Reputation Resolutions
Reputation Resolutions
AI & Tech

Brand Sentiment in AI Search: How to Measure and Improve It

How AI answers describe your brand, how to score tone and accuracy with a repeatable prompt panel, and the source work that can improve it. Template included.

Anthony WillWritten & reviewed byAnthony Will, Founder & CEOReputation Resolutions · 13+ year industry veteranUpdated November 2026 · 13 min read

Key takeaways

  • Brand sentiment in AI search is the tone, framing and choice of facts in an answer. It is separate from accuracy: an answer can be accurate and still unfavorable.
  • Measure it with a fixed prompt panel, at least five runs per evaluative prompt per engine, and a written rubric. Answers vary from run to run, so one screenshot proves nothing.
  • Any single sentiment score is a construct someone defined. Tools define it differently, so report the components (tone, warning-led rate, error rate, themes, sources) with the number of runs.
  • No one can edit or guarantee an AI answer. The effective work is at the sources: correct errors, earn accurate coverage, respond to and legitimately earn reviews, and use platform channels for rule-breaking content.
In this guide

Brand sentiment in AI search is the tone, framing and choice of facts that an AI answer uses when it describes your brand. You can measure it honestly with a fixed prompt panel, repeated runs and a written scoring rubric, and you can improve it only indirectly, by correcting and earning better sources, because no one can edit an answer or guarantee what it will say.

Buyers now ask ChatGPT, Gemini, Perplexity, Claude, Copilot and Google's AI features whether a company is any good before they visit a site. An answer can be accurate and still hurt: it can open with a complaint, hedge with "some customers report," or leave out the praise. This guide covers that layer. It is the sentiment companion to our guides on monitoring your brand in AI search, AI search visibility metrics, fixing AI hallucinations and fixing what ChatGPT says about you.

We checked every source below on November 18, 2026. Several studies come from companies that sell tracking tools, and we say so. We also sell reputation services, so we say plainly what no service can do. Every number in the worked example is invented.

What brand sentiment in AI search means

In social listening, sentiment is a label on a post. An AI answer is a paragraph a model wrote from sources, and its sentiment sits in four places at once.

Tone. The words used for you: reliable, affordable, controversial, struggling.

Framing. What comes first and how it is hedged. An answer that opens with praise and ends with one caveat reads very differently from one that opens with "mixed reviews."

Selection. Which facts and complaints the answer picks. From five hundred reviews it may surface the three about billing.

Verdict. Whether it recommends you, lists you without comment, or cautions the reader.

Sentiment is not accuracy. An answer can be accurate and unfavorable, such as a real lawsuit from 2022, or inaccurate and favorable, such as an award you never won. A false claim may be correctable. An accurate unflattering fact cannot be removed, so it needs context and better evidence. Keep the two in separate columns.

Tools define sentiment differently, which is the first reason to distrust a single number. Semrush's feature page says it classifies brand mentions as positive or general. Similarweb's guide defines it as the share of answers that are positive, neutral or negative. Peec's guide scores branded prompts grouped into themes such as pricing and support, without explaining its scale. HubSpot's free AI Search Grader gives sentiment up to 40 points inside a 100-point composite. It publishes the weights but not how each answer is scored. Four tools, four constructs. Learn the definition before you trust the trend.

Why one answer, and one score, can mislead

Answers vary from run to run. A study by SparkToro and Gumshoe.ai, reported by Search Engine Journal on January 30, 2026, ran 12 brand-recommendation prompts 60 to 100 times per platform, 2,961 runs in all. ChatGPT and Google's AI in Search returned the same brand list in under 1% of repeated runs. Cautions: it is not peer-reviewed, Gumshoe sells AI tracking, and it measured which brands were listed, not tone. We know of no published study of run-to-run variation in sentiment, so read it as a reason for caution, not an error rate. One screenshot is an anecdote.

The starting line may not be level. An EMNLP 2024 paper by Kamruzzaman, Nguyen and Kim found that language models tend to link global brands with positive attributes. The models were from 2024 and current engines may differ, but it is a reason to code competitors in the same panel and read your tone against theirs.

Where negative framing in AI answers comes from

An answer reaches its tone by one of two paths. With web search on, the engine retrieves pages and summarizes them. With search off, it relies on training data and cites nothing. Here is what each provider says about its own sourcing, as of November 18, 2026.

