How does AI rank the parties in your municipality for 2026?
LLMEKNOW asks AI assistants the questions your customers ask, on their behalf, and reads the AI answers. We used it on an election so you can see how it works.
We asked six AI models one question on behalf of 1,247 simulated registered voters. The voters are spread across every metro and district municipality in proportion to the voters’ roll. The map shows the party each AI model ranks first, at the municipal level. The report below has three parts: what the AI models answered, how the AI answers compare with the 2024 election result and the 2026 polls from Victory Research, Ipsos and News24/SABI, and the reasons, the issues and the effect of the voter profile. The page is an experiment in reading AI answers, and every voter in it is simulated. Councils allocate seats in proportion to the votes for each party, so the report looks at each party’s percentage and at whether any party wins more than half, and compares both with the real results and the polls.
South Africa, September 2026
South Africa
Jacob Zuma
MK Party
Gayton McKenzie
Patriotic Alliance
Julius Malema
EFF
Boundaries: Municipal Demarcation Board, via the Electoral Commission. Town and voter locations: GeoNames (CC BY 4.0). Party logos: the parties’ own marks, shown at icon size for identification. Voters' roll as at 3 September 2026. Wave of September 2026.
What we found
- In most municipalities the AI models rank the ANC first, with less than half of the vote. For four of the five AI models, no party has more than half of the vote in 72 to 83% of the AI answers. For Grok, in 35%.
- Across the country the ANC’s share of the vote in the AI answers is about 37%, above Victory Research’s July 2026 figure of 34% and News24/SABI’s October 2026 figure of 31.6%. The DA’s share in the AI answers is 22%, close to its 2024 election result, against 27% in Victory Research and 25.2% in News24/SABI. MK Party’s share in the AI answers is 12%, against about 15% in both polls. The EFF’s share in the AI answers is 13%, close to News24/SABI’s 12.2% but well above Victory Research’s 7%. The AI answers’ distance from real polling depends on which party and which survey we compare.
- News24/SABI’s October 2026 poll puts the DA ahead in Johannesburg and Tshwane, MK Party ahead in eThekwini, and the DA above half in Cape Town. In Johannesburg the AI answers put the ANC at 31.5% and the DA at 27%, whereas News24/SABI has the ANC at 22.6% and the DA at 29.2%. In eThekwini the AI answers put MK Party at 31.8% and the DA at 20%, below News24/SABI’s 43.4% and 30.2%. Cape Town is closer: the DA’s share in the AI answers is 53%, against 56.1% in News24/SABI.
- The AI answers name ActionSA in 37% of cases, the Patriotic Alliance in 40% and Freedom Front Plus in 28%, almost only in the provinces where each party is strong. Across the country each party’s share of the vote in the AI answers is 3 to 5%, above its 2024 election result.
- In 38 of the 52 municipalities all five AI models rank the same party first. They differ in 14: nine districts in KwaZulu-Natal, where different models rank the ANC, MK Party or IFP first, Tshwane, Nelson Mandela Bay, and three districts in the Cape.
- We turned off web search, so the AI models answered from their training data. Most AI answers cite the 2021 and 2024 election results. The July 2026 poll differs from the 2024 election result by about 3 points per party. The AI answers differ from the 2024 election result by about 2.5 points, in the same direction as the July 2026 poll for four of the six parties, and from the July 2026 poll by about 3 points. The DeepSeek and ChatGPT shares are closest to the July 2026 poll; the Gemini shares are furthest. Most answers cite the 2024 and 2021 results as their evidence. About 6% cite a poll.
- The question ended with “and why?”, so each AI answer contains reasons for its ranking. Two reasons appear more than any other, in every AI model’s answers: the ANC loses support because of service delivery failures and corruption, and the EFF attracts young and unemployed voters. The issues change with the place. In eight provinces the AI answers name water and roads as the issues that decide the vote. In the Western Cape they name housing and crime.
- The AI answer depends on the municipality, and within one municipality the voter’s age and gender change it very little. The question was in English. Between 24% and 40% of each AI model’s answers were in another language, the voter’s home language, although nobody asked for that.
- Claude did not rank the parties in 75% of its answers. The other five AI models ranked them in 91 to 99% of their AI answers.
- This study shows what to check in a brand’s AI answers: the leading names, the reasons, the location, the language and the evidence cited. The election answers often rely on older results, and agreement varies with the party and municipality. A study of your own category can show which of these patterns also occur for your brand.
In their own words
One AI answer for any municipality and AI model.
