Skip to main content

8. Findability and citation

Question: Can the data be found and cited back to the page?

Rationale​

The point of extraction is reuse. A user who searches for "undernourishment by district 2024" should find the table, open its data, and cite it as the report, the page, and the table number. An AI system that answers a question from the table should cite the same page. Findability makes the first possible; a citation that reaches the page and the bounding box makes the second verifiable.

Extracted data also age. A document is corrected, a series is revised, an extraction error is found. The record of corrections, and the monitoring of where the data are used, are what keep extracted data trustworthy after publication.

Target state​

  • Extracted tables appear in the catalog and in search under their printed titles and their subjects.
  • Every table has a citation to the document, page, and table number, and a stable URL; every value can be traced to a page and a bounding box.
  • Corrections are recorded as errata with versions, and the use of extracted data is monitored.

Maturity levels​

LevelSteps
FoundationalPublish the table titles in the document's page and the site search. Give a citation format for extracted data.
AI-readyCatalog the table records with schema.org markup and stable URLs. Cite to the page and table number, with the document's DOI where it has one. Log corrections as errata.
AI-nativeServe extracted data through the catalog API and agent interface with page and bounding box per value; register DOIs for extracted tables; monitor mentions of the tables and the document.
Running example

EX-REP-2025-03-T3.1 is found by a search for its printed title and by the subject "undernourishment". Its citation names the report, page 24, and the table number, and its tidy file traces every value to page 24 and the bounding box.

Recipes​

Recipe

8.1 Make extracted tables findable

Level
AI-ready
Skills
Catalog administrator
Indicative time
Part of publication

Result: a search for a table's printed title, or for its subject, returns the table record and its data.

  1. Publish the table records in the catalog (chapter 7), with the printed title, the subject keywords, and the series the table reports. NADA indexes tables with the other data types.
  2. Add schema.org Dataset markup to the table's page, with isBasedOn pointing to the document and distribution pointing to the tidy file, as in recipe 1.2 of the dissemination cookbook.
  3. Add the tables to the known-item question set of the dissemination cookbook (chapter 6), so that search over extracted data is scored with the rest.

Worked example. After the four table records are published, the catalog administrator searches the catalog for "undernourishment by district" and gets Table 3.1 first. A search for "hunger by region" returns nothing, because neither word is in the record, so the administrator adds "hunger" and "region" to the table's keywords, which the series record already carries. Two questions, "undernourishment by district 2024" and "which region has the most hunger", go into the known-item question set with the table as the expected result, and the next scoring run reports both found in the top five.

What this does not do. Findability by title serves users who know roughly what they want. Users who ask a question in their own words are served by search by meaning over the records, which the dissemination cookbook covers.

Recipe

8.2 Cite to the document, the page, and the table

Level
Foundational
Skills
Dissemination officer
Indicative time
One hour to set the format

Result: the organization has one citation format for extracted data, and the same format is used in the record, on the page, and by systems that quote the data.

The format names the organization, the year, the document, the table, the page, and the document's address with a page fragment that opens the PDF at the right page:

National Statistics Office (example). 2025. Food security report 2025,
Table 3.1, page 24. https://stats.example/documents/EX-REP-2025-03.pdf#page=26
  1. Set the format as above.
  2. Put the citation in the table record (citation) and on the table's page, and add the DOI of the document where it has one.
  3. For a single value, the tidy file's page and bbox columns are the citation at value level. A system that quotes the value can show the snapshot, which is what the home page's example does.
  4. Use the PDF page number in URLs (#page=26) and the printed page number in the citation text. They differ, and both are in the records.

Worked example. The citation of Table 3.1 is the one shown above. Its printed page is 24, and its PDF page is 26, because the report has two unnumbered pages of front matter, so the URL fragment reads #page=26 while the text reads "page 24". When an assistant answers "What was the prevalence of undernourishment in the Eastern Region in 2024?" from the tidy file, it quotes 9.2 percent and gives that citation. The interface shows the snapshot of the table beside the answer, because the tidy file's page and bbox columns point to it.

What this does not do. A citation to a page does not freeze the document. If the document is corrected, the citation needs the version, which recipe 8.3 records.

Recipe

8.3 Log corrections and monitor use

Level
AI-ready
Skills
Dissemination officer
Indicative time
Ongoing

Result: every correction to extracted data is dated and versioned, superseded files stay available, and the organization can see where its extracted data are used.

  1. When a value is corrected after publication (an extraction error, a document erratum, a series revision), publish a new version of the tidy file and the record, keep the old one, and log the change: table, value, old and new, reason, date.
  2. State on the table's page which version is current and list the errata, as the series pages do in the dissemination cookbook.
  3. Monitor mentions of the document and its tables with the program's Monitoring of Data Use, and run the monthly check of how AI systems answer questions the tables answer (chapter 7 of the dissemination cookbook).
  4. When an AI system quotes an old or wrong figure from the document, check first whether the table record and the tidy file are findable and current. Most fixes are on the organization's side.

Worked example. In November a reader reports that the Western Region's population in Table 3.1 is printed as 1,430 thousand in the report but as 1,340 in the tidy file. The reviewer confirms the transposition against page 24, publishes version 1.1 of the file and the record, keeps version 1.0, and adds the entry "T3.1, Western Region, Population (thousands): 1340 corrected to 1430; transposed digits; 2026-11-12" to the corrections log on the table's page. The monthly AI visibility check, which asks two assistants for the Eastern Region's prevalence, shows one assistant citing the report and page and the other quoting the national figure without a source. The table record's keywords are extended, and the check is repeated the next month.

What this does not do. Monitoring finds mentions of the document and the tables by name. Values quoted without attribution are not traceable, which is the reason the citation format and the page-level provenance exist.

Implementation options​

  • Findability: catalog records with schema.org markup; the site search index; the question set of the dissemination cookbook.
  • Citation: a fixed format with page and table number; the document's DOI; a DataCite DOI for the extracted table where the organization registers them.
  • Versioning: versioned files and records with errata; superseded versions kept.
  • Monitoring: mention extraction over publications and the AI visibility check from the dissemination cookbook.

World Bank examples​

  • The program's home page shows a value cited to its page and snapshot, which is the value-level citation of this chapter.
  • Monitoring of Data Use extracts dataset mentions from documents and harmonizes name variants, which finds where extracted tables are reused.
  • Proof-Carrying Numbers verifies numbers in generated answers against the record, which an extracted table joins once catalogued.

Resources​

Common mistakes​

  • Citing the document without the page. A 300-page yearbook is not a source for one number.
  • Correcting silently. Two users with the "same" file get different values.
  • PDF page and printed page confused. They differ by the front matter; the records hold both.
  • Extracted data invisible to search. A file on a page without a record is found only through the document.

Verification​

  • Search. A query for a table's printed title returns the table record.
  • Citation. Every table record has a citation with the page and the table number, and the URL with the page fragment opens the right page.
  • Errata. Every correction after publication appears in the log with a version.
  • Use. Mentions of the document and its tables are counted each month.

Checklist​

0 of 6 done

Ticks are saved in this browser only.