By Albrecht Lippky

Updated
11 min read · Markdown

MethodsMethodFirst-party data

How do you turn sales calls and support tickets into prompts?

The strongest prompt source many teams already own is the customer conversation: sales calls, demo notes, support tickets, chat logs and reviews.

An account executive in a headset explains something mid-call while taking notes on a paper pad
How do you turn sales calls and support tickets into prompts? | Image: Prompt Research Lab

TL;DR

  • Customer conversations are strong sources of prompt candidates because they reveal real needs, constraints and decision language, but they do not automatically measure AI-search demand.[1]
  • Use a six-step workflow: export, de-identify, extract, preserve provenance, rewrite, then group and review the prompts.[1]
  • Keep the original customer question beside every standalone rewrite, and never present an editorial variant as a verbatim quote.[3]
  • Sales calls, support tickets and reviews are biased toward particular customer groups, while qualitative saturation does not prove market representativeness.[2]
  • There is no universal ideal prompt length; preserve the buyer’s task and context rather than padding questions to meet a word target.[1]
Contents
  1. What can customer conversations reveal about AI prompts?
  2. Which conversation sources should you mine?
  3. How should you extract customer questions?
  4. How do you remove personal data before using transcripts with an LLM?
  5. How do you rewrite a customer question into AI prompt variants?
  6. How should you group prompts by persona and buying stage?
  7. What are the main biases in voice-of-customer prompt research?
  8. How should teams validate prompts after extraction?
  9. What should the final workflow look like?
  10. FAQ
  11. Sources

The strongest prompt source many teams already own is the customer conversation: sales calls, demo notes, support tickets, chat logs and reviews. These sources reveal the language, constraints and decision contexts customers actually use—but they should produce traceable prompt candidates, not be mistaken for a representative census of what everyone asks an AI system.[1]

The practical method is to export conversations, remove unnecessary personal data, isolate customer questions, preserve their original wording, rewrite each distinct intent as a standalone prompt, and then group the results by persona and buying stage.[1]

Methodology note: This article treats a “prompt” as a full-sentence question or instruction that a buyer could send to ChatGPT, Gemini or another AI assistant. Customer wording remains the evidence; the rewritten prompt is an editorial transformation.

What can customer conversations reveal about AI prompts?

Customer conversations surface questions that keyword tools and generic prompt generators often miss:

  • The problem a buyer is trying to solve.
  • The constraints they consider non-negotiable.
  • The comparisons they make between products or approaches.
  • The objections that delay a purchase.
  • The implementation questions that appear after conversion.
  • The language different personas use for the same underlying need.

Sales and support conversations are established inputs for prompt and AI-search research, alongside interviews, reviews, forums and on-site search.[2] The important distinction is between observed customer language and observed AI-search demand. A question extracted from a call is evidence that one customer expressed that need; it does not prove that the same wording is frequently sent to ChatGPT or Gemini.[1]

A useful dataset therefore keeps at least two fields:

  • original_question: the source wording or carefully marked paraphrase.
  • rewritten_prompt: the standalone version created for research or testing.

ReachLLM’s documentation makes a similar distinction: customer questions can be entered from sales calls and support conversations, but a generated prompt does not automatically inherit search counts from a related keyword.[3]

Which conversation sources should you mine?

Source What questions it surfaces How to export Bias to watch
Call recordings and transcripts Discovery questions, objections, product comparisons, budget concerns and pain points Download transcripts or retrieve them through the platform API; Gong documents both call-page downloads and date- or call-ID-based transcript retrieval.[1] Sales calls overrepresent active opportunities and may contain seller-led questions. Prospect-only question export is not established in the cited documentation.[1]
Demo notes Questions about features, use cases, integrations, implementation and proof of value Export CRM or sales-enablement notes, then label paraphrases separately from verbatim questions Notes may reflect the salesperson’s interpretation rather than the buyer’s exact wording.[1]
Support tickets Troubleshooting questions, setup problems, failure modes and post-purchase information needs Export ticket text with its issue category and relevant surrounding context Existing customers dominate the sample; prospects and non-buyers are absent.[2]
Chat logs Short, immediate questions, follow-ups, comparisons and points of confusion Export enough neighboring turns to resolve references such as “that” or “the other plan” Isolated messages can lose context, and automated chat may shape the customer’s wording.[1][4]
Reviews Objections, desired outcomes, comparisons, disappointments and reasons for choosing a product Export review text with date, product and platform provenance Reviews are self-selected; early buyers can have different preferences from later consumers.[5]
Microsoft Teams meetings Meeting questions and follow-ups tied to a transcript and chat Organizers or co-organizers can download transcripts as DOCX or VTT, subject to permissions and policy.[6] Meeting transcripts are not automatically a cross-customer question dataset.[6]
Zoom meetings Questions connected to cloud-recording transcripts or live meeting transcripts Cloud transcripts can be downloaded as VTT, while live meeting transcripts can be saved as TXT.[7] AI Companion can answer questions about meeting content, but it is not documented as a prospect-question classifier.[8][1]

The minimum useful unit is not a detached sentence. It is a question plus enough context to understand the reference, its source, its date and the speaker’s role.

