Where AI is used, and what it never decides

The Offer Risk Snapshot: Google Gemini with Google Search grounding

The Offer Risk Snapshot is the only OfferMettle tool that calls a model at all, and the only tool that talks to a server, because grounding a claim in a real, dated source requires it. It calls Google Gemini with Google Search grounding enabled, so every source URL the Snapshot returns is a page the model actually read at the time it answered — never a source the model invented or paraphrased from memory.

Every claim the Snapshot shows carries four things together, always all four, never one standing in for another:

• The claim — a plain statement about the employer, stated once.
• A source link and a date — the exact page Gemini's grounding tool read, and the date that page was published or last updated.
• A confidence label — how sure the grounded response is in this specific claim, drawn from exactly the enum high | medium | low. No other value is valid.
• A classification — drawn from exactly the enum fact | estimate | judgment, independent of confidence: something the source stated directly, something derived or estimated from what the source stated, or a judgment call OfferMettle is making about it. A claim can be a high-confidence estimate or a low-confidence fact — both labels always render together, never one alone.

A claim that arrives without a real, non-blank source URL and a non-blank source date is rejected before it ever reaches a reader. It is discarded server-side, never shown as an unlabelled fact and never silently downgraded into a lower-confidence claim — a claim either carries a dated source or it does not exist for the reader.

See something in a Snapshot that looks wrong regardless of how it was sourced? Use the corrections page to tell us what to fix.

Your Decision Brief: reads your documents, never prices the market

Once you're in the paid Sprint, your Decision Brief is built from the offer letter and any other documents you upload. It extracts what is actually written in them — pay, vesting schedules, clawbacks, deadlines, and terms worth questioning — and cites every statement back to the document and page it came from, so you can check any line yourself. It is a reading of your paperwork, not a second opinion about it.

Your Decision Brief never contains a compensation range or a market benchmark, for any role, level or location. Not one a model produced, and not one a person typed either — the service does not publish market figures at all. If you need to know what a role pays, that is a different question than the one we answer, and we would rather say so than hand you a number we cannot stand behind while you are on a deadline.

It also never asserts that a clause is enforceable and never gives legal advice. Where a term is worth a second set of eyes, your Decision Brief flags the question and tells you to raise it with an attorney — it does not answer the question itself.

AI output is never presented as a human judgement

Every AI-produced statement you see on OfferMettle is labelled as what it is: a grounded claim with its confidence and classification, or an extract from your own documents with the page it came from. Nothing is presented as though a person concluded it, wrote it, or stands behind it personally. No one at OfferMettle reads your documents — the service is software, and we would rather tell you that plainly than let a byline imply otherwise.

Nothing here recommends a course of action. Your Decision Brief separates what your documents state from what is estimated and what is a judgment, and stops there. Whether to accept, negotiate, investigate or decline is your decision, made with your offer laid out in front of you — not one we make on your behalf and not one a model makes either.

Everywhere else on the site: no model in the loop

The comparison, counteroffer, startup-equity and deadline-request tools compute entirely in your browser from the numbers you enter — no model reads your inputs, and no AI touches the arithmetic. The written guides and pages on this site are drafted following our editorial policy, which states where AI assisted a draft and where a human wrote or checked the result, and every byline follows the same standard for who is allowed to be named as an author or reviewer.

For how the Snapshot's grounding and the comparison tool's calculations work in full technical detail, see the methodology page covering the three-panel model and the Snapshot's sourcing rules. Found something on this page that's out of date or wrong? Submit a correction.