AI Reputation Desk

When an AI engine gets your brand wrong, we trace it to the source.

We red-team how a brand degrades inside AI engines, trace each false or damaging claim to a fixable public source signal, and compile the concrete upstream fixes.

See a sample deliverable How it works

How it works

Probe

Red-team the worst answers

A hostile-prompt battery spanning four attack classes — competitor-prompted, jailbreak / role-play, worst-data-surfacing, and off-voice elicitation — put to the engines under a hard per-run budget cap. The budget is spent question by question, so a run can stop before the end of the battery; every plan names the questions it did not ask. Every answer that comes back is captured verbatim.

Diagnose

Trace each claim to its source

For each false or damaging claim we map from the answer's citations and the pages we read on the brand's own site — its copy and its structured data — to the specific surfaceable signal feeding it.

Plan

Compile the AI Correction Plan

One accessible page: ranked claims, each with the verbatim model snapshot, the traced source, and the concrete upstream fix. Claims we cannot evidence are listed as unconfirmed — we never assert what we cannot prove.

The discipline

Never fabricate

A claim we cannot verify is marked unknown — never asserted. A "false" label requires a verbatim model snapshot and a traced source, both.

Human-in-the-loop

Nothing here is auto-sent, and no fix is executed by this desk. Every plan compiles to needs_review and carries that status on its face.

Accessible by default

Every rendered artifact is built with semantic structure, an OpenDyslexic toggle, and no color-only signaling, and is run through our static accessibility checks — specific criteria, not a full WCAG audit.

Scope of this work

We change surfaceable upstream signals and monitor whether the answer changes; we do not control or modify any AI model.