How we work it out

Every recommendation on this site is produced the same way, by the same code, from figures that each carry a source and a date. This page is that method, written out — including the parts that make the answer less confident.

The question

Almost every decision about a device you already own reduces to one comparison: what does each option cost per useful month? Spending on a repair buys more months of the device you have. Buying a replacement buys a different number of months for a different amount of money, minus whatever the old one is worth to somebody else. Doing nothing costs nothing and buys however long the device has left.

Putting all of them on that single axis is the whole trick. It is why a modest repair often beats an upgrade that looks far more attractive in a shop, and it is why the answer changes when the device is close to the end of its supported life — at which point the months you are buying stop being worth much.

What goes into it

Four kinds of input, and they are treated differently on purpose.

Facts about the device. What it is, when it was released, what it shipped with, what it can still run. Taken from the manufacturer’s own published pages.
Prices we have observed. Not asking prices we assume, but prices seen in licensed retailer feeds and recorded with the day they were seen. The price history is append-only: every import writes a row even when nothing changed, because storing only the changes would bias every average towards whichever shop moves its prices most.
Costs we have verified. Repair and battery-replacement prices, entered against a source and re-checked on a schedule. Each one shows how long ago it was checked, and an old figure lowers the confidence of any answer that leans on it.
What you tell us. How you use the device, what actually bothers you about it, and what you would be willing to do about it. “I don’t know” is a real answer here and a better one than a guess: it is stored as itself and it lowers confidence, rather than quietly becoming a default that makes the result look surer than it is.

Uncertainty is carried, not hidden

Nothing in the engine is a single number pretending to be exact. A cost or a benefit is a range — low, likely, high — and it stays a range all the way through the calculation, so the spread you see at the end is the real one rather than one reconstructed afterwards.

Money is handled in whole pence and rounded exactly once, at the point it is printed. A price we observed at a shop is printed as observed, because it is a fact; a figure we calculated is rounded so that it reads as the estimate it is. Presenting a computed estimate to the penny would imply a precision the inputs do not support.

Confidence is about the inputs

Every result says how confident it is, and that confidence describes what the engine was told rather than how tidy the answer came out. A narrow-looking result built on a battery health figure you did not have is a confident-looking guess, and it is labelled low confidence even though the number has few decimal places.

What lowers it: missing answers, stale costs, few price observations, a short price history.

When we cannot cost something, we say so

Cost and benefit are allowed to be separately absent. If we have no verified repair price for a particular model and fault, the repair option returns “we do not have a price for this” and takes no further part in the ranking — it does not get a plausible estimate.

That rule exists because an invented figure is indistinguishable from a real one to the person reading the page. The same applies to the engine itself: if a constant it needs is missing, it stops with an error rather than substituting a default, because a quiet default produces a confident, plausible, wrong answer.

The same inputs always give the same answer

The calculation is deterministic. There is no randomness, no personalisation you did not ask for, and no drift between two people who answer identically. Each result records which version of the engine and which published set of figures produced it, so a recommendation can be reproduced later exactly as it was given — including after the underlying prices have moved.

Every step is written down as it is calculated, not reassembled for display afterwards. Our own tests check the workings rather than only the verdict, so an answer that comes out right for the wrong reason still fails.

What artificial intelligence does here, and what it does not

AI is used to explain, never to decide. It cannot produce a figure, adjust one or round one. Explanations are templates with typed gaps, the figures dropped into those gaps come from the same database rows the page is showing, and an automatic check rejects any sentence containing a number it cannot trace back to one of them.

The site works with it switched off, and it is off by default.

What this is not

Not an inspection. We have never seen your device. A shop with it in their hand may find something we could not know about.
Not a quote. What a repair costs you is what a repairer quotes, not what we estimate.
Not a guarantee of a price. Prices move daily, and around a manufacturer launch they move sharply. We show when we last saw each one.
Not advice. It is an economic comparison. What matters to you about a device is yours to weigh.

How we score · Where our figures come from · How we make money