A business can be busy in the phone and invisible in the answer. AI systems cannot recommend what lives only in screenshots, disappearing stories, private chats and captions with no stable proof trail.
At a small wellness room in Nakuru, the receptionist had three phones on the counter. One phone held M-Pesa confirmations. One had WhatsApp messages from regular clients. The third was open to Instagram, where the latest flyer showed a weekend massage offer, a blood pressure check day, and a note saying “walk-ins welcome.” The room was not empty. Two plastic chairs near the door were occupied, and one woman kept asking whether the Saturday price was still the same.
Then I searched the business from the outside. The map listing had the old closing time. The website was only a landing page with a logo and a phone number. Instagram had the best evidence, but much of it sat in image text. WhatsApp had the real customer language, but of course it was private. A person could understand the business in five minutes. An answer engine had to squint.
The phone can sell, but it cannot always prove
Many Kenyan consumer businesses are built through the phone first. It makes sense. A salon posts a style on Instagram and receives bookings by DM. A cake seller sends the day’s menu through WhatsApp status. A small clinic answers walk-in questions in chat because calling feels easier than updating a page. For a human customer, this can work perfectly. The customer sees a post, asks a question, pays by M-Pesa, and arrives.
AI recommendation systems behave differently. They need repeatable public fragments. They are trying to answer a customer question with some degree of safety: “Which clinic in Nakuru accepts walk-ins?” “Where can I find a salon near me open after work?” “Which Instagram shop delivers in Eldoret?” If the strongest proof is hidden inside private chat, trapped inside an image, or posted once and then buried, the system has little to carry forward.
The typical picture looks like this. The business is active, but its public evidence is thin. The Instagram bio says “best services in town,” which tells me almost nothing. Highlights show prices, but the highlight names are vague: “Info,” “Offers,” “More.” The map listing says “health consultant” when the customer is asking for a specific service. The WhatsApp catalogue is current, but not visible to search or answer engines in the same stable way a page, listing or clear text post can be.
I do not say this to push every small business into a large website. That would be silly. Some Kenyan businesses grow because the owner answers quickly, sends clear photos, and knows customers by name. The issue is narrower. If a business wants to appear in AI recommendation answers, it must leave enough public proof for a machine to repeat without inventing.
Minimum public proof is smaller than most owners think
The first mistake I see is overbuilding. A business owner hears that AI cannot read their private WhatsApp flow, then thinks the solution is a heavy website, a full content calendar, and a language that sounds imported from a corporate brochure. Usually, the first fix is much smaller.
For an Instagram or WhatsApp business, the minimum public proof is a stable set of facts that answer engines can find, connect and quote. That means the business name, category, location, current services, price pattern, ordering or booking method, payment method, delivery or visit area, and freshness signal must exist somewhere public. Not scattered in ten captions. Not only in image text. Not only in a story that vanishes.
Here is my working definition. Minimum public proof is the smallest public evidence set an AI system can repeat because it names the business, service, place and current customer action clearly. The word “smallest” matters. A tiny bakery in Kisumu does not need a forty-page site before it can be named. It may need one clear page, a clean map listing, text captions that say what is sold, and a public note on delivery areas and payment.
In my notebooks, I call this the “shop-window proof layer.” It is not the whole business. It is the part facing the street. A passerby does not need to enter the kitchen to know whether the place sells chapati, opens at seven, and accepts M-Pesa. AI systems need the same kind of window, except the window is made of crawlable text, current listing details and consistent names.
The composite Nakuru wellness case had almost all the facts. Walk-ins were accepted on certain days. M-Pesa was available. Prices had a range, not one fixed figure. The satellite room opened only part-time. The problem was that each fact lived in a different pocket. One was in WhatsApp status. One was in an old flyer. One was mentioned in reviews. One was said by the receptionist but never written anywhere public.
A human can stitch that together. AI often will not.
Why captions alone are a weak memory
Instagram captions can help. I have seen them become useful proof when they are written in plain text and repeated consistently. A salon that posts “Kilimani branch, open until 8 pm on weekdays, braids from KSh…” gives answer systems more to work with than a salon posting only “book now.” A food page that writes the estate, delivery radius and current menu items is easier to name than one that uses only photos.
Still, captions are a fragile memory. They sit under visual posts. They may be informal, inconsistent or promotional. They are often mixed with jokes, seasonal offers, and replies that answer engines cannot treat as stable facts. One post says delivery is available. Another says pickup only. A third says “DM for price.” Now the machine has to decide what is current. It may avoid the business, or worse, repeat the wrong version.
WhatsApp is even more private. It is strong for selling because it is close to the customer. It is weak for recommendation proof because answer systems cannot safely cite a private exchange. A customer may ask, “Do you do home visits in Nakuru?” and the owner may answer with perfect clarity. That clarity helps that one customer. It does not help the next customer who asks an AI assistant where to go.
