Methodology

AI snapshots

ChatGPT and Gemini receive the same prompt: describe perceived brand associations in short phrases, avoid marketing copy, include unflattering associations and assign a 1–100 strength weight. Web search is disabled. We store model, provider, timestamp, prompt version, raw response and normalized tags.

Current human responses

Visitors voluntarily add one word or short phrase per brand collection wave. Counts describe this participating panel only. They are not weighted to represent a country, demographic or all customers.

Historical archive

Recovered BrandTags pages from 2008–2012 preserve raw association text and capture dates. The archived cloud view ranked tags by how many people submitted them, so what survives is that ordering, not the tallies — we publish it as a ranking and size words by it. Google is the exception: its cloud view sits behind the old tagging gate, so its ordering is reconstructed from phrase frequency and is a weaker signal. Archive rankings never enter current human totals.

Comparisons and rankings

Initial comparisons use case-and-whitespace-normalized exact matches. Model overlap is Jaccard similarity among top tag sets. Semantic clusters are only merged after review; the site does not silently decide that two different phrases mean the same thing.

Publication quality gate

• Two dated model snapshots shown independently.
• Transparent prompt and source labels.
• A substantive archive or current human layer.
• No placeholder counts, reviews or offers.
• Editorial review for high-traffic profiles.