Ask ChatGPT about your own startup before an investor's associate does
Contents
I build a tool that measures how LLMs describe companies, so at this point I've read a few thousand AI-generated summaries of startups. Most founders have never read the one about theirs. This is basically the manual version of the check my tool automates — ten minutes, no product required.
One company surprised me. They'd spent years building a proprietary dataset, but not a single model mentioned it. Two models called them a plugin. Another recommended a competitor that had already shut down. The founder had never thought to ask those questions himself.
Quick context on why I think this matters at all. When someone hears your company name for the first time — an associate doing pre-call diligence, a buyer comparing options — there's a decent chance their first move is a prompt, not your website. What comes back is a compressed version of your company, and compression throws details away. The category usually shrinks: "workflow infrastructure" comes back as "a software tool in a competitive market." The moat often gets flattened into generic differentiation language. And past a certain point you just get folded into a list — the model names two or three companies and stops, and you're either on that list or you're not.
None of this produces a signal on your end. Nobody emails you about the meeting that didn't get scheduled.
The ten-minute check
Open four models, not one. ChatGPT, Claude, Perplexity, Grok. One model's answer is an anecdote.
Don't ask "what is [my company]?" — that's a softball and you'll get a flattering answer that tells you nothing. Ask what a buyer would ask: "best tools for [the problem you solve]", "leading companies in [your category]", "[you] vs [your main competitor] for a small team." Each time, note whether you were named, at what position, and who got named instead.
Then run the one that stings: "what are the risks or weaknesses of [my company]?" This is roughly what an associate runs before your first call, so you might as well see the answer before they do.
Last thing — repeat it in a fresh session. If you show up in one run and not the next, your presence is unstable, and I'd argue that's a worse problem than being reliably third.
(If you want copy-paste prompt sets instead of improvising, I keep a written-up version of this check for ChatGPT — the Claude, Perplexity and Grok variants are linked from there.)
As for fixing what you find — nothing exotic, and honestly most of it is stuff you've been putting off anyway. Models love comparison content, so if you don't have a single "[you] vs [competitor]" page, someone else's page is answering that question for you. They lean hard on G2 and Capterra reviews, so ten real reviews quietly outperform your beautiful homepage. And say your category out loud somewhere on your site, in plain words. If you don't, the model picks a category for you, and it picks a smaller one. The general rule I keep running into: what others say about you counts, what you say about yourself mostly doesn't.
The caveat, and the plug
Five prompts is a snapshot, not a measurement. You can't tell noise from a stable pattern, and you can't see whether it's drifting. That gap is what I automated: VeritasLinks runs 500–1,000 structured prompts across seven models and turns it into a score, per-model breakdowns, and who's getting recommended instead of you. There's a public sample report (pay attention to #AI Investor Review section) if you're curious what the output actually looks like, and the first analysis is free. One thing I'll add for anyone currently raising: the page you most need to test usually isn't your homepage, it's the investor-facing version of your story — I wrote up why here. But run the manual version either way. The founders who get surprised by it usually aren't the ones with weak products — more often it's the ones whose whole story lives on their own website.
If you get a result that surprises you, tell me about it in the comments. The weird cases are how I find out where the models break in interesting ways.
