What Do the Polls Say? French Presidential Election Polls, and How They Frame the Race

As of this research edition, 29 October 2026, the Commission des sondages’ register includes a Cluster17 survey for POLITICO published on 26 October and an Ipsos bva survey for Le Parisien published on 24 October. Those are publication dates, not interchangeable fieldwork dates. Secondary reporting of the newer Cluster17 survey gives a first-round configuration with Marine Le Pen at 33 percent, Jean-Luc Mélenchon at 18.5 percent, Édouard Philippe at 16 percent, and Raphaël Glucksmann at 13 percent. It places Le Pen ahead and Mélenchon in a competitive position for the runoff in the scenarios examined. That is a conditional survey finding, not qualification for an election that has not happened.

The primary-source chart below uses Ipsos’s own published summary. Its fieldwork ran on 21 and 22 October among 1,500 registered voters, described by Ipsos as representative of adults in France. Across four tested configurations, the summary gives Le Pen 32 to 33.5 percent and Mélenchon 17 percent. Philippe ranges from 13 to 15 percent and Glucksmann from 13 to 15.5 percent. These ranges belong to different candidate configurations. They are not confidence intervals, cannot be combined into one ballot, and must not be added together as though they formed a complete result.

An intention-to-vote question usually asks respondents to imagine an election held under specified circumstances. Those circumstances include a list of names, an election round, and sometimes a choice among competing nominations. A result therefore measures an answer to a constructed situation. The construction is necessary: there is no way to ask about an electoral choice without presenting or describing one. But necessity does not make the selection politically invisible. The people included on the hypothetical ballot help determine what the answer can mean.

Ipsos’s October summary illustrates the point with Gabriel Attal. It reports 9.5 percent when he is the sole candidate of the central bloc, but 4 percent in a configuration that also includes Philippe. The comparison does not isolate a pure personal popularity score. It changes the set of alternatives respondents are considering. The scenario switch in the visual should therefore be read as a lesson in conditional measurement, not as an experiment proving exactly which individual voters moved from one candidate to another. Aggregate differences do not identify personal journeys.

A population, a recruited sample, registered voters within that sample, likely participants, and respondents expressing a candidate preference can be different groups. A headline may display only the last percentage while the methodological report describes the earlier stages. To interpret the estimate, readers need to know which denominator reaches the published table. The headline sample size does not automatically equal the number of answers used for every candidate configuration or every subgroup. Dividing a sample by age, geography, or political preference can make the relevant base much smaller.

Quota methods seek to match selected characteristics of the target population. They do not turn an online panel into a literal miniature of every relevant feature of society. People’s willingness and ability to answer, recruitment practices, and unmeasured differences remain important. That does not mean quota surveys are worthless. It means representativeness is an argued and documented methodological achievement, not a magic property bestowed by the word national. The Commission’s notice requirements give readers a place to inspect how the claim was constructed rather than simply accepting or rejecting it on trust.

Fieldwork dates identify the period during which answers were collected. Publication dates identify when the findings entered the public conversation. Confusing them can produce a false causal story: a poll published after an event may have been partly or wholly conducted before it. Even genuinely later interviews do not prove the event caused a measured difference. Many things change at once, and a survey wave is not automatically a controlled comparison. Claims that a speech, scandal, debate, or protest transformed the race need a stronger design than chronological proximity.

The sources consulted for the Cluster17 survey do not agree on the precise fieldwork window: one account reports 19–20 October, while other summaries give 19–22 October. Rather than silently choose whichever dates make a narrative smoother, this essay excludes that survey from its primary-source methodological graphic until the original notice or full report is checked directly. The results can be identified as reported findings while the documentation gap remains visible. This is not evidence of manipulation by the institute. It is a limitation in what this research edition has independently verified.

Question wording establishes categories and makes some distinctions available while leaving others out. Asking whether a measure protects public order is not identical to asking whether it restricts a liberty. Both descriptions may refer to aspects of the same proposal. The response options matter too: approval, strong approval, neither, refusal, and uncertainty can be offered or combined differently. The visual wording comparison in this essay is an explicitly invented teaching example. It is not attributed to an institute and is not evidence that a particular published survey used a loaded question.

The right response is to inspect the exact wording rather than infer it from the headline. Readers should also ask what came before the question in the questionnaire, which context was supplied, and whether a complex proposal was reduced to one attractive or alarming feature. None of these questions presupposes bad faith. They recognize that measurement has an architecture. A survey can be competently executed while still encouraging a narrow account of the political problem it asks people to judge.

Published estimates are often adjusted rather than being a simple count of identical responses. Weighting gives different statistical influence to respondents to better reflect selected characteristics or correct identifiable imbalances. Electoral adjustments may also use information relevant to past voting or participation. The analytical question is what is adjusted, on which evidence, and with what sensitivity. The existence of an adjustment is not proof that an institute changed the result to please a client. Equally, the technical vocabulary does not make every adjustment beyond criticism.

The Commission’s notice requirements include, where applicable, the criteria used to adjust raw results. That creates a practical reporting obligation for an essay criticizing a poll: identify the documented method and distinguish it from a conjecture about motives. A claim of deliberate manipulation needs evidence of something more specific than a surprising outcome. Uncertainty, ordinary methodological choices, selective media presentation, accidental mistakes, and fabricated polls belong in different categories. The question-to-headline sequence separates them so that justified scrutiny does not become a generalized conspiracy theory.

