Discovery Research for an AI Career Coach
How I conducted 20 discovery interviews in 3 weeks to invalidate my core product assumption and identify my beachhead ICP
Executive Summary
I built Grapevines, an AI career coach for senior professionals ($79/month), with an initial hypothesis that better AI-generated resumes = better job search outcomes. Three weeks of intensive customer discovery proved this assumption fundamentally wrong.
The Challenge
The Product
Grapevines is an AI career strategist for mid-to-senior professionals (10-20+ years experience) navigating career transitions. The core value proposition: strategic depth of a $300/hr coach, available 24/7 for $79/month.
The Problem I Set Out to Solve
Job seekers with strong backgrounds were applying to 100+ roles and getting nowhere. The conventional wisdom said their resumes needed optimization. AI tools could generate "perfect" resumes tailored to each job description. Problem solved, right?
My Initial Hypothesis
Better AI-generated resumes = better outcomes. If I could build AI that produced higher-quality, more tailored resumes faster than competitors, I'd win.
Why Research Mattered
I had $0 for marketing and a 6-week runway to prove the concept. I couldn't afford to build the wrong thing. Before writing a single line of product code, I needed to understand whether my hypothesis held up against real user behavior.
Research Methodology
Framework: SVPG-Inspired Discovery
I structured every interview using Marty Cagan's discovery framework:
- Problem — What's the actual pain point? (Not what they say, what they do)
- Evidence — Verbatim quotes and observed behaviors
- Desired Outcomes — What does success look like to them?
- Customer Context — Experience level, industry, urgency
- Current Journey — What have they tried? Why didn't it work?
- Assumptions to Test — My hypotheses vs. their reality
Each interview was 45-60 minutes, semi-structured, with a live product demo mid-call. I documented verbatim quotes, strategic subtext analysis, and explicit assumption stress-tests.
Participant Recruitment
20 interviews across 3 weeks (Feb 16 - Mar 7, 2026)
Sourced via LinkedIn DMs, warm network referrals, and career-focused community groups (Slack, Discord).
Screening criteria:
- 8+ years professional experience
- Active job search or completed search within 6 months
- Tech-adjacent roles (PM, engineering, design, ops, marketing)
Segmentation Approach
I intentionally recruited across a spectrum:
- Sophisticated AI users (building their own tools) ↔ First-time searchers (relied on referrals for 15+ years)
- Job seekers (primary users) + Product evaluators (UX critique, not actively searching)
- B2C perspectives (individual job seekers) + B2B perspectives (coaches, recruiters, TA leaders)
This range ensured I wasn't just validating my assumptions with friendly users.
ICP Validation & Refinement
Started with 3 Hypotheses. Validated and Sharpened Each.
I entered research with three hypothesized ICPs. The goal wasn't to "discover" new personas. It was to validate which ICP to pursue first and understand their activation triggers deeply enough to build specifically for them.
| ICP | Starting Hypothesis | Research Validation | Refinement |
|---|---|---|---|
| Prianka (The Strategic Pivoter) | Career changers (8-12 yrs exp) struggle to translate experience to new industries | VALIDATED | Vocabulary discovery is step 0. Positioning fails without it. They need direction before they need a resume. |
| David (The Plateaued Leader) | Senior ICs/Directors (15+ yrs) hitting the executive ceiling | VALIDATED | Voice authenticity is the unlock. They fear sounding like "AI slop" more than they fear a bad resume. |
| Elena (The Burned-Out Hunter) | High-volume applicants (100+ apps) with spray-and-pray fatigue | VALIDATED but DEPRIORITIZED | Highest urgency, highest churn risk, lowest willingness-to-pay. Not the beachhead. |
The Beachhead Decision
Research confirmed Prianka as the beachhead ICP:
- Highest willingness-to-pay ($100-500 for career transformation vs. Elena's $20-50 for "anything that helps")
- Clearest activation moment — The positioning conversation generated consistent "aha" reactions
- Most shareable success stories — Career pivots are inherently viral ("She went from teaching to product management")
Key Findings
What I Validated
1. High application volumes create a discoverability problem
- Average: 42 applications per interview received
- One participant: 3,000+ applicants for a single role he was hiring for
"It's a DDoS attack on recruiters. They don't have time to go through all that."
2. Referrals dramatically outperform blind applications
- 6% of applications come through referrals → 37% of hires
- Multiple participants validated: internal referrals are often REQUIRED, not just preferred
"Not one of my offers came from an online application. It was either a friend or someone reaching out to me."
