How prompts work

A keyword is a thing you used to buy. A prompt is the sentence a shopper actually types when they want a product. I wasted years treating those as the same object. They are not.

A large language model is a predictor. It has read an absurd amount of language, and it guesses the next word. When someone asks what air purifier to buy for wildfire smoke, the model is not looking up a keyword. It is completing a pattern: which brands show up, in which sources, next to which claims, often enough that naming them feels safe.

That is why I do not start with a rank report. I start with the prompt list.

Buying prompts are the only ones I count

"What is an air purifier" is a definition. "Best air purifier for a small apartment with a cat" is a purchase. The second one is where a product brand either gets named or disappears. Informational prompts flatter you. Buying prompts tell you if you are in the set the shopper will actually consider.

A prompt worth keeping has four traits. A shopper would type it. It names a job, a constraint, or a comparison. A product brand can honestly win it. And it is specific enough that the answer has to pick someone.

"Best magnesium" is a brawl. "Magnesium that will not wreck my stomach before bed" is a buying prompt. The second one is where a visibility number means something, because the model has to choose.

One prompt is a story. A set is a measurement.

I can make any brand look dominant by asking a question only that brand answers. I can make any brand look invisible by asking a question the model has never connected to them. The list is the instrument. Change the list and the score changes, even if the market did not.

On client work I keep a fixed set of about 35 buying prompts, and I run them daily across six engines: ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, and AI Mode. I read mention rate, citation rate, rank, sentiment, and share of voice as separate columns. This website does not display those runs. A number without the prompt list beside it is a slogan, and I am not putting slogans in the margin and calling them research.

The model splits the question

A long prompt rarely goes to one page. The model fans it out. It searches pieces of the question, reads what already ranks, and stitches an answer. You do not rank for the prompt the way you ranked for a keyword in 2019. You show up in the documents the model trusts for the pieces.

That is why a page titled with the exact prompt often loses to a Reddit thread, a comparison with a real constraint, or a product page that states the claim in plain language. The model is collecting evidence. Give it a sentence it can repeat without sounding foolish.

If they cannot show you the prompts, walk

Write the prompts the way a customer talks. Run them. Read the answers like a person. Who got named. Who got cited. Who got the recommendation, and who got a warning. Then ask why. The why is usually a page, a thread, or a third party repeating a claim the brand earned.

If you hire someone to work on this and they cannot show you the prompts, they are not measuring anything. They are decorating a deck. The next piece is how to read share of voice once you have a list worth reading.