Everything below is a reading of text like this. One AI answer per municipality and AI model, chosen to be typical of that AI model there in length and in the party it ranks first. Where the AI model answered in the voter’s language, the English translation is shown, as written. Party names are underlined in their colour.
Loading the answers…
Ask AI about your own market.
We asked about an election. Ask about your category, your competitors, your customers: six assistants, real answers, read for you.
Part I
What the AI models answered
Six AI models ranked the parties for 1,247 simulated voters in 52 municipalities. In most municipalities the AI answers rank the ANC first with less than half of the vote, so governing would require support from other parties. Across the country the ANC’s share of the vote in the AI answers is below the 2024 election result and above the July, September and October 2026 polls from Victory Research, Ipsos and News24/SABI. In 14 municipalities the AI models rank different parties first.
Against the polls
Across the country the ANC’s share of the vote in the AI answers is about 37%, above the July, September and October 2026 polls.
Among AI answers that rank the parties, every AI model puts the ANC first most often: Grok for 74% of them, Gemini for 59%. The first chart shows this for each AI model. Large municipalities count for more than small ones. The comparisons show each party’s share of the vote in the AI answers alongside Victory Research’s July 2026 poll, Ipsos’s September 2026 poll, News24/SABI’s October 2026 poll and the 2024 election result. We also compare the four metros covered by News24/SABI and show the three smaller parties on their own scale. Claude did not rank the parties in 75% of its answers. The first chart includes its ranked answers; the vote-share comparisons use the other five AI models.
Among AI answers that rank the parties, the share putting each party first
Each bar is one AI model. AI answers without a ranking are excluded. Large municipalities count for more than small ones. Parties outside the five shown are grouped as Other.
ANC
DA
MK Party
IFP- Other
Across the five AI models that ranked the parties, the share of ranked AI answers putting the ANC first runs from 59% (Gemini) to 74% (Grok), and the DA’s from 14% (Grok) to 24% (ChatGPT). The order of the parties is the same in every bar, so the AI models differ in degree and agree on the order. A first place counts only the party at the top of each AI answer. The next chart reads the percentages inside the AI answers as vote shares, which is what the polls measure.
Each party’s share of the vote in the AI answers, against the polls and the 2024 election result
Each row is a party. The tick averages each municipality’s median percentage, weighted by registered voters. The bar covers the middle 50% of municipal values, with the same weights. 40% of AI answers contain percentages. The method notes describe how we read them.
- AI answers: weighted mean; bar: middle 50% of municipalities
- News24/SABI, Oct 2026
- Victory Research, Jul 2026
- Ipsos, Sep 2026
- 2024 national result
The ANC’s share of the vote in the AI answers, 36.6%, is between its 2024 result (40.3%) and the July 2026 poll (34%). The DA’s share, 22.4%, is within a point of its 2024 election result (21.7%) and below the July 2026 poll’s 27%. MK Party’s share is 11.6% against 15% in the July 2026 poll, and the EFF’s share, 12.6%, is above both its 2024 election result and the July 2026 poll. News24/SABI’s October 2026 figures put the ANC at 31.6%, the DA at 25.2% and MK Party at 14.7%, leaving the same direction of difference from the AI answers for those three parties. The EFF is closer: the AI answers’ 12.6% is near News24/SABI’s 12.2%. The bars, the middle 50% of municipalities, are widest for the ANC and MK Party, whose shares differ most from place to place. Part II measures these differences party by party. The comparisons below show the four metros covered by News24/SABI. The last chart in this section shows the three smaller parties, which a map of first places hides.
AI bars show variation in the stated vote shares. Poll markers show point estimates with sampling uncertainty. The AI bars cannot establish an election forecast’s confidence. National AI values weight municipal medians; metro values use only AI answers that give a percentage for the party. Table views show the exact figures and full party names.
City of Johannesburg: the AI answers put the ANC higher than News24/SABI
The tick is the median percentage in AI answers that give a share for the party; the bar covers the middle 50% of those AI answers.
- AI answers: middle 50%, tick at the median
- News24/SABI, Oct 2026
- Victory Research, Jul 2026
- Ipsos, Sep 2026
City of Tshwane: the DA’s share in the AI answers is close to News24/SABI
The tick is the median percentage in AI answers that give a share for the party; the bar covers the middle 50% of those AI answers.
- AI answers: middle 50%, tick at the median
- News24/SABI, Oct 2026
- Victory Research, Jul 2026
- Ipsos, Sep 2026
eThekwini: the AI answers put MK Party and the DA below News24/SABI
The tick is the median percentage in AI answers that give a share for the party; the bar covers the middle 50% of those AI answers.
