Leah is a senior sourcer on a four-person talent team at a 90-person fintech startup, and she lives in the gap between a name on a shortlist and a recruiter who is ready to call that person. When a staff backend engineer role opened, her hiring manager handed her eight passive prospects and a one-line brief: "Tell me which of these is worth our time." Each prospect existed across half a dozen public surfaces, a professional networking profile, a public code repository, a conference talk, a personal blog, a developer question-and-answer history, and Leah's job was to fuse those fragments into one trustworthy picture per candidate without inventing anything, without crossing a privacy line, and without smuggling in an assumption about someone's background that had no business in a hiring file. This lesson is her method, plus the practice discipline that keeps it alive after the first enthusiastic week.

Why Synthesis Is a Distinct Skill

Gathering information and synthesizing it are different jobs, and recruiters who conflate them produce candidate profiles that are long, confident, and wrong. A single source tells you one slice: a networking profile tells you the titles a candidate chose to list, a public repository tells you what they build in the open, a conference talk tells you how they explain their thinking. None of those is the candidate. Synthesis is the work of holding several partial, sometimes contradictory slices side by side and producing one picture that is honest about what it knows and what it does not. Done badly, it manufactures a coherent story by quietly discarding the inconvenient pieces, which is how a candidate's two-line public bio becomes, three steps later, a confident claim about their seniority that no source supports.

Being Intentional About Research

The first principle underlying every effective recruiting practice, synthesis included, is intentionality. Recruiting runs on momentum more than most functions admit, and "this is how we have always done it" explains a startling amount of what a talent team does each week. Effective practice requires stopping to ask three questions of any activity: why are we doing this, what outcome are we actually trying to achieve, and is this the best way to achieve it? Applied to candidate research those questions have teeth, because most research fails not on technique but on never having been aimed at anything in particular.

Consider sourcing. Many teams source passively: post the job, wait for applications, work whatever arrives. Intentional sourcing asks different questions. What if you systematically identified the candidates who match your needs rather than waiting for them to identify you? What if you deliberately targeted underrepresented groups you are not reaching through your standard channels? What if you verified information before investing interview time rather than discovering in the second round that a claim did not hold? The same shift applies to evaluation, where research eventually lands. Many teams evaluate on interview impressions; intentional evaluation structures questions so every candidate is assessed on the same things, documents reasoning so decisions can be audited for fairness later, and measures whether the evaluation criteria actually predict on-the-job performance rather than assuming they do. Intentionality creates the conditions for improvement, because when you are clear about why you do something you can tell whether it is working and change it when it is not.

The Sources and What Each Is Good For

Leah works from a fixed mental map of what each public surface can and cannot tell her, and the discipline is never to let a single source carry more weight than it earns. When two sources agree the signal is stronger. When they conflict, the conflict itself is the finding, and it becomes a question for the screen rather than a fact she resolves by guessing.

The Privacy Boundary: Public, Minimal, and No Inference

Before Leah synthesizes anything, she works inside three hard boundaries, and they are not optional niceties. The first is public information only. She uses what a candidate has chosen to publish to a public audience, and she does not attempt to access anything behind a privacy setting, a login wall, or a connection request that would misrepresent her purpose. The second is data minimization, the principle at the heart of the GDPR for any candidate covered by it: she collects only what is relevant and necessary to assess fit for this specific role, because a hiring file is not a place to accumulate everything the internet knows about a person. A candidate's marathon times, political posts, and family photos are none of the role's business, and pulling them into a profile is both a privacy violation and a bias hazard.

The third boundary is the firmest: no protected-class inference. Leah never records, and never asks an AI tool to infer, a candidate's age, ethnicity, religion, health, sexual orientation, family status, or any other protected characteristic, whether stated outright or implied by a graduation year, a name, a photo, or a gap in employment. Under Title VII, the ADEA, and equivalent frameworks, those characteristics cannot lawfully inform a hiring decision, and the safest way to keep them out of the decision is to keep them out of the file.

These boundaries are not in tension with good sourcing; they are what good sourcing looks like. A profile built only from role-relevant public information, with protected characteristics deliberately excluded, is both more legally defensible and more useful, because it forces the brief onto the evidence that actually predicts performance. It is also where fairness stops being a separate initiative and becomes a property of the work itself, which is the only form of fairness that survives a busy quarter.

