Learning from a signal that barely exists

A rejection could mean the role was filled internally, the budget was pulled, someone was more qualified, the recruiter never opened the application, or the posting was never real. None of these is distinguishable from the outside.

Any system that tells a candidate why they were rejected is guessing. Presenting that guess as an explanation is worse than saying nothing, because the candidate may act on it.

Acting on it is the specific harm, and it is not hypothetical. A candidate told their CV is the problem will spend a fortnight rewriting a document that was fine, while the actual issue — the level they are targeting, or a market that has thinned — goes unaddressed for another month.

A job search yields tens of outcomes across different roles, companies, markets and moments. Attributing a pattern to any one factor requires holding the others constant, which is impossible here.

Fifteen rejections is not evidence that a CV is the problem. It is consistent with a difficult market, an over-ambitious target, or ordinary variance — and the agent cannot tell which.

Silence makes it worse rather than merely less informative. Most applications never receive any response at all, so the data is not a small sample of outcomes — it is a small sample of the subset of employers who bothered to reply, which is a different and non-random population.

• Different roles, companies and timings in every comparison
• No visibility into the other applicants
• Outcomes often never reported at all
• Silence is the most common result and means nothing specific

Aggregate patterns across many users are more reliable than one person’s history. If applications missing a particular hard requirement almost never receive responses, that is worth acting on, and it is visible only at scale.

For an individual, the honest scope is narrow: whether they are applying to roles well above their evidenced level, whether they are missing a requirement that recurs across their targets, whether they are applying to employers who respond to nobody.

Notice that all three are checkable without any outcome data at all. They are properties of the applications going out, which is precisely why they are reliable — the rejections are corroboration rather than the evidence.

Where a rejection includes a reason — rare, but it happens — that is worth more than a hundred silent ones. The same is true of an interviewer’s comment or a recruiter’s note.

Capture those deliberately and weight them heavily. Ask the candidate for what they learned after an interview too; their account of a conversation is far better evidence than anything the agent can infer.

Ask at the right moment and keep it short. One question the day after an interview — what did they seem unconvinced by? — gets answered while it is fresh, and it produces the single most useful row in the entire dataset.

Treat a stated reason with some scepticism too, since rejection messages are written to be inoffensive rather than accurate. "We went with someone whose experience was a closer fit" is compatible with everything, and a system reading it as a specific finding is back to inventing explanations.

The output should be observations rather than conclusions. "Eleven of your last fifteen applications were to roles one level above your current title" is a fact the candidate can weigh; "your CV is not strong enough" is an inference from nothing.

Comparative framing is more useful than absolute. Knowing that response rates in their field are generally low this quarter changes how a candidate reads their own silence, and it is information the platform has and they do not.

Tone is a design requirement here, not a nicety. This report is read by someone being rejected repeatedly, and a technically accurate summary of everything that is going badly will close the product rather than improve the search.

• State observations, never diagnoses
• Give context from the wider market, which the candidate cannot see
• Name one thing that is actionable, not five that are not
• Say plainly when there is not enough data to say anything

The obvious reaction to rejections is to lower the target, and an agent doing this automatically can talk someone out of the career they are trying to build, on the basis of noise.

Keep the target where the candidate set it unless they change it. Report what has happened, offer what can be addressed, and let the person decide whether to adjust — that decision is not the agent’s to make.

If the system does adjust anything, make it visible and reversible. A feed that quietly drifted downwards after a bad month is one the candidate cannot diagnose, and they will experience it as the market getting worse rather than as a change their tool made on their behalf.