A Kenyan business answer can look like one sentence, but underneath it may contain a registry shadow, a map listing, a directory label and a newspaper phrase stitched into a single claim.
The same business can appear in public like four different objects. On a company page it is a service provider. In a county context it is a supplier. On a map it is a pin with opening hours. In a local article it is a named example of a sector. Kijito Citation Lab sees this split often enough that the first task in a source review is not judging the answer. The first task is separating the pieces.
In a composite pass using Object A, the Nairobi home-services SME leaves a modest but believable trail: a simple site, a Google Business-style trace, a few directory echoes and occasional social posts. An answer engine describes it with a category and location. The citation points to one source, but the wording feels assembled from several. The page beside the sentence supports the name, maybe the category, but the service-area claim seems to have wandered in from somewhere else.
The answer uses sources by claim, not by business
A human asks, “What sources did the AI use for this business?” The lab often has to reframe the question. A model answer does not always use one source for the whole business. It may use different kinds of evidence for different parts of the sentence: identity, category, locality, reputation cue, availability, licence status or sector context. A citation can sit beside the whole answer while only supporting one slice.
A Kenyan business source is usable AI evidence when it can support a specific visible claim about entity identity, category, location, authority or current activity without requiring a large hidden inference.
That working definition gives the lab a practical test. The Business Registration Service may support formal existence or a registered name. A county directory may support a supplier relationship or locality. Local press may support context, history or a public activity. A Google Business trace may support place identity, address or category. A social profile may support activity, contact pattern or current product language. An international platform may support visitor-facing or marketplace presence, though often through a proxy frame.
The problem starts when one source is asked to carry a claim better suited to another. A map listing may support location, but it does not necessarily prove ownership, licence status or service quality. A press mention may show that a business was active in a sector at one point, but it may not support a current operating claim. A registration trace may show formal existence, while saying little about what the business does in practice. The answer can sound tidy because the sentence hides the labour of stitching.
Kijito Citation Lab therefore records source type and claim type together. It is not enough to note that an answer cited local press or a directory. The lab asks what the citation was asked to prove. That small discipline prevents a common mistake: treating any relevant page as supportive.
The four source roles in ordinary Kenyan answers
The lab’s Citation Source Role Typology is useful here because Kenyan business evidence often arrives as a mixed pile rather than a clean stack. A local record is Kenyan-owned or Kenya-based evidence that can speak directly for identity, locality, registration, licence, category or trade context. A local story is a Kenyan press, community or sector mention that adds context around the business or its environment. A platform proxy is an international platform, marketplace, directory or booking profile that represents the business through its own template. An unsupported echo is a repeated claim with no visible source strong enough to carry it.
These roles are qualitative labels. They do not rank the source by prestige. A local record can be weak if it is outdated or vague. A platform proxy can be helpful if the business itself keeps that profile current. A local story can be rich and still fail to support a precise current claim. The label only says how authority is being assigned inside the answer.
In the lab’s reviews, Google Business-style traces often behave like local records for place identity, though they are still platform-mediated. That makes them awkward. A map listing can be Kenya-specific and closely tied to the actual shop or operator, but the infrastructure is not a Kenyan-owned source. The lab usually treats it according to the claim: for address and category, it can function like a local record; for broader business authority, it becomes thinner.
County directories and supplier lists are another uneven class. They may be strongly local, but their formats vary. Some are easy to read. Others are trapped in PDFs, tables, older pages or lists that give a name with little context. In a visible AI answer, such sources may be ignored even when they would matter to a human evaluator. The machine may prefer a cleaner directory because the county trace is too hard to quote.
Local press works differently. It can give a business a story: participation in an event, a founder quote, a sector role, a dispute, a product launch, a community mention. But story is not the same as durable entity evidence. The lab marks it as local story when it adds context rather than proving the core identity. Sometimes that is exactly what the answer needs. Sometimes it is decoration wearing the clothes of proof.
Why official traces may sit in the background
This material does not try to answer the registry question fully; that is the focus of a separate work item. Still, official and semi-official traces appear in the broader source mix, and the lab watches how often they become visible. The pattern is uneven. A formal record can be highly relevant to entity identity and still absent from the citation path.
One reason is texture. Many answer engines seem more comfortable with narrative and directory pages than with structured records that require interpretation. A record may confirm a registered name but not explain the trading name. It may show existence but not the service category. It may require the answer to connect formal and public-facing identities. That connection is exactly where mistakes enter.
For the composite Nairobi home-services SME, a registry-style trace would help settle formal identity. But the answer’s visible claim may be about practical service: repairs, installations, areas covered, customer-facing category. A formal record cannot carry all of that. If the system needs a single citation beside a service description, a directory or company page may look more useful, even if the registry trace is more authoritative for existence.
