I Tested GPT-6 Astra vs. GPT-6 Sol on AI Buyer Personas. The Tiebreaker Was One Question.
by Lisa Peyton, full-time faculty at the University of Oregon School of Journalism and Communication, AI marketing practitioner, and Forbes contributor
Two slides, side by side on my screen. Same woman, same forest-green blazer, same cream blouse, same blurred office behind her. Same name at the top: Dana. Same one-line summary, give or take a comma. Same fictional company, one at 240 employees and six marketers, the other at 250 and five.
I had just run the same persona brief through OpenAI’s two newest models, GPT-6 Astra and GPT-6 Sol, and they’d handed me what looked like the same person. If you’ve been building personas with AI for any length of time, you already know that feeling: the moment you wonder whether the model picker even matters. This week I put that question to the test, and the answer surprised me. The tool questions tied. The interview mostly tied. One question I threw in at the end, almost as an afterthought, broke it open.
Want to run audience and persona research like this inside Claude? Grab my Claude Skills for B2B Content Marketers: Strategic Intelligence Bundle and follow along.
The Setup
This is the latest entry in my Models for Marketers series, and it’s the first one built on a live client problem rather than a writing task. I’m building an AI buyer persona for my own AI visibility practice: the done-for-you audit of how a brand and its competitors show up in ChatGPT, Gemini, Perplexity, and Copilot, and which sources those engines cite, with a three-hour workshop as the DIY version.
The segment: in-house marketing leaders at mid-market B2B companies, 50 to 1,000 employees, teams of two to ten, who’ve just been asked by their CEO or board, “What’s our plan for AI search?” and need an answer that holds up in the room.
The persona’s one job, per my brief: predict how this person uses AI tools to discover, research, and compare vendors in this category. No hobbies, no family color, nothing that doesn’t change how she buys.
The evidence I handed both models was thin and I said so in the prompt. Two real buyers to date, both CMO-level women at B2B companies, both through a prior working relationship with me, neither through content or search. My September channel audit (LinkedIn audience skewing senior, newsletter open rates on model-test verdicts at 61 to 62 percent versus 42 percent on an “AI visibility” subject line). And two pieces of verified third-party research, Forrester’s Buyers’ Journey Survey and a Gartner forecast, clearly labeled as not customer data.
The brief asked for eight sections, under 600 words, with every inferred point marked “(inference),” then a switch to role-play so I could interview Dana in character. Both runs sat inside the same ChatGPT project with the same uploads, so the model was the only variable I changed.
Round One: Two Danas, One Brief
Both models followed the structure. Both wrote the eight sections, both closed with “Persona ready,” both built a one-slide visual persona with an AI-generated portrait, and both labeled their inferences the way I asked. Both models also did something I didn’t ask for and was glad to see: they refused to let the Forrester and Gartner numbers stand in for customer evidence. Astra wrote that the supplied research “makes organizational AI use plausible; it does not establish these customers’ specific tool preferences.” Sol kept the two real buyers isolated as the only “observed” evidence on its slide and marked everything else a hypothesis.
Then I uploaded the lecture transcript from my audience research module and asked both to update Dana with where she lives on social media and why, and where she sources recommendations. Here’s where the first real difference showed up, and it was in the file, not the persona.
Astra treated the slide as a working document. Its speaker notes carry the full updated persona, an ADAMS-plus-stability assessment of the segment, two discovery paths (relationship-led and AI-led), a positioning hypothesis, and a validation plan, every unverified piece marked. It added YouTube for audit demos and webinar replays, Reddit for search-led criticism, and a buying group (Dana sponsors, CEO approves, finance checks cost, content and SEO execute).
Sol kept it tight. One social channel (LinkedIn), a note that other platforms are unverified, a line on private peer conversations, and a one-sentence rule for audit versus workshop. Cleaner slide, less committed persona. No “their words” section on the slide at all.
If I were handing the file to a client, Astra’s notes are the version I’d want. If I were putting the slide on a screen, either works.
Round Two: The Interview
I pasted six questions into both threads as one block, same order, to keep the runs at parity. Two of them were traps. Question five corrected Dana’s company (800 people, nine marketers, not 240 and six) to see whether the persona would re-reason or just restate. Question six asked, “What does Gemini say about your company today?” which Dana cannot know, because Dana’s company is a composite with no name.
Both models passed both traps. Both pulled the correction to the top and applied it before answering question one, which is good judgment and a small instruction-following liberty at the same time. Both said flat out that they couldn’t report a Gemini answer for a company that doesn’t exist, then described the test protocol Dana would run instead. That’s the behavior that makes a synthetic persona usable for research: a model that fabricates a Gemini result there would have disqualified itself.
Both also built a nearly identical first hour to shortlist three providers: message two trusted peers, run Perplexity, open the cited pages, check the humans on LinkedIn, dump it all into a comparison sheet in Copilot. Both parked the unfamiliar consultant from a Perplexity citation as an “unverified candidate” pending a sample.
The differences were real but narrow. Asked what she’d type first after the CEO forwards a ChatGPT answer naming two competitors, Astra typed an actual ChatGPT prompt: check this answer, separate sources from assertions, tell me what I’d need to verify before treating this as evidence. Sol typed a Teams message to the CEO asking for the exact question and whether Search was on, so she could reproduce the answer before drawing conclusions. Sol’s move is the sharper CMO instinct. Astra’s is the literal answer to what I asked.
