How ATS actually works

What a Resume Parser Cannot Infer

Parsers extract, they do not deduce. If a bullet shows you did forecasting but the page never says forecasting, a search for it will not find you.

Here is the one sentence to keep from this whole cluster. A parser extracts fields without comprehension. It gives no benefit of the doubt.

That sounds obvious until you look at what it rules out. It rules out the assumption every candidate makes without noticing: that the reader will join two facts together. A human reading "built and maintained the rolling 13-week cash model" knows that is forecasting. A parser reads eleven tokens, none of which is the word "forecasting", and files them under a job. A recruiter then searches "forecasting" and you are not in the results.

You were not rejected. You were never retrieved.

What is actually extracted

Vendors publish this, and the lists are short.

Oracle documents what Taleo pulls out of a resume: first, middle and last name, address, city, postcode, home phone, email and place of residence; education level, graduation date, start date, institution and programme; and for work experience, start date, end date, employer, job function and responsibility.

Greenhouse documents what its Talent Matching step extracts: "skills, job titles, years of experience, start/end dates of employment, company names from employment history, and where available, a derived industry classification associated with the candidate's past employers".

Two vendors, one shape. Fields, with words in them. No item on either list is a conclusion drawn from a sentence.

What about the systems that "understand meaning"?

Some do more than exact string matching, and it is worth being precise about how much more, because the marketing language and the documented behaviour are not the same size.

Ashby's candidate search has a mode called Similar, which "returns records with variations of the word", and the example the documentation gives is that searching "localization" may bring back resumes containing "localized" or "localize". That is morphology. It is the same word in another form.

Greenhouse's Talent Matching highlights matched terms on the resume "either as an exact match or a semantically similar match". That is a genuine step further, and it is confined to one optional feature, in one system, at one stage of review, in an add-on the employer has to have bought.

Neither of these is inference. A word-variant matcher connects "consulting" to "consultant". It does not connect "ran the monthly reporting pack" to "management information". And the plainest gate of all, a boolean search over resume text, is pure string matching: Ashby's Contains mode returns "javascript coder" as a phrase but not "javascript code".

So the safe assumption stands. If the posting screens on a term and the term is not a word on your page, treat yourself as missing from that search.

Four things a machine will file wrong, and the fix

1. The job title that is not a title

Every system stores the title as one string in a title field, and a recruiter searches that field. So a title like

Portfolio Strategy Senior Manager (promoted 4 times)

is a real and creditable fact sitting in the wrong box. Stored as written, the title matches no search exactly. Moved into a bullet under the role, "promoted four times" is one of the strongest sentences on the page. The same applies to "(12-month FTC)", "(remote)" and anything after a dash.

2. The title nobody can place

If your employer calls you a Growth Lead, a parser gets no seniority level and no function out of that string, so you are absent from searches that use either. The usual fix is honest because both halves are true: put the market equivalent beside the internal one, as "Growth Lead (Marketing Manager)".

3. The acronym with only one form

This one fails in both directions and people only ever guard against one of them.

One mention of both forms covers both searches. It costs four words.

4. The degree with no level word in it

"MBA" contains no word an education parser can file as a level. Neither does "BSc" or "LLB". This matters more than it looks, because "Master's degree required" is a knockout question answered on the application form before anybody reads the document, and Oracle lists education level as its own extracted field for Taleo.

Writing it as "Master of Business Administration (MBA)" keeps your own wording and adds the word the filter is looking for.

5. The language with no level attached

"German" on a skills line tells a machine that a string appeared. It does not say whether you order coffee in it or negotiate contracts in it, and across most of Europe the advert does say which one it wants.

A language listed with no level is at risk rather than clear, and the honest fix is to name the level the way the local market names it. Put the framework level and the plain description together: "German, C1 (fluent, professional working)". The same reasoning applies to certifications, where a name with no year leaves a reader unable to tell current from lapsed.

The pattern behind all five of these is one pattern. A person reading your page fills in a missing level, a missing expansion, a missing category. A field in a database does not, and the field is what gets searched.

The implied skill, and the line we will not cross

The largest category is the one at the top of this post: work you genuinely did, described in words that do not name it.

Our checker finds these by looking for evidence in your bullets for a skill your page never names, and it ranks the ones the posting screens on first, because an unnamed skill nobody is searching for costs you no points today.

Then it stops. It renders a question, not a button.

That is a deliberate constraint and it is the most important design decision in the product. Adding "forecasting" to somebody's skills because a bullet implied it would be inventing a claim on their behalf. The person who has to defend that word in an interview is you, and the tool that typed it will not be in the room. So the split is enforced in code: things that are a wording problem come with replacement text, because correcting wording changes no claim. Things only you can substantiate come as a question, and no tool writes an answer.

How to work through your own page in twenty minutes

1. List the two or three terms a recruiter would type to find someone for this job. Not twenty. Two or three.
2. Search your own document for each one, as a word, using find. If it is not there, it is not there.
3. For each miss, ask which kind it is. Never did it, did it and did not write it, or did it and described it differently. Only the third kind is a wording fix.
4. Fix the wording ones by rewriting the bullet, not by adding the word to a skills list. A term with evidence under it survives the follow-up question; a term in a list does not.
5. Check every acronym for both forms, every qualification for a level word, and every job title for text that is not a title.
6. Leave the genuine gaps alone. A gap you name in a cover letter is a paragraph. A gap you paper over is a discrepancy.

Run your CV against a real posting and read the questions before the score. The questions are the part that only you can answer, which is exactly why they are worth your time.

Sources: Oracle Taleo, Candidate Management, Greenhouse, Talent Matching Data Processing FAQ, Ashby, Candidate Search, Greenhouse, Search resumes for keywords