AI-representation observability — watching how assistants describe you

The blind spot

You almost certainly measure how Google sees you — rankings, impressions, clicks. You almost certainly don't measure how ChatGPT, Claude, Gemini, or Perplexity describe you. Yet for a growing share of your prospects, the assistant's summary is the first impression. If it's wrong — outdated pricing, a discontinued product, a competitor named as the market leader — you're losing deals you never see.

This is the observability half of the Agent Surface & Observability pillar, criterion AS3. It closes the loop on share-of-model: citeability is how you earn a place in the answer; observability is how you check what that answer actually says.

What to measure

Five things, in rough priority order:

1. Accuracy — is what the assistant says about you factually correct?
2. Citation — does it name and link you, or describe you anonymously?
3. Share-of-model — across your key questions, how often do you appear versus competitors?
4. Hallucinations — does it invent products, claims, or policies you don't have?
5. AI-referral traffic — how many visits now arrive from assistant links, and to which pages?

How to run a prompt-test panel

You don't need a platform to start — you need a process:

• Write the 20 questions a real prospect would ask an assistant in your category, from "best tool for X" to "does your-brand support Y?"
• Run them across the major assistants, the same way each time.
• Log the verbatim responses: what's said, who's cited, what's wrong.
• Repeat on a cadence — monthly is a sensible start — so you see movement, not just a snapshot.

That logged, dated record is your baseline and your evidence. It's also exactly the kind of repeatable check the Phase 3 automated audit is built to run for you.

Closing the loop: fix it at the source

Observability only pays off if it drives action. When you find a misrepresentation, the fix is almost never to argue with the model — it's to correct the substrate the model learned from:

• Wrong facts? Publish the correct, dated, canonical statement prominently and in structured data.
• Anonymous mentions? Strengthen your entity presence and authorship signals so you're attributable.
• Stale pricing or products? Update the source pages and their dateModified; engines weight freshness.

Fix the source, re-test next cycle, confirm the answer moved. That feedback loop — measure, correct, re-measure — is the whole discipline.

What to do this week

1. Write your 20 prospect questions and run them across the major assistants today.
2. Log every response verbatim, flagging inaccuracies and missing citations.
3. Pick the single worst misrepresentation and fix it at the source this week.
4. Diarise a re-test for next month — observability is a loop, not a one-off.