From Segments to Customer Companions: A Step-by-Step AI-Powered STP Process
By Lisa Peyton and her team of AI Super Heroes
How to use Claude, ChatGPT, Gemini, and Perplexity to build an evidence-based audience strategy — and then bring each segment to life as a 24/7 AI-powered persona you can actually talk to.
I recently ran a full AI-powered STP for a client — segmentation, targeting, positioning, and live AI personas for each segment — start to finish in about an hour and fifteen minutes.
What that process replaced would have cost a market research agency months and tens of thousands of dollars. The outputs are more evidence-based, more current, and more actionable than what the old approach produced.
Here is exactly how I did it, documented step by step so you can run the same process yourself.
Start Here: Set Claude Up as Your Home Base
Before anything else, create a project in Claude and load it with your context. This project becomes your strategic operating system for the entire process. Every conversation you start from this point will have access to everything you’ve stored there.
What to put in:
- A clear description of your business, product, or campaign and its objective
- Your current positioning — how you describe yourself, what you claim to do
- Any existing data on your current customers: survey responses, intake forms, enrollment records, client notes, CRM exports. Whatever you have. The AI pattern-matches against real data, and the difference between generic output and actually useful output is almost entirely determined by the quality of context you provide.
- Your competitive landscape: the three to five alternatives your audience might consider, and a sentence on how each one positions itself
Once that context is in place, start a new chat and use this opening prompt:
“Using the STP method, I want to identify the most promising target audience for my social media campaign. Ask me any questions to help us get the best output. I am providing you with my objective, my content and competitor audit, and the context on my business.”
Notice what this prompt does not do: it does not tell Claude what segments you think you have. That comes later. Right now, you want Claude curious and open, not confirming your existing assumptions.
Phase 1: Segmentation
Let Claude Ask Questions — Then Feed It Real Data
When you run that opening prompt, Claude will come back with questions. It will ask about your current buyers, your data, who you’ve served before, what the objective actually requires. Answer them as completely as you can.
Here is the move that changes everything: once Claude asks about existing customers, go pull your actual data and bring it in. Past survey results. Roster information. Client intake responses. Post-purchase feedback. Whatever you have.
When I did this for my own business, Claude’s initial segment recommendations were fine — useful but unsurprising. Once I uploaded my actual customer data, the recommendations changed significantly. This is not a minor difference. It is the difference between AI pattern-matching on its general training versus pattern-matching on your specific reality.
If you do not have customer data, you will still get something useful. It just will not be as validated, and you should treat those segment hypotheses accordingly — as starting points to test, not conclusions to build on.
Iterate Until You Have Two or Three Strong Candidates
Go back and forth with Claude. This is not a single-prompt exercise. You are having a working session with a very fast, very well-read research partner.
Two segments may emerge quickly. A third might feel uncertain — which is exactly when to move to the next phase.
Phase 2: Research — Put Claude’s Search on Steroids
When you hit a gap in the conversation — a segment that needs more evidence, a question about market size, a disagreement between what your data shows and what Claude is recommending — do not let Claude search the web on its own. Instead, ask Claude to write you a research prompt.
This is an important nuance. Ask Claude for a prompt that will validate the work you have done so far, not one that will confirm it. Watch for the difference.
The first research prompt Claude gave me had the problem segments already baked in — it was essentially saying, here are the segments we have decided on, go find evidence for them. That is leading the witness. I pushed back, and Claude rewrote it to something much more open: Who is buying this category? How are they buying it? What are they actually looking for? That version gave the research room to challenge the assumptions, not just support them.
Take that open research prompt and run it in all four tools:
- Claude — analytical rigor, will push back on vague framing
- ChatGPT — clean, structured synthesis; reliable baseline
- Gemini — pulls in current web data and job posting trends; strongest on market size signals
- Perplexity — research aggregator, cites sources, best at surfacing things you did not know to ask about
Run the identical prompt in all four. Save every output document.
Phase 3: Cross-Reference — And Read the Reports Yourself
Bring all four research reports back to Claude with this prompt:
“Here are the research reports. Please cross-reference all four reports against each other and against the customer data to see where they converge, where they diverge, and what surprises come up. That should give us solid evidence for our final segment decision.”
Claude will map convergence, divergence, and gaps. Where all four tools agree, you have strong signal. Where only one tool surfaces something, you have something worth investigating.
And here is the part I cannot stress enough: read the reports yourself.
I know the promise of AI is that you do not have to read everything. And most of the time that is true. But for STP specifically — for the decision about who your audience actually is — you want your own eyes on the source material.
When Claude came back and told me it had not found any mention of a particular segment in the reports, I knew that was wrong. I had read the Perplexity report. That segment was in there. I pushed back directly: I am pushing back on this — I know I saw a mention of this audience in the Perplexity report. Please scan all four reports again and look specifically for it.
Claude came back and said: You are right, I missed it.
The only reason I caught that is because I had read the source material. If I had not, that segment would have disappeared from the analysis entirely, and I never would have known.
AI makes mistakes. Your judgment is the quality control. That is not a limitation of the process — it is the process working as intended.
Phase 4: Finalize Your Segments
After the cross-reference step, you will have enough to make confident final decisions. Apply three tests to every segment before locking it in:
The reality test. Is this segment actually present in your data? Not just plausible — actually evidenced by things customers or clients have told you?
