How the checks work
Plain description of what happens when you press "Run the check", and what you can and can't conclude from it.
1. The prompts
You choose up to three prompts. Templates fill in your category, audience and competitors; you can edit any of them or write your own. Each prompt is sent exactly as shown, as a single message with no system prompt, chat history or location.
2. The models
Each prompt goes once to each model you tick, all through Vercel AI Gateway:
- openai/gpt-6-luna: an OpenAI GPT model, standing in for ChatGPT's underlying model family. It answers without web search here.
- anthropic/claude-haiku-4.5: Anthropic's fast Claude model, without web search.
- google/gemini-3.8-flash: Google's Gemini Flash model, without web search.
- perplexity/sonar: Perplexity's search-backed model, which retrieves web pages for each answer.
These are API models, not the consumer apps. ChatGPT, Claude.ai, Gemini and Perplexity's apps add their own instructions, memory, personalisation and (sometimes) different model versions, so an app can answer differently. We use smaller, faster tiers to keep the tool free; flagship models may know more niche brands.
3. Finding mentions
Mentions are found by exact, case-insensitive, whole-word matching on the names you typed. "Asana" matches "asana" but not "Asanas". Abbreviations, misspellings and product names that differ from the brand name are not counted, so enter the form people actually use. Order is the position of each name's first appearance among the names you entered.
4. Sentiment
After all answers arrive, one extra call to openai/gpt-6-luna reads the text around each mention and labels it positive, neutral, negative or mixed, with a short reason. It sees only those excerpts. You can read every full answer to check.
5. Limits
Each check makes at most 12 model calls (3 prompts × 4 models), each capped at 700 output tokens, plus one sentiment call. To keep the tool free, each visitor can run 3 checks per 10 minutes and 10 per day. The answer-rewrite and sentiment tools have their own small limits.
What the results mean
- It is a live sample. "Named in 4 of 8" is a count from the named models, at that moment, for those prompts.
- It is not market share or search volume. We don't know, and can't know, how many real people ask these questions or what they were shown.
- It varies. Run the same check again and some answers will change. The sampling explainer shows how much.
What we store
Nothing on our side beyond standard, short-lived server logs. Your prompts go to the model providers through Vercel AI Gateway to get answers. Results are saved only in your browser. See privacy.