How Premonition works
The whole method, including what it cannot tell you.
1. Profiling
We fetch your homepage and read the title, meta description and visible copy. A language model turns that into a narrow category (not “software” but “AI meeting notetaker for sales teams”), a one-line description, and a buyer profile.
Getting the category narrow matters more than anything else here. “Best CRM” and “best CRM for independent insurance brokers” return completely different vendor lists.
2. Prompt generation
We generate purchase-intent questions across six intents: best-of shortlists, alternatives to a named competitor, head-to-head comparisons, use-case specific queries, pricing and budget queries, and problem-first queries.
Your brand name never appears in a prompt. The whole point is to test whether the engines name you unprompted — priming them would invalidate the result.
3. Querying the engines
Each question goes to Perplexity Sonar and a web-connected GPT-4o. The full audit adds Google Gemini with live search. Every engine answers with live web access, so results reflect what a buyer would see that day.
We send a neutral instruction asking for specific named vendors ranked best first — the same shape as a real buyer's question. We do not tell the engine what we are testing.
4. Extraction and scoring
Each answer is parsed into an ordered list of vendors. We match your brand against its name, common aliases and domain, and cross-check the raw answer text so a mention is never missed because the extractor normalised a name oddly.
The score is the share of engine-answers that named you, weighted by rank: first place counts full, top-3 counts 0.8, top-5 counts 0.55, anything lower counts 0.3. A score of 100 means every engine named you first for every question.
5. Limitations, stated plainly
AI answers are non-deterministic. The same question can return a different vendor list an hour later. Treat the score as a directional signal, not a precise measurement — that is why the full report runs 24 questions across 4 assistants rather than a handful.
We query Perplexity, web-connected GPT-4o and Gemini. We do not query the ChatGPT consumer product directly, and results there may differ. Personalisation, memory and geography all shift what a given buyer sees.
Every verbatim answer is recorded in the full audit. If you disagree with a result, you can read exactly what the engine said and judge for yourself.