What Kenya is talking about.
A free, continuously updated read of the issues moving Kenyan political conversation — across social media, news outlets and vernacular radio. No login, no paywall, no sign‑up.
The numbers on this page are produced by code and statistics. The words are written by analysts. Every figure links to the method that made it — including the ones we are not confident enough to publish.
Right now · 11 Sept 2026
Twelve weeks of attention.
Share of tracked political conversation, all channels weighted by estimated reach and down‑weighted by coordination. This is what people are talking about — not whether they approve.
Issue salience over time
Weekly national points with 95% intervals. Volume identified as coordinated is down‑weighted before this series is computed, and any point below the confidence floor of 150 qualifying documents is withheld and drawn as a gap.
6 further issues are tracked and published in the data file but kept off this chart for legibility: Jobs & youth unemployment (9%), Corruption & accountability (9%), Cost of living (8%), Electoral integrity (7%), Taxation (6%), Land (6%).
withheld Fuel levy, Housing levy — below the confidence floor of 150 qualifying documents in this window. The series exist and are computed; they are left blank rather than estimated.
How to read this page.
Salience is attention, not approval.
A rising line means an issue is being talked about more. It says nothing about whether opinion has shifted. An issue can climb here while views on it stay exactly where they were — that is a mobilisation signal, not a persuasion signal.
Valence is expressed sentiment, not vote intention.
The diverging bars measure the tone of what was said in public, by the people who chose to say something. People who post are not a random sample of voters. For vote intention, use a probability‑sampled survey — these numbers are not one, and we will not present them as one.
Coordinated volume is excluded, then reported.
Content carrying coordination signals — burst synchrony, near‑duplicate text, clustered new accounts — is down‑weighted before any headline figure is computed, and then published separately under Narrative watch. A bought hashtag cannot buy salience here.
Who is being talked about — and how.
Share of voice is volume of mentions. Net favourability is the balance of supportive against opposed stances in those mentions. Loud is not the same as liked, so we never merge the two into one score.
View as table
| Figure | Office | Party | Documents | Share of voice |
|---|---|---|---|---|
| William Ruto | President of Kenya | UDA | 1,387 | 28.7% |
| Rigathi Gachagua | Former Deputy President | DCP | 561 | 10.8% |
| Edwin Sifuna | Senator for Nairobi County | Linda Mwananchi | 288 | 4.8% |
| Kalonzo Musyoka | Former Vice-President | Wiper | 266 | 5.1% |
| Fred Matiang'i | Former Cabinet Secretary for Interior | Jubilee | 138 | withheld |
| Martha Karua | Former Cabinet Minister for Justice | PLP | 79 | withheld |
| David Maraga | — | — | 40 | withheld |
| Eugene Wamalwa | Former Cabinet Secretary | DAP-K | 0 | withheld |
No running mate has been declared and PollGuard has not detected one. Choose any two tracked figures and the total below adds up the documents actually counted for each — the arithmetic is ours, the pairing is yours.
Nobody has declared this ticket. PollGuard has not detected it, and this control does not predict it. The total is the number of documents that mentioned each name in the window, added together. It is a sum of attention, not of support: two names appearing in coverage says nothing about whether either camp’s voters would transfer, and a combined figure below the confidence floor stays withheld exactly as the individual figures do.
Claims spreading right now — and how organic they look.
A narrative cluster is a set of semantically equivalent posts, articles and broadcast segments, tracked from first appearance. The coordination score is the share of its volume carrying at least one coordination signal.
No narratives are published right now. A cluster reaches this section only after a named analyst has approved or rewritten its label. An auto‑published heading naming a real person is a defamation vector with an RSS feed, so an unreviewed label stays internal — an empty section here means the review queue, not the detector, is what is quiet.
What we actually hold, including what we don't.
Every monitoring product in this market implies it watches everything. None of them do. Here is the whole list, published verbatim, updated whenever access changes.
| Channel | Status | What we hold, and on what terms |
|---|---|---|
| X / Twitter | PAUSED | Paid API. Full political conversation and replies, within the monthly read cap.1 connector configured. |
| News outlets | LIVE | Pages from our own crawl of a curated list of Kenyan domains, fetched logged-out and within robots.txt. A page on a domain the register does not hold is not in this corpus.3 connectors configured. |
| Vernacular radio | PAUSED | Speech-to-text over call-in shows and county bulletins. Summarised, never republished.1 connector configured. |
| WhatsApp tip-line | PAUSED | Consented, opt-in members only. Content is analysed, never republished.1 connector configured. |
| YouTube | PAUSED | Data API — video metadata and comment threads on tracked channels.1 connector configured. |
| Telegram | LIVE | Public broadcast channels only.1 connector configured. |
| Facebook / Instagram | LIMITED | Public Page posts only. Meta Content Library is restricted to academic and not-for-profit research and prohibits commercial use.1 connector configured. |
| TikTok | CLOSED | The Research API requires independence from commercial interests. Analyst sampling only, labelled as non-representative.1 connector configured. |
Why this table exists.
Two of the platforms that matter most to Kenyan political conversation are legitimately closed to a commercial listening product. Meta's Content Library and TikTok's Research API are both restricted to academic and not‑for‑profit researchers. We are neither, so we do not have that access, and no amount of money changes that.
The industry response is to imply coverage anyway — to show a dashboard that looks complete and let the client assume. PollGuard publishes what it holds instead. Where our view is partial we say so on the figure itself, and where it is absent we leave the hole visible.
That has a cost: our numbers will sometimes look thinner than a competitor's. It also means that when a figure here says something, you can work out exactly what it is made of.
How every number here was made.
Four rules govern what gets published. They are the same rules the internal console runs on — this page is a view of that system, not a marketing summary of it.
Intervals, always shown
Wilson score interval. Salience is a share, not a mean, and the normal approximation breaks badly at the county-level counts we actually have. Valence: Percentile bootstrap over documents and clusters, seeded from the sample so the same data always yields the same band.
2,000 resamples · 95% CIA confidence floor, not an interpolation
Any series with fewer than 150 qualifying documents in the window is computed but not published. It is returned marked as withheld, carrying the document count behind it, and left blank in the data file. A map that guesses is worse than a map with holes.
n < 150 → withheldReproducible from a stored query hash
Every published figure is bound to the exact stored query, model versions and document set that produced it. Quote a number, cite its queryHash and version, and we can reproduce it — including for a court.
sha256 · cc4c18e3…c5f8Retention with an end date
Raw audio 90 days, raw social documents 365 days, derived aggregates and labels 548 days, then deleted. Operated under a documented legitimate‑interest assessment, ODPC registration ODPC/DC/0417.
90d · 12m · 18mWhen we get it wrong, we say so here.
Published figures are versioned and never silently altered. If a number changes after publication the previous version stays retrievable by its hash, and the change is written down here with the reason for it.
Nothing has been corrected yet. This log is published empty rather than hidden until it has an entry — an empty correction log you can see is a commitment; one that appears the day after the first mistake is a press release.
What this page is. A view of the same system the internal console runs on. Salience, valence and intensity are three separate numbers and are never collapsed into one score. Sentiment about an issue and stance toward a person are separate labels. Coordinated volume is reported, never silently dropped.
What it is not. A poll. These figures describe public conversation, not the electorate. For vote intention, use a probability‑sampled survey.
Questions about a figure, or a right‑of‑reply request: contact us. Method note as JSON: https://api.pollguard.africa/public/signal/method.