An empty agent search needs more than a zero-result counter

An agent finishes a lead-search run and reports zero qualified companies. Before changing the prompt, establish what reached zero.

The source may have returned nothing. It may have returned companies that were all rejected. Company selection may have succeeded while contact discovery failed. The run may have stopped before it finished.

Those outcomes need different next actions. Combining them into one empty-state message makes debugging harder and can lead a team to broaden a useful search brief for the wrong reason.

Keep the original brief attached to the run

A diagnostic record should retain the offer, geography, eligibility rules and exclusions that the run actually used. Store a revised brief separately when the next run changes them.

Otherwise, a later success can make the original search appear successful even though it answered a different question. Expanding from one country to several countries changes the experiment.

Our MachGTM note on empty lead searches uses this distinction to separate a bad brief from source limitations and execution problems.

Preserve the stages where candidates disappear

For each source, I would record returned candidates, duplicates, reviewed candidates, accepted candidates and rejection reasons. I would also retain the source outcome: completed, failed or unfinished, with the budget consumed.

These are proposed diagnostics. The counters require explicit definitions. For example, a duplicate returned by two sources should not become two unique companies in an overall total.

A compact report could say:

• Source discovery completed and returned candidates.
• Deduplication reduced the unique set.
• Qualification rejected the reviewed set.
• Contact lookup did not run because no company qualified.

That report tells the next investigator where to start. “Zero leads” does not.

Inspect the dominant rejection reason

If most rejected companies were sellers of the same service, inspect examples from that group. The query may be finding supply instead of demand. If geography dominates, check whether the source actually supports the geographic constraint.

Inspect some accepted examples in non-empty runs too. A system that never returns zero may simply be accepting poor matches.

Do not treat a sample review as a measurement of recall. You need a defined relevant population to establish how many qualified companies were missed. A few plausible results cannot supply that denominator.

Separate fixtures from campaign evidence

Mach's published note includes constructed tests and a source review dated 5 November 2026. Those examples explain expected diagnostics; they are not results from a live sales campaign.

That distinction matters in an agent dashboard. A passing test can show that a report labels a known fixture correctly. It cannot show that the source covered a real market or that the qualification rules found the right buyers.

Make the next run test one explanation

Choose a bounded change based on the evidence: repair a failed source, revise one ambiguous rule, or try a query that describes the buyer's work. Keep the prior run available for comparison.

The companion guide on testing lead-search instructions provides context for reviewing the brief itself. At MachGTM, the useful outcome is a search a seller can explain—including why it stopped without a list.

Adapted with AI from MachGTM's published notes. The reporting structure is a design recommendation. This article does not report a new live search or a measured improvement in lead quality.