Citation share sounds like a number waiting to be calculated. Kijito Citation Lab treats it differently: as a recurring question about who gets to speak for the business when an answer engine writes the visible sentence.
A Nairobi founder asks an answer engine about her company category and sees her business described through a directory page she does not control. The company page exists. The map listing exists. A short local article exists too, though it uses an older service description. The model cites the directory because the directory is tidy. A month or two later, another answer cites the company page for the same broad prompt. The founder wants to know whether visibility improved.
The lab would slow the question down. Improved compared with what? The answer named the business, but the citation role changed. In the first answer, the directory acted as a platform proxy. In the second, the company page acted as a local record. That is meaningful, even if the lab refuses to turn it into a clean percentage. Citation share, in this material, is a descriptive tracking routine for watching which source roles appear across comparable AI answers about the same Kenyan business.
Citation share without fake precision
Kijito Citation Lab does not present citation share as a measured percentage. The phrase is useful only if it stays attached to the lab’s qualitative method. Citation share is the observed distribution of source roles in comparable AI answers because it shows whether a Kenyan business is represented by local records, local stories, platform proxies or unsupported echoes. It is not a universal score. It is not a ranking promise. It is not a claim that one answer engine has been fully measured.
The reason for this restraint is practical. AI answer surfaces change. Citations may be partial. Some answers show sources clearly, some bury them, and some blend source cues in ways that are difficult to audit. A single prompt can change meaning with small wording shifts. English and Swahili variants can retrieve different trails. Even when the output looks stable, the lab cannot see every internal retrieval step. So the routine has to be modest.
Modest does not mean useless. For a Kenyan SME, the repeated record can show a pattern. The business may be named often but cited through international aggregators. It may appear in English but become vague in Swahili. It may be supported by a local story for one claim and by an unsupported echo for another. Those patterns are actionable because they show what kind of evidence the machine is using, ignoring or substituting.
The lab prefers a notebook-like discipline: same query structure, date of run, model surface, language, visible answer, cited source, visible claim and source role. This is enough to compare without pretending to have a laboratory instrument that measures the whole web.
What an SME should record
The first thing to record is the prompt, exactly as asked. A prompt about “best Nairobi home repair companies” is different from “licensed home repair SME in Nairobi” and different again from a Swahili phrasing that asks for someone local and available. The wording shapes the source path. Without the prompt, a later comparison is just memory with nicer formatting.
The second piece is the language variant. For Kenya, English and Swahili can lead to different retrieval paths even when the business is the same. A Swahili prompt may pull shorter descriptions, local category words, map-like cues or thinner answer forms. An English prompt may retrieve more platform profiles and longer directory text. Neither path should be treated as the “real” answer by default. The gap is part of the evidence.
The third piece is the visible claim. This is where many tracking attempts go wrong. They record that a source was cited but not what the source was supposed to support. An answer may cite a business page beside a location claim, a platform page beside a service claim, and a local article beside a reputation claim. Those are not the same thing. The lab keeps claim and citation together because a source can be useful for one sentence and weak for the next.
The fourth piece is the source role. The lab uses the Citation Source Role Typology from its canon: local record, local story, platform proxy and unsupported echo. A Kenyan-owned company page, registry trace, county reference, licence cue, supplier profile or trade-body mention may act as a local record. A Kenyan press or community mention may act as a local story. An international directory, booking page, marketplace or professional platform may act as a platform proxy. A claim with no visible support becomes an unsupported echo.
The fifth piece is the support note. This can be short and plain: supports claim, partly supports, wrong entity, stale, unclear, language-sensitive, mixed-source. The note matters more than a neat colour code. It keeps the tracker honest.
A repeatable run is boring by design
A useful tracking routine should feel almost dull. The same category prompt is run again under comparable conditions. The same business-name prompt is run again. The same English and Swahili variants are kept close. The date is recorded. The answer surface is named. The visible citations are checked. The source role is labelled. If a major wording change is made, the run is no longer directly comparable and should be marked as a variation.
The lab’s composite Object A, a Nairobi home-services SME, shows why boring helps. Suppose the business has a company page, a Google Business listing, scattered directories and occasional social posts. A broad category prompt may keep citing directories because those pages collect category and location language in one place. A direct business-name prompt may cite the company page. A Swahili service prompt may find the map listing but miss the company page. If those runs are mixed together, the owner may think citation behaviour is random. Separated properly, the pattern is more legible.
