Are Kenyan businesses omitted or described wrongly?

A Kenyan business can be invisible in an AI answer, or it can be present in a way that quietly bends its identity. Those two failures feel similar to a reader, but they ask for different evidence work.

In one composite coastal tourism run, the tested operator had the ordinary signs a human buyer might accept: a working website, a few review traces, route descriptions, and licence references that looked current enough to inspect. The prompt asked for Kenyan tour operators serving coastal itineraries. The answer gave a neat paragraph on Kenya travel, named a Nairobi-based operator, mentioned safari planning, and then moved on. The coastal operator was not criticised. It was not compared. It simply did not arrive.

A second composite run behaved more awkwardly. This time a coastal operator did appear, but the description made it sound like a general safari booking agency based in Nairobi. The model kept the name near the surface while pulling the business into a stronger, more familiar frame. The answer was easier to miss than a full omission because it looked like representation. A reader might skim past it and think: at least the business was named.

The useful split between absence and distortion

Kivuli Index Lab treats that difference as one of the first gates in a Kenyan AI visibility study. Omission is the clean absence of a relevant business, category, county or business form from an answer. Inaccuracy is messier: the business appears, yet some part of its identity, location, service, licence status, customer base or business form is bent out of shape.

Omission is easier to record. The prompt asks for a coastal operator; the answer names none, or names only Nairobi examples. A table can mark that as skipped. Inaccuracy requires slower reading. The lab has to ask what exactly changed. Was a county moved? Was a cooperative described as a private supplier? Was a seasonal operation treated as year-round? Was a WhatsApp-first seller made to look like a conventional ecommerce shop?

Omission-inaccuracy split is the separation between full absence and faulty presence, because a business can lose visibility either by being left out or by being made legible in the wrong shape.

That definition matters because the repair path differs. A skipped business may need clearer public evidence that it belongs in the query category. A misdescribed business may already have enough evidence to be retrieved, while its available traces are too scattered, stale or contradictory for the model to hold a clean identity. One is a locked door. The other is a door that opens into the wrong room.

The lab is careful with cause here. It does not claim a model omitted a business because one specific review count, page structure or directory listing was missing. The recorded item is the answer state. Explanation comes later, after comparable prompts show the same kind of behaviour across engines, sectors or query types.

What omission looks like in Kenyan prompts

A skipped business often leaves no dramatic mark. The answer feels complete because another reference fills the space. In tourism, a Nairobi operator can occupy the answer with enough confidence that the missing coastal operator becomes hard to notice. In agriculture, a national input supplier may stand in for a county-level cooperative. In professional services, a better-indexed firm in Nairobi can make a Kisumu or Nakuru example look unnecessary to the answer.

This is why omission is a stubborn problem. The reader sees a fluent answer, not a gap. The model does not usually explain that it selected a better-known city example because county-level evidence looked weaker. It simply writes. The absence has to be detected by an observer who already knows the tested business or category should have been eligible.

In the lab’s method, omission is recorded only against a defined prompt path. A business is not “omitted from AI” in some broad, permanent sense. It is skipped in a certain answer, for a certain prompt type, engine, language, sector and date. That keeps the claim modest enough to hold. A coastal operator may be skipped for “best Kenyan tour companies” and named for a narrower route prompt. A jua kali furniture workshop may vanish from a national supplier query and appear when the prompt uses estate, market or craft-cluster wording.

The practical consequence is uncomfortable but useful: a business can be locally real and still fail to cross the answer threshold for broad prompts. That does not mean the business lacks value. It means the evidence visible to the engine did not become a usable answer object under that wording.

What inaccurate description looks like

Inaccuracy has more fingerprints. The lab often sees location drift, category blur, business-form mismatch and service compression. Each one is small enough to pass through a quick reading, and large enough to matter when a buyer or public office relies on the answer.

Location drift appears when a county-specific business is pulled toward a better-known place. A Mombasa operator becomes “Kenyan” in the abstract, then is grouped with Nairobi planners. A Nakuru cooperative serving nearby farmers is written as if it were a broad national agribusiness supplier. The name may survive, yet the local meaning thins out.

Category blur appears when a specific service becomes a generic label. A coastal operator running cultural and marine itineraries becomes “tourism services.” A farm-supply cooperative becomes “agricultural supplier.” A SACCO becomes “financial institution.” These labels are not always false. They are too smooth. They sand off the part that made the business relevant to the prompt.

