How a citation path becomes evidence

A composite search for a Nairobi service business can look tidy on the surface: the model gives a name, a category, a location cue and a confident sentence about what the business does. Then the cited source turns out to be a thin overseas profile with a similar name, or a booking page that repeats half the story and leaves the Kenyan context behind. Kijito Citation Lab begins with that kind of small mismatch. The useful question is whether the answer mentions the business and whose evidence made the answer possible.

The lab calls something an observation only when the main pieces stay attached. A prompt is recorded with its wording, language variant, model surface, visible answer, cited source and the specific claim that citation seems to support. A conclusion is treated more cautiously. It is an interpretation drawn from repeated observations under comparable conditions, with notes on where the pattern is stable and where it becomes slippery. A citation does not get credit just because it appears beside a sentence. The cited page must actually back the claim being made, otherwise the observation is marked as weak, unsupported or only partially supported.

Samples are formed descriptively, not as a scoreboard. The team builds sets around business categories, counties, language variants and source types: formal SMEs, service providers, tourism operators, social-first sellers, supplier profiles, trade references and informal-sector cases where the public trail may be uneven. A county-level enterprise may have a registry trace and a supplier listing but little press coverage. A coastal operator may have strong international platform evidence and weaker Kenyan-owned material. A social-first merchant may be visible in posts and map fragments, yet hard for an answer engine to cite cleanly. Those differences are part of the study, not noise to be brushed away.

Repeatability means another run can be compared without pretending the machine is frozen. The same query structure can be run again with notes on date, language, model surface, visible citations and major answer changes. The lab pays attention to whether English and Swahili versions retrieve the same business, choose different source types, or narrow the answer in ways that change the meaning. A Swahili phrasing can shift intent, and the team does not flatten that shift into a simple language penalty. Sometimes the path changes because the wording changed. Sometimes it changes because the available Swahili-facing evidence is thinner.

The lab also records what kind of authority the answer seems to borrow. Its qualitative citation-share frame uses four plain buckets: local record, local story, platform proxy and unsupported echo. A company page or registry trace may act as a local record. A newspaper mention can work as a local story. An international booking profile, directory or marketplace page may become a platform proxy. A repeated claim with no clear support becomes an unsupported echo. These labels are not numeric rankings. They are a way to ask which source types are allowed to speak for Kenyan businesses.

The limits are stated inside the work. Answer engines change. Citations can be partial. Local records may be incomplete, inconsistent or hard to retrieve. Trade-body references can be useful without covering the whole business. Forecasts are marked as forecasts, not findings. Uncertain cases are left with honest labels: unresolved, mixed-source, weakly supported or language-sensitive rather than forced into a clean category. That restraint is part of the method, because a messy source path is often the truer result.

Working principles

1. Observation before conclusion

   The lab records the prompt, language, cited source and visible claim before making an interpretation. A single answer may suggest a pattern, but it does not settle one.

2. Citation is checked

   A source is treated as supportive only when the cited page actually backs the claim. Proximity to an answer is not the same as evidence.

3. Language paths matter

   English and Swahili queries are compared because they can lead to different retrieval paths. The lab notes when the language shift changes source choice, entity selection or claim strength.

4. Uncertainty stays visible

   Mixed, weak and unresolved cases are labelled directly. The lab would rather leave a rough edge than force a neat story where the evidence does not allow one.