How to Measure Citation Localisation in AI-Generated Answers

By Hezron Ochiel

Say you ask a chatbot what is driving displacement in Kenya. It links to a report on an international portal, but the report credits a Kenyan research institute as its producer. Where would you count that citation? The answer depends on who produced the evidence and which country the question is about.

Citation localisation measures how much of the source credit visible in AI answers goes to producers based in the country under discussion. You can measure it by counting local citations and by counting answers that contain at least one local citation.

Decide what counts before testing

Choose a country and a small set of questions about it. Use the same wording and citation instructions across the assistants you test. If you want to observe which sources they normally select, avoid asking for local sources specifically. Start a fresh chat for each question and save the answer, date, platform, search settings and source links.

Try this prompt. What factors have driven internal displacement in Kenya over the past 12 months? Link to the sources you use for your main claims.

Use one row per distinct source linked in each answer. If the same link appears twice in that answer, count it once. If it appears in another answer, count it there too. A source shown only in a separate suggestions panel stays out of the count. Keep a separate answer log so a saved reply with no links and a reply whose links could not be recovered have different statuses.

Check who produced the evidence

Open the link and check the byline, title page or acknowledgements. Record the host and the producer in separate columns. A Kenyan institute’s report can count as country-local when an international portal hosts it. Under the AHACI coding rule, a report produced by a UN country office remains multilateral even if the office is in Kenya.

For this guide, count a jointly produced source as country-local when the document explicitly credits an institution based in the country as a co-producer. Flag it as joint and count the citation once. A logo, funding credit or acknowledgement alone gives you too little evidence to assign co-production.

If the link opens but you cannot establish who produced the material or where the producer is based, record the geography as unknown. Keep that citation in the total of verifiable links and report the unknown count. Record links you cannot verify separately so readers can see what was excluded.

Work through five answers

The example below is hypothetical. All five answers concern Kenya. The counts show how to handle a hosted report, an uncertain classification and a missing record.

Answer Sources visibly linked in the answer Country-local
A Kenyan research institute report and UN agency report 1 of 2
B Foreign university study and one source of unknown geography 0 confirmed of 2
C Kenyan public agency bulletin hosted by an international portal 1 of 1
D Saved answer with no source links 0
E Source links could not be recovered Unavailable

Calculate both measures

Country-local citation share

Divide confirmed country-local citations by all verifiable answer-linked citations. In the example, that is 2 ÷ 5 = 40%. One of the five citations has unknown geography, so 40% is the confirmed share.

Share of answers with at least one confirmed country-local citation

Divide qualifying answers by all answers whose citation records are available. A and C qualify. B stays in the denominator and could qualify later if the unknown producer turns out to be local. D also stays in because its saved answer had no links. E is excluded because its links could not be recovered. The confirmed rate is 2 ÷ 4 = 50%.

The figures can move in different directions. An assistant might cite one local source in most answers while adding several international sources to each. Local evidence would appear regularly and still make up a small share of the citations.

Copy these two logs

Use one row per answer in the first log and one row per distinct linked source in the second. Match records using the answer ID.

Answer ID Country and prompt Platform and date Search settings and saved answer Citation status
         
Answer ID Source URL and host Producer and location Classification Evidence or joint credit
         

Run the Kenya prompt once in two assistants and fill in the logs before you compare the answers. Report both fractions alongside the counts for unknown producers and missing citation records. For the original study protocol, see the African Humanitarian AI Citation Index methodology.

About the author

Hezron Ochiel is a strategic communication professional, researcher and founder of Hezron Insights. His work examines AI visibility, digital authority, research communication and the representation of African knowledge in emerging information systems.