What AI search actually asks about your company
Customers no longer type two-word queries. They paste a paragraph, ask for a recommendation, and accept whatever the model returns. Here is what those prompts really look like.
If you still picture customers typing your category into Google, you are picturing 2019. The fastest-moving teams we work with are losing pipeline to a quieter pattern: a paragraph-long prompt to ChatGPT or Perplexity that ends with "give me three vendors and tell me which to pick." The model answers in confident prose, the customer clicks the top suggestion, and your team never sees the request hit the site.
The new customer prompt
We have collected several thousand real customer prompts over the last twelve months. The vast majority share three features: they are long, they are contextual, and they expect a ranked answer. A B2B operator does not ask "best CRM." They paste five sentences about their team, their stack, and their budget, then ask for a shortlist.
This is good news for companies that have a clear story and structured proof. It is bad news for everyone else, because the model has to guess, and it guesses based on whatever public surface it can reach.
Why this breaks classic SEO
Traditional SEO optimizes for the moment someone scrolls a results page. AI search compresses that moment into a single sentence inside an answer. There is no scroll, no second click, no chance to recover with a stronger headline below the fold.
The competitive surface area also changes. You are not racing ten blue links. You are racing the model's working summary of your category, written from whatever it has indexed.
- Customers see three to five options, not ten.
- The model paraphrases your value prop. If your site is vague, the paraphrase will be vaguer.
- Competitors with cleaner public proof get cited first, regardless of pipeline size.
What models actually look for
When we sample the same prompt across ChatGPT, Gemini, and Perplexity, the patterns are remarkably consistent. Models reward category clarity, a named customer, and concrete proof. They penalize generic positioning and undated case studies.
How to respond without rebuilding the site
You do not need a redesign. You need to make the highest-traffic public surfaces answer the questions models are actually being asked. We usually start with the homepage hero, the comparison page, and one or two proof artifacts.
Once those surfaces are tight, the model has something to quote. Visibility moves quickly after that, often within a sampling cycle or two.
The takeaway
AI search is not a future channel. It is already shaping shortlists for your customers. The companies that win the next twelve months are the ones who treat their public surface as model-readable evidence, not marketing copy.