The horizontal segments in the opening chart show variation between hypothetical ballots in one Ipsos release. They do not represent sampling uncertainty. A margin of error concerns uncertainty associated with an estimate under stated statistical assumptions. A scenario range concerns a change in the question’s electoral configuration. The two may coexist, but they answer different questions. Drawing the same-looking band for both without explaining the distinction would make a sophisticated-looking chart less informative than a carefully labeled table.

Classical sampling intervals also require caution when applied to quota samples rather than simple random samples. Published approximations can be useful orientation, but they do not capture every source of error, and they do not repair a leading question or a missing segment of the population. Nor should overlapping intervals be converted into a universal rule that any difference is meaningless. Comparisons can depend on correlated responses, design, and the precise quantity being estimated. The honest default is to avoid overinterpreting small movements until the relevant methodological information supports the interpretation.

When two candidates sit close together, a chart can create an impression of clear order even though the evidence supports a more cautious description. Conversely, a substantial measured lead in a particular survey does not authorize a precise probability that a candidate will qualify months later. That probability would require a model connecting current measurements to future choices, nomination decisions, turnout, campaign events, and the final ballot. A probability from an aggregator is therefore a model output, not another way of printing the original poll.

This essay deliberately does not publish an election forecast, combine incompatible scenarios into a national average, or infer individual vote transfers from aggregate changes. It also does not treat a potential contender appearing in a survey as an officially validated candidate. These exclusions are substantive. They preserve the distinction between what a survey directly measures and the additional claims commentators often build on top of it. The intellectual value of a number increases when readers can see where measurement ends and interpretation begins.

The company conducting an investigation, the organization paying for it, and the outlet publishing it can be separate entities. They should not be compressed into the phrase a newspaper poll. In the October example, the published branding names Ipsos bva, CESI École d’ingénieurs, and Le Parisien. Those names do not all perform the same function. A newsroom can commission a survey without owning the institute; an institute’s corporate parent does not automatically dictate each editorial headline. Ownership mapping raises questions but does not answer them in advance.

Corporate history matters too. Ipsos announced completion of its acquisition of The BVA Family on 24 June 2025 and said the French operation would use the Ipsos Bva name. Treating Ipsos and BVA as wholly independent sources of corroboration after that change would require more care than simply counting brand names. The October notice register also lists surveys from Cluster17, OpinionWay, Toluna Harris Interactive, Ifop, Elabe, Odoxa, CSA, YouGov, and Verian. This is an illustrative inventory of published work, not a ranking of quality or an audited ownership census of all institutes.

One survey can support several defensible headlines: the leading candidate remains ahead; another rises; a former favorite falls; a new contender is tested; or the ranking changes under a different nomination. Editorial selection decides which of those observations becomes the opening story. It also decides what disappears when a complex release is compressed into a single line. The underlying dataset may remain unchanged while the perceived political significance shifts dramatically. Calling that framing does not require proving that the numbers themselves were falsified.

The horse-race frame can then influence which questions journalists ask politicians and which choices voters regard as viable. A candidate described principally as unable to qualify may receive less attention for a substantive policy proposal. A voter concerned with blocking another contender may attend more closely to rankings than to programs. These are plausible mechanisms of political interaction, not causal effects measured by the October survey itself. Demonstrating their size would require additional research. The essay distinguishes the mechanism it explains from empirical effects it has not established.

A chart can mislead without containing an invented number. A truncated axis can exaggerate a small change. A missing candidate can make a partial selection look like the complete ballot. A line connecting incompatible candidate configurations can imply a trend that was never measured on a consistent basis. A shaded band can look like statistical uncertainty even when it simply combines different questions. These are design decisions with intellectual consequences. The graphics in this essay therefore state their scale, their scope, their source, and the meaning of each range directly beside the display rather than delegating those explanations to a distant bibliography.

The practical aim is not to make readers suspicious of every number. It is to let them ask a better question of it. An informative chart should make the evidence easier to interrogate, not make the designer’s interpretation harder to resist. When documentation is incomplete, restraint is a visual achievement: an explicitly missing field is more honest than a persuasive invented detail. That principle also explains why the latest located survey and the survey used for the fully documented teaching chart are not silently treated as the same object.

France’s Commission des sondages publishes methodological notices and can issue corrections or warnings within the legal framework governing electoral polling. Its website describes obligations concerning publication information, methods, nonresponse, and adjustment criteria. The public register is valuable because it supplies a route from an eye-catching chart back to the documentation behind it. It is not a guarantee that any result in the register will become the election outcome, nor does it exempt a newsroom from explaining the scenario it chooses to highlight.

The Commission also reports warnings about fake polls circulating online during the municipal campaign. That is a different problem from a real institute’s contestable questionnaire. Before debating a suspicious screenshot’s margin of error, establish whether the institute actually produced it, whether the named commissioner confirms it, and whether an appropriate notice exists. The strongest practical habit is simple: follow the number backward through the chain. Candidate list, question, recruitment, fieldwork, denominator, weighting, graphic, headline. Only then ask what political story the percentage can responsibly support.