How to Turn Customer Interviews, Sales Calls and Support Tickets Into AI Search Prompts

Why use customer language instead of writing prompts yourself?

Because your team describes the product in your terms, and buyers do not. In a SparkToro test, 142 people wrote prompts for the same need. Their prompts were almost entirely different from one another, with an average semantic similarity of 0.081. Yet the assistants recognised the shared need and kept naming a similar set of brands; what changed the answers was a different need or constraint. Customer language matters for the needs, constraints and vocabulary buyers bring. Prompts written by a marketing team tend to sound like the website, which makes the brand look better than it will for real buyers. How to spot that kind of flattering selection is covered in how to detect cherry-picked prompts in an AI search case study.

What does the pipeline look like?

1. Collect questions from tickets, chat logs, call recordings, interview notes, onboarding sessions and community posts.
2. Extract the question as the customer asked it. Keep their words, including vague or wrong terms.
3. Clean it: remove personal and company names, account details and your brand name. Buyers asking assistants usually do not know you yet.
4. Rewrite lightly into how someone would type it to an assistant, such as a full sentence with the key constraint, without adding your terminology.
5. Tag each prompt: buying stage, audience or role, topic, and any competitor mentioned.
6. Group near-duplicates, keeping the most common wording and one or two variants.
7. Rank by frequency and value: how often the question came up, and in how many or how large deals.
8. Test the set across assistants with repeated runs, and keep it fixed for trend tracking. How to choose which prompts to keep is covered in how to choose the right prompts to track for AI visibility.

What customer support tickets should become AI-search prompts for a new SaaS?

Take the tickets that a buyer might ask an assistant before or while choosing a product, not account-specific requests:

• Capability questions: "Can it do [X]?", "Does it support [format or standard]?"
• Integration and compatibility: "Does it work with [system]?", "Is there a native [platform] integration?"
• Pricing and limits: "Is there a free plan?", "What happens when I exceed [limit]?"
• Setup and migration: "How do I import from [competitor]?", "How long does setup take?"
• Errors and troubleshooting: the exact error or symptom, since people paste those into assistants.
• Security and compliance: "Is it [standard] compliant?", "Where is data stored?"

Skip requests like "reset my password" or "refund invoice 1234". A new SaaS company will have few tickets, so add trial users' questions, onboarding call notes and questions from communities where your buyers gather.

Can win-loss interviews reveal prompts where competitors are recommended instead?

Yes, if you ask about research directly. G2's 2026 survey of B2B software buyers found 82% had sourced software recommendations from an AI chatbot in the previous two years. Add these questions to win-loss interviews:

• "How did you research options? Did you use ChatGPT, Gemini, Perplexity or another assistant?"
• "What did you ask it? Can you remember or find the wording?"
• "Which vendors came up, and what did it say about them?"
• "Did anything it said about us surprise you, or turn out to be wrong?"

Lost deals are the most valuable. Each one can give you a prompt where a competitor was recommended, along with the reason the assistant gave. Test those prompts, tag them with the competitor, and track them. Once you have those prompts, the first checks to run are listed in what to check first when AI recommends a competitor over your brand.

How do I turn Gong sales-call questions into an AEO prompt set?

Conversation tools such as Gong record and transcribe calls and let you find recurring themes. Gong's keyword trackers flag specific words and phrases in calls and emails. Its AI trackers look for concepts rather than exact words, and training a custom example-based tracker needs at least 500 recorded calls. To build a prompt set:

1. Set up trackers for question patterns: "do you integrate with", "how do you compare to", "what does it cost", "is it secure", "how long does implementation take".
2. Track competitor names, so you can see which comparisons come up and at which stage.
3. Export the matching snippets for a period, such as the last quarter.
4. Extract the buyer's question from each snippet, then clean, tag and group it as above.
5. Weight by deal stage and outcome. Questions raised in late-stage or lost deals deserve priority.
6. Refresh the set each quarter, keeping the core fixed and adding new questions as a separate group.

How to use sales call transcripts to find keywords for LLM content

In AI search, "keywords" are less about single terms and more about the phrases, constraints and entities buyers use. From transcripts, extract:

• Problem statements in buyers' words, such as "our schedulers spend Fridays fixing double bookings".
• Constraints and qualifiers, such as team size, industry, regulation, budget and existing tools.
• Job titles and roles involved in the decision.
• Competitor and alternative names, including spreadsheets and in-house tools.
• Objections, such as price, switching cost and security.

Compare that vocabulary with your website. Where buyers say "double bookings" and your site says "resource conflict management", your pages miss the words buyers use. Write pages that answer the exact questions, use buyers' terms in headings and first lines, and address the objections directly. Why pages with unique information win is covered in how information gain improves AI search visibility.

How should you handle privacy?

Transcripts and tickets contain personal data. Follow your consent and data-handling policies for call recording and analysis. Strip names, companies and account details before prompts leave your systems, and do not paste raw transcripts into public AI tools. Keep the prompt library free of anything that identifies a customer.

What does a prompt library entry look like?

What do people ask about building prompts from customer data?

How many prompts should come from customer sources?

Most of them. Supplement with category questions from search data, but let real questions dominate.

Should we include our brand in prompts?

Keep branded prompts in a separate group. Unbranded prompts show whether assistants recommend you to people who do not know you yet.

How often should the set change?

Keep a fixed core for trends, and review additions quarterly as new questions appear in calls and tickets.