
Measuring recruitment data quality
Measure whether your team can use the information, without treating a populated field as automatically correct or current.
Remo Vloet6 min.

A well-filled candidate profile isn't proof of correct extraction. Check each field against the source, then inspect what the system saves or changes.
Both job titles appear on the new candidate profile, so the upload looks successful. Then a recruiter notices that the first role’s end date has been attached to the second. The fields are populated, but the career history is wrong.
A CV parsing accuracy test should compare extracted fields with a pre-checked source across documents resembling those your agency receives. Assess the relationships between values as well as individual words. Then check what the ATS saves. A successful upload doesn’t establish that the resulting candidate record is correct.
Represent the formats, languages and career patterns your agency handles. Include a digital PDF, a Word document, a column-based CV and a longer history with several roles at one employer. Add difficult cases without making the whole set unrepresentative of everyday work.
Start by reviewing what arrives. An agency mainly receiving Dutch Word documents needs a different mix from one regularly processing English-language scans. Record each file’s type, language and reason for inclusion.
Use fictional documents for the initial technical exercise where possible. Before using real CVs, assess the purpose, access and supplier arrangements. Removing a name doesn’t automatically anonymise a career history. A complete production export isn’t a suitable first test upload.
Record the information a reviewer can point to in the CV, including the expected value and source location. Define when a blank is correct. Inventing a missing start date differs from failing to extract a start date clearly present in the document, and your test should preserve that distinction.
For example, use this sheet with a fictional CV for Alex:
| Check | Expected result | What to inspect |
|---|---|---|
| Two roles at one employer | Separate experience entries | Keep title, start date and end date together |
| An ongoing course | Education entry with the stated ongoing status | Don’t add an awarded qualification |
| English listed as a working language | A source-faithful entry | Don’t invent an unconfirmed proficiency level |
| No phone number in the CV | Field stays empty | Don’t copy a referee’s number |
| Only years given for an older role | Retain the source’s level of precision | Don’t imply an exact month without evidence |
These are intended test outcomes, not results from a product trial. Ask a second colleague to check ambiguous cases before assessing the parser. Sometimes the scoring rule needs clarification too.
A mistake can occur while reading the file, recognising information or placing it in a record. Keep those stages separate. A document that looks orderly on screen may produce a different text sequence when software reads it.
Parsing software often converts a file to text first and only then recognises the information. Where available, inspect the converted text after an error. An unexpected reading order may explain why a date became attached to the wrong job.
Scanned documents require optical character recognition, usually shortened to OCR. Ask whether the route you are using supports it. A file selector accepting PDFs doesn’t establish that it can process a PDF containing only images. Test one separately and check whether the outcome is understandable.
Then inspect the field mapping. A correctly recognised name can still end up in the wrong field. Correct job titles don’t make correct experience records when their dates have been assigned to another employer.
For each field, record whether the value is correct, missing, wrong or added without support. Report counts and their denominator. ‘Correct values among facts present in the source’ answers a different question from ‘correct values among everything the parser filled in’.
Break results down by document group and important field. An average dominated by simple contact details can hide problems in employment history. Consider consequences as well: a wrong optional label differs from an incorrect employer or a qualification the person hasn’t gained.
A confidence score doesn’t replace this assessment. Microsoft distinguishes confidence from measured model accuracy. A score estimates confidence in a prediction; it isn’t the result of your test set. Check which fields receive scores and whether higher scores correspond to fewer mistakes in your documents.
Don’t turn the exercise into a supplier ranking if the documents or settings differ between runs. The purpose is to understand the checking your agency will need.
Check where the information goes once processing finishes. Does an upload immediately create a record, show a review screen or produce a proposal? Use the permissions a recruiter will actually have. A different entry point may behave differently.
Process the same fictional CV again. Investigate whether that creates another candidate, fills existing blanks or changes previously checked values. Make a manual correction beforehand so you can see what happens to it.
Observe the remaining work too. A parser that fills many fields but requires lengthy repairs hasn’t necessarily produced a useful process. Note which mistakes are easy to fix, which relationships need checking again and who does that work.
Keep some documents aside for a fresh test after configuration changes. Repeating only the files used to improve the settings tells you little about documents the parser hasn’t encountered in that exercise.
With document parsing in Simply, experience, education, skills, languages and certifications have their own sections on the candidate profile. Use that structure for your assessment: do the two roles in Alex’s CV appear as separate experience entries with the correct dates? Inspect both the field and the relationship it belongs to.
Bring Alex’s fictional CV and an updated version to a Simply demo. Compare the first import with reprocessing an existing record: Simply never just overwrites what is there. A value that differs from the existing one first becomes a proposal with the old and new values side by side, and you approve or reject it. Include a manual correction in your test so you can see what the recruiter finds afterwards.
Record the decision in the review sheet. Which document groups are usable, where does checking remain necessary and which cases need a different route for now? That gives colleagues more to work with than one large percentage above a demonstration.
About the author

Remo Vloet is a co-founder of Simply, the AI Operating System for recruitment agencies: inbox, meetings, sourcing, CV parsing, search and matching, documents and automation in one system. With a background in building complex software, he contributes to the technical vision behind Simply.
LinkedInArticles by Remo VloetThere isn't a fixed number that provides enough evidence for every agency. Cover the document groups you receive and recurring exceptions; a small trial isn't a statistical guarantee.
Not necessarily. It's a model estimate that you should assess against checked examples. Establish which part of processing the score describes too.
Extraction and any enrichment should be clearly distinguishable. In a source-faithfulness test, you need to tell which values come from the CV and which lack support there.
No. The processing route needs text recognition for that. Test a scan separately and ask what happens if no usable text is found.
The first run doesn't reveal what happens to existing or manually corrected data. Reprocessing shows whether the route creates records, fills blanks or makes other changes.

Measure whether your team can use the information, without treating a populated field as automatically correct or current.
Remo Vloet6 min.

Establish whether two records describe the same person before bringing contact details or applications together.
Remo Vloet5 min.

Decide what belongs to a candidate, a job or an application before adding more fields. Your records need to make sense when those relationships change.
Remo Vloet7 min.

Find out how Simply can completely evolve your workflow. No slides, just product.