How should you extract customer questions?

Use this six-step method.

  1. Export the source material before classifying it.
    Retrieve transcripts, notes, tickets, chats and reviews into a controlled working dataset. Gong, Zoom and Teams document transcript retrieval routes, while Chorus documentation describes retrieving conversation metadata and utterances through its API.[1][9]

  2. Preserve the original wording and provenance.
    Store the exact question where available, or label a reconstructed version as a paraphrase. Keep a source ID, source type, date, account or case identifier, speaker role and a pointer to the original passage.[9]

  3. Remove unnecessary personal and confidential information.
    Delete names, email addresses, phone numbers, addresses, account IDs, order numbers, medical details, credentials, contractual terms and other information not needed to understand the intent. Pseudonymised information can remain personal data when a person could still be re-identified.[10][11]

  4. Separate customer questions from employee questions.
    A seller’s discovery question may reveal useful research themes, but it should not be presented as customer language. Likewise, separate the customer’s problem from the support agent’s answer so the rewritten prompt reflects the need rather than the solution already supplied.[1]

  5. Rewrite each distinct intent as a standalone prompt.
    Resolve ambiguous references using nearby context, but do not add requirements the customer did not express. Keep the original question beside the rewrite so an editor can audit every change.[10][3]

  6. Tag, deduplicate and review.
    Assign persona and buying stage only when the source supports them. Group equivalent needs, but retain separate prompts when constraints, use cases or decisions differ. A human reviewer should approve the final wording before it enters an AI-visibility test set.[12][13]

A practical record might contain:

source_id, source_type, date, deidentified_question, context, rewritten_prompt, persona, stage, theme, and review_status.

This structure separates what the customer said from what the research team inferred.

How do you remove personal data before using transcripts with an LLM?

Privacy checklist before sending a transcript excerpt to an AI model
  1. Copy only the smallest excerpt needed to understand the question.
  2. Remove direct identifiers such as names, email addresses, phone numbers, addresses, dates of birth and identification numbers.[14]
  3. Remove account numbers, order IDs, URLs containing personal data, credentials, private contractual details and sensitive case information.
  4. Generalise details that could identify someone indirectly, such as a rare job title, exact location, unusual employer or highly specific event.
  5. Replace retained context with neutral labels such as [customer], [industry], [region] or [product category].
  6. Keep any re-identification key separate, access-restricted and unavailable to the external model.
  7. Check the provider’s retention, reuse, subprocessors, deletion and transfer terms.
  8. Record the purpose and lawful basis for repurposing the conversation from sales or support into prompt research.[10]
  9. Escalate high-risk processing for privacy review or a DPIA where required.[11]
  10. Have a reviewer confirm that the excerpt cannot be reasonably re-identified from the remaining context.

Removing names is not the same as making a transcript anonymous. The ICO describes pseudonymisation as a process that replaces, removes or transforms identifiers while retaining a possibility of reconnecting the record to a person; the resulting data remains personal data under its guidance.[10]

The ICO also says that recording permission is not the entire privacy analysis when a third party transcribes or processes calls. Customers should understand who processes the information, why it is processed and what relevant rights apply.[11][10] A supplier that uses the material for its own purpose, such as training another model, may have a different role from a supplier acting only on the organization’s instructions.[11]

The EDPB’s 2026 anonymisation consultation adds an important EU nuance: information may be anonymous for one entity but not another, and technical, organizational or contractual restrictions do not automatically establish anonymity.[15] The reviewed EDPB page describes the guidance as adopted for public consultation, not as a separate final exemption for sending pseudonymised transcripts to AI services.[15][10]

How do you rewrite a customer question into AI prompt variants?

Start with the customer’s task, not with a preferred question template. Recent research reports different average lengths for different AI tasks: Semrush found 8.7 words for search-enabled ChatGPT prompts and 13.5 words for non-search prompts in January–February 2026.[12] A separate Semrush analysis found average queries of 7.22 words in Google AI Mode versus 4.0 words in traditional Google Search, based on U.S. desktop search sessions.[16][12]

These figures do not establish a universal ideal length, and they do not show that Gemini follows the same pattern. The useful principle is to preserve the customer’s objective, constraints and context instead of padding every prompt to a target word count.[1][3]

Illustrative worked example

The following call question is illustrative, not a verbatim customer quote or an observed AI query:

“We have a small content team. How can we monitor whether ChatGPT is recommending our competitors without adding another complicated SEO platform?”

