How AI brand recommendations are measured

Layer 1: Category discovery

When a consumer asks "what brands should I consider for X", "best running shoes for flat feet", or "what skincare brand do dermatologists trust", the AI returns a set of brands. This is the recommendation moment. PRI measures the share each brand wins at this layer.

Layer 2: Purchase discovery

Once a brand is chosen, the consumer often returns to AI to ask where to buy, which retailer is cheapest, or whether the product is in stock. PRI does not measure this. Purchase discovery is downstream and depends on retailer infrastructure, not brand recommendation infrastructure.

Each PRI run queries four AI assistants, all with web search enabled:

• ChatGPT, model gpt-4o
• Claude, model claude-sonnet-4-6
• Gemini, model gemini-2.0-flash
• Perplexity, model sonar

These four account for the overwhelming majority of consumer AI usage in the markets we cover. Web search is enabled on every run because that matches how consumers actually use these tools. API-only responses without web search measure something consumers never experience and would produce misleading recommendations.

For each sub-category in each market, AI See You constructs a structured prompt set. Prompts cover the natural shapes a consumer query takes:

• Category discovery prompts ("what are the best brands for X")
• Comparison prompts ("brand A versus brand B for Y")
• Use-case prompts ("X for Y audience" or "X for Y goal")
• Trust prompts ("which brands of X do experts recommend")
• Ingredient or feature prompts where the category warrants them

Prompts are written in natural consumer language, in the local market's English variant where applicable. The full prompt list for each sub-category is published on that sub-category's ranking page. We do this so that anyone, including the brands ranked, can see exactly what was asked.

Prompts split into two construction types. Non-branded prompts: the consumer does not name a brand. The AI returns whichever brands it considers most relevant. This is the cleanest measurement of category authority. Branded prompts: the consumer names a specific brand or compares two named brands. This measures whether AI engages with the brand when prompted by name.

Both types are run. Non-branded share is the primary signal and is the headline number on each sub-category page. Branded outcomes inform brand profiles and are reported separately.

A new canonical PRI run is produced every month. Each run is timestamped and archived. Previous months remain available at archive URLs of the form /categories/{category}/{country}/{sub-category}/{yyyy}/{mm}/.

Movement between months is shown on each sub-category page. Brands moving up are flagged with the rank change. Brands moving down are flagged likewise. Brands new to the sub-category since the previous canonical run are marked NEW.

A canonical run is the version of a month's data that has been operator-reviewed and locked. Operator review checks for prompt failures, parsing errors, brand entity collisions, and obvious anomalies. The locked canonical run is what the public site reflects.

A brand mention counts when the AI explicitly names the brand in a recommendation context within the response to a sub-category prompt. We do not count:

• Brand mentions in disclaimers, source citations, or footnotes
• Brand mentions in negative or cautionary contexts
• Mentions of a parent company where the consumer asked about a sub-brand
• Mentions in unrelated tangents within a longer response

Each unique brand mention within a single response is counted once per response. A brand recommended in five out of ten responses to non-branded prompts has a base mention rate of 50 percent. Share within the sub-category is then calculated against the total mentions of all brands across all non-branded prompts in that run.

Each sub-category page reflects a single market. The market is shown in the page title, breadcrumb, and metadata. Common markets currently covered: Australia, United States, United Kingdom. Additional markets appear as data accumulates.

Prompts are run in the language and English variant appropriate to the market: Australian English for AU, American English for US, British English for UK. Recommendation results sometimes differ by market because AI assistants tailor responses to the user's region. A brand that ranks first in AU may rank fifth in US, and that distinction is preserved on each market's page.

A brand is matched across runs by its canonical identifier (a stable internal slug), not by the display name. If AI returns "Nike" in one response and "Nike, Inc" in another, both are credited to the same brand entity. We maintain a brand profile that records the canonical name, common variants, parent company, country of origin, and competitive context.

Brand profiles are publicly visible at /brands/{brand-slug}/. New brand entities introduced after a previous canonical run are flagged NEW on the sub-category page where they first appear.

PRI stands for Product Recommendation Index. It is AI See You's public index of how often product brands are recommended by ChatGPT, Claude, Gemini, and Perplexity when consumers ask category and comparison questions.

For each sub-category, we run a structured prompt set across four AI assistants and count every brand mention in a recommendation context. A brand's share is the percentage of total brand mentions in that sub-category that go to that brand. Shares within a sub-category sum to 100 percent.

ChatGPT (gpt-4o), Claude (claude-sonnet-4-6), Gemini (gemini-2.0-flash), and Perplexity (sonar). All are run with web search enabled.