Swapping an LLM for a scored Places API lookup lifted a sales pipeline's website coverage to 81%

The author's diagnosis rests on one question: "Does this step need judgement, or does it need an index?" ^[11] In practice it becomes an assertion. If `expected == actual` is a sensible check for a step, the author wrote, the step is a lookup and a model is the wrong tool for it ^[12]. A company's website passes. It has one correct answer, it is a fact about the world, and someone already maintains an index of it ^[13].

The LLM version failed that check in ways that were hard to see early. It took about twenty minutes to wire up and looked fine for the first hundred leads ^[9]. Nothing else in the pipeline could run without a website ^[1]. Re-running a lead could change the returned domain, so there was nothing to assert against and no way to tell a regression from ordinary variance ^[6]. Nondeterminism does have one administrative upside: a step with no fixture never officially regresses. The team had filed the 34% website rate as a data problem, blamed on small forwarders with thin web presence, and the author now says it was not one ^[10].

According to the author, the Places call was the easy part and the layer around it made the swap work ^[15]. That layer runs in order:

1. Normalise the name before querying: legal suffixes such as Ltd, GmbH and SARL, punctuation, casing and transliteration ^[16]. 2. Query on name plus city plus country. Without the geography, the index returns a real business with a similar name in the wrong country ^[17]. 3. Score each candidate on token overlap with the normalised name, geographic distance and business-type agreement, and return nothing below a threshold ^[18]. 4. Cache the result on normalised name plus country, so a second campaign against the same company costs nothing ^[21].

Step three is the part I would copy first. An empty result flags the lead and routes it to manual review ^[19]. "Returning a confident wrong answer is not a state anything can handle, because nothing downstream knows to doubt it," the author wrote ^[20]. The LLM path produced exactly that state: plausible domains for companies that did not own them, found only when outreach reached the wrong firms ^[5].

The reported figures are completeness, measured on the same lead set before and after ^[23]. Website completeness rose 47 points ^[1]. The share of leads with no usable website fell from 66% to 19% ^[2]. The supplied text of the post does not include a wrong-match rate or the threshold value. So the coverage gain is measured, and the claim that answers are also right more often rests on the threshold design. The 91% phone figure has no baseline, because the LLM path often returned no phone at all and the metric went untracked [c3, c22].

For the 81% to transfer, your entities have to sit in an index someone maintains, and your records have to carry enough geography to key the query [c13, c17]. I would also expect the threshold to need tuning against your own names before the numbers look like these.

The zero applies to LLM spend on this step ^[4]. A Places lookup still costs something, which the author puts at a fraction of a cent ^[14]. The larger change is in how cost scales. The LLM step's cost grew linearly with lead volume, and the same company across three campaigns meant three paid lookups [c8, c7]. With the cache keyed on name and country, the repeat lookups drop out ^[21].