Protect · Hallucination Defense
Catch the lies before your customers do.
Engines answer confidently, and sometimes they are confidently wrong about you. Synthic crawls your site for ground truth, asks each engine about you, and checks every claim against what you actually say.
Claims found
24
Inaccurate
6
corrections ready
Unverifiable
5
no page says either way
- Inaccurate
“Synthic starts at $99/month.”
→ Actual: free first audit, Founder at $49/mo
- Accurate
“Supports ChatGPT, Claude, Gemini, Perplexity, Grok.”
✓ Matches the site
- Unverifiable
“Founded in San Francisco.”
Site is silent on location
- Inaccurate
“No longer offers schema generation.”
→ Schema & llms.txt are live features
3
verdict types
3
severity levels
live
ground truth crawl at scan time
1
correction per inaccurate claim
Anatomy
How it's built, part by part.
01Ground truth
Your site is the source of truth. Literally.
Your pages, crawled at scan time, are the only reference. Each claim gets a verdict against them — supported, contradicted, not addressed — and a contradicted claim gets a correction written from those same pages.
ground truth
Your site, crawled live at scan time
engine claims
What each engine says when asked about you
| verdict | meaning |
|---|---|
| accurate | supported by what your site says |
| inaccurate | contradicted by your site — correction written from your own content |
| unverifiable | your site is silent — a page to write, or a claim to let stand |
severity: low · medium · high — high for wrong pricing, wrong product claims, brand confusion
02Triage
Fix the brand-damaging errors first.
Wrong pricing and mistaken identity are high. A stale team size is low. Findings land in the Action Center in that order, and a new high-severity finding can page Slack or email the moment it appears.
| engine | text | which engine answered |
| model | text | the exact model id used |
| prompt | text | the question, as sent |
| response_text | text | the answer, as returned |
| mentioned | bool | did it name the brand |
| prominence | 0–100 | how early, how central |
| sentiment | enum | positive · neutral · negative · absent |
| competitors_mentioned | text[] | who else it named |
| citations | url[] | sources it cited |
| mention_basis | enum | retrieval · training |
| runs · mention_count · variance | int · int · real | multi-run agreement |
| latency_ms · capture_method · created_at | int · text · time | the evidence line |
What you get
Three things, done properly.
01
Claim-by-claim verdicts
Each thing an engine says about you is checked against your live pages: accurate, inaccurate or unverifiable.
02
Severity triage
Wrong pricing and mistaken identity rank high; a stale team size ranks low. Fix in that order.
03
Corrections you can publish
For every contradicted claim, a correction written from your own content — copy, paste, re-scan.
Specification
Inputs, outputs, cadence, evidence.
- Ground truth
- Your pages, crawled live at scan time (Firecrawl or direct fetch) — never cached from a previous scan
- Engines
- Every engine connected to the workspace; Gemini on the free plan
- Verdicts
- accurate · inaccurate · unverifiable, each with the supporting or contradicting passage
- Severity
- low · medium · high — high for wrong pricing, wrong product claims, brand confusion
- Outputs
- Findings ranked by severity; a correction per inaccurate claim written from your own content
- Downstream
- Findings feed the Action Center; an alert rule fires on new inaccurate claims
The answer is being written right now.
Run a live audit and see exactly where you stand — before your competitor does.
- Free first audit
- No credit card
- Live engine data
- Your data stays yours
- Cancel anytime