How can I monitor sentiment about my brand in ChatGPT? What we do for clients

Clients ask us about ChatGPT sentiment in one of two moods. Either someone saw a strange answer about the brand and panicked, or a board member wants "a number." Both are fair. Neither is answered by typing your brand name into ChatGPT a few times and screenshotting the result.

Here is how we handle it in practice.

Why one-off checks mislead

ChatGPT answers vary between runs, between countries, and with whatever web sources it fetches that day. If you ask once and see a bad answer, you have a sample of one. If you ask once and see a good answer, you have the same sample, just a more comfortable one.

So the first rule is that sentiment is a trend over a fixed panel of prompts, not a reading from a single conversation.

Building the panel

We start with the questions a buyer would really ask, not the ones the marketing team wishes they would ask. That usually means three groups:

Category prompts, such as "best accounting software for freelancers."
Comparison prompts, naming your brand and one competitor.
Direct brand prompts, such as "is [brand] worth it" or "[brand] complaints."

The direct brand prompts carry most of the sentiment signal. Category prompts tell you whether you appear at all. We tag prompts by type and topic so we can slice results later, and we use personas when a product serves distinct buyers, because the same question phrased by a founder and by an IT manager can return different answers.

What we measure

We run the panel daily. Promptwatch is the platform our client programs run on, and the parts we lean on for this job are specific:

Sentiment scoring. Every response that mentions the brand gets a score from 1 to 100, and there is a weekly AI summary of what moved.
Brand Book and aliases. The analysis instructions define what counts as a mention. Without this, a competitor with a similar name quietly contaminates the data.
Position in Answer and Share of Voice. A glowing mention in the last sentence of a long answer is worth less than a neutral mention up front.
A/B period comparison. We compare month against month instead of day against day.

Finding the cause of a bad answer

This is the part clients undervalue. A sentiment score tells you something is wrong. It does not tell you why.

When a score drops, we open the actual responses and then the citations behind them. Promptwatch's citation analytics list the pages and domains the model cited for each prompt, and there are dedicated Reddit, YouTube and LinkedIn citation reports. In our experience the cause is rarely your own site. It is more often an old comparison article, a forum thread, or a review page that frames you a certain way. Page Tracker lets us watch those specific URLs.

If the negative framing comes from your own pages, the next question is whether the crawlers can even read your updates. Agent Analytics shows IP-verified crawler logs for GPTBot, OAI-SearchBot and ChatGPT-User, with errors explained. Crawler logs start on the Professional plan, so a client on Essential cannot see this view.

What we do about it

The fix has three possible targets, and we pick based on the evidence:

Your own pages. Rewrite or refresh the page that should be answering the question. Content Refresh gives diff-style suggestions for existing pages.
Third-party pages. Outreach, corrections, or publishing better material that the model can cite instead. This is slow and unglamorous.
Access problems. Fix robots rules or blocked paths so the model actually reads the current version.

The Action Board keeps these as a prioritized list with owners, and a weekly Discovery agent suggests up to five evidence-backed actions per run. We still decide which to do.

What we tell clients it cannot do

Monitoring does not change what ChatGPT says. It tells you what it says now, how that is moving, and which sources are driving it. We never promise a particular sentiment score or a particular recommendation, and we are wary of anyone who does.

A single score also flattens detail. We always read a sample of raw responses each month, because "positive" can still mean "positive, but recommends a competitor for your best use case."

Getting started on a small budget

If you want to test this yourself, Promptwatch has a free Explore plan: one project, 10 prompts, ChatGPT only. That is enough for a brand-name panel and a first read on tone. Paid plans start at $95/mo for Essential, and the crawler logs you will want for diagnosis begin at Professional, $245/mo.

Start with 10 prompts, mostly direct brand questions. Read every response for two weeks. You will learn more from that than from any dashboard.