You followed the guide. You opened the GPT builder, named it "Sourcing Assistant," uploaded a few job descriptions and your best outreach templates, and told it to act as your sourcer. It writes a clean Boolean string. It drafts three candidate personas. It even spits out an outreach template in something close to your voice.
Then you ask it for actual candidates. And it hands you nothing.
This is the gap nobody mentions in the fifty "build a sourcing GPT" tutorials published this year. A custom GPT is a writing partner that happens to know recruiting. It is not a sourcer. The model is smart enough. It simply can't see anything that matters for sourcing, and no prompt fixes that.
What a custom GPT actually does when you ask it to source
A custom GPT built on ChatGPT or Claude is good at exactly the things language models are good at. It turns a messy role description into a structured Boolean query. It suggests adjacent titles and skills you might have missed. It drafts a first-pass outreach message. These are real, useful jobs, and we walk through six of them in ChatGPT for recruiters. Build a good Custom GPT and it earns its keep every week.
Notice what every one of those tasks has in common. They all produce text. None of them produces a candidate.
The model never leaves the chat window. When you ask it to "find senior backend engineers in Austin open to a move," it cannot run that search anywhere. So it does the only thing it can: it generates a plausible-looking answer. Sometimes that means inventing names and profiles that do not exist. You end up doing the real sourcing yourself, in LinkedIn and your database, exactly as before.
A model with no data is blind
Access is the wall, and a custom GPT runs straight into it. It cannot log into your LinkedIn Recruiter seat. LinkedIn sits behind a login wall and runs aggressive bot detection, and it has actively blocked automated AI agents from reading profiles. It cannot reach into your ATS or CRM to see who you have already worked, already rejected, or already placed. It cannot pull a live candidate list from a resume database. And it has no memory between sessions, so it never learns that for this client you always pass on agency-only backgrounds. Every chat starts from zero.
Stack those limits and the picture is clear. The model is articulate and completely uninformed. It can reason beautifully about a candidate pool it is not allowed to look at. That is why the output reads smart and helps you not at all when you need names, contact paths, and a read on who will actually reply.
This is also why "I built a GPT to do my sourcing" almost always quietly becomes "I built a GPT to help me write my searches." The second sentence is true and useful. The first one was never possible.
But ChatGPT has Operator now, doesn't it?
Fair objection. ChatGPT's Operator, Perplexity's Comet, and Claude's Cowork are browser agents: they drive a real browser, so they can open LinkedIn, click through profiles, and copy what they see. On paper that closes the access gap. In practice it bends the wall rather than removing it.
A browser agent works the way a very slow human does, one page at a time, inside a logged-in session you have to supervise. LinkedIn's bot detection flags that pattern fast, which is why pushing one too hard risks the account, and why outreach against the data it scrapes sits in a terms-of-service gray area. It is fine for grabbing 30 profiles off a search you already built by hand. It is not a background sourcer that works your whole market while you are in interviews.
Even when a browser agent does pull profiles, it pulls what is on the page. It still has no read on compensation, no signal on who is actually open to a move, and no memory of who you passed on last week. It automates the clicking. The judgment, the part that decides whether a name is worth your time, is exactly what it leaves on your desk.
Tool, assistant, agent: what the words actually mean
The word "agent" gets stretched to cover everything right now, so it helps to be precise. Three things are getting lumped together, and they are not the same.
A tool waits for you. You click, it does one step, it stops. A keyword search field is a tool. An assistant responds to you in a conversation and produces text or suggestions on request. A custom GPT is an assistant. An agent is different in one specific way: you give it a goal, and it decides the steps and runs them without you clicking through each one. Search, filter, enrich, surface, follow up, repeat.
The catch is that "agent" describes how something runs, not what it can see. An agent layered on top of a model with no data access is still blind, just blind on autopilot. The capability that matters for sourcing is not the autonomy. It is whether the thing has live reach into the data where candidates actually live.
| Capability | Custom GPT (ChatGPT / Claude) | AI sourcing agent |
|---|---|---|
| Turns a plain-English role into a search | Yes | Yes |
| Searches live candidate data | No | Yes |
| Surfaces who is likely open to a move | No | Yes |
| Runs in the background without prompting | No | Yes |
| Carries context between sessions | No | Yes |
| Explains why each profile fits | Guesses | Yes, from the data |
What an AI sourcing agent actually does
Strip away the hype and a real sourcing agent does five concrete things a custom GPT cannot.
First, it takes the role in plain language, the same way you would brief a colleague, and reads intent rather than matching keywords. Second, it searches candidate data your own stack cannot reach, across many sources at once, instead of generating an answer from memory. Third, it builds a view of why each person fits the brief, grounded in their actual experience rather than a guess. Fourth, it adds the signals that decide whether outreach is worth your time: a read on compensation expectations, and a read on who is actually receptive to a move right now. Fifth, it keeps running after you close the tab, and tells you when there is something new to look at.
That last point is the quiet one. A custom GPT only works while you are typing. An agent works while you are in interviews, asleep, or onboarding a different placement. The work continues without your attention, which is the entire reason the category exists.
None of this is magic. It is the boring, unglamorous part custom GPTs skip: connecting to real candidate data, and judging reachability before you spend a credit or an hour.
Where Glozo's Sourcing Agent fits
Glozo built a Sourcing Agent for exactly this gap, and it went live in June 2026, free, in manual mode. You create it from a search, and it runs in the background on the data your stack cannot reach, then emails you when there are results worth a look.
The Agent is only as strong as what sits under it, and that is the data. Glozo reads more than 10 million market signals every month and pulls candidate profiles from 30-plus sources. On top of that it runs three signals a generic model has no way to produce: a plain-language rationale for why each profile fits the brief, a Market Value estimate so you know whether a candidate sits inside your budget before you reach out, and an Open to Offers read that points to passive people who are likely receptive rather than only those who flagged themselves "open to work." Surface candidates who are both inside budget and likely to reply, before spending anything, and the math on your week changes.
A custom GPT is still worth keeping for what it is good at: drafting, rewriting, and thinking through a search. Pair it with an agent that has the data, and you stop doing the part the GPT was never able to do. If you want to see what AI does well across the rest of the recruiting stack, our guides to ChatGPT for recruiters and Claude for recruiters cover the drafting side, and how to pick AI recruiting tools by category covers the rest.