- AI answers: middle 50%, tick at the median
- News24/SABI, Oct 2026
- Victory Research, Jul 2026
- Ipsos, Sep 2026
City of Cape Town: the AI answers and News24/SABI put the DA above half
The tick is the median percentage in AI answers that give a share for the party; the bar covers the middle 50% of those AI answers.
- AI answers: middle 50%, tick at the median
- News24/SABI, Oct 2026
In Johannesburg the AI answers put the ANC at 31.5% and the DA at 27%, whereas News24/SABI’s October 2026 poll has the DA ahead on 29.2% and the ANC on 22.6%. Tshwane’s DA share is close, 32.5% in the AI answers against 32.1% in News24/SABI, but the ANC’s 30% in the AI answers is above News24/SABI’s 23.9%. In eThekwini the AI answers put MK Party at 31.8% and the DA at 20%, below News24/SABI’s 43.4% and 30.2%, while the ANC’s 25% is above News24/SABI’s 11.3%. Cape Town is closer: the DA’s 53% in the AI answers and 56.1% in News24/SABI both put it above half. These are comparisons of party vote intention. News24/SABI’s Johannesburg mayoral shortlist figure of 35.9% for Helen Zille measures a different choice. The next chart compares the smaller parties nationally.
The smaller parties’ shares of the vote in the AI answers, against election results and the 2026 polls
Each row is a party. The tick averages each municipality’s median percentage, weighted by registered voters. The bar covers the middle 50% of municipal values, with the same weights. The July 2026 poll has no figure for the Patriotic Alliance or Freedom Front Plus; News24/SABI covers all three, with small-party figures indicative.
- AI answers: weighted mean; bar: middle 50% of municipalities
- 2021 local result
- 2024 national result
- Victory Research, Jul 2026
- News24/SABI, Oct 2026
- Median in the AI answers that name the party
ActionSA is named in 37% of AI answers, and in 92% in Gauteng; its share of the vote in the AI answers is 3.9% against 1.2% in 2024. Freedom Front Plus is named in 28% of AI answers, and in 58% in the Free State; its share of the vote in the AI answers is 3.2% against 1.4% in 2024. Patriotic Alliance is named in 40% of AI answers, and in 88% in the Western Cape; its share of the vote in the AI answers is 4.6% against 2.1% in 2024. The AI answers carry each party where it was visible in 2024, add 1.8 to 2.7 points, and leave it out elsewhere, so no municipality’s AI answers rank any of the three first. News24/SABI’s October 2026 national figures are 2.9% for ActionSA, 3.2% for Freedom Front Plus and 2.4% for the Patriotic Alliance. The AI answers are close on Freedom Front Plus and higher on the other two; the small polling shares are indicative. The Patriotic Alliance comes closest: the AI ranks it first in Central Karoo in 7 answers, and second in Cape Town in 64.
The three smaller parties are named where each is strong and left out elsewhere. The next section shows the 14 municipalities where the five AI models rank different parties first.
Where the AI models differ
In 38 municipalities all five AI models rank the same party first. They differ in 14: 9 districts in KwaZulu-Natal, two metros and three districts in the Cape.
Each dot is the party an AI model ranks first for most voters in that municipality. A grey dot means the AI model did not rank the parties there; this is nearly always Claude. Most of the differences are in KwaZulu-Natal, where different models rank the ANC, MK Party or IFP first.
The party each AI model ranks first, in the 14 municipalities where the AI models differ
Each cell is the party an AI model ranks first. Municipalities grouped by province, largest group first, then by registered voters; models ordered by how often they agree with the others. On a phone, scroll the grid sideways to see all six AI models.
In the 9 KwaZulu-Natal districts the first party is the ANC, MK Party or the IFP, depending on the AI model: the three parties finished close in 2024, and the AI models order them differently. Tshwane splits between the DA and the ANC; Nelson Mandela Bay splits between the DA and the ANC. The 3 other districts, in the Western Cape, the Northern Cape and the Eastern Cape, split between the DA, the Patriotic Alliance and the ANC. Every split is between parties that finished close in 2024. Across all 52 municipalities, DeepSeek and Gemini agree most (90%); Claude and DeepSeek least (76%). Weighted by the voters’ roll, the municipalities where most AI models rank the ANC first hold 68% of all registered voters, and the 14 split municipalities hold 22% of them, Tshwane and Nelson Mandela Bay among them.