A Worked Synthesis: One Candidate, Five Sources

Here is how Leah handled one of her eight prospects, with details that are illustrative rather than drawn from a real person. The candidate, a backend engineer, appeared across a networking profile, a public code repository, a conference talk, a developer question-and-answer profile, and a personal blog. Leah's first pass was extraction. From the networking profile: a stated six years of backend experience and a current senior-engineer title. From the repository: active recent commits to two distributed-systems projects, one of which the candidate maintained rather than merely contributed to. From the talk: a clear explanation of an event-driven architecture migration. From the question-and-answer history: sustained high-quality answers on database concurrency. From the blog: two thoughtful posts on the same migration topic.

Her second pass was reconciliation. The maintainership and the concurrency answers corroborated the staff-level depth the role required, which the networking title only asserted. The talk and the blog, on the same subject, suggested genuine specialization rather than a one-off. One conflict surfaced: the profile implied the candidate led a team, but nothing in the public technical record spoke to people leadership. Leah did not resolve that by assuming. She flagged it as the most important question for the recruiter screen: "Public work shows deep individual technical contribution; leadership scope is asserted but unverified. Confirm whether they have led people, and how many."

The brief she handed her recruiter ran to about 200 words: a confidence-rated summary of relevant experience, the corroborated strengths with the source for each, the one open question, and an explicit line that no protected-class signals were collected or considered. She ran her AI assistant over the assembled public text to draft it, instructing the model to cite a source for every claim, mark anything unsourced as unverified, and refuse to speculate about anything private. A claim with no source attached did not make it into the brief.

Quality Control: Trust, but Cite

The fastest way to poison a candidate brief is to let plausible-sounding synthesis substitute for sourced fact, and AI tools are very good at producing plausible-sounding synthesis. Leah's quality control rests on a single rule: every claim in the brief points back to a specific source, or it is labeled unverified, or it does not appear. When her AI assistant wrote that a candidate had "extensive experience with high-scale systems," she checked whether any source actually said that or whether the model had inflated "active commits to two distributed-systems projects" into a grander claim. Often it had, and she pulled the claim back to what the evidence supported.

She also watches for the model resolving conflicts by quietly picking one side, which it will do unless instructed otherwise, and for it filling gaps with confident invention. The cure for both is the citation requirement: a model forced to attach a source to each statement either finds one or admits it cannot, and both outcomes are honest. A brief that says "unverified" in three places is more valuable than one that reads cleanly because the uncertainty was smoothed away.

Consistency Over Perfection

The second principle is that consistency matters more than perfection. You do not need the perfect research method or brief template; you need a good one applied consistently to every candidate. Inconsistency creates bias, because one candidate gets three sources checked and a careful reconciliation while another gets a glance at a profile and a hunch, and the difference between them is usually not the candidate. It also prevents improvement: if you research every prospect differently, you can never tell which part of your method produces the value and which part is habit.

Consistency does not mean rigidity, and the distinction matters because rigidity is the objection every experienced sourcer raises first. It means: these are our standards, we apply them to everyone, and we measure whether they work. If they are not working, we improve them, consistently, for everyone. In practice that looks like a fixed source checklist per role family, a fixed brief structure, and a fixed rule about what counts as verified, each revisable deliberately but none of which flexes silently because a hiring manager is in a hurry.

Measurement and Iteration

The third principle is measuring outcomes and iterating on what you learn. You implement a new practice, you measure results, and you see whether it improved the outcome you cared about. If it did, you keep it and build on it. If it did not, you adjust. The framing matters here: measurement is not about proving that everything you did was right. It is about learning what actually works in your context so you can improve continuously, and a negative result is often the more useful one.

Consider an example from an adjacent practice. You implement blind resume screening and measure whether it increases diversity in your interview pool. If it does, blind screening is working and you scale it. If it does not, you have learned something rather than nothing: something else is creating the disparity, perhaps your interview questions or how feedback gets interpreted, and you now know where to look. Either outcome moves you forward, and only because you measured. Applied to synthesis, the equivalent questions are whether briefs with explicit open questions produce better screens, and whether the candidates your research recommended actually performed once hired.

Knowing When the Picture Is Good Enough

Research synthesis has no natural stopping point, which is its own trap. There is always one more source, one more cross-reference, one more rabbit hole, and a sourcer can burn an hour per candidate chasing diminishing returns. Leah caps her synthesis at the point where she can answer the brief: does the public evidence support investing recruiter time, and what is the most important thing to confirm in the screen? Once she can answer those with sourced confidence she stops, even if more material exists. An exhaustive dossier is both a waste of time and a data-minimization failure, since most of what it contains is irrelevant to the role.

