ChatGPT, Claude and Perplexity cite job offers in more than nine out of ten responses. Yet, for ChatGPT, 7 out of 10 job links led to no online offer once opened. This is the finding of a survey conducted by a journalist based on 205 questions asked to several artificial intelligence assistants.
- ChatGPT, Claude and Perplexity cite job offers in more than nine out of ten responses.
- For ChatGPT, 7 out of 10 job links led to no online offer when opened.
- The likely cause is that assistants rely on pages gathered before the offers were closed.
- Employers should make their open positions easy to find and remove filled offers everywhere they were posted.
On October 16, the query “offre emploi infirmier Perpignan”, typed as on Google, was posed to ChatGPT programmatically. The answer included a list of positions, including a general nurse vacancy posted by Welljob, with its link. Less than an hour and a half later, that link opened a France Travail page where the offer was no longer online.
That day, 205 job questions were asked to ChatGPT, Claude, Gemini, Perplexity and Google. They covered 24 occupations among the most recruited, three per major family of the France Travail survey on labour needs, across 79 municipalities, from Paris to small towns. Each question was asked twice, and the assistants were queried programmatically, not via their public apps. A robot then opened, in the ensuing hours, the pages to which they pointed.
Job offers, almost every time
ChatGPT, Claude and Perplexity cite at least one specific offer—that is, a position identifiable with its employer or the job posting link—in more than nine out of ten responses. When asked for a job, they therefore propose positions.
Gemini is the exception: it rarely cites offers when the query is typed as on Google, but often when it is phrased as a sentence.
At the end of the link, often a dead end
A job link leads to the page of a specific offer, not to a list. For queries typed as on Google, 7 out of 10 ChatGPT job links led to no online offer when the robot opened them. Perplexity shows a similar finding. Most often, the page itself indicated that the offer was closed, which concerns one out of two job links for ChatGPT.
The most likely cause, according to the survey’s author, is that the assistant relies on pages accessed before the offer was closed. An offer eventually gets withdrawn, and the link provided by the AI no longer leads to the position. The author likens the situation to a GPS that directs you to a closed shop. He notes, however, that an offer shown as closed does not prove that the position no longer exists elsewhere, and that the count only concerns the links the robot was able to open.
Among the sites cited by the assistants, France Travail, HelloWork and Indeed are among the most frequent. For both ChatGPT and Perplexity, more than one out of three links provided with the offers pointed to France Travail. The survey does not compare the sites against each other and therefore does not specify which is responsible for the dead‑end links.
The employers cited also vary from one query to another. Asked twice to ChatGPT a few hours apart, the same question yields, at the median, this result: out of seven employers cited in total, only one appears in both answers. An AI response therefore reflects a specific moment and does not constitute a stable ranking.
What is within an employer’s reach
The author notes that none of the following has been measured, and nothing guarantees being cited. A candidate who follows an AI’s link and lands on a closed offer ends up at a dead end, while the employer risks losing an interested candidate. Making its open positions easy to find, starting with its careers site, can therefore be helpful. Removing an offer everywhere it has been posted as soon as the position is filled limits the risk of a candidate applying for a role that is already filled.
A legislative text could also change the situation. Bill No. 944, filed in the Senate on September 10, 2026, would prohibit publishing an offer without a salary range. As of October 26, it has not been adopted.
These results have a specific scope: high‑volume recruitment occupations, a single day of data collection, and assistants queried outside their public applications. They do not constitute a national estimate, and the observed rates cannot be directly extrapolated. The full methodology is published with the study. For a recruiter, the question raised by the survey is not limited to whether its offer is posted: it also concerns what is found at the end of the link when an AI recommends it.
