SDR Cold Email Research: The Pre-Writing Workflow That Top Teams Use
Effective SDR research produces a defensible reason to contact a particular person at a particular account. The useful output is a short evidence record: what happened, where the evidence came from, why it may matter to this role, and what remains unknown. More tabs, more scraped facts, and a longer AI summary do not automatically improve an email.
The workflow below is a repeatable operating method, not a claim that every top-performing team uses the same process or achieves a particular reply rate. It combines account selection, evidence checks, role relevance, and message review before enrollment.
Start with the decision the research must support
Define the audience and problem before looking for interesting facts. The ICP and first-signals workflow helps separate account fit from recent activity. A company may fit your market without showing a useful trigger; another may show activity while being a poor fit. Neither case should be hidden by a generic lead score.
Choose the unit of research: company, person, or buying group. Use buying-committee mapping when the problem spans several roles. A company’s announcement does not establish that every employee owns the initiative. Check the recipient’s current role before translating an account fact into a personal claim.
Create an evidence card before writing the email
| Evidence field | What to capture | What to avoid |
|---|---|---|
| Entity | Company domain, person, current role, and relevant CRM ID | Confusing similarly named companies or an old employer. |
| Observation | A narrow fact supported by the source | Converting a signal into a claim about budget or urgency. |
| Source and timing | Source URL, publication date if available, observation date | Calling an old announcement a new development. |
| Role relevance | Why the observation could matter to this function | Assuming the contact owns an initiative without evidence. |
| Unknowns | Missing, conflicting, or inaccessible information | Treating a failed lookup as proof that a fact is false. |
| Next action | Contact, research further, defer, or exclude | Sending simply because a contact address was found. |
Prefer the company’s own announcement, product page, job posting, or other direct source when it supports the claim. Read enough context to distinguish a current initiative from an archived role or a general description. Keep source URLs in the internal research record even when the email itself should be brief.
Use separate freshness timestamps for the underlying observation and the enrichment record. “Imported today” does not mean “verified today.” If sources conflict, state the conflict and hold the specific claim until it is resolved.
Use AI for bounded questions and reviewable answers
Configure research agents around questions you can evaluate: what customer segment a company serves, whether a public page describes a particular product, or which role owns a function. Select the correct company or person record type, define response types, and add guidance about terms and sources. A broad instruction to “find something compelling” is harder to test.
Test records you expect to qualify and records you expect to reject. Review supporting research, revise ambiguous questions, then inspect results in the play logs after release. Build the qualifying and non-qualifying paths deliberately in Plays. Route missing or contradictory evidence to review rather than presenting it as a confident reason to buy.
The Agents product overview explains the research capability; it does not remove the need to inspect outputs. Review entity identity, source support, timing, and the inference you intend to make. A fluent summary can still mix companies or infer a private priority from public information.
Translate evidence into a modest, relevant message
Use one supported observation and one plausible implication. Phrase the implication as a question or possibility when you do not know the buyer’s situation. The research-to-outreach workflow is useful for turning a checked finding into an appropriate next action instead of appending a decorative fact to a generic pitch.
The following are fill-in templates, not observed customer messages or claims about real recipients. Replace bracketed text only with verified information.
- Company initiative: “Your announcement describes [verified initiative]. Does [role-relevant operational question] sit with your team? We help with [specific relevant capability].”
- Hiring context: “I saw the current opening for [verified role]. If [possible workflow change] is part of that expansion, would [practical resource or short discussion] be useful?”
- Product context: “Your product page describes [verified capability or audience]. How are you approaching [relevant process question] as that motion develops?”
Do not write “I saw you visiting our pricing page” when you only identified an account. The website visitor documentation distinguishes available visitor context, and the website-to-outbound motion should use that context without inventing a named person’s behavior. Signals support prioritization; they do not prove buying intent, permission, or a current project. Compare signal-based selling tools by setup effort before automating this research workflow.
Keep review requirements consistent with the enrollment route
With agentic sequences, starting from chat or a list generates recipient drafts for review and approval. Starting from a person’s record can start enrollment without separate draft approval. Choose the route that matches your required review process; do not assume every entry point pauses automatically.
For reusable personalization blocks, smart snippets can help standardize instructions. Review generated examples and define how a message should behave when required context is missing. A blank or invented first line is not an acceptable fallback.
Before enrollment, check the sender, recipient, current role, source-backed statement, offer, and exclusions. Apply the sequence personalization audit to the full sequence, including follow-ups. A careful first email does not justify an unsupported second message.
Measure research usefulness before claiming a performance lift
Freeze a cohort of attempted research records and a review cutoff. Define an acceptable record in advance: correct entity, accessible supporting evidence, relevant finding, and an appropriate next action. Research acceptance rate is accepted records divided by all attempted records in that cohort. Count failed, pending, and unreviewed records separately and retain them in the denominator. Report N/A when no records were attempted.
Measure human time including research setup, checking, correction, and message approval. Time per accepted record is that total time divided by accepted records; if none were accepted, report the time spent and N/A for the ratio. A faster draft is not necessarily a faster approved message.
Evaluate outreach outcomes separately. Predefine the recipient cohort, observation window, deduplication rule, and outcome—such as a human reply or a qualified meeting. Do not equate opens with interest or compare a new cohort to an old one without considering audience, offer, sender, and timing changes. The sequence results documentation explains native reporting; keep any custom metric explicitly defined rather than presenting it as the same calculation.
Make the workflow reusable without making every email identical
Save approved account and contact cohorts through list building with chat. Keep the evidence card, exclusion decision, and owner available to the next rep. The Unify sales-team case study provides a concrete example of using research and signals in an operating workflow, but its reported results belong to that team and period—not a universal SDR benchmark.
The goal is a controlled path from evidence to action. Automate repeatable retrieval and formatting, preserve uncertainty, and spend human attention where it changes the decision to contact someone or the accuracy of what you say.
Frequently asked questions
What should an SDR research before a cold email?
Confirm account fit, the recipient’s current role, one relevant supported observation, and whether outreach is allowed. Record what is unknown instead of inventing a reason to contact them.
How much time should research take?
There is no universal target. Measure total research and review time against accepted, useful outputs for your own audience and workflow.
Can AI research replace human review?
AI can retrieve and organize evidence, but identity, relevance, source support, and message claims still need an appropriate review process.
Does a website visit prove a contact is buying?
No. An account-level visit does not identify every employee or establish budget, urgency, permission, or a purchasing project.
Do all agentic sequence enrollments require draft approval?
No. In Unify, chat and list routes prepare drafts for approval, while enrollment from a person’s record can start without separate draft approval.