Anti-Patterns

Implementing without foundation. The team adopts a practice without understanding why or having the system support to sustain it: structured interviews adopted because they are a best practice, with no interviewer training, no questions built into the system, and no measurement of whether outcomes changed. It happens because adopting something feels like doing something. What goes wrong is that the practice reverts, since it requires extra individual effort and nobody is measuring whether it earns that effort. The fix is to understand why you are implementing it, build it into systems so it is the path of least resistance, and measure the result.

Measuring without acting. The team collects data and tracks metrics but never converts the insight into a change. A team tracks that it hires men at 35 percent and women at 25 percent, and the number appears in a dashboard quarter after quarter without anyone investigating why or implementing a solution. It happens because measurement feels like action; a report looks like progress. What goes wrong is that the metric does not improve, because nothing was done. The fix is to close the loop: define the metric, measure it, analyze what is driving it, act on the analysis, and measure again.

Trying to change everything. The team attempts to improve sourcing, interviews, evaluation, documentation, and technology at once. It happens because everything genuinely does seem important, and choosing feels like conceding that some problems will persist. What goes wrong is overwhelm: nothing gets the focus it needs, several efforts half-land, and the visible failure makes the next attempt harder. The fix is to pick one area, implement it well, iterate, then move to the next.

Implementing without sustainability planning. The team puts a new practice in place but never builds it into systems or trains the wider team for continuity, because the initial implementation feels like completion. What goes wrong shows up months later, when a new person joins, does not know the practice exists, reverts to the old way, and a second person copies the first. The fix is to build the practice into systems, training, job descriptions, and performance expectations, so continuity does not depend on the memory of whoever led the change.

Practice

• Intention audit. Take your current candidate research practice and ask of each step: why do we do this, what outcome are we trying to achieve, and is this the best way to achieve it? Note every step that survives only on habit.
• Consistency check. Take one stage of your process, the phone screen or the pre-screen research, and document exactly what you do for each candidate. Is it consistent across candidates? If not, write down how you would standardize it without making it rigid.
• Build the source map. For one role family you hire for repeatedly, list the public sources you would consult, what each is genuinely strong evidence of, and what its limitation is. Then write the rule for what counts as verified.
• Run a synthesis with citations. Take one prospect, extract from at least three public sources, reconcile the findings, and write a brief under 250 words in which every claim carries its source or is labeled unverified. Note where your AI assistant inflated a claim.
• Measurement design. What metrics would be most valuable to track for the area you want to improve, and how would you collect them with the systems you have today?
• Implementation and sustainability planning. If you could improve one area of recruiting, which would have the biggest impact, how would you implement it, and which of the four barriers would you have to overcome first? Then think of a change you tried before that did not stick, and name what it would have taken for it to survive a new joiner and a busy quarter.

Reflection

• Which recruiting practice would have the highest impact if you improved it, and why has it not been improved yet?
• How intentional are you currently about your research practice, and how much of it is momentum?
• What would you measure to know whether your improvements were actually working?
• What barriers exist to implementing change in your recruiting function, and which one is the real blocker rather than the stated one?
• How would you ensure a new practice sustains over time once the attention moves elsewhere?

Glossary

• Synthesis. Holding several partial and sometimes contradictory sources side by side to produce one picture that is explicit about what it knows and what it does not.
• Data minimization. Collecting only what is relevant and necessary for the specific purpose, a core GDPR principle and a bias control in its own right.
• Protected-class inference. Deducing or recording a candidate's age, ethnicity, religion, health, family status, or similar characteristic from indirect signals.
• Intentionality. Purposeful decision-making about why you do something and what outcome you are trying to achieve.
• Consistency. Applying the same practices, standards, and processes to all candidates and decisions.
• Measurement. Tracking metrics related to recruiting outcomes and process effectiveness.
• Iteration. Implementing, measuring results, learning, and adjusting based on what you learn.
• Implementation. Putting a new practice into action in your recruiting process.
• Sustainability. Building practices into systems, training, and processes so they persist over time.

