Why Did I Get Few or Wrong Leads?

Most low-yield runs have one of five causes. Here's the diagnostic flow.

1. The prompt is too narrow

Symptom: Total yield < 10 candidates, ICP router classified the prompt as mixed, sources include linkedin, crunchbase, web but no high-yield source like github, osm, or news-rss.

Fix: Add specifics:

Bad: "founders in India" (yields investment firms, students, stealth-mode profiles)
Good: "Indian SaaS founders" (yields ~30 GitHub-sourced operators)
Better: "Indian B2B SaaS founders raising Series A" (yields ~40 with funding-stage signals)

2. The geography is off the beaten path

Symptom: Total yield < 20 candidates for an SMB-local prompt; OSM dominates but contributes only ~10-20 entries.

Fix: Some geographies have thinner OSM coverage. Try:

Broaden the city: "Pune" → "Maharashtra" → "India" (the system still favors specific localities but widens search radius).
Add the country explicitly: "Cebu Philippines" not just "Cebu".
For tier-3 / rural geographies, the directory hunter often outperforms OSM — but its coverage varies country-to-country.

3. Anti-bot windows on SERP-based hunters

Symptom: LinkedIn / Crunchbase / Web hunters return 0; OSM and News-RSS work fine. Conductor logs include pool_http_502 or searxng_circuit_open.

Fix: This is transient — SERPs (DDG / Bing) periodically rate-limit our outbound IPs. Wait 15-30 minutes and re-run. Or accept that GitHub + OSM + Directory carried the run, which is often the higher-quality subset anyway.

4. The ICP shape was misclassified

Symptom: SwarmActivityFeed banner shows "shape=tech-founder" but you wanted SMB-local; or vice versa.

Fix: Reword the prompt to bias the router:

For SMB: include a physical-place noun (clinic, shop, firm, agency, studio).
For tech-founder: include a tech-industry term (SaaS, fintech, B2B software).
For B2G: include tender, govt, public sector.

5. Junk candidates with placeholder names

Symptom: Candidates with name: "LinkedIn Profile" or company: "Self-employed" slip through.

Fix: This was a known parser issue fixed in May 2026. If you still see these, the candidate was probably created before the fix — they'll be filtered next swarm. If you see them in a fresh swarm, please raise a bug — the parsers should reject them outright now.

Universal sanity checks

Before you blame the system, sanity-check:

Status flipped to `complete` — if it's still scraping after 5 minutes, the swarm is still in progress. Wait.
Hunters dispatched expected sources — the SwarmActivityFeed shows a "started" event for each hunter. If you only see 2 sources for what should be a 5-source ICP, the conductor's source mix is too narrow — re-run.
No `Self-employed.com`-style fake emails — those should never appear post-May-2026; if you see one, file a bug.

When to escalate

If your team is running 10+ swarms a week and consistently getting < 30% of expected yield for known-good prompts (e.g. "Dental clinics in Berlin" should yield 80+), open a support ticket. The most common root causes for sustained underperformance are:

A fresh CFN drift event has detached IAM roles from the hunters (we've automated detection but there's a 1-2 hour false-positive window during deploys)
The scraper-pool fleet is below capacity (auto-scaling has a 5-minute lag)
A specific hunter's source has changed its HTML markup and the parser needs an update