How it works
The whole pipeline, including the parts that are unglamorous and the parts that deliberately refuse to answer. Figures below are measured on one real 1,601-connection network, so read them as an order of magnitude rather than a benchmark.
Employers are identified before anyone is judged
This is the stage that does the most work and gets the least attention. Your export gives a job title and an employer name. It does not say how big that employer is, what industry it is in, or where it is.
Which matters, because a buyer profile almost always contains a size band. “VP Operations at Halden Freight” is your buyer if Halden Freight has 400 staff and irrelevant if it has nine, and the row reads identically either way.
So every distinct employer is resolved first. That is a larger job than it sounds: the network we measured contained 1,411 different employers across 1,601 people, because most people in a long professional life work somewhere different from everyone else you know.
What the honest answer looks like
Across those 1,411 employers:
- 181 (13%) were identified with high confidence
- 335 (24%) with medium confidence
- 206 (15%) with low confidence
- 688 (49%) were not recognised at all
The low-confidence and unrecognised ones are withheld from the analysis entirely. They are not passed through with a hedge, and they are not guessed at. About a third of your employers contribute facts; the rest are analysed on job title and connection age alone, exactly as if the identification step had never run.
A confidently wrong headcount is worse than no headcount. It puts your best prospect in the discard pile because a model decided their employer was a ten-person shop, and you never find out. A tool that admits it does not recognise two thirds of your employers is telling you something true about what is possible from a file containing only a company name.
Employer facts are stored and reused across every scan, so the same employers are never paid for twice. It is the one thing here that gets better as more people use it, and it contains no personal data: headcount bands and industries, nothing about you or your connections.
Every connection is read, then the strongest are re-read
Each connection is scored against your written brief, with its employer facts attached where they exist. Not filtered, not keyword matched, scored.
Then the top slice is re-read by a larger model, and this second pass is not a formality. Running the same network through a smaller and a larger model, both correctly identified a venture partner as an investor. The smaller one ranked him 71st. The larger one ranked him 1st, and he was, unambiguously, the best contact in the network: an Index Ventures partner who had founded and sold a travel startup, at a moment when the person searching was raising for an AI travel startup.
Categorising someone is easy and every capable model does it well. Deciding that one person out of 1,601 is the one who matters is not, and that is the only question the top of a list has to answer.
The shortlist gets checked against the public web
For your strongest matches only, a web search looks for what has changed: a role move, a funding round, a public statement of a problem you solve. It reads public pages, and every finding is shown with its sources.
This is the only place in the report where a claim about recent events is permitted, and it is permitted only because a source was checked. Everywhere else the analysis is confined to job title, employer and how long ago you connected.
Identity is the risk, and it is treated as one
Common names collide constantly. Reporting one person’s news about another is worse than reporting nothing: you send a message referencing something that never happened to the recipient, and you burn a relationship you cannot get back.
So a finding is only reported when the source ties it to that person at that employer. Someone with the same name at a different company is not them. When the name is common and cannot be disambiguated, the report says so and stops. You will see reports that flag their own uncertainty, like “his title reads Head of IT Infrastructure, not CISO, so do not lead with the CISO framing”. That is the system working, not failing.
And the companies you reach nobody at
The paid report ends by inverting the question. Given the same buyer profile, which companies match it where your network reaches no one? No warm route in, which is precisely why it is worth knowing they exist.
This part is company-level only. No person from any other customer’s upload is read, surfaced or inferred, which is what keeps it on the right side of both the promise on this site and the law.
What it deliberately will not do
- Read anyone’s profile. There are no bios in the export and none are fetched. Getting them would mean scraping LinkedIn.
- Append contact details. The export contains almost no email addresses and that gap is not filled from anywhere else.
- Send anything. You get a draft. Sending it is entirely yours.
- Invent a reason. No “recently posted about” unless a source says so.
What happens to your file
It is stored in Frankfurt, deleted once a paid report is delivered, and within seven days for an unpaid scan. The report keeps its own copy of the matches so your link keeps working after the upload is gone. The full detail is here, including exactly what the AI model receives and what it does not.
See it on your own network
The free scan runs the first four stages and shows you the counts plus a worked example. No account, no card.
Scan my network, free