# How to evaluate AI job matching: an evidence-first checklist Evaluate an AI job recommendation against the job description and your actual experience. Separate supported requirements, conflicts and unknowns, especially location, work authorisation and seniority. A useful explanation cites evidence and makes missing information visible rather than hiding it behind a score. ## Judge the recommendation by what you can verify A matching tool should reduce the work of finding relevant openings without hiding the facts that determine whether you should apply. A high percentage is not enough. You need the source listing, the profile facts used, a reason for the recommendation and a clear account of what remains unknown. This guide proposes a candidate-side audit. It does not certify a product or predict whether an employer will interview you. Use it on a few actual recommendations before relying on a tool for an active search. NIST’s distinction between transparency and accuracy is useful: being able to see an explanation does not make the explanation correct. Sources: [NIST: AI risks and trustworthiness](https://airc.nist.gov/airmf-resources/airmf/3-sec-characteristics/) ## Start with the employer’s current posting Open the application link and confirm the title, location and responsibilities. The destination should be the employer or its recruiting system. A familiar logo on a job card does not establish that the posting is current or that the service has a relationship with the company. Compare the current page with the recommendation. Has the location changed? Is the role still accepting applications? Was a remote arrangement narrowed to one country? Record the discrepancy if there is one. Freshness is a property to check, not a badge to trust indefinitely. ## Separate three kinds of evidence A confirmed match has support in both the posting and your profile. A conflict has visible facts pointing in different directions. An unknown lacks enough information. These categories should not collapse into a single yes/no verdict. For example, “the role requires PostgreSQL and your profile describes PostgreSQL work” supports one part of relevance. “The role requires being based in Canada and you want to work from India” is a location conflict unless the employer offers a compatible arrangement. “The resume does not mention PostgreSQL” is missing evidence; it does not prove you have never used it. ## Work through a realistic recommendation Consider a hypothetical backend role that lists Python, ownership of production services, on-call participation and employment in the UK. Your resume describes Python APIs and an internal reporting tool. You prefer London and require visa sponsorship. The language and location preference are relevant, but the recommendation still has several unresolved requirements. First ask what the resume actually says about operations. An internal tool may have real users and operational responsibilities, but the words alone do not establish production ownership or on-call experience. Review your own work and add accurate detail if it is missing. Do not let a tool rewrite the experience into something you did not do. Next inspect the employer’s authorization language. A London office does not prove sponsorship. “Must already have the right to work” and “sponsorship available” mean different things; silence should remain unknown. If sponsorship is necessary and the listing does not establish it, ask the employer before treating the opening as a confirmed option. The useful output is therefore not “92% fit.” It is a short account of supported facts, gaps and the next question: Python work is evidenced; production ownership needs review; UK sponsorship is not established. You can then decide whether to investigate, apply or spend your time elsewhere. ### Diagram: Hypothetical Python role: separate supported language experience from unknown operational and visa requirements. 1. **Supported** — Python APIs provide evidence for the language requirement. 2. **Unknown operations** — An internal tool does not establish production ownership or on-call work. 3. **Unknown authorization** — London preference and an office location do not establish sponsorship. Ask the employer. ## Test whether corrections change the right thing Edit one inaccurate profile fact and inspect how the tool uses the corrected information in future recommendations. A responsible system should distinguish your saved profile from the original resume and explain when changes take effect. It should not quietly rewrite your employment history to agree with a role. Historical shortlists may remain unchanged because they record what was published at the time. That is not necessarily a bug, provided the product explains the distinction. What matters is whether future decisions use the corrected facts and whether you can understand which version you are viewing. Also inspect negative explanations. “No Kubernetes experience” is stronger than the evidence supports if the system merely failed to find the word. Prefer language that identifies the missing evidence and lets you correct it. The tool should not require you to accept its extraction as authoritative. ## Compare tools with a small, repeatable audit Use the same profile and preferences when comparing products. For each recommendation, record whether the original link works, whether the stated requirement is present, whether the cited profile fact is accurate and whether an important unknown was acknowledged. Do not combine these observations into an impressive-looking score unless you have a defensible reason to do so. Pay attention to the errors that cost you time: dead links, unsuitable locations, unsupported sponsorship claims or seniority requirements that contradict the recommendation. A small audit cannot estimate market-wide accuracy, but it can expose patterns you should investigate before paying. Keep privacy separate from relevance. A product can find plausible roles while having sharing defaults you dislike. Read its retention and deletion information, check whether your profile is public and identify any action it can take on your behalf. Recommendation quality is not permission to automate applications. ## Use the shortlist as a starting point for your decision Before applying, read the complete posting, choose the resume version you intend to submit and review the employer’s questions. Matching systems can help prioritize attention; they cannot know every hiring consideration or promise a response. Opened follows this candidate-controlled approach: review your facts, choose roles and locations, inspect reasons and follow direct employer links. Free emails arrive on Sundays and paid plans send daily updates. The product’s explanations should be held to the same standard described here: inspectable evidence, accurate limits and no invented hiring outcome. ## References - [NIST: AI risks and trustworthiness](https://airc.nist.gov/airmf-resources/airmf/3-sec-characteristics/) Updated 2026-11-02 by Opened editorial.