Google AI features. Google's documentation says a page must be indexed and eligible to appear in Search with a snippet to be a supporting link, with no special optimization. It also says the features may run several related searches across subtopics, which can pull in pages your test query never showed.

ChatGPT search. OpenAI's help center describes inline citations and a Sources panel, says a site must allow OAI-SearchBot to be eligible, and notes that citations can be incomplete or outdated.

Perplexity. Its help center says it searches the internet in real time and includes numbered citations in each answer.

Claude. Anthropic's help center says Claude searches the live web when that helps, and that responses include citations.

Microsoft Copilot. Microsoft Learn says Copilot Chat can ground responses in public web data. It does not say how citations are shown, so test that yourself.

What the studies say about which sources show up for brand questions is mixed. Omniscient Digital's study, published January 8, 2026, ran 240 prompts that each named a brand across five engines, through Peec's platform, and classified 23,387 cited sources. Earned media was 48% (editorial 16%, forums and social 11%, review sites 11%, directories 10%), commercial brand content 30% and owned content 23%; the total is 101% from rounding. For customer-review prompts, earned media was 82%. Peec's analysis of 30 million cited sources, published March 31, 2026, ranked Reddit first, then YouTube, LinkedIn and Wikipedia, with ranks but no percentages. Semrush's September 2025 study reported Wikipedia and Reddit among the most-cited sources, without disclosing its sample size.

Then there is the disagreement. A Profound analysis of 11.8 billion data points, reported by WARC on September 1, 2026, found brand websites were 57% of citations (47% on ChatGPT, 69% on Gemini). That sits awkwardly beside the Omniscient split. Different prompts, engines and counting rules are the likely reasons, but that is our inference, because none of the reports publishes enough to reconcile them. For your brand, your own citation log is the only evidence that counts.

Use this map to turn your log into a plan.

  • Review platforms (Google, Trustpilot, Yelp, G2, Glassdoor)

    How it shapes tone
    Star average and the wording of detailed reviews. A repeated complaint becomes "some customers report"
    How to check it
    Count recent reviews that repeat the theme
    What you can do
    Reply, earn genuine new reviews, report rule-breaking ones
  • Reddit and forums

    How it shapes tone
    Threads read as candid opinion, and one can anchor an answer
    How to check it
    Check the citations for evaluative prompts
    What you can do
    Take part openly, answer with facts, report rule-breaking posts
  • News and old press

    How it shapes tone
    An old article can become the first fact repeated
    How to check it
    Check the dates on cited articles
    What you can do
    Ask for factual corrections or updates, earn newer accurate coverage
  • Complaint boards and consumer sites

    How it shapes tone
    Unverified complaints presented as typical
    How to check it
    Search your name with words like complaints, scam and reviews
    What you can do
    Respond there, use its dispute process, report rule-breaking content
  • Wikipedia and reference sites

    How it shapes tone
    A neutral-sounding summary, including controversies
    How to check it
    Read the article and its talk page
    What you can do
    Suggest sourced edits on the talk page and disclose any connection
  • Directories, listicles and comparison pages

    How it shapes tone
    Rankings and alternatives lists shape comparison prompts
    How to check it
    Check pages cited for best and alternatives prompts
    What you can do
    Correct factual errors with evidence
  • Your own pages

    How it shapes tone
    Thin or outdated pages leave the field to others
    How to check it
    Compare your site with the answer's claims
    What you can do
    Update pages and answer the recurring complaint plainly
  • Training data with no citation

    How it shapes tone
    An older snapshot, with nothing to point at
    How to check it
    Run the prompt with search off where possible
    What you can do
    Keep sources accurate and wait for a model refresh

How to measure sentiment: the panel, the runs and the rubric

Panel basics (fixed prompts, logged-out sessions, search on and off, dated records) are in our AI search visibility metrics guide. Here is what changes for sentiment.

1. Write evaluative prompts, not just category prompts. Visibility panels ask who the best providers are. A sentiment panel asks what the engine thinks of you. Use these five types, in a buyer's words.

Prompt typeExample patternWhat it reveals
Direct evaluationIs [brand] any good?Overall verdict and lead framing
ReputationWhat do customers say about [brand]?Which complaints and praise get selected
TrustIs [brand] legit? Is [brand] safe to use?Warnings and hedges
Comparison[Brand] vs [competitor]. Alternatives to [brand]Relative tone and who is favored
Category recommendationBest [service] in [city]Whether you are named at all

2. Run each evaluative prompt at least five times per engine. The SparkToro study used 60 to 100 runs, which is not practical by hand. Five is our suggested minimum, not a validated number. Record every run.

