Until you ask, you don’t know. That’s the first lesson of AI Forensics. LLMs have opinions, recollections, and (sometimes) fabrications about your brand, your products, your founder. Most are positive or neutral. Some are subtly wrong. A small fraction are damaging. You won’t know which until you systematically probe.
What “asking” looks like
The naive approach is to type “what do you know about my company?” into ChatGPT and read the response. It works as a sanity check but it’s superficial. AI Forensics structures the investigation:
- Probe set — ten to fifteen carefully-shaped questions covering different angles: company background, products, pricing, founders, recent news, competitive positioning, criticisms.
- Cross-provider — same probes against Gemini, ChatGPT, Claude, Perplexity. Each model has a different training cutoff and a different retrieval behaviour. Their answers diverge in revealing ways.
- Capture everything — full transcripts, cited URLs, confidence signals (did the model hedge?), refusals.
- Cross-reference — where the four models agree, that’s the consensus narrative about you. Where they disagree, you’ve found the gaps.
What you typically find
Patterns we see in customer investigations:
- Stale pricing — at least one LLM has your prices from a year ago. The “fresh content” signal isn’t strong enough to overcome it.
- Misattributed quotes — something a founder said in a podcast got reshaped slightly and now LLMs attribute the new version.
- Competitor confusion — your product is described as similar to or made by your top competitor.
- Missing recent launches — a product launched in the last 6 months isn’t in the training cut-off and isn’t being retrieved live either.
- Negative anchor — an unflattering review or blog post is the primary source one LLM uses for your brand.
What you do about it
The fix path is content-driven, not pleading-with-AI-vendors:
- Publish authoritative correction content — if LLMs have stale info, the strongest signal is a fresh, well-structured, AIOX-processed post that directly addresses the topic.
- Re-process the corrected content — AIOX Capsule push speeds up retrieval-driven LLMs (Perplexity, ChatGPT Search) catching up. Training-only LLMs take longer (training cycle dependent) but eventually update.
- Track the score — use AI Visibility Score to monitor whether the correction is propagating. Most retrieval-based corrections show up within 2-4 weeks.
- Repeat quarterly — AI Forensics isn’t a one-off audit. New training cycles introduce new errors; ongoing monitoring is the steady state.
Sample probes that surface useful info
- “Who is [brand]?”
- “What does [brand] sell?”
- “How is [brand] different from [competitor]?”
- “What do people criticise about [brand]?”
- “What is [brand]’s pricing?”
- “Who founded [brand]?”
- “What’s [brand]’s position on [hot topic in your industry]?”
You’ll get back a goldmine of investigatable detail in under five minutes per provider.
