Research Patterns

This reference defines two research contexts:

1. Self-Research (Step 1) — Deep research on the user’s own company to build a strong ICP foundation
2. Target Research (Step 6) — Research each discovered company using Plan→Research→Synthesize

Both use the same 3-phase pattern but with different sub-questions and goals.

Self-Research (User’s Company)

This is the most important research in the pipeline. Every downstream decision depends on it.

Sub-Questions

• “What does {company} sell and what specific problem does it solve?”
• “Who are {company}’s existing customers? What industries, company sizes, and use cases?”
• “Who are {company}’s competitors and what differentiates them?”
• “What pricing model does {company} use and who is the typical buyer persona?”
• “What use cases and pain points does {company}’s marketing emphasize?”

Page Discovery

Discover site pages dynamically — do NOT hardcode paths like /about or /customers:

1. Fetch browse cloud fetch --allow-redirects "{company website}/sitemap.xml" — primary source, has ALL pages
2. Scan sitemap URLs for keywords: customer, case-stud, pricing, about, use-case, blog, docs, industry, solution
3. Optionally fetch browse cloud fetch --allow-redirects "{company website}/llms.txt" for page descriptions
4. Pick the 3-5 most relevant URLs from the sitemap and fetch those
5. Sitemap is the source of truth. llms.txt is bonus context but often incomplete.

External Research

• Search: "{company} customers use cases reviews"
• Search: "{company} alternatives competitors vs"
• Fetch 1-2 of the most informative third-party results (G2, blog posts, comparisons)

Synthesis Output

From all findings, produce a company profile:

• Company: name
• Product: what they sell, how it works, key capabilities (2-3 sentences, specific)
• Existing Customers: named customers or customer types found
• Competitors: who they compete with, key differentiators
• Use Cases: broad list of use cases the product serves (NOT tied to one vertical)

Do NOT include ICP, pitch angle, or sub-verticals in the profile. Those are per-run targeting decisions made in Step 2 after the profile is confirmed. The profile is a general-purpose company fact sheet that works regardless of which vertical you target next.

Why This Matters

A thin profile produces generic search queries, weak lead scoring, and cookie-cutter emails. A rich profile with specific customers, competitors, and use cases produces targeted queries, accurate scoring, and emails that reference real pain points.

Target Company Research (Step 6)

Sub-Question Templates

Generate sub-questions from these categories based on the ICP and enrichment fields requested. Not every category applies to every company — pick the most relevant.

Priority 1 (Always ask)

• Product/Market: “What does {company} sell and who are their customers?”
• ICP Fit: “How does {company}’s product/market relate to {sender’s ICP description}?”

Priority 2 (Ask in deep/deeper)

• Tech Stack: “What technologies, frameworks, or infrastructure does {company} use?”
• Growth Signals: “Has {company} raised funding, launched products, or expanded recently?”
• Pain Points: “What challenges might {company} face that {sender’s product} addresses?”

Priority 3 (Ask in deeper only)

• Decision Makers: “Who leads engineering, product, or growth at {company}?”
• Competitive Landscape: “Who are {company}’s competitors and how are they differentiated?”
• Customers/Case Studies: “Who are {company}’s notable customers and what results do they highlight?”

Search Query Patterns

For each sub-question, generate 2-3 search query variations:

Finding Format

Each finding is a self-contained factual statement tied to a source:

Confidence levels:

• high: Directly stated on the company’s own website or official press
• medium: Inferred from job postings, third-party articles, or indirect signals
• low: Speculative based on industry/category, or from outdated sources

Research Loop Rules

1. Process sub-questions by priority — Priority 1 first, then 2, then 3
2. 3-5 findings per sub-question, then move on — Don’t exhaust a topic
3. Use parallel tool calls — Search multiple queries simultaneously when possible
4. Rephrase, don’t retry — If a search returns poor results, try different keywords
5. Fetch selectively — Don’t fetch every URL from search results. Pick the 1-2 most relevant based on title and URL
6. Stop at step limit — Respect the depth mode’s step budget per company
7. Homepage first — Always fetch the company’s homepage before branching to other pages
8. Deduplicate findings — Don’t record the same fact twice from different sources

