How picks are made

Trained knowledge only

Every page is written by an AI model (claude-opus-5-5, generation opus-5-5_g1) from what it learned in training. It has no web search, no browsing and no retrieval while it writes: the generation runs with every tool switched off by configuration, and a run that reports any tool is discarded. Its knowledge stops at 2026-06, which every page states, so a product released later cannot appear.

Two AI models, not one

On the main categories, the page shows what Claude and GPT choose. Each model answers once and the page gives the date. Neither sees the other's answer.

Claude answered when the page was generated, with the full instructions and with the neighbouring categories in view, so as not to name the same product twice. GPT was asked afterwards, a short question under the same rules: no web, no tools, only products sold in this country, never a price. That difference can explain part of a disagreement.

Code compares the answers and writes the verdict: the same product, the same line, the same brand, or no agreement. Across 1046 categories, the models choose the same brand in 48% of them and the same product in 17%.

Agreement is not proof. The models learned from much of the same web and can be wrong together. We do not say which one is right, because nobody has compared these products.

The “Where to find” list is the one part that comes from a web search. It runs later and separately, and it never enters the choice.

The product photo comes from that same search: the image a shop or the brand publishes on the product page, shown from their site with its name underneath, never copied here. It can show another size or colour, and it is never from Amazon.

The price and the specifications come from those same store pages, read by code and by a small AI model (Claude Haiku) with no tools, after the pick was made: they never change it. A specification is shown only when its words and numbers are in a sentence copied word for word from the page. The price, excluding shipping, is read again every day, always with the store and the date, and disappears after three days without a new reading; a pack, a marketplace seller or a sold-out product gives none. It never comes from Amazon. Health and baby pages show no price and no specifications.

Since November 10, 2026, this site shows no link to any shop.

A prediction, not a test

On the other categories one model answers, once. Either way the page shows the pick, why, who should not buy it and what matters when choosing. Like any AI answer it is a prediction from what the model learned, not a test in a lab. The model's own confidence is recorded but never shown, because it is not a reliable signal.

What the pick levels mean

Parts and accessories that depend on what you already own get a guide instead of a pick. Books, music, films and game titles are not covered: there, "best" is taste, not a purchase decision.

No stars, no scores

There are no star ratings or scores on this site. The model never writes links or prices. A price shown comes from the page of a store that sells the product, with the store's name and the date, never from Amazon: an Amazon price will only appear if it comes directly from Amazon's API.

Independence

Every pick, criterion and brand on this site comes from the AI models. No brand, seller or marketplace has any say in them, and the models never see prices or commissions. The site shows no links to shops for now and earns nothing from its pages.

Data

Every page has a machine-readable version at the same address with .json appended.

Contact

hello@intelligencepick.com