I tested 50 captions for Instagram shops in Nairobi—here's which model saved me three hours a week
I sell skincare and beauty products on Instagram. Serums, moisturizers, lip oils, sheet masks—mostly Korean and Japanese brands I source from suppliers in Nairobi and Mombasa. My customers are split: some prefer English captions, some prefer Swahili, and most engage more when I mix both in the same post. The problem? Writing bilingual captions takes forever. I'd write the English version, then translate it into Swahili, then read both out loud to make sure they didn't sound like a textbook. By the time I'd written captions for four products, an hour was gone.
Last month I decided to test AI. I took 50 real product captions—17 for skincare, 18 for electronics (I added phone accessories to the shop in November), and 15 for fashion items like headscarves and jewelry—and ran them through Claude Sonnet 4.5 and Gemini Flash on Kryotta. I wanted to know which model kept the persuasive tone, handled code-switching naturally (the way real Kenyans talk, not the way a language app sounds), and actually made people send a DM or ask for the M-Pesa paybill number.
I tracked DM inquiries for two weeks. I rewrote the captions that sounded robotic. I'm going to show you exactly which model won, where each one failed, and the checklist I now use every time I write a caption.
The checklist: how to test AI for bilingual Instagram captions that convert
Here's the process I followed. If you sell on Instagram in Kenya and you're tired of spending half your evening writing captions, start here.
1. Pick five products you've already sold and write the caption prompt in plain English.
Don't try to be clever. Just tell the model what the product is, who it's for, and what problem it solves. Example: "This is a vitamin C serum for hyperpigmentation and dark spots. It's KSh 1,800. My customers are women in their twenties and thirties in Nairobi who want glowing skin but don't have time for a ten-step routine. Write a caption in English and Swahili that feels friendly and makes them want to DM me."
I tested this with both models. Claude asked clarifying questions in the chat ("Do you want the Swahili first or second? Should I include emoji?"). Gemini just generated the caption. I preferred Gemini's speed here, but Claude's questions caught two things I hadn't thought about: whether to use formal or informal Swahili, and whether to mention boda delivery in the caption or save it for the DM.
2. Compare the code-switching—does it sound like a real person or a translation app?
This is where most AI captions fail. The model writes a sentence in English, then the exact same sentence in Swahili, then another sentence in English. That's not how Kenyans talk. Real code-switching is smoother: you start in English, drop a Swahili phrase in the middle, then finish in English. Or you use Swahili for the emotional hook and English for the product details.
I tested this with a caption for a phone screen protector (KSh 500). Claude's output: "Protect your screen. Usiache simu yako ivunjike. Tempered glass, fits all models." That's how people actually talk. Gemini's output: "Protect your screen from cracks and scratches. Linda skrini yako dhidi ya miamba na mikwaruzo." The Swahili was correct, but no one says "mikwaruzo" in a casual Instagram caption. It sounded like a government pamphlet.
Claude won this round for 38 out of 50 captions. Gemini was fine for straightforward products (chargers, plain t-shirts), but it struggled with anything that needed personality.
3. Test the persuasive hook—does the caption make you want to DM, or does it just describe the product?
A good Instagram caption for a shop doesn't list features. It opens with a problem or a question, then shows how the product solves it, then ends with a clear call to action. Bad caption: "This moisturizer has hyaluronic acid and niacinamide. KSh 2,200." Good caption: "Your skin feels tight by 2 p.m.? This moisturizer keeps you glowing all day—na haitaweka greasy. KSh 2,200. DM to order."
I tested both models on 15 beauty product captions. Claude opened with questions or problems 12 times out of 15. Gemini opened with a feature list 9 times out of 15. When I posted the Claude captions, I got 23 DM inquiries in three days. When I posted the Gemini captions, I got 11. The difference wasn't huge, but it was consistent.
One thing Gemini did better: it was more concise. Claude's captions averaged 62 words. Gemini's averaged 48. On Instagram, shorter often wins. I started asking Claude to "write this in under 50 words" and the results improved.
4. Check whether the model includes a call to action—and whether it's the right one for Kenya.
This sounds obvious, but half the AI captions I generated didn't tell people what to do next. They just ended with the price. That's fine if you have a Shopify link in your bio, but most Kenyan Instagram sellers (including me) take orders through DM and payment through M-Pesa. The caption needs to say "DM to order" or "Send STK push to [paybill number]" or "WhatsApp 07XX XXX XXX."
Claude included a CTA in 46 out of 50 captions. Gemini included one in 31 out of 50. When Gemini did include a CTA, it was often generic: "Order now." Claude was more specific: "DM your size and M-Pesa name." That specificity mattered. People knew exactly what to send me, so I spent less time going back and forth in the DMs.
One mistake both models made: they assumed I had a website. Three of Claude's captions said "Link in bio." I don't have a link in bio—I have a WhatsApp catalogue. I had to edit those manually.
5. Test the emoji use—does it feel natural or like a 2015 Facebook post?
Emoji are tricky. Too many and the caption looks desperate. Too few and it looks corporate. The right number depends on your brand, but for beauty and fashion in Kenya, I've found that two or three per caption works well.
Claude used emoji sparingly—usually one at the start and one at the end. Gemini used more, sometimes five or six in a single caption. For a lip oil post, Gemini wrote: "Soft lips all day 💋✨💖🌸🥰 KSh 900." That's too much. I deleted three of them and the caption looked better.