The underlying intent contains three elements:

  • The team is small.
  • It wants to monitor competitor visibility in ChatGPT.
  • It wants a simple workflow rather than a broad SEO suite.

Possible prompt variants:

  1. Discovery variant:
    “How can a small marketing team monitor whether ChatGPT recommends competitors instead of its brand?”

  2. Comparison variant:
    “Which tools help a small marketing team track brand and competitor mentions in ChatGPT without adopting a full SEO platform?”

  3. Implementation variant:
    “What is a simple way for a small content team to measure competitor visibility in ChatGPT and turn the findings into content actions?”

These are three research prompts, not three claims that the customer used all three phrasings. The source question should remain attached to every variant, and each variant should be tested separately rather than treated as having inherited demand from the original wording.[3]

How should you group prompts by persona and buying stage?

Persona and stage are useful metadata, but they should be evidence-based. A job title in a CRM record may support a persona label; a call topic may support a buying-stage label. If the source does not establish either attribute, use “unknown” or “not recorded” rather than inventing a profile.[17][3]

A practical grouping model is:

  • Problem owner — discovery: questions about a problem, risk or desired outcome.
  • Evaluator — consideration: questions about features, alternatives, proof and fit.
  • Procurement stakeholder — decision: questions about price, contracts, implementation risk and approval.
  • Administrator — post-purchase: questions about setup, workflows, troubleshooting and maintenance.

Search Engine Land’s funnel guidance identifies discovery calls, demos and procurement conversations as useful inputs for mapping prompts to the buying process.[2] It also recommends organizing by persona, intent, use case, geography, industry, product and branded or non-branded status.[2]

Do not let the grouping overwrite the source. A support ticket asking how to configure an integration is usually a post-purchase implementation question. It should not silently become a generic acquisition prompt simply because the product team wants more commercial content.

What are the main biases in voice-of-customer prompt research?

Customer conversations are valuable precisely because they are situated. That also limits what they can prove.

  • Existing-customer bias: Support tickets describe people who bought, activated or contacted support. They cannot represent people who never purchased or never raised a ticket.
  • Sales-pipeline bias: Sales calls overrepresent active opportunities and may reflect the salesperson’s agenda.
  • Small-sample bias: A narrow set of accounts can create the impression that a theme is widespread when it is only common within one segment.
  • Politeness bias: Buyers may soften criticism when speaking directly to the vendor. Clozd recommends independent win-loss interviewing as one possible countermeasure, but this is vendor guidance rather than a published effect-size estimate.[18][11]
  • Review self-selection: People who leave reviews may differ from people who do not. Research on online product reviews describes how early buyers’ preferences can differ from later consumers’ preferences.[5]
  • Frequency confusion: Ten similar tickets may show repeated support friction, but they do not show ten independent buyers searching for the same prompt.
  • AI-demand confusion: Customer wording is not automatically a measured ChatGPT or Gemini query. Exact-prompt demand may remain unknown.[3]

Qualitative saturation should not be turned into a quota. A systematic review of 23 articles found that empirical studies often reached saturation within 9–17 interviews for homogeneous populations and narrow objectives, but broader designs and more complex populations required different sample sizes.[19] That finding helps plan qualitative work; it does not show that a small group of customers represents the whole market.

A stronger sampling plan keeps groups separate:

  1. Current customers.
  2. Lost opportunities.
  3. No-decision opportunities.
  4. Churned customers.
  5. Prospects who never became opportunities.
  6. Independent or non-customer interviews.
  7. Reviews and public discussions.

Compare their themes rather than blending them into one undifferentiated frequency table.

How should teams validate prompts after extraction?

Validation should happen at three levels.

1. Source validation

Check that the rewrite preserves the customer’s actual task and does not add unsupported assumptions. Confirm whether the question was verbatim, lightly edited or reconstructed from several turns.

2. Research validation

Test whether the prompt is natural, distinct and useful for the intended persona and stage. Remove duplicates that differ only cosmetically, but retain meaningful differences in geography, budget, compliance, use case or product constraints.[1]

3. AI-visibility validation

Run the prompt in the target engines and record the date, market, language, answer, cited sources, brand mentions and competitors. Treat the result as a dated observation: AI answers and source selection can vary by engine, location and time. Finseo’s own analysis found that Google AI Mode and AI Overviews share only 13.7% of cited URLs on the same query, so even two Google surfaces are worth testing separately.

Keep these measurements separate:

  • Number of source conversations.
  • Number of distinct accounts.
  • Number of distinct customer intents.
  • Number of rewritten prompts.
  • Number of prompts tested in each AI engine.
  • Visibility or citation results for those tested prompts.