3. Add two or three competitors as controls. Run the same prompts about them, so your tone is read next to theirs.

4. Code each answer with a written rubric, frozen before the first run. Changing the rubric mid-year breaks your trend line.

  • Mention

    Codes
    0 not named. 1 named in a list. 2 main subject
    How to decide
    Does the answer name you, and how prominently
  • Tone

    Codes
    -2 to +2 (scale below)
    How to decide
    Read the whole answer, not one adjective
  • Lead framing

    Codes
    Praise, neutral, hedged, warning
    How to decide
    Classify the first sentence only
  • Accuracy

    Codes
    Accurate, inaccurate, outdated, unverifiable
    How to decide
    Check each claim against your records and the cited page
  • Verdict

    Codes
    Recommend, list only, caution
    How to decide
    Does it tell the reader what to do
  • Complaint themes

    Codes
    Free tags such as billing, delays, service
    How to decide
    Tag each theme once per run
  • Sources cited

    Codes
    Domain and source type
    How to decide
    Log every link shown

Tone scale. +2: clear recommendation with specific praise. +1: favorable overall, possibly with a minor caveat. 0: neutral description, or evenly balanced. -1: unfavorable overall, or a notable caution weighted above the praise. -2: leads with a warning or complaint, or advises against.

5. Have a second person code a sample. Give another reader 20% of the answers, without your codes, and compare. Frequent disagreement means the rubric is vague. If an AI model does the coding, spot-check its labels the same way.

6. Trace every complaint to a source. For each theme, note which cited page carries it. That turns a score into a to-do list.

A worked example with invented numbers

Alder & Pine Movers is a fictional regional moving company. The table is illustrative: made-up results for two unnamed engines, five runs per prompt, search on. It is not data about any real company or engine.

PromptEngineMean tone (5 runs)Warning-led runsRuns with an errorSource types cited
Is Alder & Pine Movers good?A+0.41 of 51 of 5Review site, own site
Is Alder & Pine Movers good?B-0.83 of 50 of 5Complaint board, forum thread
What do customers say about Alder & Pine Movers?A0.01 of 51 of 5Review site, news article
What do customers say about Alder & Pine Movers?B-1.24 of 51 of 5Complaint board, forum thread
Is Alder & Pine Movers legit?A+0.60 of 50 of 5Own site, directory
Is Alder & Pine Movers legit?B-0.42 of 51 of 5Forum thread, directory

The arithmetic uses definitions we chose. Mean tone is the five run scores added and divided by five. Warning-led rate is runs whose first sentence is a warning, divided by runs. Error rate is runs with any inaccurate or outdated claim, divided by runs. Across the three prompts, Engine A averages +0.3 tone with 2 of 15 runs warning-led (13%). Engine B averages -0.8 with 9 of 15 runs warning-led (60%). A competitor control, also invented, scores +0.6 on Engine A and +0.2 on Engine B.

What it shows: the problem is concentrated on Engine B, and it traces to two cited sources, an old complaint-board thread about a damage claim and a forum thread. The errors trace to an outdated directory listing. So the work is specific: reply factually on the complaint board, correct the listing, publish a plain page on how damage claims are handled, and ask recent customers, all of them, for reviews. Then re-run the panel. Whether Engine B changes is not in the company's control.

With five runs per prompt, a move from -0.8 to -0.4 could be noise. Look for movement across several prompts and rounds, and report components with the number of runs. If someone demands one sentiment score, say how you defined it and show the parts beside it.

If one engine is the problem, our pages on ChatGPT reputation management, Gemini reputation management and Google AI Overview reputation cover how each reads its sources.

