How to analyze your LinkedIn connections with ChatGPT

Step 1: get the file

On desktop: Settings & Privacy → Data privacy → Get a copy of your data → Connections → Request archive. Tick Connections only; the full archive takes up to a day, Connections alone usually arrives in about ten minutes.

Inside the zip is Connections.csv. It has six columns: first name, last name, profile URL, email address (blank for most people), company, position, and the date you connected. No bios, no headlines, no company sizes. That is the whole of what you have to work with, and every limitation below follows from it.

Step 2: clean it

The file opens with a Notes: line and a paragraph of prose before the real header row. Open it straight into Excel or Sheets and everything lands in one mangled column, because the importer reads that prose as your header.

Delete the first three lines. Then delete the Email Address column, which is empty for roughly 90 to 98 percent of rows and is not useful for this. You want four columns: name, company, position, connected-on.

Step 3: write the brief

Before pasting anything, write down what you are selling and who buys it. Four or five sentences. This is the single biggest determinant of output quality, and it is the step people skip.

Be specific about the buyer: job titles, company size band, industry. "Anyone in operations" produces a list you cannot act on. "Heads of Operations or VP Ops at logistics firms with 100 to 1000 staff" produces one you can.

Step 4: the prompt

Paste your rows in batches of about 40. This is the prompt, and it is the same shape as the one running in production here, minus the parts that only matter at scale.

You are reviewing my LinkedIn connections to find people worth contacting about what I sell. For each person you get only: name, job title, employer, and the year we connected. You do NOT have bios, post history, or company size. Reason from job title, employer and connection age, and nothing else. Never state or imply a fact you cannot derive from those four fields: no "recently posted about", no "just raised". WHAT I SELL: [your brief]. WHO BUYS IT: [your buyer]. For each row return: score 0 to 100 for how high they should rank if I could only send twenty messages; one of customer / partner / investor / hiring / skip; and one line of at most 18 words explaining the call. Most people are skip. A network of 3,000 typically yields 100 to 200 worth any contact. Being generous wastes my time and my reputation.

The two instructions doing the most work are the one forbidding invented facts and the one insisting most people are skip. Without the first you get plausible-sounding fiction about real people, which is worse than no answer. Without the second you get 60 percent of your network marked interesting.

Where this breaks

All four of these are things I hit building a tool that does this, measured on a real 1,601-connection network. None of them make the manual approach wrong. They are just the point at which it stops being an afternoon.

1. Ranking does not survive batching

This is the big one. To rank 3,000 people you have to compare them against each other, and no assistant can hold 3,000 rows in one conversation. So you batch, and each batch is scored in isolation.

Scores drift between batches. An 85 in your third paste is not calibrated against an 85 in your thirty-eighth, because the model has no memory of the earlier ones. Sort the merged results and the top of the list is partly an artefact of which batch someone landed in.

I have a hard number for how much this matters. Running the same network through a smaller model and a larger one, both correctly identified a venture partner as an investor. The weaker ranking put him 71st. The stronger one put him 1st, and he was, unambiguously, the single best contact in the network. Same data, same categories, completely different answer to the only question that mattered.

2. Company names do not normalise themselves

Acme Ltd, ACME Limited, Acme Ltd. and Acme are one employer and four strings. So are Société Générale and Societe Generale. A pivot table treats them as four companies and two companies respectively.

On a 1,601-row export, normalising properly collapsed about 1,400 apparent employers to roughly 1,000 real ones. Doing that by hand is a genuine afternoon, and doing it badly quietly splits your best accounts.

3. The row does not say how big the company is

Your buyer profile says 100 to 1000 staff. The row says VP Operations | Halden Freight. Nothing in the file tells you whether Halden Freight has twelve people or twelve thousand, and that single fact decides whether this is your buyer or a waste of a message.

An assistant will know the large, recognisable employers and not the rest. Asked to identify 1,411 employers from a real network, a strong model returned a confident answer for about a third and honestly admitted it did not recognise the other two thirds. That is the correct behaviour, and it means most of your rows go into the analysis with the most important qualifier missing.

4. Consistency, over 40 pastes

Forty batches is forty copy-pastes, forty responses to collect, and forty opportunities for the output format to wander. You will get a stray label, a row scored on a different scale, an answer that quietly drops three people. At batch nine you are still reading carefully. At batch thirty-one you are not.

When doing it manually is the right answer

• Under about 500 connections. One or two pastes, no batching problem, no ranking problem. Just do it.
• One specific question. "Which of these people work at agencies?" is a filter, not a ranking, and filters batch fine.
• You know your network well. If you can recognise the right answer when you see it, you do not need a system to rank it for you.
• You want to understand the problem before paying anyone to solve it. Run two hundred rows by hand. You will learn more about your own buyer profile than any tool will tell you.

What a tool actually adds

Not intelligence. The same models are available to you directly, and the prompt above is most of the work.

What it adds is the unglamorous half: normalising employers so they group, identifying company size so your buyer profile means something, keeping scoring consistent across the whole network rather than within a batch, re-ranking the top slice with a stronger model so the ordering can be trusted, and checking the web for the handful of people worth checking. That is the part that is tedious rather than clever, and it is the part that takes the afternoon.

If that sounds like something you would rather not spend a Saturday on, that is what this site is. If it sounds like an afternoon well spent, the prompt above is genuinely the whole method and you should use it.