Stop AI From Hallucinating Your Cold Email Personalization: The Firecrawl + Evidence-Rule Workflow

The Pipeline

1 CRAWL    Firecrawl: home, /about, /product,
           /blog (index), /pricing
           + reviews source if local/B2B-reviewed
2 EXTRACT  LLM, strict JSON:
           facts: [{claim, quote, url}]
           review_themes: [{theme, count}]
           sufficient: true/false
3 BRANCH   insufficient → fallback line
           (segment-pain template, honest)
4 WRITE    input = ONE selected fact/theme ONLY
           "restate as one natural observation,
           no additions, no flattery"
5 QA LOG   line + its evidence stored side by side;
           weekly sample: does the quote support
           the line? override rate → prompt fixes

• •Crawl narrow, extract hard: whole-site crawls cost more and bury the extractor in navigation soup; five pages carry nearly all the signal. Selection of WHICH fact to use is its own small call: prefer recent, specific, and offer-adjacent over impressive-but-irrelevant.
• •Ban flattery in the writing prompt explicitly: "loved", "impressed by", and "amazing" are the house style of hallucinated openers, and prospects have learned the smell (the saturation problem). Observation beats compliment: "You switched pricing to usage-based last quarter" outperforms "Love your innovative pricing!" everywhere.
• •The review-theme angle is the sleeper: for local and reviewed-B2B niches (pairs perfectly with Maps pipelines), aggregated review themes are pain statements written by the prospect's own customers, the most relevant evidence that exists, as long as it stays aggregated.
• •Run the split test honestly: grounded-with-fallback vs fully-generated arms, 200+ sends each, judged on positive replies. The grounded arm wins ours consistently, and the failure analysis of the generated arm is always the same: a handful of fabrications that turned warm prospects cold.

Why Architecture Beats Prompting

Every "just tell it not to make things up" attempt fails eventually because instruction-following degrades exactly when retrieval is thin, the model fills gaps, confidently, at row 4,000 of a quiet batch. Structural separation cannot degrade that way: the writer literally does not possess facts beyond the evidence payload, so its worst output is a clumsy sentence, not a false one. This is the same principle as the classifier's strict output contract and the scoring rubric's justification requirement: at volume, you trust structures, not intentions.