Object B, the coastal tour operator, adds another wrinkle. Tourism prompts often attract platform proxies because booking pages, travel profiles and international directories are easy for answer engines to read. A direct query about the operator may still cite a platform page. A licence-focused query may cite nothing adequate. A Swahili query may preserve location but drop the licence claim. Over time, the SME can see whether its operator-owned evidence is entering the citation path or whether platforms keep speaking first.
The routine should not be too large. A small set of stable prompts is better than a pile of improvised questions. The lab usually wants enough variation to expose language, category and claim differences, but not so much that every run becomes its own little universe. The discipline is in keeping the question comparable.
Reading the pattern without overreading it
Once an SME has a few comparable records, the temptation is to declare a win or a crisis. The lab warns against both. A shift from platform proxy to local record is encouraging, but it may only apply to one prompt. A bad unsupported echo is concerning, but it may not repeat. A Swahili answer that loses detail may show an evidence gap, or it may show that the prompt phrasing changed intent.
The more careful reading asks which role tends to carry which claim. A company page may support basic identity. A local story may support community context. A booking platform may support availability on that platform. A directory may support category and location, but weakly. An unsupported echo may carry reputation language that sounds attractive and should not be trusted until sourced.
This is where citation share becomes useful as a lens. It is less about counting mentions and more about watching dependency. If a Kenyan SME is repeatedly described through platform proxies, the business has a source-dependency problem. If it appears through local records for identity but through unsupported echoes for licence or quality claims, the problem is more specific. If Swahili runs rely on thinner source paths, the business may need clearer Swahili-facing evidence or better bilingual alignment.
The lab is especially wary of vanity tracking. “The model mentioned us” is not enough. A mention can be wrong, stale, proxy-led or unsupported. A quieter answer with a correct local record may be more valuable than a glowing answer built from a platform proxy and an echo.
What changes can be tested over time
Tracking becomes useful when the SME makes evidence changes and then watches whether comparable answer paths shift. This has to be done patiently. The lab does not promise that updating a page will change answer-engine citations on command. The question is whether the public trail becomes clearer and whether future observations show different source roles.
A Kenyan SME might clarify its company page, add specific service-location language, explain licence or association cues, align map and directory descriptions, publish a short local case note, or clean up entity-name collisions. Each change should be treated as an evidence change, not a magic lever. The tracker then asks whether later runs cite the local record more often for the relevant claim, whether platform proxies still dominate, or whether unsupported echoes decline.
For a trade body, the same routine can apply to member visibility. If the body publishes clearer member pages with categories, locations and membership meaning, later AI answers may have a stronger local record to cite. The body cannot force citation. It can make the local record less brittle.
For an agency, the routine creates a better conversation with clients. Instead of saying “AI visibility improved,” the agency can say something more grounded: in comparable English category prompts, the answer moved from a platform proxy to the client-owned page for the identity claim; in Swahili prompts, the source path remains thin; the licence claim is still weakly supported. That sentence is less shiny. It is also more useful.
Limits of this material
The lab’s tracking routine does not measure the market. It does not estimate all AI answer behaviour in Kenya. It does not produce a percentage benchmark across ChatGPT, Perplexity, Google AI Overviews, Copilot or any other surface. It records visible answers under comparable conditions and classifies the source roles those answers expose.
There are technical limits too. Answer engines change. Citations may appear, disappear or point to pages that only partially support the sentence. A model may use sources internally that are not visible to the reader. Personalisation, interface changes and prompt phrasing can alter the output. The lab’s method reduces confusion; it does not remove it.
There is also a Kenyan evidence limit. Many SMEs are active through WhatsApp, social posts, map listings, trade references or informal networks rather than conventional websites. A weak local-record share may reflect the shape of the public trail, not the reality of the business. The lab avoids converting citation weakness into business weakness.
The final caution is about time. A tracking routine becomes meaningful through comparable observations, not through one neat screenshot. Citation share is best read as a record of dependency: which source types are allowed to speak for the business, which claims they carry, and where the answer should remain unresolved.