Your Matching Engine Scores Against the Posting You Wrote

What is your matching engine comparing?

Two blocks of text, yours and the candidate's. A matching feature builds a representation of the requisition text, compares each application against it, and orders the review queue by closeness. Nothing in that operation knows what the job is. It knows what you typed, which caps the quality of the ordering at the specificity of the posting.

That is why boilerplate is expensive here. A benefits paragraph, a values statement and a line about a growing team compare about equally well against every application in the pool, so they contribute noise to the ordering while taking up the space a task description could have used. A requirements list of twelve bullets where three actually matter produces a queue shaped by the nine that do not.

The comparison also runs against whatever your parser extracted from the upload, which means two failure modes stack. A skill that never made it into the record cannot match anything, however well the posting is written, so it is worth knowing what your parser drops before you tune the text.

One more thing has changed underneath this. When a large share of applications are drafted with the posting open in another window, similarity to your posting measures how the application was drafted. The loop closes: your job description becomes the candidate's prompt, and the ordering rewards whoever mirrored it most completely.

Write the three tasks the role does in a normal week

Open with the work. Three sentences, each naming a task the person does most weeks, the thing they produce, and who receives it. That paragraph is the highest-value text in the posting: it gives a qualified reader something to recognise, and it gives your matcher concrete language to compare against instead of adjectives that fit every applicant equally.

The difference is easy to see side by side.

> Before. Seeking a detail-oriented marketing professional with strong communication skills and a track record of success in a fast-moving environment.

> After. You will write and ship two lifecycle email sequences a month, brief a designer on each, and report open and reply rates to the head of growth at month end.

What the second version adds is specificity, which is what a matcher has to work with. It also does something the first version cannot: it lets a strong candidate rule themselves out, which costs you nothing and saves them a week.

Then state the minimum requirements separately from the tasks, and keep that list to things a person either has or does not. Anything you would not disqualify a candidate for belongs under preferences, and putting it in the requirements block trains both the reader and the software to treat it as a hard minimum. If you are drafting the posting with an assistant, this is the section to write yourself, because a generated requirements list defaults to the genre average and the genre average is exactly the text every other posting already contains.

Should the posting ask for AI skills?

Ask for it only when you can say what AI does in the role. A line reading familiarity with AI tools names no task and no output, so it sits at the same distance from every application in the pool and adds nothing to the ordering or to a reader's understanding of the job. Naming the task and the expectation attached to it does both at once.

The language is genuinely spreading, which is why the vague version has stopped working. Indeed Hiring Lab's August 2026 snapshot puts AI-related job postings at 6.3% of US postings, past their prior peak of 3.3% in 2022 ¹. That counts postings that mention AI, which measures employer language rather than adoption or requirement. The spread is uneven by function: as of December 2025, nearly 45% of data and analytics postings carried AI-related terms, against about 15% in marketing and 9% in human resources ².

It is also priced. Lightcast reports that advertised salaries in postings mentioning AI skills run 28% higher than in postings that do not, roughly $18,000 more a year, and that in 2024 over half of postings requesting AI skills sat outside IT and computer science ³. Read that as a raw comparison between two groups of postings rather than a like-for-like premium, since AI-mentioning postings skew senior and urban, and Lightcast sells skills data. The useful part is directional: the phrase has a price attached, so writing it in without meaning it invites applications you did not budget for.

The fix is the same move as the tasks paragraph. Write the sentence that says what the assistant is for: drafting first-pass copy that a human edits, reconciling two data sources before a person checks the exceptions, summarising calls into a template. That sentence tells a candidate what the job is, and it gives your matcher something narrower than a category. Working out what AI actually does in the role before the posting is written is the step that makes the sentence writable, and there is a separate craft question in how to phrase the requirement without drowning in unqualified applicants.

Read your own top twenty before blaming the tool

Post the requisition, wait a week, then open the first twenty applications in the order your system presents them and score each one against the three tasks. If the ordering tracks your judgment, the posting is doing its job. If it does not, look at what the top twenty have in common, because that shared thing is what your text asked for.

Usually it is a phrase. A tool name so common in the field that listing it tells you nothing. A seniority word that pulls in people managing teams for a role with no reports. A degree requirement copied from a template that was written for a different function. Each of those is one line to edit, and editing the line is faster and more durable than adding a filter downstream to undo it.

Edit the posting rather than the filter, as a standing rule. A filter is invisible to candidates, accumulates without an owner, and has to be maintained by whoever inherits the requisition. Posting text is public, self-documenting, and fixes the applicant pool and the queue ordering in the same edit. It also gives you something to point at later when someone asks what the role required.

One caution worth holding. A feature that orders your review queue may sit inside the definition of an automated employment decision tool. New York City's Local Law 144, in force since January 1, 2023 and enforced since July 5, 2023, reaches a tool whose output substantially assists or replaces discretionary decision-making about employment, so the classification follows from what the ordering does in your process, whatever the vendor calls the feature ⁴. Before you rely on the queue order, settle whether the feature counts as an AEDT for the places your candidates sit, write down the reasoning while it is fresh, and put the classification to counsel if a real decision rests on it.

See a sample report >