A gauge with the seven sentiment levels, its needle on very positive, and an AI answer with a highlighted passage about AccuLLM scored 92.

How AccuLLM Scores Brand Sentiment in AI Answers

Emil Martin Vork Heunecke

November 16, 2026

From 23 November 2026, AccuLLM scores brand sentiment in AI answers with a new, more accurate model. This post goes into detail about how the sentiment is scored, with examples.

From 23 November 2026, AccuLLM scores sentiment with a new model for all accounts. The scores are more accurate across all seven sentiment levels.

Your sentiment history will show a step on that date. In general, expect fewer negative scores, more neutral ones, and all seven levels used more consistently. If you want the new model sooner, or your earlier answers scored with it too, contact support.

Sentiment is a measure of how favourably an AI answer speaks of your brand. In this blog post, we explain how we get from a prompt to a sentiment score and what the seven levels mean. We then go through examples, from the nuances of scoring a brand mention in a single sentence to mentions whose meaning depends on the headings and structure around them.

How a Prompt Becomes a Sentiment Score

From prompt to score. The prompt "Which tools track brand mentions in AI answers?" gets an AI answer. In the answer we find the text about your brand, which is the heading "Best tools for AI visibility" and the sentence "AccuLLM tracks brand mentions in ChatGPT, AI Overview, AI Mode and more…", and the text about each of your competitors. For your brand and for each competitor, that text is the passage, and it gets a score. Here your brand's passage scores positive, 76 of 100.
  1. We send each of your prompts to the AI engines you track: ChatGPT, AI Overview, AI Mode, and more.
  2. In each answer, we find your brand and each of your competitors by the spellings you have set up.
  3. Each brand we find gets its own passage: the text about that brand, along with the headings it sits under.
  4. We read each passage and give it a score from 1 to 100, which places it on one of seven levels.

Each brand gets one score per answer. Praise or criticism of another brand does not count toward yours, unless the answer compares the two.

The Seven Levels of Sentiment

The seven sentiment levels with score ranges and an example each. Very positive, 85 and up: "AccuLLM is the best choice on this list". Positive, 71–84: "AccuLLM is a good fit for agencies and enterprises". Slightly positive, 57–70: "Other tools worth a look include AccuLLM". Neutral, 43–56: "AccuLLM tracks brand mentions in ChatGPT, AI Overview, AI Mode and more…". Slightly negative, 29–42: "AccuLLM lacks certain features that competitors offer". Negative, 15–28: "AccuLLM is a disappointing tool". Very negative, 14 and below: "AccuLLM is the worst tool ever". A handwritten note beside the three negative examples says "lies!".

Positive, neutral, and negative alone would be too coarse. "Worth a look" and "the best choice" are both positive, but one is much stronger than the other. With seven levels, mild and strong praise get different labels, and so do mild and strong criticism.

Each level is a range on one scale, so two scores on either side of a border are still close. A 56 is neutral and a 57 is slightly positive, and the sentiment behind them is nearly the same.

Passages on Their Own

Some sentences can be judged from their own words. Even then, the wording can be tricky.

Negation

Some sentences use negative words to say something positive. Going by the words alone would get them wrong, so we score what the sentence means.

"You won't find AccuLLM slow, expensive or hard to use." scores positive, 76 of 100.

"Slow", "expensive", and "hard to use" are all criticisms. But "you won't find" negates them, and the sentence becomes praise.

Mixed sentiment

An answer often says something for a brand and something against it. When the two weigh about the same, they balance out.

"AccuLLM is easy to use." scores slightly positive, 69 of 100. "AccuLLM is overkill for a hobby project." scores slightly negative, 36 of 100. Together, "AccuLLM is easy to use, but overkill for a hobby project." scores neutral, 47 of 100.

One point is for AccuLLM and one is against it. They cancel each other out, and the score lands in the middle.

Close calls

When a passage supports two readings equally well, we choose the less negative one.

