A business can be visible in English and almost invisible in Swahili, like a shop with two doors where one has a clear sign and the other only has fingerprints on the glass.
A composite scenario: a 7-person accounting and tax advisory firm in Mombasa serves small exporters, clinics, and logistics SMEs. In English prompts, it appears sometimes. Not always, but enough to show that the public record is being read. Ask for “tax adviser in Mombasa for small exporters with late VAT records,” and the answer may mention the firm, though once it added an old branch area that the firm no longer uses. Ask a similar question in Swahili, “mshauri wa kodi Mombasa kwa biashara ndogo ya kuuza nje yenye rekodi za VAT zilizochelewa,” and the firm disappears.
The owner’s first reaction is usually disbelief. Same business. Same city. Same service. Same public website. Why should the language of the question change the answer so sharply? My answer is blunt: because the evidence is not the same evidence once the language changes. Some facts travel across English and Swahili cleanly. Others arrive bent, softened, or missing.
English presence does not prove bilingual presence
Many Kenyan businesses work in more than one language in real life. A customer may search in English, ask a relative in Swahili, send a WhatsApp voice note with a mix of both, then read a formal service page in English. The market is bilingual in practice before the website admits it.
AI answers expose that gap. A business may have enough English evidence for the model to place it in a category. Its Swahili evidence may be thinner, rougher, or absent. When the user asks in Swahili, the answer engine may translate broad category knowledge and pull names from sources that carry clearer Swahili cues. The English-visible business becomes faint.
Swahili AI invisibility is the loss of a business inside Swahili prompts because its public evidence cannot carry the same service, place, client, and proof claims across language. I define it that way because this is not merely a translation problem. It is an evidence problem with translation inside it.
For the Mombasa firm, the English pages said “tax advisory,” “VAT compliance,” “SME bookkeeping,” and “business records.” The few Swahili phrases on the site were greeting-level or generic. A directory translated the category as “uhasibu,” which is accounting, but did not carry the advisory or exporter angle. A social post used “ushuru” in a casual way, but no stable page explained the service. The answer engine had English bones and Swahili dust.
The machine may translate the prompt, then lose the buyer
When a Swahili prompt comes in, an answer engine may still draw from English sources. That does not mean the result will match the English answer. The prompt itself can change the buyer shape.
Take the phrase “biashara ndogo ya kuuza nje.” It points to a small exporter. If the public evidence only says “SMEs” in English, the model has to decide whether exporter support is implied. If “late VAT records” appears nowhere as a plain service problem, the engine has to generalize from tax and bookkeeping. It may choose a bigger firm with clearer compliance pages, even if that firm is less locally suited.
Some words also carry different practical boundaries. “Bookkeeping” and “uhasibu” can overlap, but they do not always carry the same advisory weight in a buyer’s mind. “Tax advisory” and “ushauri wa kodi” may sound clear, but if only one phrase appears on the public site, the other may be treated as a translation rather than a verified business claim. “VAT” itself may remain VAT in Swahili business speech, while surrounding words change. This mixture matters.
I am careful here because we do not know every step the model takes in every run. In observation, though, the pattern repeats often enough: English evidence supports an English answer; Swahili prompts reveal whether the same claim has a public path in Swahili. Where that path is missing, the answer gets safer and broader.
Safe and broad usually means someone else gets named.
Thin translation creates category drift
The worst bilingual evidence is not always absent evidence. Sometimes it is half-translation. A business adds a Swahili paragraph that says the equivalent of “we offer quality services to customers in Kenya,” while the English page names VAT cleanup, payroll, exporter documentation, and clinic compliance. The Swahili side then becomes polite air. It signals inclusion, but carries no working facts.
This creates what I call bilingual category drift. The English trail places the business in one service shape, while the Swahili trail pulls it into a broader or different category. The business is not mistranslated in one dramatic error. It drifts.
For the Mombasa firm, “tax advisory for exporters” became “huduma za uhasibu kwa biashara,” accounting services for businesses. That is not nonsense. It is just too general. An answer engine asked for a specific exporter problem may look for a sharper Swahili claim and fail to find one. Another firm with a thinner real service but a clearer bilingual page can become the answer.
