By KYN AI Advisory Team — AI implementation specialists, Singapore
To test a data vendor, send it 100 of your own records, verify every returned field by hand against a primary source, and score three numbers: how many records it returned, how many returned values were correct, and how recent they were. Do the test before you sign. A sample of 100 gives a rough read, accurate to about 10 percentage points at worst, which is enough to rule a vendor out or in for a first purchase.
Vendors publish headline figures for coverage and accuracy. These measure their data on their sample. Your leads, in your markets, may differ. The test measures the vendor on your data.
Why test on your own records?
Because accuracy varies by country, company size and field. A provider that is strong for US software firms can be thin for a Jakarta manufacturer or a Hong Kong trading company. Coverage for Southeast Asian small and mid-size companies is the usual weak point, so check it directly.
A test also gives you a baseline. Re-run it every six months, and you will know if quality is slipping.
How do you build the sample?
The sample decides whether the result means anything.
- Use records you can verify. Choose 100 companies where you can confirm facts from a primary source: the registry, the company's own site, or a customer you know.
- Mirror your target mix. If 60% of your targets are in Singapore and 20% each in Malaysia and Indonesia, the sample should look the same. Include small and mid-sized firms in proportion.
- Include some you know well. Add 10 to 20 records where you already know the correct values, as controls.
- Do not tell the vendor which records are controls. Send all 100 together.
- Freeze the input. Record the date you sent the sample, the fields requested and the identifiers supplied (name, domain, registration number).
If the vendor allows only a limited trial, ask for the sample size you need in writing. A vendor unwilling to be tested on your data is telling you something.
What fields should you request?
Request the fields you will use, not everything on offer. A typical set for B2B outreach:
- legal name and registration number;
- website domain;
- industry;
- employee count band;
- country and city;
- a business contact: role, name where lawful, and email;
- any trigger the vendor claims, such as hiring or funding.
Each extra field adds cost and a second place for errors. If you will not act on a field, do not buy it.
How do you verify the results?
Verify against primary sources, one field at a time.
| Field | Primary source | |---|---| | Name and registration number | Registry. See company data sources for Southeast Asia | | Website | Visit it; check the footer for the legal name | | Industry | Company site; registry activity code | | Employee band | Careers page, public statements; treat as approximate | | Contact role | Company team page | | Email | Source of publication; send-free verification where allowed | | Trigger | The cited URL |
Mark each returned value as correct, wrong, or cannot verify. Keep "cannot verify" separate; do not count it as correct.
Budget the time. Verifying 100 records across six fields is roughly 4 to 6 hours for one person. That is cheap against a year's contract.
Which three scores matter?
Compute them per field and overall.
- Fill rate. Records where the vendor returned a value, divided by 100. A vendor that returns nothing for 40 records has a coverage problem.
- Accuracy. Correct values divided by values you could verify. This is the number to watch. Report it per field, because the mix hides problems. A vendor can be 95% right on domain and 60% right on contact email.
- Freshness. For contacts and triggers, the share still valid today. Compare against the company's current team page and careers page.
Then compute usable yield: records with every field you need, correct. That is the number of leads you actually get. A vendor with a high fill rate and low accuracy can have a lower usable yield than one with the opposite profile.
How reliable is a sample of 100?
Not precise, but useful. If a vendor is correct on 90 of 100 verifiable values, the true rate probably lies within about 6 percentage points either side. At 50%, the margin is about 10 points. This is the standard sampling error for a proportion, and it holds if the sample is random.
Use it as a gate, not a ranking:
- Under 70% accuracy on a field you need: reject the field.
- 70% to 90%: usable with checks in your workflow.
- Over 90%: usable with sampling.
Choose your own cut-offs based on cost of error. An email field at 80% accuracy means one in five sends bounces, which hurts your sending reputation; Google asks senders to keep the spam rate in Postmaster Tools below 0.10% in its email sender guidelines. See out-of-office, bounces and unsubscribes: agent rules.
What else should you check before buying?
- Source. Where does the vendor get its contact data? Ask for the lawful basis in your market. See how to find business emails legally in Singapore.
- Terms. Can you store the data after the contract ends? Can you use it in an agent?
- Cost per usable record. Total fee divided by usable records, not by records returned.
- Waterfall behaviour. If a tool queries several providers in turn, ask which one answered each field, so you can see where errors come from.
- Refund or credit policy. Some vendors credit failed lookups. Read the terms.
How does this fit a custom research agent?
If the vendor's accuracy is poor in your market, a custom agent that reads registries and company sites may give better results for those fields, with a source link on each claim. If the vendor is strong on some fields, use it for those and your agent for the rest. See Clay vs a custom research agent for the comparison.
Where KYN fits
KYN builds research agents in the client's own cloud account and can run this test as a script: the same 100 records go through each source, results are stored with the source and time, and a person marks correct or wrong in a simple review screen. Every action is logged with the actor.
The first 14 days are a review of the vendors you already pay for and the sample your team would trust. A full build ships in under 90 days. See sales systems and integrations.
Frequently asked questions
Is 100 records enough to test a data vendor?
It is enough for a first decision. The margin of error is up to about 10 percentage points. If two vendors are within that range of each other, run a second sample of another 100.
What accuracy should I expect from a data vendor?
It varies by field, market and company size. Do not rely on a published figure. Measure on your own records and set a cut-off by field, based on the cost of an error.
How often should I re-test?
Every six months, and before any renewal. Contact data decays as people change roles, so a result from last year does not describe today.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
