GenPilot research notes
What our lead-search tests show, and what they do not
More search results do not necessarily mean more useful prospects. On October 24, 2026, we tested a public business-research brief in an isolated production QA account. Here is what happened, including the limitations.
Reviewed by the GenPilot team · Updated October 24, 2026
What matters
- Search pages and identified businesses are different counts
- Deeper search was incremental, not a controlled A/B benchmark
- Published phone numbers are not ready-to-send email addresses
- A duplicate found during review reduced the apparent shortlist
The brief and test method
The test looked for independent wine retailers in Tours, France, with public contact routes, for an inventory-software offer. It excluded supermarkets, producers without a retail outlet, restaurants and online-only sellers. Target fit was not treated as evidence of active buying intent.
We ran standard search, then deeper search on the same QA list. Previously seen candidates were excluded from the second run, so deeper results represent additional records, not an independent total or a fair speed comparison. No outreach emails were sent.
Observed results before manual duplicate review
Standard search completed in 153 seconds: 24 source observations, seven entity records assessed, three shortlisted. Two had draftable contact routes; another still needed contact information.
Deeper search completed in 308 seconds across three rounds: 47 source observations, nine additional entity records, five initially shortlisted. These are observations from one brief, not expected yields for other searches.
The duplicate that changed the interpretation
Manual review identified the same business represented in two languages with differently written locality information. Counting those records as separate prospects overstated the useful yield. We deployed a repair that flags that exact identity pattern for review rather than silently merging records.
Replaying the affected records through the repair left three automatically shortlist-eligible records from that deeper run. This was a deterministic replay, not a second live search proving a new conversion or accuracy rate.
Contact attribution is not contact verification
Contacts included published phones, a directory listing and a provider-returned professional profile/email. We did not independently prove current ownership, delivery or phone reachability. We did not measure responses or sales.
A separate narrow search around Eidsvoll, Norway, produced 57 source observations and two identified businesses, but no shortlist: the requested recent-registration and missing-website conditions could not be established. More retrieval did not supply the missing evidence.
How to evaluate your own search
Start with one explicit brief and inspect a small sample. Count distinct relevant businesses, supported contact routes and unresolved criteria—not just rows or a high fit score. Change an unrealistic constraint deliberately rather than lowering standards invisibly.
- Can the source establish the business and geography?
- Is the named contact and role attributable to that business?
- Is a buying signal actually supported, or only inferred?
- Which criteria are unknown, and are any records duplicates?
- Would you contact this prospect after reviewing the source yourself?
Questions
Does deeper search guarantee more leads?
No. It spends a larger, bounded research budget looking for additional candidates. Source coverage, duplicates and your evidence requirements still determine what can be shortlisted.
Is this an independent performance benchmark?
No. It is a first-party QA observation from a small number of briefs. The two modes ran sequentially, not on equivalent independent samples. It does not establish accuracy, deliverability or revenue performance.
Put the workflow to work