AI brand sentiment analysis: tracking what assistants say about you

What AI brand sentiment analysis is

AI brand sentiment analysis is the practice of reading what AI assistants say about your brand and sorting it into favourable, unfavourable and neutral. It answers a different question from mention tracking. Mention tracking asks whether you were named. Sentiment asks how you were described.

Both matter. An assistant can name you and still steer a buyer away, by calling a rival the better choice for teams or calling your pricing hard to understand. A buyer who reads that sentence leaves with an opinion, and a mention count would score the answer as a win.

Think of it as the difference between being invited to the party and what people say about you once you are there. Mention tracking checks the guest list. Sentiment listens to the conversation.

A sentiment score tells you the mood. The sentence tells you what to fix. Always keep the sentence.

How AI brand sentiment tools score it

Most tools use a language model to read each answer and judge the tone around your brand name. They then report the result as a score, as counts of positive and negative mentions, or as both.

Peec AI is the most specific. Its documentation says the score reflects language from words like trusted, reliable and leading through to critical language or negative associations (Peec AI docs). Conductor's documentation describes tracking negative and positive brand mentions (Conductor docs). Semrush's knowledge base says its Brand Performance reports track sentiment and share of voice, and that this data is updated weekly (Semrush knowledge base).

Before you rely on any of these, ask the vendor four things. Which model makes the judgement? Are neutral sentences counted, and how? Can you see the sentence behind each verdict? And are questions that contain your name kept apart? A tool that can answer all four is giving you something you can check. A tool that cannot is giving you a mood ring.

Why a single sentiment score misleads

A single score averages praise and criticism across every topic into one number, so it can stay flat while something important changes underneath. It also inherits the limits of the model that produced it.

The model limit is documented. A study of large language models on sentiment tasks found they performed well on simpler tasks but lagged on more complex ones that need deeper understanding or structured sentiment information (Zhang and others, arXiv preprint). Telling whether an answer praised your support but criticised your pricing is that harder, structured kind of task. A single overall label flattens it.

The table shows the problem with made-up numbers. Two months give the same overall balance, but the reason a buyer would hesitate has moved from one topic to another.

Two more things distort a score. Most of what an answer says about a brand is neutral description, such as what it does or what it costs. If neutral sentences count toward the average, they pull every score toward the middle. And questions that name your brand get answers about you specifically, which can be more detailed and more favourable, so mix them in and the score drifts upward.

How to track brand sentiment in AI answers by hand

You can track sentiment yourself by collecting answers to a fixed set of questions and recording each claim about you with its sentence. It takes an hour or two a month for a modest question list.

1. Use the same buyer questions you track mentions with, so sentiment and mentions share one denominator.
2. Ask each assistant every question, and save each full answer with the date and the assistant.
3. In every answer that names you, find each sentence that makes a claim about you.
4. Skip plain facts, such as what you sell. Keep only sentences that praise or criticise.
5. File each kept sentence under a topic, such as pricing, support or ease of use, and mark it favourable or unfavourable.
6. Do the same for your two or three main rivals, so you can see where they are praised and you are not.
7. Count by topic each month, over the same answers, and keep the sentences beside the counts.

Keep questions that contain your name in a separate column. They answer a different question: what an assistant says when someone already knows you, rather than what a new buyer hears. Our guide to tracking brand mentions in AI search has a sheet layout you can extend with these columns.

Treat any change of one or two sentences with caution. Assistants rarely give the same answer twice: a SparkToro and Gumshoe study found less than a 1 in 100 chance that ChatGPT or Google's AI gave the same list of brands in any two responses (SparkToro). A sentence that appeared once may simply not appear next time.

Which questions surface sentiment

Sentiment shows up most in comparison and choice questions, where a buyer asks which product to pick. Simple what is questions mostly produce neutral description, which tells you little about how you are judged.

So weight your question list toward choice and comparison questions if sentiment is what you want to learn about. Keep them in the words a buyer would use. Our guide to finding buyer questions shows where to look for real wording.

Do not write questions that lead the assistant. Asking what is wrong with your brand will produce criticism because you asked for it, and that criticism tells you about the question rather than about what buyers hear.

When the assistant is simply wrong about you

Sometimes the problem is not tone but accuracy: the assistant states something about you that is not true. Sentiment analysis will often file that as negative, but the fix is different, so separate wrong claims from unfavourable opinions.

There is good public evidence that assistants get facts wrong often enough to check. In a study coordinated by the European Broadcasting Union and led by the BBC, journalists from 22 public service media organisations reviewed more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity. 45% of the answers had at least one significant issue, and 20% contained major accuracy issues (EBU). That study was about news, not brands, so it does not tell you how often assistants are wrong about companies. It does show that a confident answer is not proof of a correct one.

For a wrong or out of date claim, look at what the answer cited. If it links to a page, that page is often the source of the claim, and fixing or replacing it is the most direct lever you have. Our guide on why AI recommends your competitor covers the common causes.

What to do with what you find

Act on topics, not on the score. Pick the topic with the most unfavourable sentences, read them together, and decide whether each one is fair, wrong or out of date.

• If the criticism is fair, pass the sentences to the product team, because they are a buyer's view of a real weakness.
• If the claim is wrong, publish the correct fact clearly on a page an assistant can read.
• If the claim is out of date, update the page or listing the answer cited.
• If a rival is praised on a topic where you are silent, write the page that covers that topic properly.

Then keep asking the same questions. A fix shows up as fewer unfavourable sentences on that topic over several weeks, not as one good week. If nothing changes after a couple of months, read the cited pages again, because the assistant may be relying on a source you have not found yet.

Where Agentcites fits

Agentcites reads every answer for what it claims about you and the rivals it tracks, and files each claim under a topic as favourable or unfavourable. It keeps the exact sentence, and it throws a verdict away if the sentence it quotes cannot be found in the answer. Neutral description is not recorded. The brand perception page shows how it works.

It also compares the newest weekly scan with the one before, as two counts per assistant and topic beside the number of non-branded answers, with no arrow, because a difference of one answer is not a trend. Its limits: it only judges brands it already tracks, the judgement is made by a model and can be wrong, and it does not yet check claims against your own pages to say which are false. It asks ChatGPT, Perplexity, Gemini and DeepSeek, for $149 a month per website.