Analyze how a major AI search product handles citation and grounding.

Glassform is a search product. Type a question and it writes a paragraph answer with small numbered footnotes, each one a link to a real page. Rukshana Patel researches auto-insurance claims and uses it to look up repair-cost benchmarks and injury-claim precedent before she signs off on a payout.

How to answer this, stage by stage

The interviewer isn't grading whether you've used Glassform. They're grading whether you know the difference between a citation and a checked citation.

Let's learn

What happens when a citation looks real, clicks through to a real page, and the page still doesn't say what the sentence next to it claims?

Glassform is a search tool. Ask it a question and it writes a short paragraph answer, with small numbered footnotes linking out to the pages it drew from.

For a long while, Rukshana clicked through two or three footnotes out of every ten Glassform answers, confirmed the number matched, and moved on. That habit alone was saving her something close to forty minutes a claim file compared to searching manually.

Then Glassform's writing got smoother. The footnotes multiplied, four or five per paragraph, each one a clean little pill with a favicon and a domain name. Rukshana, like the rest of her team, stopped clicking almost entirely. The pills looked thorough. Thorough felt like checked.

Glassform's pill shows one of these four. The other three are exactly what makes a citation checkable instead of just clickable.

At its worst: a $38,000 payout, approved on a footnote that linked to a real, current repair-cost guide, for the wrong vehicle class entirely. The claim read smoothly. Nobody caught it until an appeal, three months later, forced a re-check.

What I would leave alone: a citation backing a low-stakes definitional line, like "a standard policy typically covers collision damage," linked to a glossary page, doesn't need the same scrutiny as a specific dollar figure or a precedent. Forcing a quote onto every single footnote, including the harmless ones, just adds noise for no real gain.

The lesson: a citation that merely proves a real page exists is not the same feature as a citation that proves the sentence next to it is true. A search product that treats those two things as one feature will feel completely trustworthy right up until someone actually checks.

Now here is the same thing as a story

The short version above is what you'd say defending this answer live. Read this one for how Rukshana actually found the gap.

Rukshana can tell a straightforward fender-bender claim from a disputed total-loss claim within the first paragraph of an adjuster's note, a skill six years of reading them built into her without her noticing.

Glassform arrived on her team in the spring, and for months it was simply faster. She'd ask it for the going repair-cost range on a given part for a given vehicle class, read the paragraph, glance at the footnote pills, click one or two to be sure, and move to the next file.

Both pills render the exact same way. Nothing in the product tells you which one you're looking at.

By autumn she wasn't clicking at all. The answers had gotten longer, the footnotes more numerous, and clicking four or five links per file just to spot-check felt like undoing the whole point of the tool.

The pipeline has four steps. Only one of them was ever a real check, and it quietly stopped happening.

Then her review director ran an audit: ten Glassform-cited claims, pulled at random, every footnote checked by hand against the actual source page.

The dangerous quadrant is top left: polished, confident, and wrong. That's exactly where the vehicle-class citation landed.

One of the ten was the $38,000 claim: a footnote to a real, current repair-cost guide, for a different vehicle class than the one in the file. The citation existed. The claim it supposedly backed did not.

A fluent answer and a correct one are not the same test. Glassform had only ever been graded on the first.

With the quoted line shown under every footnote, the same claim now gets caught on the first read: the quote under the citation names a different vehicle class than the one on Rukshana's screen, in plain text, no click required.

The old pill asked Rukshana to trust that a link existing meant a claim was checked. The new one shows her the actual sentence and lets her judge it herself.

We built the clean pill because five footnotes with quotes under each one looked cluttered in an early demo, and nobody had pushed back on it yet. It took watching a real payout get reopened, and 1,200 prior approvals get flagged for re-review, to see that the clutter was actually the evidence.

AUDIT, run on one footnote pillNot a design framework. AUDIT exists to judge whether a claim, or a citation, has earned your trust.

Five checks, one page. The one Rukshana's team skipped for five months was the last one: testing it themselves.

The recap, one line per letter: ask who ran the check is Rukshana's team, not Glassform, and only rarely by month five; uncover the eval set is the 50-citation hand sample, not a vendor score; demand the version pin is the missing date and model version on every pill; isolate what's missing is the quoted line itself; test it yourself is the director's ten-claim audit.

And if you want to be sure it really works, try it somewhere elseSame five letters, a factory floor instead of a claims desk. This time the missing piece isn't a quote, it's a manual revision number.

Torqueline is an AI assistant a mid-size auto-parts manufacturer built for its maintenance floor. A technician asks it for a torque spec on a press machine, and it answers with a citation to the equipment manual.

Mapped onto AUDIT: ask who ran the check finds that nobody at the manufacturer had ever confirmed Torqueline's citations pointed at the manual's current revision, not a superseded one still sitting in the same file share. Uncover the eval set finds there wasn't one; the vendor's "manual accuracy" claim was based on whether a citation resolved to any manual page, not the right revision. Demand the version pin finds the actual fix: showing the manual's revision number and date next to every citation, the same missing piece as Rukshana's missing quote, just wearing a different name. Isolate what's missing is that revision number. Test it yourself is a technician cross-checking one torque spec against the physical rating plate bolted to the machine, the factory-floor version of clicking through.

Two manual pages, one live and one three revisions old. Nothing in the citation told the technician which one he had.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "the pill proves a link, not a claim. Show the quote." and stop.
Cost: there's no budget to redesign the footnote UI this quarter. Say so, and start by logging a hand-checked sample of 50 citations weekly, so the team at least knows the real rate instead of assuming click-through still happens.
The model gets better, for real: if Glassform's underlying model gets meaningfully more accurate, that's still not a reason to skip the quote. A better model that's wrong once in 50 citations does just as much damage as a worse one, if nobody can tell which citation is the wrong one.

Where people run it wrong.
They treat "the link resolves" and "the claim is grounded" as the same fact.
They ask for a single "grounding score" instead of asking what that score was actually tested against.
They assume more citations per answer means more trustworthy, when it can just mean more surface area nobody is checking.

How to use it live. When an interviewer hands you an artifact, footnotes, a benchmark claim, a vendor report, ask yourself one question before anything else: if I clicked through right now, would I be checking a fact, or just confirming a link isn't dead? Say that difference out loud first.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0