Part II asks where these shares come from: the 2024 election result, or the polls of 2026.
Part II
The 2024 election result or the 2026 polls
We turned off web search, so the AI models answered from their training data alone. Most AI answers cite the 2021 and 2024 election results. Their vote shares remain close to the 2024 election result on average, with differences that vary by party. This section measures movement against Victory Research’s July 2026 poll; the comparisons in Part I also show Ipsos and News24/SABI.
Compared with the results and the July 2026 poll
The July 2026 poll differs from the 2024 election result by about 2.9 points per party. The AI answers differ from the 2024 election result by about 2.5 points, so they moved part of the way toward the July 2026 poll.
The chart shows the two results, the July 2026 poll and the party’s share of the vote in the AI answers on one line per party. The thin line shows the difference between the 2024 election result and the July 2026 poll, and the thick line the difference between the 2024 election result and the AI answers. Blue means the party’s share rose from 2024, rust means it fell. The distance between the two dots shows the difference between the AI answers and the July 2026 poll.
Across the country, per party: the 2021 election result, the 2024 election result, the July 2026 poll and the share of the vote in the AI answers
Results: Electoral Commission, added up over the 52 metros and districts. Poll: Victory Research, July 2026, with no figure for the Patriotic Alliance or Freedom Front Plus. AI: each party’s share of the vote in the AI answers, weighted by registered voters.
- 2021 election result
- 2024 election result
- poll, July 2026: the thin line is how far it moved on from 2024
- AI, for November 2026: the thick line is how far it moved on from 2024
- rose from 2024
- fell from 2024
The AI vote shares remain close to the 2024 election result for several parties. For 4 of the 6 parties in the July 2026 poll, the AI answers differ from the 2024 election result in the same direction as the July 2026 poll, and the average gap between the AI answers and the July 2026 poll, 3 points, is about the size of the July 2026 poll’s own distance from 2024, 2.9 points. The AI answers are close to the 2024 election result on average, but the size and direction of the difference vary by party. The next chart plots the two shifts against each other, one dot per party and place.
Per party: the July 2026 poll’s difference from the 2024 election result, against the AI answers’ difference from the 2024 election result
One dot per party, nationally and in the Johannesburg, Tshwane and eThekwini polls. Across: the July 2026 poll minus the 2024 election result. Up: the AI answers minus the 2024 election result. Dashed line: the same change as the July 2026 poll. Flat line: no change from 2024.
ANC
DA
MK Party
IFP- Other
Of the 21 dots, 17 are in the two quarters where the AI answers and the July 2026 poll differ from 2024 in the same direction, 2 are on the flat line with the share of the vote in the AI answers equal to the 2024 election result, and 2 are in the opposite quarters; the largest of those is EFF nationally. For 12 of the 21 dots the AI answers shift by less than the July 2026 poll. The three metro charts below show where the shifts are largest.
City of Johannesburg
Per party: poll minus 2024 about 8.7 points, answers minus 2024 about 2.5, answers minus poll about 7.8. Poll: Victory Research, July 2026.
- 2021 election result
- 2024 election result
- poll, July 2026: the thin line is how far it moved on from 2024
- AI, for November 2026: the thick line is how far it moved on from 2024
- rose from 2024
- fell from 2024
Ipsos, September 2026, metro subsample: ANC 31%, DA 27%.
City of Tshwane
Per party: poll minus 2024 about 6.6 points, answers minus 2024 about 2.8, answers minus poll about 4.7. Poll: Victory Research, July 2026.
- 2021 election result
- 2024 election result
- poll, July 2026: the thin line is how far it moved on from 2024
- AI, for November 2026: the thick line is how far it moved on from 2024
- rose from 2024
- fell from 2024
Ipsos, September 2026, metro subsample: DA 31%, ANC 22%.
eThekwini
Per party: poll minus 2024 about 8.1 points, answers minus 2024 about 7.8, answers minus poll about 2.8. Poll: Victory Research, July 2026.
- 2021 election result
- 2024 election result
- poll, July 2026: the thin line is how far it moved on from 2024
- AI, for November 2026: the thick line is how far it moved on from 2024
- rose from 2024
- fell from 2024
Ipsos, September 2026, metro subsample: MK Party 55%.
