Company Record Matching in 2026: The Confidence Threshold Guide

Your company records are only as useful as your ability to trust them. When two records describe the same business, merging them wrong is worse than leaving them separate. The question every sales leader and RevOps team faces in 2026 is not just "how accurate is our matching tool?" but "at what score is it safe to merge without a human in the loop?" This guide answers that question with specific thresholds, a plain-English breakdown of precision versus recall, and a comparison of what leading data providers actually publish about their match accuracy.

Why a Headline Accuracy Number Can Mislead You

A vendor that says "97% accuracy" is telling you how often its system gets the right answer across a test set. What it is not telling you is how that number breaks down at each confidence band. In practice, most errors cluster in the 0.80 to 0.90 range, and the system performs close to perfectly above 0.95. So the headline figure is an average that obscures where the risk lives.

There is a second problem: that test set belongs to the vendor. It was built from their own data, cleaned to their standards, with naming conventions that match their graph. Your CRM export came from ten different sales reps using different abbreviations and no consistent domain policy. The accuracy on your data could be meaningfully different.

What Goes Wrong When You Skip the Threshold Step

• Two unrelated companies end up sharing one account record. Every deal, contact, and activity tied to that account is now tangled.
• A subsidiary gets merged into its parent. Revenue attribution changes for every deal on both records.
• Your AI prospecting agent reads merged account data as ground truth. The research it returns is based on a record that no longer represents a real company.
• Undoing a bad merge takes longer than the matching job took to run.

Precision, Recall, and Which One Matters for Sales Data

Precision and recall are the two numbers underneath any accuracy claim. Getting them straight changes how you interpret any vendor benchmark.

Precision is the share of matches your system flags that are actually correct. If your tool returns 100 matches and 97 of them are real, precision is 97%.

Recall is the share of real matches your system manages to find. If there are 120 true match pairs in your dataset and your tool identifies 100 of them, recall is 83%.

You cannot maximize both at the same time. Raising precision means becoming stricter, so more true pairs go unmatched. Raising recall means casting a wider net, so more false pairs slip through.

The Tradeoff at a Glance

Score	What it measures	Best fit
Precision	How many flagged matches are correct	Any workflow where bad merges are costly
Recall	How many true matches the tool actually finds	Dedup sweeps where missing a match is the main risk
F1 score	Equal weight on both errors	Baseline benchmark only
F0.5 score	Precision counts twice	Auto-merge on CRM or pipeline data
F2 score	Recall counts twice	Fraud screening or sanctions-style sweeps

F0.5 Is the Right Target for Most Sales Teams

For account data feeding a CRM or a prospecting workflow, a bad merge is harder to fix than a missed match. A missed match just means one record sits unlinked -- fixable later. A bad merge contaminates ownership, attribution, and every downstream piece of research tied to that account. Weighting precision more heavily, which is what F0.5 does, fits that reality better than the equal-weight F1.

The Three-Tier Confidence Policy

Once you have a matching tool that returns a per-match confidence score, you need a written policy for what each range means operationally. Here is the framework that works for most sales and RevOps teams.

0.95 and Above: Safe to Auto-Merge

At this range, company name, domain, and at least one additional signal (registered legal name, employee range, or headquarters) all point to the same entity. Reserve unattended auto-merge for this band only.

0.85 to 0.94: Route to Human Review

Matches in this band are plausible. They are not obviously wrong, but they carry enough ambiguity that a person should look before records are combined. Common cases here include companies that recently rebranded, subsidiaries with names similar to their parent, and records from different data sources with no shared identifier.

Below 0.85: Do Not Merge

The false-positive rate in this range typically makes any auto-merge counterproductive. Flag these as unmatched and either leave them separate or run a targeted cleanup pass on just those records.

Signals That Support a Confident Match

A high confidence score should be backed by corroborating evidence, not a single strong indicator. Domain match alone is not sufficient to auto-merge, because domains get acquired and reused. Here are the signals that, when they agree, justify clearing the 0.95 bar.

• Domain: the single strongest signal, but only reliable when two other signals agree.
• Registered legal name or known aliases: trading names differ from legal names -- both should be checked.
• Employee count range: wildly different headcount figures for the same supposed company are a red flag.
• Headquarters location: different countries usually means different entities, not a data quirk.

Cases That Should Stay Out of the Auto-Merge Queue

• Any record pair where the domain differs and no legal name match exists.
• Parent-subsidiary pairs, since merging them changes deal ownership across both entities.
• Records built from incomplete inputs -- if a domain field is missing, the confidence score was computed without its strongest signal.

