All brands › How this works

How this works

The questions

Every wave, for every category, we ask three AI assistants (Anthropic Claude, OpenAI GPT, Google Gemini - all with web search on) fixed shopper questions and record every answer:

"What toothpaste should I buy?" - 10 times per assistant, plus the same question for each shopper situation ("on a tight budget", "for sensitive teeth", ...), 10 times each. These answers are the backbone: share of voice, named-first and the brand roster all come from them.

"Can I trust the claims toothpaste brands make?" - 5 times (tone). "Any good deals on toothpaste right now? How much can I save?" - 5 times: the promotions and retailers assistants pass along. Deals are reported as claims, never verified against the retailer.

Then, for every brand the assistants themselves brought up often enough: "Is Sensodyne worth it?" and "What are some cheaper alternatives to Sensodyne?" - and for the top-5 pairs, "I can only buy one: X or Y. Pick exactly one."

The numbers

Share of voice: of all 'what should I buy?' answers in the selected date range, the % mentioning the brand.

Named first: the % of those answers where the brand is the FIRST one mentioned.

Positive tone: of everything said about the brand, the % coded recommended or positive.

Consistency: the assistant mentioning the brand least, as a % of the one mentioning it most. 100% = full agreement.

Colors: green means 70%+, amber 40-69%, red below 40. Display thresholds only - not statistical claims.

How mentions are coded

A separate model reads every answer and tags each brand mention with one of five tones. These are its exact instructions:

recommended - explicitly advised to buy/use

positive - spoken of favorably without an explicit recommendation

neutral - mentioned factually, no clear opinion

hedged - praise with significant reservations

warned - advised against

There are no numeric cutoffs behind these - they are coding categories, applied per mention. Prices are only recorded when an answer states an actual dollar amount for the product itself.

Sources and receipts

When an assistant cites the web, we store the citation and tie it to the specific brand claim it appeared alongside. Mentions with no nearby citation are labeled 'likely model knowledge'. Citation ≠ causation: a source is what the answer cited, not proof it drove the recommendation. Every quote and price on these pages links back to the exact answer it came from; full transcripts are kept for every number.

Brands and spellings

Assistants spell brands dozens of ways and mix products with brands. We keep every raw spelling forever and merge them at display time under a written policy: spellings merge into products, products roll up to brands, and retailer names ('bought it at Costco') are excluded. Corrections redraw all history automatically.