The passage "AccuLLM is rarely used for hobby projects." has two readings. "A caution to hobby users" would be slightly negative. "A note on the tool's main audience" scores neutral, 43 of 100, and is the one we use.

Criticism is uncommon: about 1% of brand mentions in our data read as negative. A negative score should therefore point you to something an answer really said. A wrong one costs you more than a missed mild criticism, because you would spend time looking for criticism that was never really there.

Facts and figures

AI answers often state a fact about a brand, such as a price, a feature, or a figure. A fact can sound good or bad to you, but we score only what the answer says about it.

"AccuLLM lets you track up to 50 competitors per brand." scores neutral, 52 of 100. With ", which is plenty for most markets." added, it scores slightly positive, 64 of 100.

The first answer does not say whether 50 is many or few, so we score it neutral. The second calls the figure plenty, so we score it slightly positive.

Comparisons

AI answers often compare your brand with others. Being named beside other brands does not count for or against you. What counts is what the answer says about your brand in relation to the others.

"AccuLLM is easy to use." scores slightly positive, 69 of 100. With "But so is every other tool on this list." added, it scores neutral, 49 of 100. With "None of the other tools on this list are." added instead, it scores positive, 82 of 100.

The praise is the same in all three. When every other tool shares it, it no longer sets AccuLLM apart, and the score is neutral. When no other tool does, it counts for more. The second sentence never names AccuLLM, but it belongs to the same paragraph, so we read it as part of the passage.

"AccuLLM is easier to use, but another tool does more." scores neutral, 53 of 100. "Another tool does more, but AccuLLM is easier to use." scores slightly positive, 64 of 100.

Both sentences say the same two things. The weight falls on what comes after "but", and the score follows.

Passages in Context

A passage is the text in an answer that is about your brand. We find your brand by the spellings you have set up, matched strictly, so words that only resemble it are not picked up. Around each mention we include the sentences around it that are about your brand, the points listed under it on a list, the headings above it, and, in a table, the row or column that holds the headers.

Headings and paragraphs

Often when a sentence mentions your brand, the rest of the paragraph is relevant too. So we read more than that one sentence: the rest of the paragraph that is about your brand, and the heading above it.

The answer has the heading "Best tools for AI visibility", a line of other text, and the paragraph "AccuLLM tracks brand mentions in ChatGPT, AI Overview, AI Mode and more. It is easy to use." Reading only the sentence that names your brand scores neutral, 50 of 100. Reading the whole paragraph scores slightly positive, 69 of 100. Reading the paragraph with its heading, as we do, scores positive, 77 of 100.

On its own, the sentence that names AccuLLM is a plain fact. The rest of the paragraph adds mild praise. The heading makes the whole paragraph a recommendation.

Lists and tables

When your brand is on a list, we include the list's heading. If the heading does not say the list is ranked, even first place is neutral.

A bullet list headed "AI visibility tools" with AccuLLM as the first of five bullets scores neutral, 50 of 100.

When the list is ranked, first place is positive and last place is neutral.

A numbered list headed "Ranked AI visibility tools". AccuLLM at number 1 of 5 scores positive, 73 of 100. AccuLLM at number 5 scores neutral, 52 of 100.

If the heading also carries praise, that praise counts too, and even last place is positive.

A list often names your brand on one line and describes it in the points below. We read those points too.

A list of tools, each with points under its name. Under AccuLLM: "Tracks ChatGPT, AI Overview, AI Mode and more", "Clear view of brand and competitor mentions", "Detailed sources and citations", "Accurate sentiment scores", "Very flexible filters", "Hierarchical static and dynamic tags" and "Unlimited API and MCP access". The item with the points under it scores positive, 73 of 100.

The line that names AccuLLM says nothing on its own. What the answer thinks of it is in the points beneath.

In a table, the header row does the same job as a heading.