Depth Mode Behavior

Quick Mode (100+ leads)

• Skip Phase A — No sub-question decomposition
• Phase B: Fetch the company homepage. Run 1-2 supplementary searches if homepage data is thin.
• Phase C: Extract available data, score ICP, write email from what’s available
• Budget: 2-3 total tool calls per company
• Trade-off: Fast and cheap, but emails may be less personalized

Deep Mode (25-50 leads)

• Phase A: Decompose into 2-3 sub-questions (Priority 1 + selected Priority 2)
• Phase B: For each sub-question, run 2-3 searches + fetch 1-2 URLs. Target 3-5 findings per sub-question.
• Phase C: Synthesize from all findings. ICP reasoning references specific evidence. Email uses the most specific/compelling finding.
• Budget: 5-8 total tool calls per company
• Trade-off: Good balance of depth and scale

Deeper Mode (10-25 leads)

• Phase A: Decompose into 4-5 sub-questions (Priority 1 + 2 + selected Priority 3)
• Phase B: Research exhaustively. Fetch multiple pages per company (homepage, about, blog, careers, product pages). Target 3-5 findings per sub-question.
• Phase C: Synthesize with cited evidence. ICP reasoning is detailed. Email references multiple specific signals.
• Budget: 10-15 total tool calls per company
• Trade-off: High quality intelligence, but slow and expensive

Synthesis Instructions

After the research loop completes for a company, synthesize findings into the output record:

ICP Scoring

Score 1-10 using ALL accumulated findings as evidence:

• 8-10: Strong match. Multiple high-confidence findings confirm right industry, company stage, and clear pain point alignment. The pitch angle directly addresses a visible need supported by evidence.
• 5-7: Partial match. Some findings suggest relevance but key signals are missing or low-confidence. Adjacent industry or unclear pain point.
• 1-4: Weak match. Findings indicate wrong segment, too large/small, or no apparent connection to sender’s product.

Write icp_fit_reasoning referencing specific findings: “Series A fintech (from Crunchbase), uses Selenium for scraping (from job posting), expanding to EU market (from blog) — strong fit for browser infrastructure.”

Email Personalization

Use the richest, most specific findings for email context:

• Opening: Use the most concrete finding (a specific product feature, a recent launch, a job posting)
• Bridge: Connect a finding about their challenges/stack to the sender’s pitch angle
• If only low-confidence findings exist, keep the email shorter and more general — don’t fabricate specificity

Enrichment Fields

Map findings to enrichment fields:

• product_description → from Product/Market findings
• industry → inferred from Product/Market
• employee_estimate → from LinkedIn search or careers page findings
• funding_info → from Growth Signals findings
• headquarters → from company homepage or about page
• target_audience → from Product/Market findings
• key_features → from product page findings

If a field has no supporting findings, leave it empty rather than guessing.

Anti-Hallucination Rules

Apply these at synthesis time. They exist because the failure mode — especially on Framer/Next.js landing pages with little server-rendered copy — is for the subagent to pattern-match visual cues onto the sender’s ICP and fabricate a plausible-sounding description:

1. Typography is not a product. Never infer product_description, industry, or target_audience from fonts, design system, framework choice (Framer, Next.js, React), or site polish. “Framer-built” and “uses Geist Mono” are observations about tooling, not signals of what the company sells.
2. No ICP leakage. If the homepage is thin and external search turns up nothing, do NOT default the target’s description toward the sender’s ICP. Manufacturing AI ≠ browser automation just because both use AI.
3. Quote, don’t paraphrase from memory. product_description must quote or closely paraphrase a specific phrase from extract_page.mjs output (TITLE / META_DESCRIPTION / OG_DESCRIPTION / HEADINGS / BODY) or from an external search result. If no such phrase exists, write Unknown — homepage content not accessible.
4. Cap scores on thin evidence. If product_description is Unknown, set icp_fit_score ≤ 3 and icp_fit_reasoning: Insufficient evidence — homepage returned no readable content. Do not justify a higher score on inferred signals alone.

• references/research-patterns.md(opens in a new tab)
• Raw file(opens in a new tab)
• browserbase/skills(opens in a new tab)