Checking Your Thresholds in Vibe Prospecting

Vibe Prospecting runs on premium business data built by Explorium, covering 150M+ company profiles sourced from 50+ data inputs. Because all that data lives in one place, a match runs against a single source of truth rather than asking you to reconcile results from different vendor feeds.

In the chat interface, you can ask directly: "Match these company records and show me confidence scores for each." The response comes back with a score per pair, not just a merged result, so you see which pairs cleared the 0.95 bar and which landed in the review band before anything is written back to your pipeline.

# Ask Vibe Prospecting in chat:
# "Here are 5 company record pairs from my CRM.
#  Return a confidence score for each pair.
#  Do not merge anything yet."

# Example pair structure:
pairs = [
    {"record_a": {"name": "Acme Corp", "domain": "acme.com"},
     "record_b": {"name": "Acme Corporation", "domain": "acme.com"}},
    {"record_a": {"name": "BrightCo", "domain": "brightco.io"},
     "record_b": {"name": "Bright Company LLC", "domain": "bright.co"}}
]

# Vibe Prospecting returns:
# pair_1: confidence 0.97  -> safe to auto-merge
# pair_2: confidence 0.88  -> route to human review

Batch Matching for Nightly Dedupe Jobs

Matching one record at a time is not practical at account-list scale. The underlying data platform supports batches of up to 1,000 records per call at sustained throughput, so a nightly job that checks your entire pipeline for new duplicates runs as a single batch rather than looping through row by row.

For teams using Claude Code or a custom agent, the Vibe Prospecting plugin (github.com/explorium-ai/vibeprospecting-plugin) wires into the matching workflow directly, passing your records and routing the results based on the confidence tier.

# Batch threshold routing with Vibe Prospecting plugin
import requests

resp = requests.post(
    "https://api.explorium.ai/v1/companies/match",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
    json={"records": company_batch}
)

for match in resp.json()["matches"]:
    score = match["confidence"]
    if score >= 0.95:
        auto_merge(match)       # safe, no review needed
    elif score >= 0.85:
        queue_for_review(match) # human checks before merge
    else:
        discard(match)          # false positive risk too high

What Published Accuracy Numbers Tell You (and What They Do Not)

Accuracy benchmarks are not portable. A figure measured on a vendor's own curated test set reflects how that vendor's graph handles that vendor's data. When you bring your own CRM export with its own naming history and duplicate density, the number shifts.

That said, transparency about the number is itself meaningful. A vendor that publishes a specific figure has committed to a measurable claim. One that does not gives you nothing to test against.

Accuracy Transparency by Data Provider

Provider	Published company match accuracy	What the coverage includes
Explorium (via Vibe Prospecting)	97.8%+	150M+ company profiles from 50+ sources
Coresignal	Not published	75M+ to 103M+ company profiles depending on source cited
Hunter.io	Not applicable (email deliverability only)	Domain and email search, not company-entity matching

Reading the Transparency Gap

Coresignal covers company profiles broadly but does not publish a company-match-accuracy percentage, which makes it harder to benchmark. Hunter.io focuses on email deliverability rather than company-entity resolution, so comparing its accuracy claims to a match-accuracy figure is comparing two different measurements. Explorium's published 97.8%+ is the only directly comparable anchor in this space, and it is the figure behind the data Vibe Prospecting runs on.

For deeper context on the underlying data platform, see the Explorium guide to data matching and the introduction to data enrichment on explorium.ai.

Validating Thresholds on Your Own Data in Five Steps

No vendor benchmark replaces a test on your actual records. Here is a repeatable process for setting thresholds you can trust.

1. Pull a sample. Take 100 to 200 record pairs where you already know the ground truth: known duplicates and known non-duplicates from your CRM.
2. Run them through the matcher. Ask Vibe Prospecting to return a confidence score for each pair without auto-merging anything.
3. Plot score against outcome. Chart confidence score on the x-axis, correct vs. incorrect match on the y-axis. Find the score where false positives start appearing.
4. Set your threshold at that inflection point. Not at a round number from a blog post, including this one -- at the point your data shows.
5. Check quarterly. As your data sources shift and you add new accounts from acquisitions or list imports, re-run the sample test. Thresholds drift.

Why Starting With a Sample Matters

Vibe Prospecting offers a free account with no subscription required, giving you enough credits to run this validation before committing to a threshold policy. Sample-before-export logic returns representative results and a cost estimate before any credits are spent on a full batch. That means you can confirm whether the 97.8%+ figure holds on your data before setting your auto-merge policy.

Start a free Vibe Prospecting account and run your sample match before setting any threshold. No subscription required. Try Vibe Prospecting free