Blog · updated 20 October 2026

Building a B2B prospecting engine, in public

Everything here comes from running the workflows ourselves: measured token counts, real CSV output, and the failures we hit before shipping. No “10x your pipeline” promises — the numbers are in the tables.

• The only part of your prospecting stack that can bill you while you sleep Creating and fencing Google Places (New) and DeepSeek keys before importing: per-API restriction, daily quota cap, budget alert, the real Places free tier (1,000 Enterprise requests/month) read from the official list today, and the 3-request test that tells “billing missing” apart from “bad field mask”.
• The 5 checks that tell you if a workflow JSON will actually import Node types against the instance catalog (and the trap in /types/nodes.json), what a typeVersion above your maximum really does, pinned data that fakes a working flow, hardcoded keys, and how many nodes are actually executable. With the numbers from our own runs on 2.43.6 and 1.60.0.
• The 429 is not where you think it is: DeepSeek concurrency vs Google quota DeepSeek limits concurrent connections per account (2,500 on flash, 500 on v4-pro), not a requests-per-minute quota. The kit's 4-item batches with a 1.2 s pause use 0.16% of that ceiling — so where the limit actually comes from, and which error codes should never be retried.
• Why a Places search returns 60 businesses and never more pageSize caps at 20, the token stops the loop, and 60 is the documented ceiling across all pages. What each extra page costs, and the four mistakes that silently stop you on page one.
• Google Places fieldMask: the one field that quadruples your lead bill The API bills the highest SKU present in the response and the mask decides which one. Real prices per SKU, what we request, and the exact monthly volumes where Places is free.
• n8n lead generation: the five stages of a pipeline that runs itself Google Places field-mask pricing, native pagination, batched LLM scoring and a measured cost of USD 1.81 per 1,000 leads.
• AI lead scoring for local businesses: a rubric that survives real data Six weighted signals and one veto, why the model must return an id, and how to calibrate minScore using the discards instead of your intuition.
• “It worked once”: why your n8n workflow returns 0 items on the second run A post-mortem of the deduplication node's memory — the evidence in the instance database and the one-click remedy.
• The AI node returns empty: max_tokens, reasoning and the JSON that never arrives Green node, HTTP 200 and an empty content. The tokens measured with reasoning on (351–800 per lead, one returned empty) and the three fixes that work.