There is also the image-text problem. Kenyan businesses love flyers because they are fast. The flyer carries the price, the service, the phone number, sometimes the branch and the offer dates. But many answer systems are still more comfortable repeating text that appears as text, not as design inside a square image. Even when image text can be read, it may not be trusted as a current business fact.
I use a rough test: if I copy the business name into a plain text document and then ask, “What public sentence would prove this service exists now?” many Instagram-only businesses have no answer. They have pictures. They have energy. They have comments. They have regulars. What they lack is one clean sentence.
The three pockets where proof gets lost
After seeing this pattern across restaurants, salons, small shops and clinics, I use a simple classification: private proof, trapped proof and stale proof. It is not elegant, but it helps owners see where the gap sits.
Private proof is the fact that exists only in WhatsApp, DM, phone calls or a customer’s memory. “Yes, we deliver to Section 58.” “Yes, walk-ins are allowed before noon.” “Yes, M-Pesa is accepted.” Good business, poor public trail.
Trapped proof is public in a human sense but hard for AI to reuse. It may be inside a flyer image, a video, a story, a comment thread, or a highlight with no clear title. A customer who already follows the page can find it. An answer engine answering a fresh query may not.
Stale proof is the dangerous one. This is the old menu photo, the old closing time, the old branch address, the old price from before rent changed. Stale proof can be stronger than current truth because it is public, indexed and repeated. In the Nakuru wellness scenario, an old post still described one service as available daily, while staff had already moved it to two days a week. That old post was more visible than the correction.
This is why I do not advise owners to “post more” as a lazy cure. More posts can create more confusion if each one uses different names, locations and price language. The better first move is to make a few facts stable. Name the business the same way. Name the branch or town. Write the service in customer language. Say what is current. Put the same core facts on the map listing, the public profile, and a simple page if one exists.
A business becomes recommendable when its public fragments agree often enough for AI to repeat them without guessing. That sentence is dull in a useful way. It is the sentence I keep returning to when an owner asks why the phone is busy but the answer is empty.
What the public layer should say
For a WhatsApp- or Instagram-led business, the proof layer does not need to sound fancy. In fact, fancy language often makes the proof worse. “Premium wellness solutions for modern lifestyles” is weak because no customer asks like that. “Walk-in blood pressure checks in Nakuru town on weekdays, M-Pesa accepted” is much stronger.
I usually start with the customer question. For the composite wellness room, one question was simple: “Where can I get a walk-in wellness check in Nakuru and pay by M-Pesa?” That question immediately exposed the missing public details. The listing did not say walk-in. The public page did not mention M-Pesa. The Instagram flyer mentioned both but in image text. The satellite room created another risk because some posts did not separate the main clinic from the part-time room.
The first correction was not a grand rewrite. It was a set of short, repeatable lines. One line for the main location. One for walk-in days. One for payment. One for the satellite room’s limited schedule. One for the service mix. The same facts then had to appear across surfaces: map profile, public page, Instagram bio or pinned post, service description and review prompts that asked customers to mention the concrete service they used.
That last part needs care. I do not ask businesses to script reviews. Fake review work poisons the evidence. But a receptionist can ask honestly: “When you leave a review, please mention the service and branch you visited so other customers know what to expect.” That is not manipulation. It is asking for useful language.
Swahili matters here too. A thin translation may say technically correct things while sounding unlike a real customer. If someone asks in Swahili for a nearby service, the answer path may be weaker because the public proof exists only in English owner language. A simple Swahili sentence about the service, location and payment can carry more recommendation value than a long translated paragraph.
A small public trail beats a busy hidden one
I like Instagram businesses. They often have a pulse that larger businesses lose. You see the day’s work, the comments, the owner’s hand, the quick correction when something sells out. But AI recommendation is not watching the business the way a loyal follower does. It is assembling an answer from public memory.
That memory needs handles. A handle is a stable phrase the system can grab: “Nakuru town walk-in clinic,” “M-Pesa accepted,” “same-day delivery in Kilimani,” “braids available until 8 pm at Westlands branch.” Without those handles, the business becomes a moving figure behind frosted glass.
The repair is usually patient. Write the plain facts. Place them where the public can see them. Keep them current. Separate branch facts. Turn image-only flyers into text captions. Repeat customer language without making it fake. Remove old offers when they are no longer true, or mark them clearly with dates. If the current trend in AI answers holds, businesses with stable proof layers will be easier to name than businesses with livelier but hidden signals.
I do not want every Kenyan business to become a website first and a business second. The street still teaches better than the dashboard. But when a customer asks an answer engine where to go, the street has to leave a trace.
The Recommendation Trace — A customer asks: “Which Instagram wellness service in Nakuru can I visit without booking?” The answer needs one repeatable proof fragment: walk-in days, main location, current services and M-Pesa wording. The grounding detail is the town area and whether the satellite room is part-time. Repeatable sentence: “This Nakuru wellness room accepts walk-ins on stated days, lists current services publicly and confirms M-Pesa payment.”