Several URLs may repeat one story

Shared wording can reveal distribution, but it can also be generic text that a website uses on many pages. It does not demonstrate independent confirmation, additional victims, audience size or increased public awareness. A case connection needs more than similar words.

Not reported is not absent

A short description may omit emotional harm, reporting barriers or a later outcome. An unassigned theme means the available text did not establish it. It is not evidence that it never happened.

• Uneven coverage. GDELT, source availability and the vocabulary select what can be found. Languages, regions and smaller sources can be missed.
• Mostly short source descriptions. Metadata can be incomplete, misleading or internally inconsistent. Selective deeper reading is recorded separately and does not make the entire collection full-text verified.
• Repeated and promotional material. Syndication and user-generated blogs can dominate a window. The source comparison filter exposes one observed concentration; it does not endorse other sources.
• Unverified allegations. Source statements remain attributed. Charges are not convictions; requested compensation is not money returned.
• Limited validation. Automated tests establish software behavior. Same-system contextual checks can repeat the same interpretive error, especially across languages.
• No population or loss totals. Do not add reports as incidents, sum repeated loss amounts, or use theme percentages as victim prevalence.

When comparing periods, inspect available files, source composition and review scope first. More records can reflect changed collection or repeated coverage rather than a change in offending.

March reporting announced a change in Japan’s fraud statistics. A later police bulletin explains that the revised categories also apply to figures for earlier months in 2026. SNS investment and romance fraud now sit within the broader “special fraud” total. Comparing that total with an older series therefore requires checking what each series includes.

This is a selective deeper-reading example, separate from the title-and-description assessments. It does not make every media record independently verified. Source: Japan’s National Police Agency, bulletin of 20 April 2026. Summary and interpretation by RSRC; agency information used under its public-data reuse terms.

Australia’s March 2026 taskforce summary reports 377 online assets referred for removal and 1,004 suspected transactions referred for investigation or blocking. These are referrals, not counts of confirmed removals or recovered payments. It also describes a support trial with few referrals; that alone does not establish effectiveness.

RSRC interpretation based on ACCC content, published 6 March 2026, under CC BY 4.0. No agency endorsement is implied. This selective source check is separate from the metadata assessments.

2026 year-to-date coverage with expanded story review. Method RSRC-QCA-2026-10-08.2; codebook 1.1.0.

Each included record is assessed against these 24 themes. A theme is assigned only when the saved evidence supports it. Several themes can apply to the same record.

Data from The GDELT Project, using its GDELT Article List (GAL). Collection, selection, coding and interpretation are by RSRC; GDELT does not endorse these assessments.

GDELT’s published terms of use permit use and redistribution of its datasets without a fee, with attribution and a link to GDELT. We checked these terms on 9 October 2026. The GAL documentation describes using its metadata to help readers find and link to coverage.

This permission supports sharing our metadata-based research and aggregate findings. It does not establish permission to republish publishers’ full articles, photographs or other separately protected material. This site links to sources and presents RSRC’s assessments; it does not distribute the collected article bodies.

The graphs count distinct included URLs within each observation month. The chart builder applies the same filters as the record search. Its percentage view divides matching URLs by all included URLs observed within the same month and selected dates, before theme, text, language, country or coverage-type filters. Line and bar views show the same values. Clicking a point preserves the filters and narrows the dates. A record with several themes counts once in each applicable theme. Months without any included records have no percentage estimate. These measures describe the collection, not worldwide offending or victim prevalence.

Earlier years and later weeks will be added to the existing record. Stable identifiers and retained revisions let researchers distinguish new coverage from a correction. A new year will not replace an old one.

Possible additions include saved searches, clearer timelines of reported developments, language-specific retrieval improvements, and comparisons across years with coverage controls.

A downloadable research dataset

A future download could contain stable record identifiers, source URLs, reviewed themes and a data dictionary. It would need a versioned citation, correction policy, privacy review and clear reuse terms. GDELT attribution would accompany the download. Separately protected third-party material would require its own permission; article bodies and images would not be included.

This would help other researchers inspect decisions and build on the work. A downloadable dataset is a future option; this release provides a searchable research collection.

• Herrera, L. D., & Hastings, J. (2024). The trajectory of romance scams in the U.S. In 2024 12th International Symposium on Digital Forensics and Security (ISDFS) (pp. 1–6). IEEE. https://doi.org/10.1109/ISDFS60797.2024.10527224
• GDELT Article List documentation.
• RSRC Research Insights brings together findings from research publications; this media collection is a separate evidence source.