The pipeline nobody explains end to end
Most advice about "beating the ATS" focuses on one stage of a longer pipeline — usually parsing, sometimes keyword matching — and treats it as the whole story. It isn't. A real application, at a company running a modern screening stack, moves through several distinct stages before a human ever opens it, and each one can independently cost you a shot at getting read.
Stage by stage
- Upload. Your file arrives as a PDF, DOCX, or occasionally plain text. Nothing has happened to it yet — this stage just gets bytes into the system.
- Parse. The document gets read: text extracted, tokenized, split into sections (Experience, Education, Skills), specific entities identified within each (a company name, a date range, a job title). See our ATS guide for exactly what breaks here.
- Structure. Extracted fields get written into the ATS's own database — the format a recruiter's search tool actually queries later.
- Lexical match. The crudest layer: does your CV's text literally overlap with the job posting's? Keyword presence, nothing more. Easy to game, easy to pass, and on its own it tells a recruiter almost nothing about fit.
- Semantic match. A layer built on embeddings compares the meaning of your experience to the posting's requirements, not just shared words, so "led sprint planning" can register as Scrum Master experience without using that exact phrase.
- Score and rank. Whatever survived the previous stages gets a number, and candidates get ordered by it. The threshold for "a human looks at this" varies wildly by company and by how many applications came in.
- Human review. The stage every previous one was gatekeeping access to. Everything before this point is infrastructure deciding who gets read. This is where someone actually reads.
Where candidates actually lose ground
Stage 2 (parsing) and stage 4 (lexical matching) are where the most damage happens, and they're also the two stages candidates can most directly control: a clean, single-column, real-text document survives parsing, and specific, detailed achievements survive both lexical and semantic matching better than vague ones do. Stages 3, 6, and 7 are largely out of your hands — you can't control an ATS's database schema or a recruiter's specific bar for "worth reading."
What this means for you
You can't optimize for every stage equally, and you don't need to. Get the document structure right, and write with real, specific substance instead of padded keywords, and you've addressed the two stages actually shaped by what you submit. The rest of the pipeline is the system's problem, not yours to solve by guessing at a black box.
Related guides
- How an ATS actually reads your CV — stage 2, in full
- The résumé keyword myth — stages 4 and 5, and why passing one isn't passing both
- The agentic problem — more on the semantic layer specifically