Measuring what LLMs say about a brand, not just whether they mention it

The short version

Problem: Buyers use LLMs to evaluate vendors, and brands can't see what the models say about them.
Insight: AI visibility tools track whether you appear, not which associations come with you.
What I did: Built Konote, and ran a 1,000-prompt study on HubSpot across Claude and GPT-4o.
Result: Associations proved stable enough to track, and not always the intended ones.

Abstract painting: an inked eye at the centre, with coloured shapes on rings around it, the near ones tied on, the far ones on broken lines, on a plum field

The full story

Everyone tracked visibility, nobody tracked what was said

The talk was all about AI visibility: does your brand show up? But 6sense research suggested buyers use AI less for discovery and more to evaluate platforms. What the models say about you matters as much as whether they mention you.

I had just read Ulli Appelbaum's The Science of Brand Associations, on brands living as networks of associations in buyers' minds. Those associations now live in LLMs too.

First, prove it can be measured

Rand Fishkin's SparkToro research found AI brand recommendations are highly inconsistent: a few brands appear reliably, the rest is noise. I wanted to know if associations behaved the same way.

I ran 1,000 prompts on HubSpot across Claude Sonnet 4.5 and GPT-4o, split into brand, category, and competitor prompts, and tracked five associations. When I suspected a handful of SMB-angled prompts were skewing results, I removed them and ran 400 more. The scores barely moved.

Associations held across 1,000 prompts

Associations were stable, and a reliable metric. Ease of use held everywhere, inbound marketing barely surfaced in category questions, and the two models disagreed on 'all-in-one'. I published the study on LinkedIn, and it got good feedback.

Then the method became a tool

Konote turns the method into a tool. You set the associations you intend to own, it scans both models, and shows which land, which are missing, what the models add, and how competitors compare. Peec.ai has since added a similar feature.

The work

Konote overview: the associations LLMs attach to Salesforce, scored against its intended positioning — Overview: intended against actual associations (click to enlarge)

Attribute scores by model (click to enlarge)

Try it on your brand →Free scan. Directional, not definitive.

The study

HubSpot, 1,000 prompts, Claude Sonnet 4.5, and GPT-4o

60%+Ease of useThe one association that holds everywhere: above 60% for both models and all three prompt types.
14–17%Inbound marketingHubSpot's founding story almost vanishes when buyers ask generic CRM questions.
Claude ≠ GPT-4oAll-in-one platformClaude attaches it across every prompt type. GPT-4o mostly doesn't.
≈ 0 changeRobustness checkRemoving the SMB-angled prompts and running 400 more barely moved any score.

Read the full write-up →

What I'd do differently

I never marketed Konote properly. It did its job for the study, but I'd have given it a real launch.