Prompts, engines and cells
A prompt asked of one engine is a cell. Twenty prompts on six engines make 120 cells. Each segment has its own prompts, so a second segment adds cells of its own. A cell collects several answers a week. ChatGPT, Gemini, Google AI Overviews and Google AI Mode answer every prompt on six daily updates and once more on report day, so their cells hold up to seven answers by the weekly report and might read “Named 5 of 7 days”. Perplexity and Microsoft Copilot answer on report day and three days later, so their cells hold up to two. Claude answers once a week. A Snapshot asks each cell twice.How an answer is read
A language model reads each answer and answers a fixed set of questions about it. Each question is about one brand: does the answer name it, and how strongly does it treat it? So that the reading recognises every name buyers use for you, add product names and other spellings under Names that count as you in your brand settings. The reading also checks how the answer matches your profile and brand facts, frames your brand and competitors, and uses its cited sources. Your tracked competitors receive the same presence and strength readings. On a weekly report, Heralded also grades untracked discovered brands named in an answer using the same strength criteria. Each answer lands on one rung:
Two top choices tied for first are both Recommended, not Leading. If the model says an answer names your brand but grades its strength as absent, the answer counts as Mentioned.
The rungs and the figures share two names, so keep them apart. The Mentioned figure counts every answer that names you, from Weak up. The Recommended figure counts answers on the Leading or Recommended rung, so a Leading answer is read as Leading and still counts toward Recommended.
Which answers count
- Not measured counts for nothing. When an engine errors, returns nothing, or gives a reply the reading can’t decide even after a retry, the answer is not measured. It sits in no numerator or denominator, and Heralded never scores it as zero. Google AI Overviews is not measured more often than the other engines.
- Prompts that name your brand are left out of Mentioned, Recommended and the Heralded Score. An answer to “Is Acme a good CRM?” names Acme whatever the engine thinks of it, so it says nothing about whether buyers find you. These prompts still count toward how AI describes you and share of voice.
From answers to figures
- Mentioned and Recommended are shares of answers.
- The Heralded Score combines the two on a calibrated scale.
- Mentioned, Recommended and the score each carry a 95% range.
- A change is only called real when a paired test says the figure moved.
- Share of voice compares you with your tracked competitors.
- Perception reads what the answers say about you.
- Verifying actions is how Heralded judges whether finished work paid off.
- When a figure is empty covers the cases where there is nothing to show.