3. ~30% of job postings are ghost jobs
"About a third of the postings out there are probably not real... We put jobs out there to drive interest."
4. Human-in-the-loop is non-negotiable
"I still have myself in the loop... because I want the humanized piece of it and not a bot."
Even sophisticated AI users (Claude Max Pro subscribers, $200/month) explicitly rejected fully automated outputs.
5. Sophisticated users WILL pay for leverage
I assumed power users would just DIY with ChatGPT. Wrong.
"What I see here is a whole lot easier than what I've been doing."
— From a participant using Copilot, Gemini, and NotebookLM simultaneously
What I Invalidated
- "Resume + JD in a chatbot" is sufficient — Users need guided flow, not raw analysis
- Pure chat interfaces work — Structured conversation dramatically outperformed open-ended prompting
- Generic AI output satisfies users — Explicitly rejected as "sanitized"
- Volume applications are efficient — 3% of applicants reach interview stage (down from 15.25% in 2016)
The Surprising Finding: Generic CV Anomaly
Two participants had run A/B tests on their own applications. The finding that changed everything:
Generic, keyword-focused resumes OUTPERFORMED tailored, metric-driven ones in ATS systems.
One participant's data: his "optimized" resume with quantified achievements got a 1.3% response rate. His generic keyword-stuffed version performed better.
This single finding reoriented my entire product thesis:
Access and channel matter more than resume quality.
I inverted my product workflow: positioning and outreach FIRST, resume generation LAST.
Research → Product Decisions
Direct Traceability: Every P0 Feature Cites a Customer Quote
| Research Finding | Verbatim Quote | Product Decision |
|---|---|---|
| Initial hook differs from positioning output | "I felt like there was a bit of a bait and switch. What hooked me wasn't addressed in the output." | P0: Persist differentiator throughout flow |
| Running analysis twice produces different results | "The AI chose a different differentiator when I ran it again. How do I know which one is right?" | P0: Lock differentiator after first run OR show top 3 options |
| Rubric scores without explanation | "Why did I get a 'Stretch' score? What does that even mean?" | P0: Transparent 100-point breakdown with evidence |
| Contact identification is manual and slow | "I'm building a Python scraper just to find who to reach out to." | P0: Stakeholder mapping with peer-first ordering |
| Navigation dead-ends | "How do I get back to that initial analysis? I can't find it." | P1: Persistent "My Read" access throughout flow |
| Long wait times feel unproductive | "It's taking forever. I don't know what it's doing." | P1: Progress indicators with stage labels |
Strategic Pivot: Job Search as B2B Sales
One participant, a Director with 20+ years experience, repeatedly framed his job search as enterprise sales:
- Target accounts — "I'm not applying everywhere. I have 15 companies I actually want to work for."
- Multi-threading — "You don't just email the hiring manager. You reach their peers, their skip-level, their internal champions."
- Qualification criteria — "I score every opportunity before I spend time on it."
His framing crystallized my product thesis. Job search isn't about optimizing applications. It's about running a sales process.
Impact & Learnings
Measurable Outcomes
- 3 major product pivots directly from research
- Workflow inversion (positioning first, resume last)
- Voice calibration made mandatory (not optional)
- Stakeholder mapping added as core feature
- 15 P0/P1 features with direct customer attribution
- 1 beachhead ICP selected with 3 sub-variants driving prioritization
- 100% of backlog items traceable to specific user quotes
What I'd Do Differently
- Recruit from "non-obvious" segments earlier — The clinical pivoter (healthcare → tech) was my most insightful participant. I almost didn't recruit her because she was "too different."
- Record sessions for team alignment — I took detailed notes, but lost nuance. Verbatim quotes in a doc don't capture the frustration in someone's voice.
- Run more competitive product tests — Only 2 of 20 participants had tried direct competitors.
The Principal PM Takeaway
The most valuable research insight wasn't a feature request. It was an invalidation.
Killing my "better resume = better outcomes" hypothesis early saved months of building the wrong thing. The Generic CV Anomaly didn't tell me what to build. It told me what NOT to build, and that was worth more.
User research isn't about collecting feature requests. It's about stress-testing your assumptions with people who have no incentive to agree with you. The participants who pushed back hardest taught me the most.
This case study documents discovery research conducted February-March 2026 for Grapevines, an AI career strategist for senior professionals.