For the complete documentation index, see llms.txt. Every page on this site is also served as Markdown: append `.md` to any URL, or send `Accept: text/markdown`.
Skip to content
Attensira LogoAttensira
Book a demo

Blog · Analytics & Measurement

Brand sentiment tracking in AI answers: How to measure it honestly

Brand sentiment tracking in AI answers measures whether the answers that judge your brand judge it well. Read each answer for its stance on you, positive, neutral or negative, and leave out answers that name you without judging. Score only the answers with a stance, show the count beside the score, and keep the sentence that carries each stance.

Start freeBook a demo

7 days free · card required

Three answers about Attensira, scored by hand

  • Perplexity, “what is Attensira”

    Stance
    Positive
    What carries it
    “an agent that ships the fix”, “not just a dashboard”

  • ChatGPT, “Attensira vs Profound”

    Stance
    Neutral, with a reservation
    What carries it
    “I'd treat its comparative claims as vendor claims”

  • ChatGPT, “Attensira vs Profound”

    Stance
    No stance: not scored
    What carries it
    Our starting price in a table row
Analyst agentAnalyst · reads this workspace's answersattensira.com workspace, read over the Attensira MCP · · 3 stored answers
Karl-Gustav KallasmaaKarl-Gustav Kallasmaa, Founder & CEO

Last updated

Key facts

Social listening taught marketers to track sentiment: count the positive and negative mentions, watch the ratio. Applied to AI answers, that method produces a number that looks precise and means very little. The reason is simple. Most AI answers that name a brand do not have an opinion about it.

This post explains how to track sentiment in AI answers so the number is about opinions, shows three real answers about us scored by hand, and covers what to do with the result.

What is different about sentiment in AI answers?

A social post about a brand is usually written because someone has a view. An AI answer names brands because the question asked for names. "What does Acme cost?" gets "Acme starts at 49 dollars a seat": your brand, no opinion. "Best tools for X?" gets a list of five names, often with no judgement on any.

So an AI answer can relate to your brand in three ways:

  • Takes a position

    Example
    "Acme is the better fit for small teams"
    What it is
    A stance: positive, neutral or negative

  • Names without judging

    Example
    "Acme starts at 49 dollars a seat"
    What it is
    No stance

  • Does not name you

    Example
    Anything else
    What it is
    Not in the sentiment question at all

A neutral stance is a real opinion: "Acme is one of several reasonable options" is a balanced verdict. No stance is the absence of one. They look alike and mean opposite things.

Why does counting "no stance" as neutral ruin the number?

Suppose you track 50 questions. Twenty are pricing and feature questions, whose answers state facts. Thirty are "which is best" questions, whose answers judge.

If you count fact answers as neutral, your score is a blend of thirty real opinions and twenty non-opinions. Now add ten more pricing questions to your tracking. Your score moves towards neutral. Nothing about how engines view you changed; you changed the mix.

A number that moves when you edit your settings is not a measurement. The fix is to leave no-stance answers out of the score entirely, and divide only by answers that took a position. That is how our own sentiment measure works: the shares and the score divide by scored answers, never by all answers read.

What do real answers about us look like, scored?

Here are three answers about Attensira, all read on 2026-10-03 in our own workspace, scored by the rule above.

Positive stance. Perplexity, "what is Attensira": it described us as "an agent that ships the fix", "not just a dashboard". That is a judgement in our favour, and the sentence is lifted from our own About page.

Neutral stance, with a reservation. ChatGPT, "Attensira vs Profound": after a fair side-by-side, it added that "the most detailed comparison I found is published by Attensira itself, so I'd treat its comparative claims as vendor claims rather than independent validation." That is a balanced verdict with a caveat about our evidence. It is fair, and it tells us exactly what to fix: independent sources.

No stance. Another ChatGPT answer to the same question put our starting price in a table row. It names us and judges nothing. It belongs in share of voice and in claim checking, not in sentiment.

Three answers, three different treatments. A tool that counted all three as "mentions with neutral or better sentiment" would tell us nothing we could act on.

How do you calculate an AI sentiment score?

  1. Collect answers that name you, for the questions you track, on each engine.
  2. Read each for its stance on you: positive, neutral, negative, or none. Keep the sentence that carries it.
  3. Drop the no-stance answers from the calculation, but keep counting them.
  4. Compute the shares: positive, neutral and negative, each divided by answers with a stance. They sum to the whole of the scored answers.
  5. Compute a score, if you want one number: the average stance on a scale from -1 to +1. In our own product, above +0.15 reads as mostly positive and below -0.15 as mostly negative; zero means opinions cancelled out.
  6. Report three counts beside it: answers scored, answers read, answers not scored yet.

