Tools that analyze how LLMs talk about my company: how we audit brand mentions for clients
How we analyze what LLMs say about a client's company: mentions, sentiment, framing, competitors and sources, and the tools we use to run the audit repeatedly.
The first thing most clients ask is some version of "what does ChatGPT actually say about us?" It is a good question, and an uncomfortable one, because the answer is often a mix of accurate, outdated and wrong. We run a structured audit to answer it, and we run it repeatedly, because a single screenshot proves very little.
What we look at
Presence. For each buyer prompt, is the brand named? We group prompts into category questions, comparison questions and direct brand questions, since the answers differ a lot between them.
Position. First in the answer, or the fifth bullet? This changes the commercial value of a mention.
Framing. The attributes attached to the brand. A mention that says "affordable but limited" does a different job than "the standard for large teams." We read raw responses for this, because framing hides inside sentences.
Neighbors. Which competitors are named alongside, and which are named when the client is not.
Sources. The pages the model used. This is where fixes live. If an answer repeats an outdated pricing claim, the cited page is the thing to correct.
The tools we use for it
We run client programs on Promptwatch. For this audit it gives us what we need in one place.
Sentiment is scored per response on a 1 to 100 scale, with a weekly AI summary of how engines describe the brand. A Brand Book holds analysis instructions that decide what counts as a mention, and brand aliases catch name variants, which stops us undercounting a client who is known by two names. Position in Answer and Share of Voice give the numbers, and competitor heatmaps show neighbors across models, topics and countries.
For sources, citation analytics break references down by page, domain, Reddit, YouTube and LinkedIn, and show offsite mentions where the brand is named without a link. When an engine says something wrong, we usually find it traced to a third-party page rather than the client's own site.
We also look at whether models can read the client at all. Agent Analytics shows IP-verified crawler logs for bots like GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot. A brand that is described thinly may simply be one the crawlers could not fetch. Crawler logs start on Professional at $245 a month, and not on Essential.
A caveat we always give
Promptwatch monitors ChatGPT Search, Google AI Overviews, Google AI Mode, Copilot, Gemini, Perplexity, Alexa and Meta AI through the real interfaces. Claude, DeepSeek, Mistral and Grok are tracked via API. We say this up front, because an API response is not identical to what a person gets in the app, and clients should know which they are looking at.
Other options clients bring us
Some arrive with AthenaHQ, Scrunch AI or Profound already shortlisted. AthenaHQ rolls results into a GEO score and an action list, with Starter at $295 a month. Scrunch AI includes sentiment analysis and AI bot traffic reporting, with Core at $250 a month for four LLMs. Profound is an enterprise product, with an earlier estimate of about $40k a year. Clients who already use Ahrefs or HubSpot often start with the AI features in those products.
What we check for each is whether it lets us open the raw answers and trace them to sources. A tool that gives only a score makes the audit harder to defend.
Turning findings into work
An audit that ends in a slide is wasted. We turn findings into three kinds of task:
- Correct the source. If a third-party page is wrong or outdated, ask for a fix or publish a better one.
- Fill the gap. If the brand is absent where competitors appear, write the page the engine would cite. The Answer Gap report lists those prompts.
- Clarify the framing. If the brand is described by a limiting attribute, give engines a clear, honest page about who the product is for.
The Action Board collects these as a prioritized list, and its Discovery agent suggests up to five evidence-backed actions per week.
How often to rerun it
Weekly, on a fixed prompt set, with the same competitors. Compare periods rather than days. Models change answers constantly, so a movement of one prompt is noise and a movement across a topic is a signal.
Limits
No tool can control what a model says, and none can guarantee a recommendation. Sentiment scoring is a tool's judgment, so we always read a sample of raw answers. To try the audit yourself, Promptwatch has a free Explore tier for 10 prompts on ChatGPT, and paid plans from Essential at $95 a month.