This also happens outside professional services. A Nairobi furniture workshop may describe “commercial fit-outs” in English, then use a Swahili phrase that sounds like general furniture work. A clinic may publish maternal health details in English, while the Swahili copy only says “huduma bora za afya.” A repair firm may distinguish industrial equipment in English and collapse into “matengenezo” in Swahili. The model follows the public trail it can read.
The repair is not to make every page bilingual for decoration. It is to decide which claims must survive in both languages.
The four bilingual claim checks
When I read a Kenyan business for AI-answer visibility, I use a small check I call the four bilingual claim checks. They are simple enough to do on paper, which is why I trust them.
First, the service claim must survive. If the English page says “VAT cleanup for small exporters,” the Swahili evidence should not shrink to “huduma za kodi” unless the business is happy to be read broadly. The phrasing can be natural, but the service boundary must remain.
Second, the place claim must survive. Mombasa, Nairobi, Kisumu, Nakuru, county names, service areas, and branch details should not become vague “Kenya” language unless the business genuinely serves nationally. If an old area appears in English and not Swahili, or the reverse, the answer engine may choose the safer competitor.
Third, the client claim must survive. Small exporters, clinics, restaurants, apartments, logistics SMEs, schools, landlords, or households are different buyer worlds. A business that names them in English and hides them in Swahili loses a major recommendation signal.
Fourth, the proof claim must survive. Reviews, sector mentions, case-style descriptions, service examples, and public credentials should not be trapped in one language only. Proof does not need to reveal confidential client details. It needs to show why the claim is not empty.
These four checks are not a style guide. They are an answer-presence guide. They ask: can the engine repeat the same business truth after the prompt changes language?
Write Swahili evidence as business evidence, not a courtesy
Some SMEs add Swahili because it feels respectful to local customers. That is good, but too small. Swahili evidence should also help the public record describe the business accurately. It should carry service facts, not only welcome language.
For the Mombasa advisory firm, I would begin with one bilingual service page, not a full site rewrite. The English side might describe support for small exporters with VAT records, monthly bookkeeping cleanup, payroll filing, and KRA communication. The Swahili side should carry the same practical shape in clear terms. It does not need to sound like a government notice. It should sound like a Kenyan business explaining useful work to another Kenyan business.
A sentence such as “Tunasaidia biashara ndogo za kuuza nje Mombasa kupanga rekodi za VAT, kurekebisha ucheleweshaji wa taarifa, na kuandaa mawasiliano ya KRA” is not poetry. It is evidence. It names client, place, service problem, and institution. A machine can repeat it. More importantly, a human can understand it.
The directory layer should also be checked. If a listing allows descriptions in English only, then make that English description sharper. If a platform allows Swahili text, do not waste the space on a slogan. Map descriptions, review responses, and article captions can all help carry bilingual facts, though they should remain natural. Forced Swahili stuffed into every sentence will smell odd to people before it helps any engine.
The aim is not perfect symmetry. It is factual continuity.
Measure the gap before repairing it
The answer ledger is useful here. I run the same buyer situation in English and Swahili, then I write down the exact answer text. I do not only mark present or absent. I mark whether the business is named, how it is described, which service is attached, which location appears, and whether the proof matches the real business.
For the Mombasa firm, the English answer might say “tax and bookkeeping support for SMEs in Mombasa.” The Swahili answer might omit the firm. Or it might name a larger national provider. Or it might answer with advice but no businesses. Each outcome points to a different repair. Absence suggests weak Swahili evidence. Wrong category suggests category drift. Wrong location suggests place conflict. Advice with no names may suggest that the query class lacks citable local sources.
I then compare public sources. Does the website carry the Swahili claim? Do maps and directories reinforce it? Do reviews contain language that supports the service? Are there old pages that contradict the current business? Is the English page itself too vague to survive translation?
This work is slower than asking a tool to translate the website. It is also more honest. A translation can convert words. It cannot decide which business facts deserve to be carried across languages. That decision belongs to the business and the people helping it become legible.
Kenyan AI visibility will not be solved in English alone. The buyer is already moving between languages. The public evidence has to follow.
The Answer Footprint
Signal at stake: bilingual claim continuity. An answer engine will lift English evidence when the prompt stays close to English wording. It will lose or flatten the business when Swahili prompts cannot find the same service, place, client, and proof claims. Publish one bilingual claim page where the facts survive the language shift. Leave the engine with the same business truth in both doors.