How to Audit What an Assistant Says About Your Competitors
Run the same fixed prompt set you use for your own visibility and record what is said about each competitor, not only whether you appear. Capture four things per run: who was named, what capability each was credited with, which source was cited for it, and whether anything said about any vendor was factually wrong. The useful output is rarely 'we were not mentioned' — it is finding that a competitor is credited with something they do not do, or that a claim about you traces back to a page you control and can fix. Accept upfront that a large share of findings are not actionable, and decide before you start which categories you will act on.
This is the competitive version of a visibility audit and it is more useful than the self-focused one, because your own absence has many causes and someone else's specific attribution usually has a traceable source.
What to record per run
| Field | Why it matters |
|---|---|
| Vendors named | The competitive set as the assistant sees it |
| Capability credited to each | Where positioning is landing or not |
| Source cited per claim | The only actionable part |
| Factual errors, about anyone | Includes errors in your favour |
| Run number and date | Same-prompt variation is large |
Sample repeatedly, not once
The same prompt produces different answers across runs, so a single run tells you almost nothing about what is typically said. Use the same repeated sampling protocol as any other visibility measurement, and treat a claim appearing once in five runs differently from one appearing five times in five.
The three findings worth acting on
A claim about you sourced from a page you control and can correct. A competitor credited with a capability they do not have, where the cited source is a third party you could reasonably ask to correct it. And a capability comparison where the cited source is your own page making a claim you can strengthen with evidence. Everything else — models being wrong for reasons you cannot trace, competitors mentioned more often than you — is information rather than a task.
What not to do with this
Do not write content designed to correct an assistant's opinion of a competitor. Beyond being a poor use of effort, publishing content whose purpose is to depress a named company in generated answers is a reputational risk that outlives any visibility gain. Correct claims about yourself; report factual errors about others through whatever channel the source offers.
Our own interest here
We sell measurement of exactly this, so the recommendation to do it repeatedly and systematically is one we profit from. The honest framing: a team can run this manually with a spreadsheet and a fixed prompt list, and for a small competitive set that is genuinely sufficient. What tooling buys is consistency and history, not a different answer.
FAQ
How many prompts and how many runs? — Enough prompts to cover the buying questions that matter, and enough runs that you can distinguish a consistent claim from a one-off. Three to five runs per prompt is a common starting point; fewer and you are reading noise.
Is it worth auditing competitors we rarely lose to? — Only if they appear in answers to your buying questions. The competitive set an assistant names is often not the one your sales team recognises, and that mismatch is itself the most interesting finding this audit produces.
What if an assistant says something false about us? — Trace the citation. If it points at a page you control, fix the page. If it points at a third party, contact them. If there is no citation, there is usually nothing to act on, and saying so honestly is better than inventing a remediation.
Put this to work on your own website.
Fenn finds what your customers ask, drafts the articles and site fixes, and measures what ChatGPT, Claude, Gemini, Perplexity and Grok say about you — with every change waiting for your approval.