Methodology
How worth. estimates pay, and how well it does.
worth. estimates what similar roles advertise. That's not the employer's actual budget, and published market pay isn't the same thing as fair pay. It's a reference point to help you ask better questions.
Where the data comes from
Real UK job ads that state a salary, collected from Reed, plus public ads people check with worth. that state one. We use the figure the employer published, never anything self-reported, and we drop duplicates and reposts so one job can't count twice. The model the site is using now learned from 6,604 ads collected between 18 September and 4 October 2026, and was trained on 4 October.
How an ad is read
An AI model (Jev, from TypeSafe) reads every ad with the salary hidden and answers the same 25 questions about it: the type of work (one of 25 ONS occupation groups), seniority, scope, people management, how rare the skills are, and yes/no questions such as budget responsibility or shift work. Those answers, plus the region, become about 70 inputs to the model.
How the estimate is made
A gradient-boosted model learns how those inputs relate to advertised pay and predicts a midpoint. The ranges come from how far its predictions missed on ads it hadn't seen: the headline range is the middle half of those misses (it should hold the real salary about half the time) and the wider range is the middle 80%.
How we test it
Tested on 4 October 2026 with 5-fold cross-validation: every ad scored by a model that never saw it. "Median error" means half of the estimates were closer to the real advertised salary than this, and half were further away.
| Across all ads | Result |
|---|---|
| Ads tested | 6,604 |
| Median error of the midpoint | 11% |
| Midpoint within 10% of the real salary | 46% |
| Headline range held the real salary | 50% (aim: 50%) |
| Wider range held the real salary | 80% (aim: 80%) |
| By type of work | Ads tested | Median error | Wider range held it |
|---|---|---|---|
| Business, media and public service professionals (legal, finance, HR, marketing, social work) | 1,266 | 13% | 75% |
| Science, research, engineering and technology professionals (incl. software, IT) | 924 | 16% | 69% |
| Skilled metal, electrical and electronic trades | 811 | 9% | 89% |
| Business and public service associate professionals (sales executives, recruiters, buyers, account managers) | 624 | 13% | 75% |
| Corporate managers and directors | 512 | 14% | 73% |
| Skilled construction and building trades | 319 | 9% | 88% |
| Administrative occupations (admin, finance assistants, office administration) | 257 | 8% | 88% |
| Process, plant and machine operatives | 256 | 6% | 93% |
| Science, engineering and technology associate professionals (technicians, IT support) | 212 | 9% | 87% |
| Transport and mobile machine drivers and operatives | 211 | 10% | 90% |
| Other managers and proprietors (hospitality, retail, services) | 201 | 12% | 80% |
| Teaching and other educational professionals | 177 | 12% | 76% |
| Caring personal service occupations (care workers, teaching assistants) | 140 | 7% | 91% |
| Culture, media and sports occupations (design, journalism, arts) | 127 | 12% | 68% |
| Health professionals (doctors, nurses, pharmacists, therapists) | 103 | 10% | 87% |
| Secretarial and related occupations (PAs, executive assistants, receptionists) | 97 | 10% | 84% |
| Textiles, printing and other skilled trades (incl. chefs) | 92 | 9% | 90% |
| Elementary administration and service occupations (cleaning, warehouse, hospitality) | 79 | 7% | 91% |
| Customer service occupations (call centre, customer service) | 66 | 6% | 92% |
| Elementary trades and related occupations | 42 | 8% | 100% |
| Skilled agricultural and related trades | 26 | 5% | 85% |
| Sales occupations (shop and retail sales) | 23 | 4% | 87% |
| Health and social care associate professionals | 18 | 11% | 89% |
| Protective service occupations (police, fire, prison officers) | 7 | 12% | 71% |
| Leisure, travel and related personal service occupations | 7 | 5% | 71% |
| Community and civil enforcement occupations | 7 | 16% | 71% |
Groups with fewer than 20 ads tested can't give high confidence, whatever their error. Groups not listed had too few ads to test.
What "similar ads" and confidence mean
- Similar ads are ads in our data for the same ONS occupation group, at a seniority within three-quarters of a level of this one. Every result shows the count.
- Confidence uses both that count and how accurate the model was in testing for that occupation group. Fewer than 10 similar ads, or a group where its median error is above 25%, means low confidence and no verdict. High confidence needs 40 or more similar ads and a group tested on at least 20 ads, with a median error of 18% or less and a wider range that held the real salary at least 70% of the time.
- Percentiles in the detail view compare an ad against similar ads, or against every ad we've read where it says so. They describe where an ad sits, not whether it's fair.
Public-sector pay scales
NHS Agenda for Change bands, teachers' pay scales, school and council grades, Civil Service grades and the university pay spine are set by the grade, not chosen for each ad. Comparing them with market pay wouldn't say whether an offer is fair, so worth. doesn't give a verdict on them. It names the scale and grade, shows what similar private-sector ads advertise for reference, and suggests questions about whether the duties match the grade. Ads like these are also left out of the model's training data, so it learns from market-set pay only.
What it can't tell you
- Ads that hide the salary. The model only learns from ads that publish pay. Ads that say "competitive", often senior roles, may pay differently, so for those the estimate could run low.
- Mostly one job board. Most of the data is from Reed, so roles that are rarely advertised there are thinly covered.
- Base salary only. Bonus, equity and benefits aren't counted.
- How the job is really done. Workload, support and expectations depend on things an ad rarely says. That's why worth. suggests questions to ask instead of drawing conclusions.
The homepage numbers
They're read live from our data: ads collected, ads the current model learned from (after removing duplicates), and its median error from the test above. "Reed ads that state a salary" is the share of every Reed search result our collector has looked at since the date shown, salary or not. It describes our sample, not every Reed listing.