Read the scored count before the score. Five scored answers out of three hundred read is five opinions.

What does an answer that never names you look like?

Step 1 collects answers that name you, and most do not. ChatGPT on 4 October 2026, asked to “check if ChatGPT mentions my brand”: no brand named, no stance, nothing to score. It belongs in share of voice, not in sentiment.

Answer · ChatGPTasked

Question: check if ChatGPT mentions my brand

ChatGPT answered: I can help check that, but I need your brand name (and ideally your website URL or product category). … I can then help assess your visibility. Note that ChatGPT responses can vary by prompt, model, account, and whether web search is enabled, so a proper check usually tests multiple customer-style questions rather than one query. Tools that monitor this type of AI visibility typically track brand mentions, recommendation position, citations, and competitor appearances across AI answers.

Sources cited

  1. 1ahrefs.com
  2. 2siftly.ai

Not namedAttensira is missing from this ChatGPT answer.

Analyst agentAnalyst · reads this workspace's answersattensira.com workspace, read over the Attensira MCP · · 1 stored answer
How we read these numbers

Should you score by engine and by question type?

Yes, separately. Engines lean on different sources, so a brand can be described warmly on one and with reservations on another. Question types differ too: comparison questions produce more stances than definition questions. Each cut should divide by its own scored count.

What you should not do is compare a sentiment score between two tools. Unless both use the same stance rules and the same denominator, the numbers answer different questions.

What do you do with negative sentiment?

Read the sentence first, then the page it cites. Negative stances come from three places:

  • A wrong fact. "Acme does not support SSO" when it does. That is a misrepresentation; fix the source page. See fixing what AI gets wrong about your brand.
  • A fair criticism. "Acme is expensive for small teams." If true, it is product feedback. You can answer it on your own pages, for example with a clear plan for small teams, but you cannot argue it away.
  • An old criticism. A complaint about something fixed long ago, repeated from an old review. Update the record: a changelog entry, a dated note on the review site if you can respond there, current pages that state the fix.

The reservation in our own example is the second kind. ChatGPT is right that our comparison page is vendor-written. The fix is to give it independent evidence to cite: listings, reviews and third-party pages that describe us accurately.

How is sentiment different from share of voice and claims?

They answer different questions, and mixing them causes most of the confusion in this area.

  • Share of voice

    Question
    How often are we named?
    Unit
    Answers naming you ÷ answers read

  • Claims

    Question
    Is what they say about us true?
    Unit
    Each factual statement, checked against facts you confirmed

  • Sentiment

    Question
    When they judge us, is it good?
    Unit
    Answers with a stance, scored

An answer can name you, state your price wrongly and recommend you warmly, all at once. That is three readings, not one. For the first measure, see how to calculate share of voice.

What should a sentiment report look like?

One line per engine, with the counts that let a reader judge it:

  • Engine A

    Answers read
    Count
    Scored
    Count
    Positive
    Share of scored
    Neutral
    Share of scored
    Negative
    Share of scored
    Score
    -1 to +1

  • Engine B

    Answers read
    Count
    Scored
    Count
    Positive
    Share of scored
    Neutral
    Share of scored
    Negative
    Share of scored
    Score
    -1 to +1

Under the table, quote the most consequential sentences: the strongest negative stance and the most common reservation, each with the engine, the date and the page the answer cited. Those quotes are what people act on. The score tells them whether to look; the sentences tell them what to do.

Leave out a total row that blends engines unless every engine has a similar scored count. A blend dominated by the engine with the most answers hides exactly the differences the report should show.

Does one engine name us more than the other?

The same 61 questions, read in ChatGPT and in Perplexity, counted apart. The two name attensira.com at about the same rate; a gap this small is inside the noise, so it is not a finding.

of ChatGPT answers name attensira.com
4.2%
n=2,346 answers
of Perplexity answers name attensira.com
4.5%
n=2,172 answers
link to it, ChatGPT / Perplexity
6.6% / 3.9%
n=4,518 answers
Analyst agentAnalystattensira.com workspace, Attensira MCP, ChatGPT and Perplexity · n=4,518 answers · to

Which questions produce the most useful sentiment?

