Ranking Is Not Citation

Which sources does AI cite about jobs? I measured 100 queries. The answer depends entirely on what you ask.

Tan Truong  ·  Organic Search Lead  ·  September 2026

I asked ChatGPT 100 real employment questions — 20 in each of five intent categories — and recorded every source it cited.

Two further results are worth more than the headline:

• The U.S. government is tied with Indeed as the most-cited source in the whole study — bls.gov 35%, indeed.com 35%.
• Every job board other than Indeed, combined, was cited 7 times. LinkedIn managed 3, ZipRecruiter 2, Glassdoor 1, Monster none.

No client or employer data was used. Every figure is a public observation, reproducible by anyone with a ChatGPT subscription. Measurement cost: $0.

1Job board visibility swings from 0% to 95% by intent

Read the first and last rows together. On “how do I write a resume”, a job board is almost certain to be cited. On “is nursing a good career”, across 20 queries, not once.

This is not a content-quality gap. It is a category assignment: the model treats career questions as requests for official statistics and routes them to the Bureau of Labor Statistics. Opinion content — however good — is not invited into that answer.

2It is not job boards versus government. It is Indeed versus everyone.

The government cluster runs wider than bls.gov alone: dol.gov, irs.gov, eeoc.gov, consumer.ftc.gov, usa.gov, careeronestop.org, onetonline.org, and agency sources such as faa.gov and fmcsa.dot.gov. It appeared in every intent category, never below 30%.

The definitions category makes the point sharply. Asked what a 1099 contractor or a signing bonus is, the model reached for irs.gov and law.cornell.edu. In the United States, employment definitions are treated as tax and legal questions, not dictionary questions.

3Answer length tells you whether there is room to be cited

Salary answers ran a median of 76 words — a third the length of other categories. The model returns one BLS figure and stops. There is no second paragraph to be cited in.

This reframes the usual advice. On short-answer intents, publishing longer and more thorough content is the wrong move, because there is no space in the answer. The only way in is to be the source of the number itself.

4What follows from this

• Stop buying blended AI visibility scores. Ask any vendor for the query set. If it is not segmented by intent, the score measures their sampling, not your site.
• Choose fronts on base rates, not ambition. At a 5% base rate you need roughly 318 queries per group to detect a change. At 40% you need about 44. Same budget, seven times the resolution. A front an incumbent holds at 90–100% should be conceded, not contested.
• On statistical intents, become a source rather than a publisher. The BLS has authority and lacks one thing: current data from live job postings. A job marketplace has exactly that. Published as real data — methodology, sample size, collection date, stable URL — it is citable in a way an opinion article is not.
• Measure with a control group. Anything else leaves you reading a chart that was moving anyway.

5Method

6Limitations

Stated plainly, because a study without them is marketing.

• One run per query. Model outputs vary between runs, so these percentages carry real sampling noise. Differences of a few points mean nothing. The 95-versus-0 contrast is far too large to be noise; the middle rows should be read as approximate.
• One engine, one country, one day. Citation behaviour changes with model updates. Re-measure before quoting these numbers as current.
• 20 queries per category. Enough to establish the shape, not enough for precise rates.
• Matching is by domain pattern, so a site counts as cited regardless of how prominently.

The study is designed to be repeated. Anyone with a ChatGPT subscription and a weekend can reproduce it.

---

Tan Truong — 14 years in organic search, running acquisition for a national employment marketplace. This study uses no data from that role. Measurement stack: ChatGPT via Codex CLI, Python, bootstrap resampling. Total cost: $0.