Inside the multi-LLM valuation council: how Home prices a property with AI

Most instant valuations online come from a single model: one set of weights, one output, presented as a number. Ours works differently. When you ask Home to estimate a property's value, the request goes to what we call a valuation council: three independent systems that each produce an output, then a consensus step that reconciles them.

This post explains how that works and why we think it produces more honest results.

The three inputs

GPT-4o-Vision and the listing photographs. Photographs carry information that address-level data does not. A kitchen that was last updated in 2003, a conservatory added without planning permission, a garden that backs onto a railway embankment: these things affect value and they show up in images before they show up anywhere else. GPT-4o-Vision reads the photographs alongside the listing description and returns a qualitative assessment: condition, apparent layout, features that add or subtract value.

Algorithmic comparables. This is the part most familiar to anyone who has used an automated valuation tool before. The system pulls recent sold prices from Land Registry data for comparable properties within a defined radius, adjusts for size, type, and time of sale, and returns a price-per-square-foot range for the postcode.

Cross-model consensus. The outputs from the two steps above do not always agree. A property might sit in a postcode where sold prices suggest one range, but the photographs suggest the specific unit is in better condition than the local average. The consensus step, which runs across models including Gemini and Grok in addition to the two above, looks at the spread of outputs and arrives at a reconciled range with an explicit confidence interval.

Why a council rather than a single model

A single model will always produce an output. That is not always a virtue. If the comparables data is thin because the postcode rarely transacts, or if the photographs are low quality, a single model will still return a number. It just will not tell you how much uncertainty surrounds it.

The council approach means disagreement becomes visible. When the inputs diverge significantly, the confidence interval widens and we say so. When you're buying or selling in an area with limited recent sales, a range of £380,000 to £440,000 with a note about data limitations is far more useful than a single headline figure like £412,000 with no explanation. That distinction matters most in exactly the cases where an online tool and an agent's own read of a property are likely to disagree.

The same logic applies across the roughly 1,249,531 live listings Home tracks at any one time: a 3-bed house in Fringford, Oxfordshire has a thin comparables pool because rural postcodes transact rarely, so the photographs carry more of the weight than they would for a 1-bed flat on Henry Street in Manchester, where sold prices turn over fast enough that the algorithmic step alone gets close.

Speed

Running three independent systems in sequence would take too long to be useful. In January 2026 we parallelised the requests using an HTTP connection pool, which reduced the average valuation response time by roughly 60% compared to the sequential version. The result is fast enough to feel like a live query rather than a batch job.

We have written before about what AI can and cannot tell you about the value of a house; the council structure is our answer to the "cannot" half of that piece, and to the hallucination problem that a single unchecked model is prone to.

You can explore house prices in your area at house price data.

All valuations on Home are estimates. They are not formal appraisals and should not be used as the sole basis for an offer or listing price. Sold price data is sourced from HM Land Registry; coverage and recency vary by area.

Further reading: the ONS UK House Price Index.