CV parser vs CV formatter: one fills a database, the other fills an inbox

A CV parser is software that reads a CV and turns it into structured data: name, job titles, employers, dates and skills, each stored in its own field in your ATS or CRM. A CV formatter takes the same file and produces a document, the candidate's CV in your agency's template, for a client to read. This guide explains how parsing works, where it breaks, how to test the parser you already own, and which of the two tools a given problem calls for.

What a CV parser is

CV parsing (resume parsing, in the US) is the step that happens when you drag a file onto a candidate record and the fields fill themselves in. The parser reads free-form text written by a person, decides which words are a job title and which are an employer, and writes each one into the right box.

The point is search. A folder of 40,000 PDFs can only be searched as loose text. A database of 40,000 parsed records can answer "finance managers within 30 miles of Leeds with SAP and ACCA, moved jobs in the last three years". Everything your ATS does afterwards, including matching, filtering and reporting, depends on how well those fields were filled in.

What comes out is not a document. It looks like this:

{
  "name":      { "first": "Priya", "last": "Nair" },
  "email":     "priya.nair@example.com",
  "location":  { "city": "Leeds", "country": "GB" },
  "experience": [
    { "title": "Finance Manager", "employer": "Halden Foods",
      "start": "2021-03", "end": null, "current": true },
    { "title": "Financial Analyst", "employer": "Orrin Logistics",
      "start": "2017-09", "end": "2021-02" }
  ],
  "education": [ { "degree": "BSc Accounting", "year": 2016 } ],
  "skills":    [ "SAP S/4HANA", "IFRS", "Cash forecasting" ]
}

Nobody sends that to a client. Hold on to that thought; it is the whole of the parser-versus-formatter distinction, and we'll come back to it.

How a CV parser works

Whatever the vendor calls its technology, the pipeline has the same four stations.

1. Text extraction. A Word file or a "real" PDF carries a text layer the parser can read directly. A scan or a phone photo has none, so the parser runs OCR first to recognise the characters. Either way the parser has to decide the reading order: which piece of text comes after which.
2. Sectioning. The text is cut into blocks and each block gets a label: contact details, profile, experience, education, skills. Headings help ("Work History"), but candidates write "Where I've been" and "My toolbox", so parsers also look at what's inside the block.
3. Entity extraction. Inside each block the parser tags the individual facts. In "Finance Manager, Halden Foods, Mar 2021 – to date", three separate things have to be recognised and tied together as one job.
4. Normalisation. Values are cleaned and mapped to a fixed schema so records are comparable: dates to one format, "Fin. Mgr" to "Finance Manager", two spellings of a skill to one. Many parsers also map titles and skills to a taxonomy, which is what makes "accountant" find "finance manager" in a search.

Rules, machine learning, LLMs: three generations of parser

Vendors rarely tell you which approach sits under the bonnet, and most commercial products blend all three. It still helps to know how each one behaves, because they fail differently.

That last failure deserves a second look. A rule-based parser that can't find an end date leaves the field empty. A language model may fill it with something plausible. For a database you search, a blank is annoying. A confident wrong value is worse, because nobody goes looking for it.

Where CV parsing fails, and why

A CV is designed to be looked at. A parser doesn't look, it reads a stream of text. Most parsing errors come from the gap between how a page appears and the order its text is stored in the file.

The classic case is the two-column CV. On the left is what the candidate sent. On the right is what a parser gets if it reads straight across the page, joining each line of the main column to the sidebar line that happens to sit beside it.

# reading order: straight across the page
PRIYA NAIR CONTACT
Finance Manager priya.nair@example.com
EXPERIENCE +44 7700 900123
Finance Manager Leeds, UK
Halden Foods · 2021 – present SKILLS
Led month-end close across 4 entities SAP S/4HANA
Financial Analyst IFRS reporting
Orrin Logistics · 2017 – 2021 LANGUAGES
Built the weekly cash forecast English, Hindi
EDUCATION CERTIFICATES
BSc Accounting, Univ. of Leeds, 2016 ACCA (2019)

Feed that stream into stations 2 to 4 and the record comes out like this:

In fairness to today's products: column detection is now a headline feature. Textkernel's parser page lists automatic detection of single- and multi-column layouts, and Bullhorn's notes on its 2026 parser upgrade say multi-column designs are read more accurately than before (sources in the vendor section). Good parsers get the simple sidebar right most of the time. The same mechanism still catches them on the stranger layouts below.

Ten causes worth knowing by name

How to test CV parser accuracy yourself: the 10-CV torture test

Every vendor publishes an accuracy figure, and every figure is true for the CVs it was measured on. Affinda, for example, states "over 95% accuracy on core candidate fields" on its resume parser page. Notice the careful wording: core fields, scored their way, on their sample. The only number that describes your desk is one you measure on your own candidates. It takes about an hour.

Step 1: pick ten real CVs

Pull them from your own inbox so the sector and languages match your work. One of each:

1. The controlPlain single-column Word file, standard headings. If this one fails, stop here.
2. Two columnsA PDF with a sidebar for contact details and skills.
3. Table layoutBuilt in a Word table, dates in the left-hand column.
4. Header contactName, phone and email only in the page header or footer.
5. Scan or photoA scanned PDF or a phone picture of a printout.
6. Designer templateIcons, skill bars, a timeline graphic, a photo.
7. Another languageOne you really recruit in, ideally with accented characters.
8. Awkward datesNumeric day/month dates, "to date", an overlap and a gap.
9. Long contractor CVTwelve or more assignments, some through an agency.
10. LinkedIn PDF exportThe "Save to PDF" file, which has its own two-column layout.