This prevents a common reporting error: presenting the number of extracted questions as if it were the number of people asking the same thing in an AI system.

What should the final workflow look like?

The defensible chain is:

lawfully process → de-identify → extract → preserve provenance → rewrite → segment → review → test

Conversation platforms help with retrieval and search. Gong documents transcript downloads, transcript API retrieval, phrase search and an assistant that can answer questions about call context.[20][1] Chorus documents meeting-term search, natural-language AI search and API access subject to permissions.[1] Zoom and Teams provide transcript retrieval and AI-assisted questions about meeting content.[4][21]

None of those capabilities, on the cited evidence, establishes a universal automatic export of only prospect questions. Use the tools to locate evidence, then apply a separate extraction and editorial review process.[1][11]

The result should be described precisely: a privacy-reviewed, traceable set of prompt candidates derived from customer language. It is not a representative map of private AI conversations, a universal prompt-length formula or proof of future visibility improvement.[22][3]

Frequently asked questions

Are sales calls a good source of prompts for ChatGPT and Gemini?
Yes. Sales calls reveal discovery questions, objections, comparisons, constraints and buying-stage language that may not appear in keyword data.[1] They are still a biased source because active opportunities and seller-led conversations are overrepresented, so preserve provenance and compare them with support, lost-deal and non-customer sources.
How do I turn a support ticket into an AI prompt?
Separate the customer’s problem from the support agent’s response, retain enough context to resolve ambiguous references, and rewrite the need as a standalone sentence.[1] Keep the original ticket wording beside the rewritten prompt so reviewers can verify that the rewrite did not add unsupported requirements.
Should customer questions be copied word for word?
Preserve the original wording wherever possible, but create a separate standalone rewrite when the original depends on missing context or contains personal data.[3] Label paraphrases and redactions clearly; never present an editorial rewrite as a verbatim customer quote.
How long should a prompt from a sales call be?
There is no universal target length. Semrush reports different averages for search-enabled and non-search ChatGPT prompts, while its Google comparison also shows different wording lengths in AI Mode and traditional Search.[16][12] Preserve the buyer’s task and constraints instead of padding the prompt to reach a word count.
Can I upload raw sales transcripts to an LLM?
Do not assume that recording permission or removing names makes raw transcripts safe to upload. The ICO recommends considering purpose, lawful basis, third-party processing, supplier roles, retention, security and reuse, while pseudonymised data can remain personal data.[11]
How many customer conversations are enough for prompt research?
There is no universal quota. A systematic review found saturation often within 9–17 interviews for homogeneous populations and narrow objectives, but that does not establish market representativeness or cover absent customer groups.[19] Track which personas, stages and sampling groups are represented instead of treating a conversation count as proof of completeness.

Sources

  1. [1]GEO Research Inputs From Support Tickets and Sales Calls — lseo.com
  2. [2]How to map AI search prompts to every stage of the sales funnel — searchengineland.com
  3. [3]Prompt Research | ReachLLM Docs — docs.reachllm.com
  4. [4]Prompt research: The next layer of SEO and GEO strategy — searchengineland.com
  5. [5]Self-Selection and Information Role of Online Product Reviews | Information Systems Research — pubsonline.informs.org
  6. [6]Start, stop, and download live transcripts in Microsoft Teams meetings | Microsoft Support — support.microsoft.com
  7. [7]Using Meeting transcripts — support.zoom.com
  8. [8]Asking in-meeting questions with AI — support.zoom.com
  9. [9]Migrating Chorus recordings to Zoom Revenue Accelerator - Zra - Zoom Developer Docs — developers.zoom.us
  10. [10]Pseudonymisation | ICO — ico.org.uk
  11. [11]Previously asked questions | ICO — ico.org.uk
  12. [12]ChatGPT traffic analysis: Insights from 17 months of clickstream data — semrush.com
  13. [13]Voice Of The Customer Surveys – Useful Or Fundamentally Flawed? — Greenbook — greenbook.org
  14. [14]Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models — edpb.europa.eu
  15. [15]Guidelines 02/2026 on Anonymisation — edpb.europa.eu
  16. [16]Google AI Mode’s Early Adoption and SEO Impact — semrush.com
  17. [17]Using smart recording with AI — support.zoom.com
  18. [18]Win-Loss Analysis Best Practices for a Successful Program | Clozd — clozd.com
  19. [19]Sample sizes for saturation in qualitative research: A systematic review of empirical tests - PubMed — pubmed.ncbi.nlm.nih.gov
  20. [20]Use Gong Assistant On A Call — help.gong.io
  21. [21]Use Copilot in Microsoft Teams meetings - Microsoft Support — support.microsoft.com
  22. [22]Gong on Gong: How we expedite the call review process with Ask Anything — gong.io