How It Works
The complete methodology: what the AI actually analyzes, the scoring rubric, the rewrite framework, and the validation layer that catches AI-authorial language before it reaches your report.
The audit in three steps
Step 1Public LinkedIn profile scrape. You paste your profile URL. Apify's Maestro actor reads the publicly visible profile page — headline, About, experience, education, certifications, skills, featured, recommendations, honors, banner and profile images — and returns structured data. No login. No password. No LinkedIn API. We never touch authenticated or private data.
Step 2ICP context via a short questionnaire. Eight questions gather the signals the AI needs to personalize every recommendation: your role and industry, target audience (ICP), what's holding your profile back today, the tone and positioning you want, how often you post, the topics you want to be known for, your biggest recent impact, and anything to avoid mentioning. This is the single source of truth the AI references in every bullet it produces.
Step 3Single-call AI audit. One consolidated AI request (OpenAI GPT-5) receives the full scoring rubric, full skill instructions, full scraped profile, and full questionnaire in one pass. This replaces the older approach of splitting sections across parallel calls (which produced drifting, inconsistent output). One call ≈ 120–180 seconds and ≈ 14,000 input tokens / 5,000 output tokens.
What the AI produces
- Overall score (0–100) — sum of the 10 per-section scores.
- Per-section scores (0–10 each) — profile picture, banner, headline, About, URL, tone & cohesion, recommendations, experience, honors & awards, education & certifications.
- Three-dimension breakdown — clarity, authority, conversion potential, each scored 0–10.
- One-sentence diagnosis — the overall framing of the profile.
- Three Quick Fixes — the highest-impact changes, each with "what's wrong", "why it matters", and an exact copy-paste rewrite.
- Per-section strategy (opt*) — 2–5 strategic bullets per section explaining how to think about the section, grounded in your ICP.
- Per-section suggestions (sug*) — 2–5 action-only bullets per section (≤25 words each), every bullet containing a concrete hook (exact text, tool, placement, dimension, or step).
- Three headline options — 120–220 characters each, first-person, copy-paste-ready.
- Full About rewrite — 1,500–2,000 characters, flowing first-person paragraph.
- Featured section strategy — three slots typed as proof, lead magnet, and call-to-action, each with a recommendation and rationale.
The scoring rubric
Ten sections are evaluated against three underlying dimensions:
- Clarity — does a reader understand who you help and what you deliver within five seconds? Is the ICP named specifically (not "businesses" or "professionals")? Is the offer stated up front?
- Authority — are there credibility signals (client results, specific numbers, named companies, recognizable brands, a named methodology)? Do the visuals (photo, banner) reinforce authority?
- Conversion potential — is there a clear call to action? Does the profile guide a reader toward the next step (DM, book a call, lead magnet)? Are there identifiable conversion barriers (default banner, generic About, missing Featured)?
The rubric was refined across 100+ real profile rebuilds for founders, fractional executives, B2B agency leaders, and enterprise operators. It is stored in versioned skill files (SKILL.md, scoring-framework.md, rewrite-playbook.md) that the AI reads at inference time.
The rewrite framework
- Headline (5 formulas). The AI picks the formula that best matches your ICP and offer, then generates three 120–220 character variations.
- About (5-block structure). Hook opener → proof/story body → named ICP → offer + method → call-to-action close. First person, 1,500–2,000 characters.
- Banner. 1584×396 px, brand-aligned colors, one bold value statement, optional credibility signals. Specific dimensions and font sizes provided.
- Featured (3 slots). Slot 1 = proof (case study, testimonial, article), Slot 2 = lead magnet (checklist, template, free resource), Slot 3 = CTA (book a call, newsletter, download).
- Experience. Benefit-driven summaries with quantified outcomes, strong action verbs, media attachments as proof.
Voice-safety validation
After the AI produces output, a validation layer catches AI-authorial language that shouldn't appear in a user-facing rewrite:
- Banned AI phrases — "I recommend you", "Here's what I'll do", "My approach is", "Let me walk you through", "I analyze your", "I write your", "I'm the world's best".
- Schema-identifier leaks — internal field names like
sugAbout,optHeadnever reach user-facing text. - Count rules — headline options must be 3 (tolerated 3–5); About rewrite must be 1,500–2,000 characters; bullets must be 2–5 per field.
Violations trigger soft warnings logged alongside the audit. Hard schema errors (missing fields, invalid scores) trigger a retry before the report is saved.
AI model & infrastructure
| Primary model | OpenAI GPT-5 |
| Fallback model | Claude Sonnet 4.6 (toggleable via AI_PROVIDER) |
| Input tokens per audit | ~14,000 (with prompt caching after first call) |
| Output tokens per audit | ~5,000 |
| Wall-clock time per audit | 115–180 seconds |
| Scraper | Apify Maestro (public LinkedIn profile reader) |
| Storage | MongoDB Atlas |
| Hosting | Bare-metal VPS (AlmaLinux 8, Docker, Nginx) |
Version history
- 2026-04-22 — Moved opt*/sug* fields from newline-dash strings to native JSON arrays; eliminated downstream regex splitters.
- 2026-04-21 — Switched default provider from Claude Sonnet 4.6 to OpenAI GPT-5; added provider dispatcher.
- 2026-04-21 — Moved report-email delivery entirely server-side (previously fire-and-forget from browser).
- 2026-04-20 — Single-call consolidated AI audit replaced the prior 3-parallel-call architecture.