The capability test. Can you genuinely serve this segment given what you offer and how you operate? A large segment you are poorly positioned to serve is a distraction.
The differentiation test. Does your positioning create meaningful separation from alternatives in the eyes of this specific segment? If they cannot feel the difference between you and a competitor, targeting them is expensive and inefficient.
Start with two to three segments maximum. Every additional segment creates a separate content track and a separate set of decisions. Two strong, well-evidenced segments will serve you better than four fuzzy ones.
Phase 5: Build Persona Prompts with Claude
Once your segments are finalized, ask Claude to build a detailed persona prompt for each one. Be specific about where you plan to use it:
“Now I want to create personas for each of these segments. Please write a persona prompt for each one that I can use in ChatGPT. Ground each prompt in both the customer data and the research findings. Make each prompt detailed enough that ChatGPT can convincingly roleplay the persona for testing content, positioning, email copy, and objection handling.”
Telling Claude which tool you plan to use is worth doing — it calibrates the output format. What comes back will be a detailed character brief for each segment: a specific name, age, location, role, motivations, pain points, goals, communication preferences, and behavioral patterns — all grounded in your actual data and the four research reports.
Phase 6: Create Your AI Customer Companions
This is the step that will change how you work.
Take each persona prompt Claude generated and bring it into ChatGPT. Create a project folder — name it something tied to the campaign or product. Create a separate conversation for each persona, paste in the prompt, and use the best model with extended thinking enabled.
When ChatGPT responds, it will step into the persona and introduce itself. Rename that conversation after the persona’s name. What you now have is a 24/7 AI-powered version of your target customer, available whenever you need it.
(I made a detailed tutorial video walking through exactly how to set this up in ChatGPT — you can watch it here: youtu.be/kvL7MoiEiyQ. I recorded it a couple of years ago. ChatGPT now has proper projects, which makes the organization cleaner, but the core method is the same.)
Now start asking questions. Test everything before it goes live.
Ask where she spends time on social media and why. Ask what would make her click on a LinkedIn post versus scroll past it. Share a piece of content you are considering and ask for her honest reaction. When I showed my persona a recent article headline, she came back with: “the hook works, but it creates a trust problem — you would need to earn my trust almost immediately.” I would not have caught that from my own perspective alone.
These personas are not decorative. They are pressure-testing tools. Use them before you publish, before you write copy, before you decide on channels. Ask them about objections. Ask them where they would realistically discover you. Ask them to react to your competitors’ messaging.
The accuracy is grounded in the depth of context you provided — your real customer data plus four deep research reports plus the model’s training on countless examples of people just like your persona. It is not perfect. It is also significantly more useful than guessing.
And because this now takes an hour instead of six months, update your personas regularly. Quarterly at minimum. The research reports you run in three months will be different. So will your personas.
Phase 7: Create the Visual Persona Card
The final step is making your persona visible to a team or stakeholder group.
Open a new ChatGPT conversation — separate from your persona companion chats. Paste in the persona prompt Claude wrote and add this instruction:
“Create a visual persona card based on this information. Include a portrait-style image of the persona, their key demographics, primary motivations, pain points, preferred channels, and a characteristic quote.”
ChatGPT will generate a PowerPoint slide with an AI-rendered portrait and all the key data points laid out — and because it comes out as a PowerPoint, the text is fully editable. If you want a flat image version, run the same prompt in Gemini, which tends to produce stronger image output, though the result will not be editable.
The Full Process at a Glance
- Load context into Claude — business description, existing customer data, positioning, competitive landscape
- Open with an STP prompt that asks Claude to ask you questions — do not lead with your existing assumptions
- Answer Claude’s questions with real data — surveys, intake forms, roster, CRM notes, whatever you have
- Iterate collaboratively until two or three segments feel solid
- Ask Claude to write an open research prompt — watch for leading the witness; the prompt should validate, not confirm
- Run that prompt in Claude, ChatGPT, Gemini, and Perplexity — save all four reports
- Bring all four reports back to Claude and ask it to cross-reference for convergence, divergence, and surprises
- Read the reports yourself — your human judgment is the quality control; AI makes mistakes
- Push back when something is wrong — if you know a segment is in the data, say so and ask Claude to look again
- Apply the three tests to finalize segments — reality, capability, differentiation
- Ask Claude to write detailed persona prompts for ChatGPT, grounded in your data and research
- Build each persona as a ChatGPT project conversation — your 24/7 AI customer companions
- Test everything against your personas before it goes live
- Generate visual persona cards in ChatGPT or Gemini
- Refresh the entire process quarterly — it takes an hour now; there is no excuse not to
The research that used to cost a market research agency months and tens of thousands of dollars now takes less time than your average strategy meeting. The outputs are not just faster — they are genuinely better, because they draw on more evidence than any single person or team could manually synthesize.
What AI cannot do: read your data with your experience, catch its own mistakes, or make the final judgment calls about which segments are worth your time. That part is still yours.
AI does the heavy lifting. You do the thinking.
Lisa Peyton is an AI marketing consultant, Fractional Chief AI Officer, and faculty member at the University of Oregon. She is the creator of the LEAD Framework and publishes the Weekly AI Marketing Brief at aimarketinginsights.beehiiv.com.
