The number English tools report, and the one they don't
Across 300 comments collected between late April and mid-May 2026, English accounted for 69% of everything written. Run a standard English-only sentiment tool on that corpus and you get a number: 65% negative. Uncomfortable, but manageable. A brand team sees it, logs a ticket, moves on.
What that tool doesn't surface is the Naija Pidgin layer underneath. Twenty-two comments — 7.3% of the sample — written in Nigerian Pidgin came in at 91% negative, with zero positive sentiment recorded. Not low positive. Zero. That is a 26-percentage-point gap below the English bucket's already-high negative rate, and it is the kind of signal that disappears when you aggregate into a single "overall sentiment" score.
This article is about what that gap costs.
What Pidgin actually says
The Pidgin comments directed at @AirtelNigeria are not vague dissatisfaction. They are specific, named grievances delivered in a register that carries emotional weight precisely because it is not the register of a customer service form.
"@AirtelNigeria E no go better for Una as una ignore me o since 6hr"
— @GanganTmw62650 (twitter)
"E no go better for Una" is not a complaint. It is a curse — a social sanction, the kind of thing said when a person has exhausted every polite option. An English-only classifier reads it as unstructured text and either skips it or miscategorises it as neutral. A human analyst who knows Pidgin reads it as the loudest thing in the room.
@teeban441 went further:
"@AirtelNigeria @AirtelNigeria , God will punish you , punish your worker, person wey get ahm self, make I Dey suffer with network connection despite all d stealing , nah mtn wey no too good for my area cause ahm o😭😭💔🥲 it's not really you guys fault, don't tell me to send my n"
— @teeban441 (twitter)
The comment trails off — Twitter's character limit cut it — but the structure is already complete. The writer invokes divine punishment, names the theft ("all d stealing"), then pivots to a competitor (MTN), then walks back the blame slightly. That emotional arc — rage, accusation, resignation, partial retraction — is a five-beat sequence that no keyword-based English sentiment model is built to parse.
@knewgods compressed it to three lines: "change your bio, please. / you guys have the worst network coverage on earth. / i no fit curse una since e no go work... like your network." The Pidgin punchline — "i no fit curse una since e no go work" — is a joke built on the logic that even a curse would fail on Airtel's network. It is also, structurally, a review.
Sheng and the moral register
Sheng — the Nairobi street mix of Swahili, English, and local dialects — made up 12.3% of comments, the second-largest non-English bucket. Its 57% negative rate sits below the Pidgin floor but above the Swahili bucket's 44% negative. The more important difference is qualitative: Sheng commenters don't just report problems. They assign moral blame.
@charleschalo8 wrote to @SafaricomPLC:
"@SafaricomPLC Safaricom you can't even be nyinyi ni wezi , you are imposing okoa jahazi whenever one is subscribing for offer moto upon reporting all am told is its a system error, it's indeed a created system error because you want to meet your daily target"
— @charleschalo8 (twitter)
"Nyinyi ni wezi" means "you are thieves." The commenter doesn't stop at the accusation — they construct a theory: the system error is deliberate, engineered to hit a revenue target. That is not a service complaint. That is a corruption allegation, embedded in a code-switched sentence that an English classifier would partially read and largely misfile.
@khantopcooler took a different angle — competitive threat rather than moral judgment: "Si mtupee Free bundles za ku stream. Hata hamwezi promote loyal Customers Kama sisi inabidi tuhame Safaricom tuende Airtel." Translation: "Just give us free streaming bundles. You can't even promote loyal customers like us — we'll have to leave Safaricom and go to Airtel." The threat is real and specific. It names the competitor. An English-only tool sees "Safaricom" and "Airtel" and possibly flags it as a brand mention. The competitive churn signal is gone.
Swahili: the quieter split
Swahili comments — 5.3% of the sample — split evenly: 44% neutral, 44% negative. That near-parity is itself a finding. Swahili commenters on @SafaricomPLC posts were as likely to be making a transactional request as lodging a complaint. @anjelo_kir74235 wrote: "Hi safaricom jana nilituma pesa kimakosa kwa till number mnaweza nisaidia kureverse hio pesa" — "Yesterday I sent money by mistake to a till number, can you help me reverse that money." That is a service request, not a complaint. It reads as neutral because it is neutral. But it also represents a customer who chose Swahili over English to communicate with a brand that operates primarily in English, which is its own data point about where the language barrier sits.
@Maliksteamy08 was less patient: "Hamjai niongeza fuliza limit I wonder what's up jameni" — "You haven't increased my Fuliza limit, I wonder what's going on." Fuliza is Safaricom's mobile overdraft product. The comment is short, the frustration is mild, and it would almost certainly be classified as neutral by an automated tool. Whether it signals a broader Fuliza ceiling complaint from a segment of users who communicate in Swahili is exactly the kind of question English-only tools cannot answer.
What the brands posted, and what came back
The context matters. @Vodacom's highest-engagement post in this window was a football prediction contest — "Got a scoreline in mind? Share it with #VodacomGiganathi and stand a chance to WIN a @KaizerChiefs jersey" — which pulled 258 likes, 437 comments, and 19,685 views. @AirtelNigeria's Workers' Day post — "You've answered the calls, sent the emails, met the targets" — drew 372 comments on 19 likes, a ratio that signals the comment section was not aligned with the brand's tone.
The brand posts are in clean, corporate English. The replies that came back — in Pidgin, Sheng, Swahili, and mixed registers — are in the language people use when they are not performing for a brand. That gap in register is where the intelligence lives, and where English-only tooling goes dark.
@Derahh14 made the stakes plain in English, which is worth quoting because it shows the same frustration pattern appearing across language buckets: "It's been 4 good hours since Airtel asked me to send a DM with my details so they could look into scamming me of N30,000, and there's still been no response. Airtel refund my money or activate my subscription!!!!!!!!!!!!" The English is unambiguous. The Pidgin equivalents are equally unambiguous — just in a language the tool doesn't read.
The measurement problem
The honest framing is this: Naija Pidgin at 7.3% of comments is a minority bucket. It does not dominate the conversation by volume. But 91% negative with zero positive is not a marginal signal — it is a floor, and it belongs to the users most likely to have exhausted every other channel before going to a public post.
The "other code-switched mix" bucket — 3% of comments, 56% negative — shows the same pattern at smaller scale. @californiashep mixed Yoruba ("Kilode gangan na") with English and Pidgin in a single complaint about data bundles depleting without use: "Kilode gangan na daily I wan Dey sub ni Abi why is my gb now lasting even without using it the next thing I am receiving is your have 500mb left Kilode gan na." A tool that reads English catches "why is my gb now lasting even without using it." A tool that reads nothing else misses the emotional escalation that the Yoruba framing carries.
The aggregate sentiment for this dataset — 75% negative across all 51 posts and mentions — is accurate. It is also incomplete. The distribution of that negativity across language buckets, and the intensity variation between them, is where the operational intelligence sits. Brands that measure only the aggregate are counting the right total and missing the right problem.