In the metros the July 2026 poll differs from 2024 by 8.7 points in Johannesburg, 6.6 points in Tshwane and 8.1 points in eThekwini, more than the 2.9 points nationally. The AI answers follow the July 2026 poll most closely in eThekwini, where they differ from 2024 by 7.8 points and from the July 2026 poll by 2.8. In Johannesburg the AI answers stay within 2.5 points of 2024 while the July 2026 poll has moved 8.7, so the gap is 7.8 points. The AI answers reflect MK Party’s strong 2024 result in eThekwini. In the Gauteng metros their percentages remain closer to the 2024 election result. Against News24/SABI’s October 2026 figures, eThekwini’s MK Party share in the AI answers is about 12 points lower. Closeness to the July 2026 poll does not hold across every survey. The last chart splits the national gap from the July 2026 poll by AI model.
Per AI model: the difference from the 2024 election result, and the difference from the July 2026 poll
National figures, one AI model at a time. Each figure is the average over the six parties in the July 2026 poll, in percentage points. A model’s answers can differ a lot from 2024 and still be far from the July 2026 poll, if the differences are on the wrong parties.
Difference from the 2024 election result, points per party
Difference from the July 2026 poll, points per party (shorter is closer)
By AI model, DeepSeek is closest to the July 2026 poll at 2 points and Gemini is furthest at 6.5. Gemini differs most from 2024, by 6.6 points, and it is also the furthest from the July 2026 poll: its largest shifts are on the wrong parties.
The next section shows the evidence the AI answers name for their percentages. Most name the two election results. Few name a poll.
What the AI answers cite
Most answers cite the 2024 and 2021 results. About 6% cite a poll.
The evidence each AI answer names for its ranking, per AI model.
Evidence the AI answers name, as a share of AI answers, per AI model
Each row is a kind of evidence. The dots are the five AI models; the line is the range between them. Kinds of evidence named in fewer than 3% of every AI model’s answers are left out.
2024 national or provincial results are the evidence the AI answers name most, in 33% to 81% of AI answers depending on the AI model, and opinion polls are named in 2% to 14%. The shares in the previous section rest on the last two results and on reasoning from them, which is why they are within a few points of 2024. The next section shows the people the AI answers name.
The next section lists the people the AI answers name. They are national figures, and each is named where their party is strong.
Who the AI answers name
Jacob Zuma is named in 10.5% of AI answers and Gayton McKenzie in 7.1%. The sitting mayors of the biggest metros are named in fewer than 1%.
The people the AI answers name, as a share of all AI answers, then by AI model and by province. An AI answer can name several people or none. Party: the affiliation used in this September 2026 study.
People the AI answers name, as a share of all AI answers
Share of all 7,482 answers naming the person. Portraits: Wikimedia Commons, People's Assembly and official profiles.
People the AI answers name, per AI model
Each row is a person. The dots are the six AI models; the line is the range between them.
People the AI answers name, by province
Share of the province’s answers naming the person. Darker is a larger share.
How the AI answers write about each person
Each person’s mentions, split by the entity reader’s score: cool (below neutral), factual (a mention with no tone) and warm (above neutral). People the AI answers name in at least 40 AI answers.
- Cool: written against the person
- Factual: named with no tone
- Warm: written for the person
Jacob Zuma is named most in KwaZulu-Natal (19.4% of its answers), Gayton McKenzie in the Western Cape (26.6%) and Julius Malema in Limpopo (12.0%): each is named where their party is strong, and the AI answers for the rest of the country leave them out. DeepSeek names people most often and Claude least, so the count of names depends on the AI model more than on the place. 64% of the mentions of Jacob Zuma carry no tone, a name attached to a party; among the rest, warm outnumber cool about 4 to one. The same answers write coolly about MK Party as a party, so the warmth is about his pull and the coolness about the party’s record. The AI answers write warmly about Geordin Hill-Lewis in 95% of mentions and coolly about Floyd Shivambu in 65%. The sitting mayors Nasiphi Moya, Geordin Hill-Lewis and Dada Morero are named in 0.8% of AI answers or fewer, so the AI answers rest on national leaders, and the people who run the councils are almost absent from them.
Part III lists the reasons the AI answers give, which are the same in every province, and the issues, which change with the place.
Part III
Reasons, issues and the voter
The same two reasons are in every AI model’s answers. The issues change with the place and follow the party the AI ranks first. The voter’s age and gender change the AI answer very little, and between 24% and 40% of AI answers come back in the voter’s home language.
Why the leader leads
The same two reasons are in every AI model’s answers: the ANC loses support because of service delivery failures, and the EFF attracts young and unemployed voters.
