How recommendations work
Our methodology is applied consistently without favouring either party. Our assumptions and thresholds are published here so you can assess them for yourself.
Which parties are assessed
Our model compares the Conservatives and Reform UK — the two centre-right parties with a national presence in Great Britain. Other parties appear in the projections because they determine who the centre-right is competing against but they are not the subject of our recommendation. Northern Ireland has a different party system and is outside our scope.
Data sources
The baseline is the verified result of the July 2024 general election in every constituency in Great Britain, at candidate level, on the 2024 boundaries (as published in the House of Commons Library results collection; our copy is derived from Democracy Club candidate-level data). On top of that our model applies a rolling average of published national and nation-level (Scotland and Wales) Westminster polling and seat-specific overrides for material events: by-elections, defections, confirmed candidates or withdrawals, formal stand-down arrangements and credible constituency polling. See details of current and past polls included for each update.
How the projection is made
Each party's 2024 share in a constituency is scaled by the ratio of its current nation-level polling to its 2024 nation-level result and the respective shares are then rescaled to sum to 100. This is a proportional-swing model: a party that doubles nationally is projected to roughly double everywhere, rather than gaining the same number of points in every seat. Independents and minor parties hold their 2024 share. It is a deliberately simple, checkable method — it does not attempt seat-level demographic modelling and it will be least reliable where local circumstances have changed in ways which national polling cannot identify. That is what the overrides and the correction process are for.
The decision rules
- If the data for a seat is unreliable or a material change is unresolved, we publish no recommendation.
- The route test asks whether consolidation of the centre-right vote could actually result in a win in any given constituency. We take the stronger centre-right party's projected share and add a proportion of the weaker centre-right party's vote — because not every voter would transfer. If, even under an optimistic assumption (80% of the weaker party's vote transferring), the combined total still falls short of the projected leader, we state that there is no realistic route to a centre-right victory at present. Where the combined vote overtakes the leader in a constituency only under the optimistic assumption but not under our central one (60% transferring), the result is shown as a marginal opportunity rather than a safe one.
- If the Conservatives and Reform UK are within seven points of each other, the headline is too close to call, with the party fractionally ahead shown as a current lean — a lean, not a recommendation. Where the two are dead level, no lean is given.
- Outside that band, we recommend the stronger centre-right party.
The three parameters — the 80% and 60% transfer assumptions and the seven-point margin of uncertainty or error — are judgements and reasonable people could set them differently. They are held in versioned configuration, applied identically to every seat and any change to them is recorded in the changelog.
Explore the thresholds yourself
Move the sliders to see how the seat counts would change under different judgements. This is exploratory only — published recommendations always use the stated parameters (80% / 60% transfer, seven-point band).
Confidence
Confidence measures the reliability of the evidence behind a result — not how decisive the result is. A firm recommendation can rest on stale data and deserve low confidence; a knife-edge lean can rest on fresh data and deserve high confidence. Each release grades the polling for each nation by age and specificity, and every seat in that nation inherits the grade until fresher polling arrives. A credible constituency-specific poll raises a seat to high; missing material candidate information lowers it. A recommendation for a party that is not standing is treated as a defect and never published.
Updates
In ordinary periods our full model is recalculated and published monthly. Material events — a by-election, a defection, a confirmed candidate, a stand-down arrangement, significant constituency polling — trigger an interim update to the affected seat(s). Once a general election is called, publication moves to weekly, then twice-weekly in the final fortnight leading up to that election, with one clearly-labelled final recommendation in the last 72 hours before that election. Every release is recorded in the changelog.
Known limitations
Our use of a uniform proportional-swing model misses genuinely local effects unless they are entered as overrides. Nation-level polling for Scotland and Wales is sparser than Great Britain-wide polling. Candidate selection is incomplete this far from a general election, which is why many seats carry an unconfirmed-candidate status and a tempered confidence level. Our model projects toward the next general election, whenever it is called; the words “at present” are doing real work.