Step 2: score ten fields per CV

Upload each file the way your team normally would, then compare the record with the original. Score each field 1 if it is exactly right, 0 if it is blank, and mark it W if it is filled in but wrong. Ten CVs times ten fields makes 100 checks, so your total is already a percentage.

PARSER TESTED: [ATS / vendor]        DATE: [..]      TESTER: [..]
Score per field: 1 = exactly right · 0 = blank · W = filled in but wrong

Field                         | CV1 | CV2 | CV3 | CV4 | CV5 | CV6 | CV7 | CV8 | CV9 | CV10
------------------------------|-----|-----|-----|-----|-----|-----|-----|-----|-----|-----
1  Full name                  |     |     |     |     |     |     |     |     |     |
2  Email                      |     |     |     |     |     |     |     |     |     |
3  Phone                      |     |     |     |     |     |     |     |     |     |
4  Location (city)            |     |     |     |     |     |     |     |     |     |
5  Current job title          |     |     |     |     |     |     |     |     |     |
6  Current employer           |     |     |     |     |     |     |     |     |     |
7  Current role start date    |     |     |     |     |     |     |     |     |     |
8  All roles found, in order  |     |     |     |     |     |     |     |     |     |
9  Highest qualification      |     |     |     |     |     |     |     |     |     |
10 Skills (no junk, none lost)|     |     |     |     |     |     |     |     |     |
------------------------------|-----|-----|-----|-----|-----|-----|-----|-----|-----|-----
Correct (count of 1s)         |     |     |     |     |     |     |     |     |     |
Wrong (count of Ws)           |     |     |     |     |     |     |     |     |     |

TOTAL CORRECT: [..] / 100      TOTAL WRONG (W): [..]      TOTAL BLANK: [..]
Contact fields (rows 1–4):     [..] / 40
Career fields  (rows 5–8):     [..] / 40   <- the ones your searches depend on
Worst CV type: [..]            Worst field: [..]

Step 3: read the result

• Look at rows 5 to 8 before the total. A parser can score 85 overall by acing emails and phone numbers while misfiling a third of job titles. Searches run on titles, employers and dates.
• Count the Ws separately. Blanks get noticed and fixed. Wrong values sit in the database and quietly remove good candidates from search results.
• Find the pattern by column. If CVs 2, 6 and 10 account for most of the damage, you have a layout problem, and the fix is at intake. If CV 7 is the outlier, you have a language problem, which is a vendor question.
• Re-run it when something changes. Parsers are updated behind the scenes. Keep the ten files and repeat the test after an ATS upgrade or before a renewal.

Once you know the share of records that need a manual fix, the cost is simple arithmetic. Put your own numbers in:

Resume parsing software: who makes it

The market splits into four groups, and recruiters mostly meet the second one.

• Specialist parsing engines. Companies whose product is the parser itself, sold as an API to software vendors, job boards and large employers.
• The parser inside your ATS or CRM. Most recruitment systems embed one, frequently licensed from a specialist. Bullhorn's knowledge base, for instance, says its upgraded parser, rolling out through 2026, incorporates Textkernel technology alongside its existing framework.
• General-purpose language models and document AI. Tools not built for CVs that a technical team can prompt to return CV fields.
• Open-source libraries. Free to run, yours to maintain. A fit for engineering teams, rarely for an agency.

These are five of the long-standing specialists. The descriptions are taken from each vendor's own site on the day we checked (October 2026) and are their claims, not our test results. We haven't listed prices: most are usage-based or quoted, so check each vendor's pricing page.

If you run an agency, the practical takeaway is short: you are probably not shopping for one of these. You are choosing an ATS, and the parser comes with it. Ask the ATS vendor which engine they use, then run the torture test during the trial.

CV parser vs CV formatter: the distinction most buyers miss

Recruiters often arrive at parsing software while trying to solve a different problem: "candidate CVs look terrible and I need them in our template before they go to the client." Parsing sounds like the answer because it also involves software reading a CV. It solves a neighbouring problem.

One honest complication: a formatter has to read the CV too. Any tool that moves a candidate's experience into a new template must first work out where the experience section is, so there is parsing-like work inside every formatter. The difference is the destination, and it changes what counts as an error. A parser can normalise "Fin. Mgr" to "Finance Manager", because a database wants consistency. A formatter that did the same would be altering the candidate's CV, and a client document should say what the candidate said.

Some parsing vendors sell both, as separate products, which is itself a clue. Textkernel's parser page mentions a separate Resume Formatter product next to the parser.

Which one you need: a decision rule

• "I can't find candidates I know are in our database." Parsing. Run the torture test, then talk to your ATS vendor.
• "Consultants retype CVs into the CRM." Parsing. Your ATS probably has a parser that is switched off, or one that nobody trusts.
• "We're building a job board or a careers portal." Parsing, and this is the case where you buy an engine directly.
• "Our shortlists go out in five different layouts." Formatting.
• "Consultants spend their evenings copy-pasting CVs into the Word template." Formatting.
• "A client's portal mangles the CVs we submit." Both ends meet here. Your client's ATS is parsing your document, so send it something a parser can read: single column, real headings, no text boxes. Our ATS-friendly CV formatting guidelines cover the details.

That last case is where the two tools help each other. A CV that has been rebuilt into a clean single-column template gives any parser, yours or the client's, far less to trip over than the two-column original. Formatting does not fix a parser, but it removes most of the ten causes above from the documents you file and send.

Questions recruiters ask about CV parsing

Where to go next

If the formatting side is your problem, our comparison of the best CV formatting software for recruitment agencies lists the tools in that category, ours included, with the trade-offs. If you are choosing the system the parser lives in, start with the recruitment agency software overview. And if the torture test turned up skills nobody can see on the page, read how to spot AI-manipulated CVs before you trust the skills field again.