What an LLM Perception Check Tells You About Your Brand

Buyers no longer start every search on Google. A growing share of them open ChatGPT, Claude, Gemini or Perplexity and ask a question in plain language: who makes a good one of these? is this company any good? what should I use instead?

Whatever those models answer is, for that buyer, the truth about your brand. Most companies have never checked what it is.

The check itself

An LLM perception check is a step in our client onboarding process, and it happens before we write a single deliverable. We query the major models about the brand and record what comes back. That includes the flattering answers, the wrong ones, and — often the most revealing result — the models having nothing to say at all.

That output is saved as a baseline. It sits alongside the rest of the research we do before we ever meet a new client: their website and product pages, a social media audit, a review-site check across places like G2 and Trustpilot, and an SEO snapshot covering domain authority, keyword rankings and backlinks.

What the answers usually reveal

Three patterns come up repeatedly.

Silence. The model does not recognise the brand and either says so or describes the category instead. This is the most common result for younger companies, and it is not neutral — a competitor is being named in that answer instead.

Stale information. The model describes a product that has since changed, a positioning the company has moved away from, or pricing that no longer exists. Models learn from what has been written publicly, and if nothing recent has been written, they answer from what is old.

Someone else's framing. The model repeats how a review site, a forum thread or a competitor's comparison page characterises the brand. That framing was written by someone with no stake in getting it right.

Why we do this before the first meeting, not after

We bring the findings to the onboarding meeting as a deck and walk the client through it. Then we ask them to confirm, correct or add to what we found.

That ordering is deliberate, and it does two things. It shows the client we understood their market before they had to explain it. And it produces a corrected, client-approved picture of the brand — which becomes the single client context document everything we write for them is drawn from.

That last part matters more than it sounds. Large volumes of unorganised context are exactly what produce AI-generated content that is confidently, plausibly wrong about a client's own product. Organising the context first is how we prevent it.

What changes afterwards

The baseline is not a report that gets filed. It defines what the content plan has to accomplish: which questions the models answer badly, which topics have no credible public discussion behind them, which corrections need to exist somewhere a model can find them.

You cannot improve how AI describes you without first knowing what it currently says. That is the whole reason the check comes first.