Standards

Methodology

How evidence is selected, weighed, and disclosed here — including what happens when the evidence is bad, contradictory, or missing.

Published 21 August 2026

Job-search advice has a specific failure mode. A number is measured once, on one platform, in one year, and then circulates for a decade with its origins stripped off until it is quoted as a general truth. This page describes what is done here instead.

The source hierarchy

Sources are ranked by how directly they observed the thing being claimed. A study that measured applications outranks an article describing that study, which outranks a blog post citing the article from memory.

Tier 1 — preferred: Government agencies and official labour statistics, official datasets, peer-reviewed studies, academic papers, primary company documentation, first-party data, and original Jobsearch.ing research.
Tier 2 — used with attribution: Reputable research organisations, credible industry reports, established institutional surveys, and major publications reporting findings they produced themselves.
Tier 3 — last resort: High-quality secondary analysis, used only where primary evidence does not exist. Always labelled on the page, with an explanation of why nothing better was available.

Community sources — forums, subreddits, comment threads — are genuinely valuable for discovering what job seekers are actually struggling with, and for qualitative texture. They are never used as statistical evidence. If a page draws on community observation, it says so and describes what that observation can and cannot support.

Anonymous SEO content is not used as a source for any numerical claim, regardless of how many other pages repeat it.

How a statistic is selected

Before a figure is quoted anywhere on this site, the primary source is located and read. Not the article about the source — the source. The following are recorded for each one:

the publishing organisation, title, URL, and publication date;
the date it was last accessed and checked;
the specific passage or table the claim rests on;
what the source supports, stated in our own words;
what the source explicitly does not support, so a narrow finding cannot later be widened by an edit.

If the primary source cannot be located, the claim is not made. “Widely reported” is not a provenance.

When sources disagree

Disagreement between credible sources is shown, not resolved. Two reputable studies producing different application-to-interview rates are usually not in conflict at all: they measured different populations, in different years, counting different things.

Where that is the case, the page presents both figures with their definitions side by side and explains what accounts for the gap. What will not happen is the two being averaged into a single tidy number that neither source supports and no reader could trace.

Where the evidence genuinely cannot settle a question, the page says so. An honest “nobody has measured this well” is more useful than a confident number with nothing behind it.

Evidence statuses

Claim reviews in the Evidence Observatory carry one of six statuses. They are deliberately not true and false: most circulating job-search statistics fail in more specific ways, and collapsing those into a binary would lose the part that matters.

Two boundaries do most of the work. Unverified is a statement about the evidence and never about the world — it means no source could be traced, not that a claim is false. Contradicted is the much stronger statement that good evidence points the other way, and it is used only when that evidence can be named on the page.

Contradicted: Identifiable evidence of reasonable quality points the other way. Reserved for cases where we can name that evidence, not for claims that are merely unsupported.
Source withdrawn: The claim did have a traceable source, and that source has since been retracted, corrected, taken down, or disavowed by whoever published it.
Unverified: We could not trace the claim to any identifiable primary source. This is a statement about the evidence, not about the world: an unverified claim is not thereby false, it is unsupported, and we say so rather than guessing.
Partially supported: A real source underlies the claim, but the claim as circulated overstates it — by generalising a narrow population, dropping a qualifier, rounding a range to a single number, or restating a measure as something it did not measure.
Context dependent: The claim is true or false depending on a condition the claim itself omits — the industry, the seniority, the labour market, or which of several different quantities is being counted. There is no single correct answer to report.
Supported: An identifiable primary source supports the claim as it is normally stated, and the claim does not overreach that source's population, period or measure.

A review is published only when the claim has been traced as far toward its original source as we can get. Many researched claims stay unpublished because that trail is still open, and an absent review says nothing about the claim.

Original calculations

Some figures on this site are derived rather than observed — conversion arithmetic, expected volumes, confidence intervals. For every one of these:

the formula is published on the page that uses it;
the inputs are named, including which are the reader’s own;
the assumptions are stated, particularly where they are known to be imperfect;
the derivation is distinguished from measurement, so a reader never has to guess whether a number was counted or computed.

The calculators are a worked example of this. Each one shows its intermediate arithmetic, and each states the assumption it is least comfortable with rather than hiding it.

Uncertainty

Sample size is treated as part of a finding, not a footnote to it. Where a rate rests on a small number of observations, the uncertainty is quantified and displayed alongside the estimate rather than acknowledged in passing.

The same applies to language. “Suggests”, “is associated with”, and “causes” are different claims requiring different evidence, and are not used interchangeably to make a sentence read better.

Future original research

These are commitments, not a description of work already done. Jobsearch.ing has not yet published original research.

When it does, each study will:

publish its methodology before collection, so the research question cannot be fitted to the results;
state its sampling frame and recruitment method, and describe who the sample cannot speak for;
report sample sizes for every breakdown, not just the headline;
collect only from participants who explicitly and specifically consented to research;
publish its limitations as part of the findings rather than as an appendix.

Dates and updates

A publication date reflects when a page was actually published. An updated date appears only when the substance of the page changed — never because the site was rebuilt or a typo was fixed. Rotating dates to appear fresh is a form of lying to the reader, and to search engines, and it is not done here.

Corrections

Errors are corrected quickly and visibly. Where a correction changes a consequential claim, the page carries a dated note describing what was wrong and what changed. Findings are not silently revised.

Corrections to research findings are treated as part of the published record, on the principle that anyone who cited the original should be able to see what changed.

How automation is used

AI tools are used here for drafting, structuring, editing, source organisation, and writing the software this site runs on. That is ordinary, and disclosing it is ordinary too.

What AI is never used for:

as a source of fact. A language model asserting a statistic is not evidence that the statistic exists;
to generate a citation. Every source is a document that was located and read;
to mass-produce pages. Volume is not the goal, and scaled thin content is the failure mode this publication is defined against.