Supernova for Contacts: Benchmark Results

Why DiscoLike can build this

Customers tell us every day which companies they want and which people they want to reach there. Supernova was built on customer ICP searches and company fit validations: the companies customers search for and the criteria that decide whether a company fits. Our contact wizard, typed contact requests, and title lists supplied by users add the other half, the people customers want at those companies. Together, that record shows what a good person-at-company result looks like for a given request, and what a bad one looks like.

Getting the company right is not enough. Plenty of contacts look right on a keyword match and are not the buyer:

• An assistant to the CEO is not the CEO.
• A recruiter for engineers is not an engineer.
• A former VP is not a VP.
• Someone who sells to hospitals does not work in one.

We trained the model to separate these cases. In the top 50 results, these wrong-role traps fell from 7.2% to 0.3%.

How it works

Our previous contact search matched a request against a written summary of each person. Supernova for Contacts replaces that summary with a compact profile: the person’s title, plus their company’s industry, sub-industry, size, and revenue.

A small model trained specifically for contact fit turns each profile into an embedding. It is multilingual: in our evaluation, titles in other languages separated wrong-role traps nearly as well as English titles. It understands seniority ladders and about 3,700 job-title abbreviations, and it reads ambiguous abbreviations such as CPO, CSO, or MD in the context of the company’s industry.

At search time, Supernova scores the company of every candidate contact, and that company fit is combined with role fit from the contact model into one score, so both the person and the company decide a contact’s position.

The gain holds for every kind of request

Role fit in the top 10, by where the request came from:

Request type	Previous contact search	Supernova for Contacts
Contact wizard requests	62.0%	88.3%
Requests typed by users	62.1%	83.2%
Title lists supplied by users	73.0%	98.4%
Derived from company searches	62.1%	88.2%
Clients held out from all tuning	60.6%	89.3%

Fit stays high deep into the list, which matters because contact audiences are large: 72% at ranks 101 to 1,000 and 62% at ranks 1,001 to 10,000.

What the model learned

Measured on job titles the model never saw in training, compared with the previous version of the model:

Held-out titles	Previous version	Supernova for Contacts
Ranks the right seniority above the wrong one	96.6%	98.6%
Matches an abbreviation with its meaning (CTO and Chief Technology Officer)	91.6%	97.2%
Reads an ambiguous abbreviation by industry (CPO, CSO, MD)	53%	71%

Evaluation methodology

All evaluations ran on October 28 and 29, 2026, and compare Supernova for Contacts with DiscoLike’s previous contact search on the same requests and starting people. An automated language model judge assessed every result; these scores are automated assessments rather than human qualification decisions.

Contact fit uses 386 requests built from real company searches, each with a target company profile, a country, and a persona to reach. A result counts only when the role and the company both fit. Results still spread across about 53 different companies in every 100 contacts. The single weight that combines role fit and company fit was tuned on 313 requests and checked on the remaining 73, which came from different clients. The 91.8% company fit is close to the 93.1% that Supernova company search alone returns in its top 10 on the same requests. This evaluation applied the country filter, but not employee or revenue filters. Role fit in this test (77.8%) is lower than in the role fit test (89.5%). The role fit test ranks job titles on their own, while this test searches each contact together with their company profile and the judge assesses role and company in one pass.

Role fit uses 2,011 contact requests: 785 real requests from the contact wizard, typed requests, and title lists supplied by users, plus 1,226 requests derived from real company searches and run through the same contact wizard customers use. Real requests are capped per client so that no single customer dominates, and about one client in five was held out from all tuning.

For lookalikes, we picked 100 random people as starting points and retrieved the most similar people for each. A result counts as a lookalike when it has the same function and seniority at a similar company. The judge saw the same fields the new contact profile uses, which somewhat favors the new search.

Results at ranks 1 to 100 were all judged. Deeper bands were sampled and weighted by their size, the same depth reporting as the Supernova benchmark. In total, the evaluation used about 1.5 million automated judgments.

The judge reads each person’s title and company. It does not verify that the person still holds the role today. These benchmarks measure the share of returned contacts that fit, not the share of all matching people found, and time or cost savings were not measured.

Where this is going

Supernova made company discovery about ICP fit. Supernova for Contacts extends that to the people at those companies, so a single request for a person at a company goes from the market you describe to the people you need to reach there. It is one step in a broader change we are making across DiscoLike: optimizing lead discovery end to end, from account to contact, around fit.

Start with your target market

Start with the company criteria in the DiscoLike ICP generator. For agent workflows, see the MCP setup guide or company search API.

---

Related reading:

• Introducing DiscoLike Supernova: A Purpose Built ICP Fit Model
• Semantic Contact Search vs Title Matching
• Beyond Lookalikes: ICP Fit and Full TAM Search
• Title Search Alone Finds the Wrong Companies