Can an AI app tell you what something is worth?

Point a phone at almost anything now and software will name it: brand, model, often the exact edition. That part genuinely works, and it works well enough that it is no longer worth arguing about.

Then it shows you a price, and the price is a different kind of claim entirely. Identification is a question about the object in front of the camera. Value is a question about other people: what somebody actually paid for one of these, recently, in roughly this condition. No amount of looking at your photo answers that, because the answer is not in the photo.

Two numbers that look identical on screen

A language model asked "what is this worth" will produce a figure. It will be fluent, specific, and formatted exactly like a researched answer. It is an impression formed from text the model absorbed during training, which is to say an average of things people have written about prices, frozen at some point in the past and unattached to any particular sale.

A researched number is the middle of a set of completed sales you can go and look at. Those two things render as the same bold figure in the same card, and nothing about how they look tells you which one you got.

This is not a reason to distrust AI in the loop. It is a reason to ask one question of any tool that gives you a price.

The question: can it show you the sales?

If a tool can show you the actual completed listings behind its number, with dates and prices you can click, the number is checkable and you can judge it yourself. If it cannot, you are being asked to take a confident figure on faith, and a confident figure with nothing behind it is the most expensive kind of wrong: it feels like research, so you stop researching.

Everything else worth asking follows from that one question. Does it distinguish what sellers are asking from what buyers actually paid, which are routinely different numbers in both directions? Does it admit when the evidence is thin instead of averaging two sales into a confident median? Can you tell it that it got the item wrong?

Where we stand, since this is our guide

Price×Point uses a model for the part models are good at: looking at your photo and working out what the item is. It does not ask the model what the item is worth. The price comes from searching completed sales for that specific item and reporting the median, alongside what sellers are currently asking and what the thing costs new, in three separate lanes so they cannot blur into one another.

Every comp links out to the listing it came from. You can open the sales the median was built from, and if one of them is not your item you can strike it out and the price is recalculated without it. Your judgement wins over ours, the original number is kept beside the new one so the disagreement is visible, and striking is free, because charging someone to correct our work would be charging them for our mistake.

There is also a screen for the listings we rejected: the search returned them, we judged they were not your item, and you can read them and see whether you agree. A tool that is confident in its filtering can afford to show you what it threw away.

It says when it does not know. When the identification is uncertain the card says "couldn't confirm the exact product, this is our best description" rather than picking something plausible. When it cannot identify an item well enough to search, it says so and does not search: "we couldn't tell what this is well enough to search for a price, so we haven't looked." A price resting on a single completed sale is labelled as resting on a single sale. None of that is flattering, and all of it is information you need.

You can argue with it. A wrong guess is a normal event, not a dead end, so there are five ways to fix one: type what the item actually is, pick the right one out of the look-alike listings it considered, draw a box around the object you meant in a crowded photo, add a close-up of the label or maker's mark, or edit the search text itself. There is an undo if your correction turns out worse than the original, and the first fix of any item costs nothing.

What it still will not do

Thin data stays thin. For something genuinely obscure there may be two completed sales in the world, and no tool can conjure a reliable median out of two sales. We would rather show you the two and say so.

It does not judge condition. The photo shows the side of the object facing the camera, and on most used goods condition moves the price more than the model name does. The number is a starting point that assumes a condition you have to supply yourself.

It is also not a marketplace. It tells you the number; you take it wherever the item actually sells.

What to do with any price an app gives you

Ask where the number came from. If the app will show you completed sales, read two or three of them and check they are really your item, your generation, your variant. If it will not show you anything, treat the figure as a starting guess rather than a price, and go look at completed sales yourself before you list.

The useful distinction is not human against AI. It is a number with receipts against a number without them. See sold price vs asking price for why the two most-quoted numbers disagree, and why Google Lens can identify it but not price it for the same split one step earlier in the process.