The short version
We collected every job advert posted on ReliefWeb from June 2011 to date (currently September 2026): 516,428 adverts. A large language model read each one and reviewed it against a list of 65 humanitarian themes and topics. The results are shown as a post count, or a share of posts.
What this cannot tell you
Read these before citing a number.
- Adverts are not hires. A post may be filled internally, readvertised or never filled.
- ReliefWeb is not the whole sector. Large UN agencies often recruit through their own portals; UNICEF, for example, has posted almost nothing on ReliefWeb since 2021. We also expect that most national NGO hiring happens on country-specific job boards.
- Total volume moves a lot. Adverts went from 52,965 in 2022 to 19,254 in 2025. Topic trends are therefore shown as a share of all adverts, so a shrinking market does not look like a shrinking topic.
- Small themes and topics don't have many hits. Eight topics have fewer than about 160 core adverts in fifteen years. They sit on an Exploratory themes list, shown as counts only and left out of rankings.
Source
The ReliefWeb API, operated by OCHA, including expired adverts. The current, incomplete month is always left out, so the latest month shown is September 2026.
An advert posted jointly by several organisations counts once for each of them. ReliefWeb, OCHA and the advertisers did not produce and do not endorse these figures.
How topics are tagged
Each advert was read by a language model (OpenAI GPT-6 Luna), given the 65 topic definitions in the glossary. It reads adverts in English, French and Spanish. A topic counts when the job itself involves it, not when the word appears in the organisation's description of itself or in standard recruitment wording.
For each relevant topic the model assigns a level:
- Core: the topic or theme is central to the role. This is the default measure used on this site.
- Duty: the topic is a substantial part of the work but not its main subject. Counted together with core as "core or duty", which is broader and noisier.
- Skills asked for, and passing mentions, are not counted.
Some topics contain others. A few topics are specialisms of a broader one, and every post tagged with a specialism is also counted under the broader topic:
- Health includes sexual and reproductive health, epidemics and disease control, and MHPSS.
- Cash and voucher assistance includes multi-purpose cash, cash coordination and group cash transfers.
So an MHPSS post appears under both MHPSS and Health. Adding Health and MHPSS together would count that post twice: the Health figure already includes it.
Adverts with under 200 characters of text, about 0.3%, were not read and carry no topic.
How good the tags are
To check the tagging, a second, stronger model (OpenAI GPT-6.1 Sol) tagged a sample of 300 adverts on its own, without seeing our tags. We then compared the two in both directions.
- When our model tags a topic, is it right? Almost always. Our model made 188 core tags on the sample. The check model agreed with 185 of them: 98%. So a post counted under a topic on the dashboard is very likely to be about that topic.
- Does our model find every topic? Not quite. The check model found 249 core topics in the same 300 adverts. Ours had found 217 of them: 87%. Our model misses about one in eight, so counts run a little low.
- Duty tags are looser. About three in four were confirmed.
What this means in practice: trends are more reliable than absolute numbers. If the miss rate stays steady over time, a topic's rise or fall is real even when its count is slightly understated. This is one model checking another, not a hand count by people.
ReliefWeb's own fields
Job type, experience and career category are ReliefWeb's own fields, filled in by the advertiser. We do not alter them. Career category is patchy before 2018.
Organisation types
ReliefWeb gives each organisation one type but does not separate national from international NGOs. We split NGOs using their home country and where they advertise.
High or low income. The home country is the one ReliefWeb records for each organisation. We look it up in the World Bank's income classification, which groups countries as high, upper-middle, lower-middle or low income by gross national income per person. "Low or middle income" means any of the last three. We use the current classification (October 2026) for every year, so a country that has moved group since 2011 is treated by its current group throughout. An organisation whose home is not a single country, such as "World", counts as international.
- International NGO: home country is high income.
- Locally based NGO: home country is low or middle income and at least 80% of its located adverts are there.
- Regional NGO: home country is low or middle income but it advertises across several countries.
This is a proxy. It counts some local affiliates of international networks as locally based, and counts national NGOs in high income countries as international. The 45 or so largest organisations in the two smaller classes were checked by hand; the rest follow the rule, and a few known misfits are corrected by hand (for example Oxfam and ActionAid, global confederations whose secretariats sit in Nairobi and Johannesburg). NGOs based in Türkiye or Syria are treated as local to the Syria response, because many Syrian NGOs are registered in Türkiye and work across the border.
How to cite
Corrections and changes
If a number looks wrong, tell us at contact@curinta.com. Changes to the method are logged here.
- No changes since launch.