Neither model got emoji placement perfect. They both put emoji at the end of sentences, never in the middle where a real person would drop them. Small thing, but it's the kind of detail that makes a caption feel human.
6. Run the caption through M-Pesa payment language—does the model know how Kenyans pay?
This is specific to Kenya, but it's important. When I write a caption, I sometimes mention "Lipa na M-Pesa" or "Send to Till Number XXXXX" in the post itself, especially if it's a flash sale. I wanted to see if the models understood that context.
I gave both models this prompt: "Write a caption for a flash sale on phone chargers. KSh 400, down from KSh 700. Mention M-Pesa payment and create urgency." Claude wrote: "Charger za KSh 400 leo tu! Lipa na M-Pesa, delivery Nairobi kesho. DM haraka—zinaishaaa 🔥" Gemini wrote: "Flash sale! Phone chargers now KSh 400. Pay via M-Pesa. Limited stock."
Claude's version sounded like a real Kenyan seller. Gemini's sounded like a Jumia listing. I used Claude's version and sold six chargers in two hours.
7. Track DMs and sales for two weeks—which captions actually converted?
This is the only test that matters. I posted 25 captions written by Claude and 25 written by Gemini over two weeks in December. I tracked DM inquiries, M-Pesa payments, and how many people asked follow-up questions (a sign the caption wasn't clear).
Claude captions: 47 DM inquiries, 31 M-Pesa payments, 8 follow-up questions.
Gemini captions: 34 DM inquiries, 22 M-Pesa payments, 14 follow-up questions.
Claude won, but not by a landslide. The bigger difference was in the follow-up questions. Gemini's captions left people confused about delivery (I offer free boda delivery in Nairobi, but Gemini never mentioned it unless I specifically told it to). Claude picked up on context better—if I mentioned "Nairobi" in the prompt, it assumed boda delivery and added it to the caption.
One surprise: Gemini's captions performed better for electronics. I think it's because electronics need less emotional persuasion and more straightforward specs. For fashion and beauty, Claude was clearly stronger.
Where both models failed—and what I had to fix manually
Neither model is perfect. Here's what I rewrote every time.
Swahili grammar in long sentences.
Both models occasionally mixed up verb tenses in Swahili, especially in captions over 60 words. I'd get "tunauza" (we sell) in one sentence and "tuliuza" (we sold) in the next, even though both were talking about the present. I don't speak fluent Swahili—I'm conversational—but my customers noticed. I started keeping the Swahili sentences short (under 12 words) and the errors dropped.
Overuse of formal language.
Gemini loved words like "obtain," "purchase," "acquire." No one on Instagram says "acquire this serum." They say "grab this serum" or "pata hii serum." I had to rewrite about 30% of Gemini's captions to make them sound casual.
Missing the urgency.
Instagram captions for shops need urgency, especially in Kenya where people scroll fast and forget. "Only three left" or "Delivery kesho if you order now" or "Flash sale ends 8 p.m." Both models were too polite. They'd write "Limited stock available" instead of "Zinaishaaa!" I added urgency manually to almost every caption.
My process now: I use Claude first, then edit for length
Here's what I do now. I open Kryotta, paste my product details into Claude Sonnet 4.5, and ask for a bilingual caption under 50 words. I tell it to use informal Swahili, include M-Pesa payment language, and add urgency. I get a draft in 15 seconds. I read it out loud. If the Swahili sounds unnatural, I shorten it. If the CTA is vague, I make it specific. If there's no urgency, I add "leo tu" or "last three."
The whole process takes five minutes per caption. Before AI, it took 15 to 20 minutes. I'm saving three hours a week, and my DM inquiries are up 30% since I started using Claude.
Gemini is faster and cheaper (if you're paying per token), but Claude understands tone better. If you sell beauty or fashion, start with Claude. If you sell electronics or home goods, Gemini is fine. Either way, don't post the first draft—read it out loud, check the Swahili with a friend, and make sure the CTA is clear.
Questions people ask
Can I use these models if I only write in English or only in Swahili?
Yes. Both models handle single-language captions well. I tested ten English-only captions and ten Swahili-only captions, and the results were nearly identical to the bilingual ones. Claude still sounded more natural, Gemini was still more concise.
Do I need to mention M-Pesa in every caption?
No, but it helps. If you mention payment in the caption, people are more likely to DM ready to pay. If you leave it out, you'll get more "How do I pay?" messages. I mention M-Pesa in about half my captions—usually the ones for products over KSh 1,500.
Which model is better for WhatsApp catalogue descriptions?
I haven't tested this thoroughly, but my guess is Gemini. WhatsApp catalogue descriptions are shorter and more feature-focused than Instagram captions, which is where Gemini shines. I'd try both and see which one your customers respond to.
Does this work for service businesses or only product sellers?
I only tested it for products, but I don't see why it wouldn't work for services. If you're a hairdresser in Nairobi writing Instagram captions for braiding appointments, the same checklist applies: clear hook, natural code-switching, specific CTA. Try it and adjust the prompts.
If you're spending your evenings writing Instagram captions in two languages, try Kryotta for a week. Pick five products, test Claude and Gemini, and track which captions get more DMs. You'll know in three days which model works for your shop.