A table with the columns Tool, AI engines and Weak points. AccuLLM's row says "ChatGPT, AI Overview, AI Mode and more" and "None found". Read alone, the row scores neutral, 47 of 100. Read with the header row, it scores positive, 75 of 100.

Read alone, "None found" says nothing about AccuLLM. Under the header "Weak points", it is praise.

Reading twice

Sometimes even that context is not enough. Without the whole answer, a passage can seem to say the opposite of what was meant.

Top row: the passage "What to avoid. AccuLLM does all three." read alone, with the lines between shown as grey placeholders, scores negative, 20 of 100. A negative first reading is read again with the whole answer. Bottom row: the whole answer. Under "What to avoid" it says "Avoid tracking prompts that are outdated, irrelevant or too narrow. It helps to use a tool that: researches your brand to get up-to-date information; identifies the relevant categories; suggests many prompts for each category and search intent. AccuLLM does all three." The second reading scores positive, 75 of 100.

Read alone, the passage seems to say that AccuLLM does three things you should avoid. In the whole answer, "all three" turns out to mean the three things a good tool does, which happens to describe AccuLLM's prompt suggestions.

So when the first reading is negative, we read the passage again together with the whole answer. A passage is then scored negative only when it reads as negative both alone and in the whole answer.

You may wonder why we do not always read the whole answer. In our tests, that gave no better scores overall and more than four times as many wrong negatives: praise for another brand, or criticism of something else in the answer, pulled your own brand's score down (even though the answer made no comparison). The focused passage is also what we highlight in the app, so you can see the exact part of the whole answer your sentiment score is based on.

Accuracy

We measure accuracy by testing the scores against a large set of carefully labelled passages from real AI answers. We labelled them with state-of-the-art language models over many rounds. Each round was reviewed and corrected by hand before the next, with close attention to the nuances described in this post. Any measure of accuracy is only as good as the labels behind it, so we put a lot of time into getting them right and into making the set large enough.

99% of scores are within one level of the label, up from 80% with the previous model. About four in five are on exactly the same level. The scores that are one level off are mostly among neutral, slightly positive, and positive. Most of them are close calls, where the score could have gone either way. Almost none are two levels off, and none are more than two.

A chart of how far scores are from their labels. About 80% are on the same level as the label, about 8% are one level more negative and about 12% one level more positive, so 99% are within one level. Under 1% are two levels off in either direction.

Grouped into a negative side, neutral, and a positive side, the score and the label agree about nine times in ten. When they differ, it is usually over whether a plain fact should count as praise or criticism, or when there is mixed sentiment. None cross between the negative and the positive side.

Three charts of how often the label agrees with the score. When the score is on the negative side, the label is too 91% of the time and neutral 9%. When the score is neutral, the label is neutral 88% of the time, on the negative side 2% and on the positive side 10%. When the score is on the positive side, the label is too 90% of the time and neutral 10%.

The scores are not fully deterministic, so there is slight noise, usually a point or two. That is only enough to change which of the seven levels a passage lands on where it was already a close call.

Across languages

About half the test passages are in English. The rest are in other European languages, such as German, Spanish, and Danish, and the scores are as accurate in each of them.

A bar chart of how close scores are to the label, for English, German, Dutch, French, Italian, Spanish, Portuguese, Danish, Norwegian, Swedish, Finnish and Polish. Between 73% and 84% are on the same level, and between 16% and 26% are one level off. In every language, 2% or fewer are two levels off.

We also tested languages with non-Latin scripts, such as Chinese, Hindi, Arabic, and Russian, using passages translated from English. About four out of five translations score on the same level as the original, and nearly all of the rest are one level off. None cross between the negative and the positive side.

A bar chart of how close translations score to the English original, for Chinese, Japanese, Korean, Thai, Hindi, Arabic, Hebrew and Russian. Between 73% and 86% are on the same level, and between 14% and 27% are one level off. In Arabic, Hebrew and Russian about 2% are two levels off.