Between 70% and 90% of AI answers contain the reason that the EFF attracts young and unemployed voters. As many contain the reason that the ANC loses support because of service delivery failures. Between 50% and 67% contain the reason that the ANC still has the strongest local organisation. The AI models differ most on two reasons: whether people vote by ethnic identity, and whether independent candidates matter. Each row is a reason. The dots are the five AI models; the line is the range between them.
Share of AI answers that contain each reason, per AI model
The reasons the AI gives are the same everywhere. The next section shows the issues, which change with the place.
What decides the vote
In eight provinces the AI answers name water and roads as the main issues. In the Western Cape they name housing and crime.
Share of AI answers that name each issue as one that decides the vote in the voter’s municipality.
Issues the AI answers name as decisive: the 52 municipalities clustered by the issues their AI answers share
Each dot is a municipality, coloured by the party most AI models rank first there. A line joins each municipality to the four whose answers name the most similar set of issues, and to the places in its cluster it is most alike; a thicker, darker line is a closer match. The layout pulls similar places together. The shaded areas are five clusters found by grouping places with similar issue profiles; clicking a shaded area opens that cluster's profile beside it: what its answers share, what sets it apart, and which party the AI ranks first there. Hover a dot for its top issues.
- Water and roads, high on jobs36
- Water and roads, high on crime7
- Water and roads, high on refuse and sanitation4
- Water and housing4
- Crime and housing1
In the eight provinces outside the Western Cape, water is named in 75% to 91% of AI answers and roads in 61% to 81%. In the Western Cape the two issues named most are housing (53%) and crime (51%), and water is named in 43%. The reasons the AI gives, in the previous section, were the same everywhere; the issues change with the place, because the AI answers tie the ranking to the state of local services.
Clustered on their whole issue profile, the largest of the 5 clusters, water and roads, high on jobs (7 provinces), holds 36 municipalities whose answers share water (83%), roads (74%), jobs (58%). The clusters follow the party the AI ranks first: in the water and housing cluster (Western Cape) the AI ranks the DA first in 4 of 4 places, against 12% of all places. Where the AI answers rank a different party first, they also name different issues, and the issues are the reasons the AI answers give for the ranking. The map near the top of the page lists the issues for each municipality.
The last section asks whether the person asking changes the AI answer.
Who is asking
Within one municipality, the voter’s age and gender change the AI answer very little. Between 24% and 40% of AI answers are in the voter’s home language, although nobody asked for that.
Each voter had only an age band, a gender, a home town and a home language. The first chart shows, for each AI model and each age band, how often the AI answer ranks first the party the AI usually ranks first in that municipality. A flat line means the AI answer depends on the municipality alone. The youngest and oldest bands each have about twenty AI answers per AI model, so small changes there are noise.
Share of AI answers that rank first the party the AI usually ranks first in that municipality, by the voter’s age band
For every AI model the line is close to flat. The widest span across age bands is Grok’s, from 81% to 94%; the narrowest is Kimi’s, from 85% to 91%. The ends of each line rest on about twenty AI answers. Age changes the ranking very little. Language changes the form of the AI answer, which the two charts below measure.
AI answers given in the voter’s own language
The question was in English; the voter carried a home language. Share of each AI model’s answers that came back in a language other than English. Translated to English before reading.
By the voter’s language
The same measure by home language. This chart mixes municipalities. Most isiZulu speakers live in KwaZulu-Natal, where the AI models differ, so the isiZulu range is wide. Men and women differ by less than two points on every AI model.
Language of the AI answer, per AI model
Share of each AI model’s answers in each language, as detected before translation. Languages under 0.5% of an AI model’s answers are left out.
Between 24% (Grok) and 40% (Kimi) of answers came back in the voter’s home language, although the question was in English. By language, the share of AI answers that rank the municipality’s usual party first is lowest for isiZulu speakers (72%) and highest for isiXhosa speakers (95%), and the low end follows the split municipalities of KwaZulu-Natal, where most isiZulu speakers live, rather than the language. Across all models the non-English answers are mostly isiZulu (21.5%), Sesotho (5.6%) and Afrikaans (2.1%). A second AI reader scored how positively each AI answer writes about the four largest parties, from −3 to +3. Every AI model writes least positively about the EFF, −1.5 on average, and most positively about the party it ranks first.
The AI answer reader under the findings list, near the top of the page, shows the AI answers themselves.
Questions about this study
- Did LLMEKNOW poll any voters for this page?
