004 — case study

Grounds

Every AI answer about you, cross-examined against its own sources.

roleDesign & engineering
platformWeb
stackTypeScript, Next.js, SerpApi
year2026
Grounds hero
Landing page2026
148
searches in the recorded audit
8
markets compared
8
verdict types
context

The case that started it

In 2025 a Google AI Overview told searchers that a Minnesota solar installer was being sued by the state Attorney General. It wasn't. The answer cited four sources, and none of them contained the claim. The company reported losing a $150,000 contract.

Google corrected that answer, which is the point: generative answers change underneath you, so this has to be measured continuously, not checked once.

research

What I found

  • 01
    Existing tools measure visibility, not truth

    AI-visibility tools track whether a brand is mentioned and with what sentiment. None check whether the cited sources actually support the sentence they're attached to.

  • 02
    Answers differ by geography

    Generative answers diverge across regions, and a company can't see what searchers in other markets are shown. So location is treated as a variable.

  • 03
    Suppression is data

    Google often hides AI Overviews on adverse queries. Grounds records that as a measurement instead of an error.

  • 04
    An unreadable source must never become an accusation

    Paywalled or blocked sources get UNVERIFIABLE, which is excluded from scoring.

system

How it works

  1. Docketing: the entity, its aliases, locales and look-alike names.
  2. Line of questioning: factual, operational and adverse probes, like real searches.
  3. Deposition: probes run across engines and locations, stored as append-only observations.
  4. Claim decomposition: answers split into single claims, each bound to its citations.
  5. Cross-examination: each claim checked against cited pages and independent searches.
  6. Scoring and remedy: attribution integrity, plus the one source to fix first.
design

Design decisions

  • 01
    A legal vocabulary

    Docket, deposition, cross-examination, dossier. The metaphor tells you what the tool does before you read a line of docs.

  • 02
    Red is reserved for defects

    Only claims counted against the entity are red. Grounded is black, excluded is grey.

  • 03
    The dossier prints

    It's designed to be saved as a PDF and handed to someone, so it reads top-down like a report.

close-ups

Design, up close

Screens from the live site and the recorded Wolf River Electric dossier.

Dossier header and the four scores
Dossier header and the four scores
Verdict distribution: red only for defects
Verdict distribution: red only for defects
A live capture, re-checkable in SerpApi's archive
A live capture, re-checkable in SerpApi's archive
Headline: serif italic for the verb
Headline: serif italic for the verb
Deposition: replaying a recorded audit
Deposition: replaying a recorded audit