The lab therefore avoids the phrase “ignored” too casually. A source can be absent because it was not retrieved. It can be absent because it was retrieved but not cited. It can be absent because it did not support the particular claim the answer chose to make. Those are different failures. The visible answer rarely lets researchers separate them cleanly.
A better research posture is to ask where a source class is strong. Official traces support formal existence and identity. County references can support local institutional context. Local press supports story and public context. Maps support place and category. Social profiles support activity signals when they are public and clear. Platform proxies support presence inside a marketplace, travel, booking or professional template. The question is whether the answer uses each source for the claim it can actually bear.
The patchwork becomes risky at the seams
The most interesting errors tend to appear at the seams between source classes. A model may connect a formal name from one place with a trading category from another. It may attach a county cue from a supplier list to a business with a similar name. It may cite a local article about a sector and then imply that a specific business holds a role the article does not name. These are not wild hallucinations. They are seam errors.
A seam error is harder for a casual reader to catch because each piece may be plausible. The name exists. The category exists. The county exists. The cited page is related. The mistake is in the join. The lab has found that this kind of error is especially important in Kenya, where business names can collide across counties and where formal, social and platform identities may not align exactly.
Object B, the composite coastal tour operator, gives a different version. A booking page supports tour products. A local tourism mention supports regional context. A licence cue, if visible, supports formal permission or sector membership. A WhatsApp route supports how enquiries happen. An answer may compress all of those into one smooth recommendation. The compression is useful for the reader but risky for evidence. One citation cannot carry the whole bundle unless the cited page truly contains it.
This is where the lab’s support labels become practical. Strong support means the cited page backs the specific claim. Partial support means it backs part of the claim but not all. Weak support means the connection is topical or indirect. Unsupported echo means the claim appears without a source that can carry it. Mixed-source means the claim appears to require several sources, even if only one is shown. Language-sensitive means English and Swahili variants alter the source choice, entity match or claim strength.
Those labels are deliberately plain. They help a reader slow down and ask a better question: which part of the answer is supported, and by what kind of source?
How readers should inspect a source mix
The lab does not turn the research into a checklist of fixes, but it does suggest a useful inspection order. Start with the visible claim, not the citation. What exactly is the answer saying about the business? Is it naming the entity, describing a service, locating it, recommending it, or implying authority? Then read the cited page with that claim in mind. A page can be relevant and still fail the support test.
For Kenyan SMEs, agencies and trade bodies, this order matters because different sources serve different public functions. A business page may need to clarify category and location. A registry trace may need consistent naming. A county or trade reference may need enough context to be machine-readable. A social profile may show activity, but it should not be the only place where core identity lives. A map listing may help with place identity while leaving broader claims exposed.
The lab’s strongest source-path reviews are usually not dramatic. They read like careful accounting. This sentence is supported by the company page. This location cue is supported by a map trace. This authority claim is only supported by a platform page. This current-activity claim is unclear. That slow separation is the point. A Kenyan business answer becomes less mysterious once it is broken into claims and source roles.
There is a useful rough image for this. An AI answer about a business is often a matatu signboard painted from spare parts: one colour from a directory, one slogan from a platform, one route cue from a map, one old phrase from a local article. It can still get the passenger close. It can also send them to the wrong stage.
Limits of this source review
The lab cannot see every internal retrieval step behind an answer. It can only inspect the visible route between prompt, answer, cited page and claim. A source may influence the answer without appearing as a citation. A citation may be selected for readability rather than full evidentiary strength. Answer engines also change their citation behaviour, so an observation has to be dated and treated as repeatable only under comparable conditions.
The review also does not declare one Kenyan source class universally superior. The Business Registration Service, county directories, local press, maps, social profiles and platform pages each support different claims. A formal record can be authoritative for identity and weak for live services. A local article can be rich in context and thin on current operations. A map trace can be current and still incomplete. The lab keeps those distinctions because flattening them would make the answer easier to summarise and less true.
There is also the issue of incomplete public evidence. Some Kenyan businesses, especially smaller and social-first operators, may be real and active while leaving only fragmentary public traces. A weak AI answer may reflect a weak retrieval path, a thin public record, or a mismatch between language and source availability. The lab usually cannot isolate one cause from a single answer.
The cautious finding is that AI answers about Kenyan businesses use a patchwork of sources, with each source class carrying a different kind of claim. The risk is not merely that an answer uses the wrong page. The deeper risk is that a page suited to one claim is quietly asked to support another. That is where local records, local stories, platform proxies and unsupported echoes begin to blur.