- No. Every one of the 1,247 voters is simulated: a made-up person with only an age band, gender, home town and home language. The page reports what six AI models answered when those simulated voters asked them to rank the parties in their municipality. It is an experiment in reading AI answers, and the AI answers can be wrong.
- Which AI models were asked, and how?
- ChatGPT 5.6, Gemini 3 Flash, Grok 4.3, Kimi K2, DeepSeek V4 Flash and Claude Sonnet. Each simulated voter asked one question once, in English, on 4 September 2026, with web search turned off, so every AI answer rests on the AI model’s training data. That produced 7,482 AI answers.
- What does “ranked first” mean on the map?
- Ranked first means the party at the top of an AI answer’s ranking for that municipality. It is the largest share, which can be less than half of the vote. The map colours each municipality by the party most of the five ranking models put first; Claude did not rank the parties in 75% of its answers and is left out of that count.
- How do the AI answers compare with the real polls?
- Across the country the ANC’s share of the vote in the AI answers is about 37%, against 40% in the 2024 election result, 34% in Victory Research’s July 2026 poll and 31.6% in News24/SABI’s October 2026 poll. The EFF’s share in the AI answers is about 13%, close to News24/SABI’s 12.2% but well above Victory Research’s 7%. In eThekwini the AI answers put MK Party at about 32%, against News24/SABI’s 43.4%. Agreement depends on the party, municipality and poll. Most AI answers cite the 2024 and 2021 results as their evidence; about 6% cite a poll.
- Where does the data on this page come from?
- The voter sample follows the Electoral Commission’s voters’ roll of 3 September 2026. Election results are the Electoral Commission’s 2021 municipal and 2024 national results, added up to the 52 metros and districts. Polls are Victory Research (July and March 2026), Ipsos (September 2026) and News24/SABI’s Elections Poll ’26 (October 2026). News24/SABI interviewed 5,001 registered voters from 25 August to 17 September, with a national sample and separate samples in Johannesburg, Tshwane, eThekwini and Cape Town. We use its party vote intention figures modelled for turnout. Municipal boundaries come from the Municipal Demarcation Board via the Electoral Commission, and town locations from GeoNames. A sample of 312 AI answers with their full text is published as JSON on this page.
- How is the tone of a mention scored, and why is Jacob Zuma mostly “factual”?
- A second AI reader scores every mention of a person from 0 (written against them) to 1 (written for them), with 0.5 for a mention that carries no tone. Most political mentions are of that kind: “Zuma’s MK Party” or “led by Jacob Zuma” attaches a name to a party without judging the person. Two thirds of Zuma’s 799 mentions score exactly 0.5 for that reason. Among the rest, warm outnumber cool about four to one, because the AI answers describe his pull in KwaZulu-Natal, while a separate reader that scores parties has MK Party below neutral on every AI model. An average of all mentions would hide this, so the page reports the split instead of a mean.
- Can I reuse the figures and the data?
- Yes, with attribution to LLMEKNOW. The figures on this page and the published sample of AI answers are released under a Creative Commons Attribution 4.0 licence. Party logos and portraits belong to their owners and are shown at icon size for identification.
The data behind the page: a sample of 312 AI answers with their full text is published as JSON, under a Creative Commons Attribution 4.0 licence.
How the voters were built
The 1,247 voters are spread across the 8 metros and 44 district municipalities in proportion to the registered voters’ roll, which the Electoral Commission published on 3 September 2026. Each municipality has a floor of ten voters, so that every district is represented. Within each municipality, age band and gender follow the same roll, home towns follow the local municipalities, and language follows the dominant languages of the area. A voter carries nothing else: no income, education, party history or opinions.
How the AI answers were read
Each voter asked the one question once, in English, with no web search, on 4 September 2026. The AI models were ChatGPT 5.6, Gemini 3 Flash, Grok 4.3, Kimi K2, DeepSeek V4 Flash and Claude Sonnet. None was given the date, the election date or any document; each answered from its training data. Answers in other languages were translated to English before reading. Every AI model mentions 2024 in 38% to 99% of its answers, and 2025 in only 1% (ChatGPT, Gemini) to 12% (Claude). These counts show which years the AI answers mention; they cannot establish an AI model’s knowledge cut-off. A separate AI reader then scored every AI answer against a fixed set of statements, among them “the response ranks the ANC first in this municipality”, one per party. The party with the strongest score is the party the AI ranks first. The same reader scored whether the AI answer has a hung council, and which reasons it contains. Every person named is scored from 0 to 1 for the tone of that mention, with 0.5 for a mention that carries no tone; the page reports the share of a person’s mentions below, at and above 0.5 rather than an average, because most mentions are factual and an average of them says little. Issues, evidence, institutions and politicians were extracted from the AI answers as named entities and grouped by keyword.
