Stop asking ChatGPT what you should earn
Ask ChatGPT what a junior backend developer earns in Denver and you'll get a number. It sounds sure of itself. What it rarely tells you is how many job ads that number rests on, because it isn't counting job ads. It's wr
Ask ChatGPT what a junior backend developer earns in Denver and you'll get a number. It sounds sure of itself. What it rarely tells you is how many job ads that number rests on, because it isn't counting job ads. It's writing the most plausible-sounding answer.
I'm building SkillXray, a tool that reads your resume against real job ads and shows which skills you're missing and what they pay. Early on, I had to decide what to show when there isn't enough data. I chose to show nothing, and I want to explain why, because it's the opposite of what most career tools do.
The problem with a confident number
Salary sites have a quieter version of the same problem. A figure for a rare role in a small city might rest on three ads, and it's printed in the same font, the same size, with the same confidence as a figure that rests on three hundred.
When the number is wrong, you pay for it: you ask for too little, or you learn the wrong skill because someone told you it pays more.
The rule: a number carries its sample, or it doesn't exist
Every figure SkillXray shows has to travel with four things: where it came from, how many ads it rests on, when it was read, and whether it was measured from ads or taken from a published benchmark. In TypeScript, that's a type, so a bare number can't reach the page without them:
export function measuredOrNull<T>(
value: T,
meta: MeasuredMeta,
floor: number = MIN_SAMPLE
): Measured<T> | null {
if (meta.sampleSize < floor) return null;
return measured(value, meta);
}
MIN_SAMPLE is 10. Below ten ads with an advertised salary, the function returns null, and the page shows a blank where the figure would be. Not an estimate or a range. A blank.
Ten isn't magic. It's a line below which a typical value is mostly noise, and the comment on that function says what we believe: a missing figure is strictly better than a wrong one.
A stricter rule for "what a skill is worth"
The tool also says what a missing skill tends to add to pay. That's a comparison, so it needs two samples, not one: ads that ask for the skill, and ads for the same role that don't. Each side needs at least five ads with a salary, or the skill gets no price at all.
This means a lot of skills show up unpriced, especially outside big markets. That was uncomfortable to ship. A product page full of blanks looks less impressive than one full of dollar signs. But a premium computed from two ads against three is a coin toss presented as a fact.
Advertised pay only
Some job APIs fill in a salary when the employer didn't post one, using their own estimate, and mark it as predicted. We drop every one of those. If an ad didn't say what it pays, it doesn't count towards what ads pay. It's a small rule, but without it, "typical advertised pay" would quietly include someone else's guess.
What it costs
Honesty here has a real price:
- Fewer pages. A salary page for a role in a city with too few priced ads isn't published, and it stays out of the sitemap.
- Blanks in the product. Some readers see a gap with no price on it and assume something is broken.
- Smaller claims. We can't say "millions of jobs analysed". We say how many ads, from where, and when.
I think it's worth it. If you're deciding what to learn next or what to ask for in an offer, a number you can check beats a number that sounds good.
Try it
SkillXray is free to start: pick a role or paste a job ad you want, and it shows which skills you're missing, what to learn first, and a short test for each. Every pay figure shows how many ads it's based on, or stays blank.
I'm looking for honest feedback from people who are job hunting right now. If you try it and send me your thoughts, good or bad, the first 25 people get 3 months of Pro free. skillxray.com
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.