Comparison and "which is best for" questions. Those are the answers where an engine has to weigh brands against each other, so they carry most of the stances, and they are the answers buyers read just before a decision. Brand questions such as "what is [brand]" mostly produce descriptions with no stance; pricing and feature questions almost never produce one. If your sentiment score rests on a handful of scored answers, the cure is usually more comparison questions in the tracked set, not more days of data.

How many answers to a generic question can be scored at all?

For the question above, none about us. Over 30 days we were named in 0 of 66 answers, so there was no stance to read; Ahrefs was named most. That is why the scored count sits beside every score.

answers to “check if ChatGPT mentions my brand” naming Attensira
0 of 66
n=66 answers
named Ahrefs, the most-named brand
21 of 66
n=66 answers
answers with a stance on Attensira to score
0
n=66 answers
Analyst agentAnalyst · reads this workspace's answersChatGPT and Perplexity answers, attensira.com workspace, Attensira MCP · n=66 answers, 24 of 31 days read · to

How often should you look at sentiment?

Monthly is usually enough to act on, because stances change slowly. Daily data still matters, because it builds the count you need for a trustworthy number; but a day-to-day sentiment line moves mostly on sampling noise. Look at changes only when they clear a significance test for the number of scored answers involved.

What can go wrong with automated stance scoring?

Stance is read from the text by a model, and it can be wrong. Sarcasm, hedged recommendations and long comparisons are the usual trouble. Two safeguards keep it honest: show the sentence beside every score so a person can check it, and let people flag a score they disagree with without it silently changing the number. Our own product does both: every scored answer is listed with its sentence, and a flag is an annotation, never an edit.

How does Attensira track sentiment?

Attensira scores stance the way this post does, on every answer about you from the last 90 days, across ChatGPT, Perplexity, Claude and Google AI, with Gemini from Growth. The sentiment page shows the score and the positive, neutral and negative shares with the scored count beside them, by engine and by topic, with each scored answer and its sentence. Answers that name you without judging are left out of the score and counted separately. When a negative stance traces to a wrong fact, the crew drafts the fix to the source page for your approval.

That is the service of an SEO agency at the price of a tool: reading what the engines think of you, and doing the work when they are wrong. To see how the engines describe you today, run the free audit.

What else do people ask?

What is brand sentiment in AI answers?
It is the position an AI answer takes on your brand when it takes one, such as praise, a recommendation, a reservation or a criticism. Many answers that name a brand take no position at all, for example when they only quote a price, and those should be left out of the score rather than counted as neutral.
Why not count answers with no opinion as neutral?
Because then the score moves whenever the mix of questions moves. Track more pricing questions, which produce answers that state facts without judging, and your score drifts towards neutral although no opinion about you changed. Leaving no-stance answers out keeps the score about opinions only.
How many scored answers do I need before the sentiment score means anything?
Enough that one answer does not swing it. Five scored answers are five opinions, not a verdict. Always read the scored count next to the score, and treat a change as real only when it is larger than the noise for the number of answers involved.
Can I improve negative sentiment in AI answers?
Sometimes. Read the sentence that carries the negative stance and the page it cites. If it repeats a wrong fact, fix the source. If it repeats a fair criticism, the fix is in the product or in answering the criticism openly on your own pages. Negative sentiment that is accurate is information, not a bug.

Where do these facts come from?

2 sources, 4 statements: each statement with the page it came from and the day it was checked.

  1. Attensira leaves answers that name a brand without judging it out of the sentiment score instead of counting them as neutral.

    docs.attensira.com · retrieved

  2. Attensira's sentiment score is the average stance of scored answers from -1 to +1, and a score of null means no answer in the window took a position.

    docs.attensira.com · retrieved

  3. attensira.com workspace, read over the Attensira MCP on 2026-10-03: a Perplexity answer to "what is Attensira" took a positive stance on Attensira, a ChatGPT answer to "Attensira vs Profound" a neutral stance with a reservation that the comparison was vendor-published, and another ChatGPT answer to the same question named Attensira's starting price without judging it.

    attensira.com workspace, read over the Attensira MCP on

  4. Attensira reads stance for answers from the last 90 days, and each scored answer is listed with the sentence that carries the stance so it can be flagged.

    docs.attensira.com · retrieved

Want software that does your AI-search visibility work, with you in control of every change?

Start freeBook a demo

7 days free · card required