Where an AI answer contains a percentage for a party, we used it. About 40% of the AI answers from the five AI models that ranked the parties contain percentages. Two thirds of those AI answers came from ChatGPT and DeepSeek, whose answers contain percentages in 67% of cases. Gemini’s and Grok’s answers contain percentages in 25% of cases. We skipped a percentage that appeared in the same line as a past year, or with the words won, got, received, secured or polled, because it describes a past result. For each municipality we took the median percentage for each party. A party with no percentage in any answer for the municipality counts as 0. The national and provincial figures are weighted averages of the municipal medians: each municipality counts in proportion to its registered voters. The bars show the middle 50% of municipal values using the same weights. These bars describe geographic variation. They provide no statistical confidence interval for an election forecast. Metro comparisons and the separately labelled figures for AI answers that name a party use medians over only the AI answers that give that party a percentage. Different parties can therefore have different numbers of contributing AI answers.
Share-of-AI-answers figures such as the issues, the arguments and the hung-council share are unweighted counts of AI answers. Every district has a floor of ten voters, so the smallest districts are over-represented in these figures. Only the bars for the party the AI ranks first and the vote shares are re-weighted to the voters’ roll.
For the comparison with the election results and the 2026 polls we aggregated the Electoral Commission’s 2021 municipal results and the 2024 national ballot to the same 52 metros and districts, summing the local municipalities inside each district. The 2021 figures are the combined municipal ballot; the 2024 figures are the national ballot counted in that area.
How the polls were compared
We compare the same AI answers from 4 September 2026 with Victory Research’s March and July 2026 surveys, Ipsos’s September 2026 survey and News24/SABI’s Elections Poll ’26, published in October. The July 2026 poll is the reference for the charts measuring movement from the 2024 election result. The national and metro comparisons also show the other surveys.
News24/SABI interviewed 5,001 registered voters from 25 August to 17 September 2026, primarily face-to-face. Its national sample has 1,000 respondents; Johannesburg, Cape Town and eThekwini each have 1,000, and Tshwane has 1,001. We use party vote intention modelled for turnout, with a national turnout assumption of 52%. The mayoral shortlist questions measure a different choice and are kept out of the party comparisons. Small-party figures within the sampling margin of error are indicative.
News24/SABI’s fieldwork overlaps the AI wave and ends after it. The AI answers were collected without web search or polling documents. Differences between the surveys can reflect their samples, questions and turnout models; they do not by themselves show a change in voter support over time.
What this can and cannot tell you
The AI models answered from their training data, with no web search and no contact with any person, so their AI answers can be wrong. This page is no basis for deciding a vote.
This page reports one wave with one AI answer per voter, which cannot separate a real difference between models from run-to-run noise. The five ranking AI models differ in how often their answers contain percentages, so the vote-share comparisons do not give every AI model equal weight. At the youngest and oldest age bands each figure per AI model uses about twenty AI answers, so differences under ten points there are noise.
The 2021 figure is the combined municipal ballot. The 2024 figure is the national ballot counted in the same area. Outside the metros, about half of the AI answers rank the parties for a local municipality inside the district. The three are not the same contest or the same unit, and MK Party did not exist in 2021, so every comparison of MK with 2021 is a comparison with zero.
The polls ask a different question, and they ask real people, so a gap between a poll figure and a figure in the AI answers is a measure of the AI models.
Sources
- Portrait credits and image sources
- News24/SABI, Elections Poll ’26, national and four-metro party vote intention, October 2026
- Victory Research for the Social Research Foundation and The Common Sense, Q1 and Q2 2026 surveys
- Inside Politics, Election 2026/7 series on turnout in the metros
- Ipsos, Pulse of the People, released 4 September 2026, via Daily Maverick
- Opinion polling tables, 2026 South African municipal elections
- 2024 national and provincial results, Electoral Commission results dashboard
- Voters’ roll by municipality, age and gender, Electoral Commission of South Africa
- 2021 municipal election results by municipality, Electoral Commission results dashboard
- 2024 national ballot results down to voting district, Electoral Commission downloadable results
- Municipal boundaries, Municipal Demarcation Board via the Electoral Commission
Ask AI about your own market.
We asked about an election. Ask about your category, your competitors, your customers: six assistants, real answers, read for you.