We asked five AI assistants what magnesium to buy, 500 times
1 September 2026 · Vladimir Fedorov, co-founder and CTO, ShelfSignal
This analysis was measured and published with ShelfSignal at shelfsignal.pro. A shopper asking ChatGPT what magnesium to buy gets three or four brand names. We wanted to know which ones, how stable the list is, and what the brands that make it have in common. So we asked twenty magnesium buying questions on ChatGPT, Claude, Gemini, Perplexity and Grok, five times each, and read all 500 answers.
Two hundred and twenty-eight brands were named at least once. Thorne took 44% of the shelf, Pure Encapsulations 35%, and the tenth brand on the list took 7%. The full ranking, the per-assistant split and the citation data are on the ShelfSignal AI Shelf Index for magnesium, which we re-run every month.
Three things came out of the run that we did not expect.
No two assistants recommend the same shelf
Thorne is named in 66% of Grok's answers and 25% of Claude's - same questions, same day. Nature Made takes 38% on ChatGPT and is never named on Claude at all. That is not noise: it holds across five runs per assistant.
It matters because losing on one platform and being absent from another are different problems with different fixes. A low share means the assistant knows you and ranks you below someone else. A zero means it never encounters you, which usually traces to which sites that assistant reads. A single blended visibility score hides both.
Assistants read a brand's own pages and still name someone else
This was the finding that changed how we build the product. For one brand in the top ten, assistants opened four pages on that brand's own domain across thirty answers. They quoted those pages. They named the brand in none of the thirty.
The pages were reachable, indexed, relevant and useful. They cleared every bar except the last one. What the assistant could not lift out of the text was the sentence saying who made this and what it is for - because the page spoke as "we", and the brand name appeared only in the header, the logo and the tab title.
No rank tracker surfaces that. It looks like a ranking problem and it is a product-data problem, which is the kind a content team can fix in a week. This is the sort of diagnosis ShelfSignal reports are built around: not whether a brand appears, but what stands between the brand and the answer.
The answers come from a handful of domains
Answering twenty questions, the assistants opened 4,957 links. healthline.com alone accounted for 5% of them, and the top ten domains carried most of the rest - consumerlab.com, examine-style review sites, PubMed Central, Amazon listings.
We also counted, for each domain, how often reading it ended in an answer that named a given brand. A domain read constantly with a low number beside it is a page the assistants already trust and the brand is missing from. That is a target list, and it is nothing like a backlink list.
Why we published this
ShelfSignal measures how AI assistants recommend brands in a category, and diagnoses why the assistants pick who they pick. We started with magnesium because supplements are decided on attributes an assistant can read - form, dose, third-party testing, certification - which makes the reasoning legible. Nothing in the pipeline is specific to supplements: a category is a set of buyer questions and the brands that come back.
The index is free and public, and we will publish a new category as each one is measured. If you want your own brand's position, the questions you lose, and the pages of yours the assistants read, tell us your brand and category.
Method: 20 non-branded buyer questions × 5 assistants × 5 runs = 500 answers, collected 1 September 2026 through official model APIs with web grounding enabled. Every run stored with model version, prompt, run index and timestamp. Brands are shown as measured from public AI answers; none are ShelfSignal customers or partners.