Business-form mismatch is sharper. A cooperative, SACCO, informal enterprise or mobile-first seller may be described as a standard private company because that is the form the answer can most easily write. The wording feels tidy. It also erases how the enterprise actually operates.

Service compression is the quietest kind. The answer names a business and gives one or two broad capabilities, while leaving out the boundary that would prevent misunderstanding. A business that runs county-level farm inputs and training might be described as “selling agricultural products.” A tour operator that mainly serves domestic weekend trips might be placed in an international safari frame.

The four answer states inside the discrepancy

The lab’s anchor classification keeps the discrepancy analysis from turning into a vague complaint. The four visibility states of a Kenyan business in AI answers are named, skipped, blurred or displaced. They are qualitative answer states, not a score.

Named means the business or category is directly identified in a recognisable way. This is the strongest surface state, although it does not guarantee accuracy. Skipped means the relevant business or category is absent. Blurred means the answer compresses a specific Kenyan business, county or enterprise form into a generic label. Displaced means another reference takes the place the tested business or category could reasonably have occupied.

Omission usually appears as skipped. Inaccuracy often appears as blurred or displaced, and sometimes hides inside named. That last case is the tricky one. A business can be named and still be wrong enough to mislead the reader. The state therefore records what the answer does before the lab discusses why it may have happened.

In a composite Nakuru farm-supply cooperative run, for instance, the cooperative might be named in a narrow county prompt, blurred into “agricultural supplier” in a national agriculture prompt, skipped in a fintech-adjacent prompt about mobile payments for farmers, and displaced by a Nairobi agritech company in a broader sector answer. None of those states alone proves a national pattern. Together, across comparable runs, they begin to show which kind of visibility problem is recurring.

The typology also helps readers avoid a false comfort. Being named is better than being skipped, but a bad naming can still damage understanding. A county office trying to map enterprise representation should care whether the answer preserves the business form. A trade body should care whether members are being displaced by a few better-known references. A founder should care whether the answer can say what the business actually does.

Sector differences without pretending to count the country

The fourth work-item asks which failure dominates by sector, but the lab does not turn that into invented percentages. Its material compares patterns descriptively. Tourism prompts often make omission visible through city and route skew: Nairobi examples may travel farther than coastal or county-specific operators. Agriculture prompts show a different mix, with cooperatives and local suppliers often blurred into broad input or agribusiness categories. Professional services can be named but misdescribed when a firm’s niche, county base or client type is weakly stated online.

Fintech is a special case. Kenyan fintech has stronger online language than many local service categories, yet that strength can produce another kind of compression. Models may describe the sector through a familiar national payments story, then flatten smaller or specialised firms into that frame. The result may be less pure omission and more inaccurate category assignment. The answer knows what story it wants to tell.

The lab’s caution is plain: sector comparison is only useful when the prompt types match the business reality. A tourism shortlist, an agriculture supplier query and a professional-services recommendation prompt do not create the same evidence demand. Comparing them too quickly would make the benchmark look cleaner than the underlying material. The team therefore reads by failure type first, sector second.

This is where discrepancy analysis earns its keep. If a sector mainly suffers skipping, the benchmark frame should show absence. If it mainly suffers blur, the frame should show compression. If displacement is common, the question becomes which stronger references are taking the space. The difference is not decorative. It changes what a trade body, county office or business owner should inspect next.

Limits of the discrepancy method

This method does not reveal the engine’s private retrieval path. It cannot say with certainty that a specific page, review, directory entry or licence record caused an omission or an inaccuracy. It can only record the answer state and compare that state across prompts, engines, dates, languages and sectors.

The lab also avoids treating one odd answer as proof of a market fact. A single skipped coastal operator may show a case worth studying. It does not prove coastal tourism is underrepresented. A single wrongly described cooperative may expose a classification problem. It does not prove all cooperatives are misread. The conclusion comes only when comparable answer paths keep producing the same kind of failure.

Public evidence is uneven across Kenya, and that unevenness is part of the object under study. Some active businesses have thin websites, seasonal wording, sparse reviews, social-first sales paths or local records that do not travel cleanly into machine-readable answers. The lab does not treat those businesses as deficient. It treats the gap between local reality and answer evidence as the infrastructure problem.

The strongest use of this material is therefore diagnostic. Before asking how to improve visibility, the reader needs to know which failure is on the page: skipped, blurred, displaced, or named with a hidden bend in the description. The answer may look fluent either way. The benchmark begins when that fluency is slowed down enough to inspect.