Related Lessons

• AI-Assisted Sourcing: Boolean Search Optimization covers the search step that produces the prospects this lesson then synthesizes.
• Verifying Candidate Information: Spotting Hallucinations and Inaccuracies goes deeper on the citation discipline this lesson treats as one rule.
• Data Minimization: Collecting Only What's Necessary works the second privacy boundary into concrete decisions about what enters a hiring file.
• Diversity and Bias in Sourcing: How AI Can Help and Harm extends the fairness material into the sourcing stage itself.
• Interview Preparation: Candidate Research & Structured Question Generation is where the open questions in your brief become interview questions.
• Hands-On Project: Design a Sourcing Workflow with AI and Guardrails is where you assemble this method into a workflow of your own.

Closing

Your recruiting system is only as good as your practices. Good sourcing lets you interview better candidates, fair evaluation produces better hiring decisions, clear documentation makes fairness audits possible, and effective communication produces a better candidate experience. Research synthesis sits near the front of that chain, which is why the discipline it demands, sourced claims, explicit uncertainty, and firm privacy boundaries, pays out downstream in decisions that hold up. Every practice in this lesson shares one shape: intentionality about the goal, consistency in application, improvement through measurement, and survival only when built into systems and given continued attention.

Key Takeaways

• Synthesis is a distinct skill from gathering, and each source has a ceiling. A networking profile is self-authored marketing, public code is strong evidence but undercounts private work, a talk is a single moment. Never let one source carry more weight than it earns, hold the slices side by side, and treat conflicts between them as findings rather than problems to guess away.
• Stay inside three privacy boundaries. Public information only, data minimization so you collect only what the role needs in line with GDPR, and no protected-class inference of age, ethnicity, health, family status, or anything else under Title VII and the ADEA. Keep protected characteristics out of the file to keep them out of the decision.
• Reconcile, then flag what you cannot verify. When a profile asserts something the technical record does not support, that gap becomes the single most important question for the screen, not an assumption you resolve yourself.
• Cite every claim or label it unverified, and aim for a brief rather than a dossier. AI synthesis produces confident, plausible inflation unless you require a source for each statement. Produce a short, source-backed summary with strengths, open questions, and an explicit note that no protected signals were collected, then stop once it can answer whether to invest recruiter time and what to confirm in the screen.
• Intentionality is the foundation. When you are clear about why you do something, you can evaluate whether it is working and improve it. Momentum is the default, and momentum is not a method.
• Consistency creates fairness and enables improvement. Inconsistent practice produces bias and makes measurement meaningless. Consistent practice, revised deliberately for everyone at once, produces both fairness and the ability to learn.
• Measurement enables iteration, but only if you act on it. You cannot improve what you do not measure, and a negative result is informative rather than a failure. Measuring without acting is its own anti-pattern.
• Start with one area and do it well. Assess your current state, prioritize on impact, feasibility, and leverage, pilot, measure, then scale. Trying to change everything at once produces overwhelm and nothing else.
• Build change into systems, not individual effort. Practices embedded in systems, training, and job descriptions survive busy quarters and new joiners. Leadership attention and publicized measurement keep them alive, and what you measure is what teams will optimize for.

Frequently Asked Questions

How long should a synthesis take per candidate? Long enough to answer two questions with sourced confidence, and no longer: does the public evidence justify investing recruiter time, and what is the most important thing to confirm in the screen? There is no fixed number, but the stopping rule is the same for a rich public record and a sparse one. If you are still researching after those questions are answered, you are building a dossier, which wastes time and works against data minimization at once.

What if the sources genuinely contradict each other? The contradiction is the most valuable thing you found. Do not resolve it yourself and do not let an AI assistant resolve it quietly, which is what it will do unless you instruct it otherwise. Record both readings, state plainly that the point is unverified, and convert it into the specific question the recruiter should ask on the screen. A brief whose single sharpest contribution is a well-framed open question is doing exactly what a brief is for.

Is it acceptable to send a connection request to see more of a profile? Not if the request misrepresents your purpose. The boundary is public information only, meaning what the candidate chose to publish to a public audience. Anything behind a privacy setting, a login wall, or a connection obtained under false pretext is outside it, and the fact that it would be useful is not an argument. If you need information a candidate has not made public, ask them for it in a screen, where they can decide what to share.

Our hiring managers do not want a new brief format. Is it worth pushing? Push through involvement rather than mandate. Ask what they actually need from a candidate brief, incorporate it into the design so the format is partly theirs, and pilot it on one search rather than announcing it for all of them. Then show the result: the open question that changed a screen, the candidate the old format would have oversold. Resistance responds to a concrete success in the manager's own pipeline far better than to an argument about best practice, and a manager who helped design the format will defend it to their peers.