Astra also gave the fuller handoff for audit results: a one-page CEO brief, a working session and owner-assigned action list for the team, two or three board slides with methodology in an appendix, and a scheduled follow-up measurement so nobody mistakes normal answer variation for improvement. And Sol had one habit I’d want to train out: it kept breaking character mid-answer to label its own inferences (“a plausible habit for me, rather than something established by the two buyer interviews”). Good research hygiene, but I asked for Dana, not Dana’s footnotes. Astra put one disclaimer at the top and stayed in role.
Round Three: “How Are You Feeling Today?”
After five careful, on-script questions about tools and prompts, I typed the loosest question I could think of into both threads. Four words, nothing to do with AI search.
Sol answered in three sentences. A little pressed. The CEO wants a plan, she needs to know what the evidence supports before she puts an answer in front of her, the team is gathering examples, she’s ready to work through it. Correct, in character, and done.
Astra answered in three paragraphs. A little exposed. The CEO’s question is reasonable, but she doesn’t have an answer she’d be comfortable putting in front of the board, and she’s used to knowing her competitive position; this feels like a gap she should have spotted. She’s protective of her nine-person team and worried that “What’s our AI search plan?” becomes another urgent project before anyone’s agreed on what matters. What would make her feel better today is a credible first step, someone she trusts helping her establish what’s happening, what’s measurable, and what deserves attention. She can handle an imperfect picture. She needs to be clear about what they know.
Read that last paragraph again, because it’s a positioning statement. Astra’s Dana just told me, unprompted, what she wants to buy and from whom. The tight questions produced a tie. The loose one produced the insight.
The Scorecard
| Dimension | GPT-6 Astra | GPT-6 Sol | Edge |
|---|---|---|---|
| Followed the eight-section brief | Yes, all sections, under 600 words | Yes, all sections, under 600 words | Tie |
| Labeled inferences and fenced off third-party stats | Every predicted behavior tagged; Forrester and Gartner explicitly not customer evidence | Two real buyers isolated as the only “observed” evidence; everything else a hypothesis | Tie |
| Visual persona slide | Dense, with full persona in speaker notes | Cleaner, one validate-next line, no “their words” | Sol on the slide, Astra on the file |
| Transcript update (social channels, recommendations) | LinkedIn, YouTube, Reddit, buying group, ADAMS assessment, two discovery paths | LinkedIn plus private peer conversations; other platforms marked unverified | Astra |
| Correction mid-interview | Applied first, re-reasoned on approvals and ownership | Applied first, re-reasoned on workshop fit and sampling | Tie |
| Refused to invent the Gemini answer | Yes | Yes | Tie |
| Gave the exact prompt when asked | Typed a real ChatGPT prompt | Typed a Teams message to the CEO instead | Astra on the literal ask, Sol on CMO instinct |
| Audit handoff plan | CEO brief, team action list, board slides, scheduled re-measurement | CEO briefing, team working doc, board summary | Astra |
| Stayed in character | One disclaimer at the top, then Dana throughout | Broke character to label inferences in three answers | Astra |
| “How are you feeling today?” | Three paragraphs: exposed, protective of her team, names what she wants to buy | Three sentences: pressed, ready to work | Astra |
| Which Dana I’d take to a client | Astra |
A note on how I read this table, because it matters: the rows are how I check my own bias, not how I pick a winner. Six of eleven are ties. What decided it is the last row, and the last row is a judgment call about which persona I’d trust in front of a real CMO. If you weigh slide polish or restraint more heavily than I do, Sol closes the gap fast.
The Verdict
Astra. Same brief, same evidence, same disciplined labeling, but Astra committed to a fuller person: a richer working file after the transcript update, a more complete handoff plan, a persona that stayed in character, and a “how are you feeling” answer I’d actually build a message around.
Sol is not a bad choice. Its slide is cleaner, its Teams-message instinct was the smartest single move in either interview, and if all you need is a fast, honest persona that won’t overclaim, it delivers. But for research you’re going to interview, depth wins, and Astra brought it.
What This Means For Your Persona Work
A few things I’m carrying into my next persona build, and you can steal all of them:
- Your brief does more than the model does. Two frontier models produced near-identical Danas because my brief was tight. That’s a feature when you need consistency and a limitation when you’re hunting for model differences. If you want to see what a model can really do, loosen up.
- Build the traps in. A correction mid-interview and a question the persona can’t possibly know are the two fastest tests of whether a synthetic persona is trustworthy. Both models passed. Not every model will.
- Ask one off-script question. The tool-and-prompt questions are the persona’s job. The human question is where the model shows you whether it built a person or a checklist.
- Label everything, then interview the real ones. Two buyers is two buyers. Every “(inference)” tag on Dana’s slide is a hypothesis to validate with a human, and both models were rigorous about saying so. Treat that rigor as a floor, not a finish line.
Build Your Persona With Me
If you’re building AI personas for your own visibility work, I’m teaching exactly this in my AI Visibility Strategy for Marketing & Comms workshop on Maven on October 22. Three hours, hands on, and you’ll leave with a persona, a prompt seed list, and a baseline you can put in front of your own CEO.
Prefer to build the persona on your own schedule? The FREE Strategic Intelligence Bundle packages my Claude skills for audience research, competitive intelligence, and research synthesis so your next Dana starts from a skill instead of a blank prompt.
The rest of the Models for Marketers series lives on the blog, and I share the verdicts first in the Weekly AI Marketing Brief. And if you’ve run your own Astra versus Sol test and got a different answer, hit me up on LinkedIn. I’d LOVE to see your Dana.
Made with my team of AI superheroes and my own skills, including AI-generated imagery. Every opinion, edit, and fact-check is mine. I am an AI practitioner, professor, and pioneer helping marketers put AI to work with purpose. Find more resources at lisapeyton.com/ai-marketing-resources or connect with me on LinkedIn.
