Mechanic 01 · How LinkedIn Recruiter Searches Profiles | The Hire-Me Playbook
Mechanic 01 · Findable · Under the Hood

How LinkedIn Recruiter searches profiles

Most job-search advice says to add keywords to your profile. Keywords matter for the first cut, but they don't decide the ranking. LinkedIn Recruiter uses a machine-learning model — a Gradient Boosted Decision Tree running on dozens of signals — trained on one thing: whether the candidate replies to the recruiter's InMail1.

Since 2020, most of those signals have been deep-learning embeddings — patterns the model learns from data — not lists of titles, skills, and seniority levels picked by hand2. That's five years ago, long before the 2024 AI wave.

The Recruiter search pipeline Sources: [1], [2]
Recruiter search pipeline: query → Galene retrieval → Pensieve re-rank → ranked candidates, trained on InMail Accept. STEP 1 Recruiter query keywords, filters STEP 2 · RETRIEVE Galene (Lucene) candidate set from index STEP 3 · RE-RANK Pensieve (deep-learning) semantic embeddings, since 2020 TRAINED ON InMail Accept — did the candidate reply?
Two stages. Galene finds candidates by matching specific profile fields (title, past titles, company). Then the ranking model — GBDT using Pensieve embeddings — re-orders them. It is trained on one signal: did the candidate reply to the recruiter's InMail?

lightbulb The big idea

Recruiter search isn't scoring your keywords. It is predicting whether you would reply to a recruiter's message — the one target the ranking model is trained on1. Every field on your profile feeds that single prediction.

What's actually running

Two stages. The first is Galene, LinkedIn's search platform. Galene pulls an initial set of candidates by matching the recruiter's search against specific profile fields — current title, past titles, current company, past companies. Each field is scored on its own, not merged into one big text blob1. Free-text fields on the profile are searched alongside them.

Then Pensieve takes over. It is LinkedIn's platform for machine-learned representations, and it feeds those learned patterns into the ranking model — replacing most of the hand-picked lists of titles, skills, and seniority levels that used to matter most2. In practice, "backend engineer" and "server-side developer" no longer need the same exact words to score as similar. LinkedIn launched this in 2020 — five years before the current AI wave.

What that means for your profile

LinkedIn has never said how much weight each profile field carries in the model. Any guide claiming "your headline counts 3× your experience" is inventing a number. This card won't.

What LinkedIn has said: profile completeness itself helps you get found. The Profile-level meter checks seven fields — photo, location, industry, education, position, skills, and summary — and LinkedIn's help center states a complete profile "helps improve the discoverability of your profile in search results"3. An empty field can keep you from showing up at all.

One more disclosed signal: Recruiter marks candidates who have a "Have company connections" tie — a 1st-degree connection to an employee at the hiring company4. It's a filter recruiters can turn on, not a hidden rank boost. Chapter 7 treats it as a visibility lever.

The 2024 change most guides missed

LinkedIn shut down Skill Assessments in 2024. The "Verified" badges that career blogs still tell you to chase — the ones recruiters could once filter for in Recruiter search — no longer exist on the platform5. If a guide tells you to take the Excel assessment for the badge, it is describing a feature that is gone.

The recruiter's own patterns

Three things LinkedIn has published about how recruiters send messages and which candidates reply:

InMails under 400 characters get a 22% higher response rate than the average; those over 1,200 characters get 11% below6. Personally written messages beat bulk-sent ones by roughly 15%.

Candidates with Open-to-Work turned on reply to InMails at a 37% higher rate6. The 37% figure comes from the same LinkedIn dataset as the character-count numbers above — same year, same method.

AI-Assisted Search shows an 18% higher InMail acceptance rate than manual filter-based search7. That's LinkedIn's own comparison for the AI layer that now shapes a growing share of recruiter searches. Being found is not the same thing as being ranked well.

Two things to act on today

Two places to start:

  • Rewrite the first field a recruiter sees on the search results. That's your headline. Chapter 1 walks through three formulas for fitting a searchable phrase, a skill, and a number into 220 characters.
    Read Chapter 1 arrow_forward
  • Score your profile on the fields that actually matter. The Scorecard checks the ones this card named: completeness, structured fields, company connections. Four minutes, no signup.
    Take the Scorecard arrow_forward

Notes

  1. LinkedIn Engineering — Qi Guo, Sahin Cem Geyik, Cagri Ozcaglar, Ketan Thakkar, Nadeem Anjum, Krishnaram Kenthapadi. "AI Behind LinkedIn Recruiter Search and Recommendation Systems," 2019. Primary source for the Galene retrieval architecture, Gradient Boosted Decision Trees as the ranking model, structural-field matching (current title, past titles, current company, past companies as separately weighted textual-similarity features), and InMail Accept as the training signal. linkedin.com/blog/engineering/… ↑ ↑ ↑

  2. LinkedIn Engineering — Benjamin Le, Daniel Gmach, Aman Grover, Roshan Lal, Jerry Lin, Austin Lu, Qingyun Wan, Leighton Zhang. "Pensieve: An embedding feature platform," 2020. Describes deep-learning embeddings replacing hand-crafted title/skill/seniority/location features across LinkedIn Talent Solutions ranking models, including Recruiter search. linkedin.com/blog/engineering/ai/pensieve ↑ ↑

  3. LinkedIn Help Center — "Your Profile level meter" (undated, current at access, 2026). States a complete profile "helps improve the discoverability of your profile in search results, and to increase profile search appearances." Lists the seven fields the meter scores: photo, location, industry, education, position, skills, summary. linkedin.com/help/linkedin/answer/a594698 ↑

  4. LinkedIn Recruiter Help — "Spotlights in Recruiter and Jobs." Officially names Have company connections as a filterable spotlight: 1st-degree connections to employees at the hiring organization. linkedin.com/help/recruiter/answer/a414283 ↑

  5. LinkedIn Help Center — "LinkedIn Skill Assessments." Confirms Skill Assessments and the associated "Verified" badges were fully discontinued in 2024. LinkedIn's stated reason: examples of applied skills became more valuable than a standardized test score. linkedin.com/help/linkedin/answer/a507663 ↑

  6. LinkedIn Talent Blog — "How to Improve Your InMail Response Rate, According to LinkedIn Data," 2022. Analysis of tens of millions of corporate-recruiter InMails, May 2021 – April 2022: messages under 400 characters got 22% higher response than average; over 1,200, 11% below. Individually sent messages ~15% higher. Open-to-Work candidates responded at a 37% higher rate. Methodology (sample size, date range, exclusions) disclosed. linkedin.com/business/talent/blog/… ↑ ↑

  7. LinkedIn Talent Solutions — "New AI-Assisted Search and Projects in Recruiter" (product page, 2024/2025). LinkedIn reports an 18% higher InMail acceptance rate for AI-Assisted Search sessions vs. manual filter-based search. Sample-size and date-range for the 18% figure are not itemized on the page. business.linkedin.com/talent-solutions/… ↑

verified
Evidence

Seven primary-source citations, all Tier 1 by the project's evidence bar. Every claim on this page maps to LinkedIn's own engineering blog, product pages, or help-center documentation. Methodology notes, sample sizes, and the specific gaps we did not claim to know live in the underlying research bundle.

arrow_backAll mechanics Next: Mechanic 02 · Open to Work