What you can and cannot change

  • The exact wording of an answer

    Can you change it?
    No
    Realistic route
    We know of no control that edits an answer for everyone
  • Which pages an engine cites

    Can you change it?
    Not directly
    Realistic route
    Make accurate pages indexable and useful. Google says there is no special optimization
  • Accurate but unflattering coverage

    Can you change it?
    Not by removal
    Realistic route
    Add context, earn newer accurate coverage
  • Wrong facts on your own pages and profiles

    Can you change it?
    Yes
    Realistic route
    Fix them and date them
  • Wrong facts on someone else's page

    Can you change it?
    Sometimes
    Realistic route
    Ask the publisher, with evidence. The publisher decides
  • Reviews or posts that break a platform's rules

    Can you change it?
    Sometimes
    Realistic route
    Report through the platform. The platform decides
  • Personal data about a named person in an AI answer

    Can you change it?
    Sometimes
    Realistic route
    OpenAI's removal request, Gemini's report option; reviewed case by case
  • Claims that exist only in training data

    Can you change it?
    Slowly, if at all
    Realistic route
    Only when a newer model is trained, on the provider's schedule

Vendor guides quote timelines from weeks to months. The ones we read cited no verifiable source for their figures, so we give none. Search-connected answers can reflect new pages sooner than training-only answers, but no provider publishes a schedule.

Source-level work you can do

1. Correct what is wrong at the source. Start with pages you control: your site, Business Profile and directory listings. Then write to third-party publishers with the URL, the error and your evidence. Some will correct it. For Wikipedia, the conflict-of-interest guideline discourages editing about yourself or your employer, points you to talk-page edit requests, and requires paid editors to disclose. Our guide to Wikipedia and AI visibility goes further.

2. Earn accurate independent coverage. Engines often cite sources other than your own site, so a newer accurate article gives them something else to read. It is slow and may not be cited. See how to get cited by AI. Done properly, LLM seeding means publishing accurate material where models read.

3. Respond to reviews and earn new ones within the rules. Google's contributed content policy bars incentives for reviews and says businesses may not selectively solicit positive reviews. It allows asking customers for genuine reviews. So ask every customer, reply to criticism specifically and politely, and keep records. See how to respond to negative reviews, our review management service, and, for threads, Reddit for AI search and what Reddit Reviews means for brands.

4. Clarify your own pages around the complaint themes. If delays or refunds are the recurring theme, publish a plain page stating your process, timelines and policy, with dates. Google says no special schema is needed; the page must be indexable and useful. For ChatGPT search, OpenAI says the site must allow OAI-SearchBot.

5. Remove what breaks a platform's rules, through that platform. Google's policy lists fake and misleading content, harassment, personal information, off-topic content and advertising or solicitation. Google's Search removal routes cover narrow categories such as government ID numbers, bank details and doxxing, and Google generally will not remove accurate news. Honest criticism usually stays. See how to remove a negative Google review and our negative review removal page. If a falsehood is repeated on many pages, see AI-generated defamation.

6. Report wrong personal information to the engines. OpenAI's help center says that under some privacy laws, including the GDPR, you can ask for personal data to stop appearing in ChatGPT responses, with or without an account, and that review weighs accuracy, public role, the age of the information and public interest. In the Gemini app, Google's help says to select More, then Report legal issue. These routes are per engine, not guaranteed, and do not remove anything from the source website.

A 90-day plan

  • 1 to 14

    Work
    Write the rubric and 25 to 40 prompts across the five types. Choose engines and competitors. Build the sheet
    Output
    Frozen rubric and prompt list
  • 15 to 30

    Work
    Run a baseline twice, two weeks apart, five runs per evaluative prompt. Code, with a second reader on 20%. Build the source map
    Output
    Baseline with ranges, a theme list, a source map
  • 31 to 60

    Work
    Fix pages you control. Send evidence-based corrections to publishers. Start fair review requests to all customers. Publish pages for the top two or three themes. File rule-violation reports
    Output
    A dated change log
  • 61 to 75

    Work
    Re-run the same prompts on the same engines. Compare components, not one composite
    Output
    Round three results
  • 76 to 90

    Work
    Decide what to continue, change or stop. Set a monthly cadence
    Output
    A short decision memo

Expect that some engines will not move, and write that down.

Tactics we would not use on AI sentiment

Fake or planted reviews and sock-puppet posts. The FTC's rule on consumer reviews and testimonials took effect October 21, 2024. Per the FTC's guidance, it covers fake reviews, incentives conditioned on a particular sentiment, undisclosed insider reviews, and suppressing reviews through intimidation or unfounded legal threats. Google's policy calls paid reviews and multiple accounts fake engagement.

Hidden text or hidden instructions aimed at AI systems. Google's spam policies name hidden text, such as white text on a white background or zero font size, as abuse. Hiding instructions for assistants in a page is the same idea. We found no provider document that endorses it.

Undisclosed participation on Reddit or Wikipedia. Wikipedia requires paid editors to disclose, and Reddit communities enforce their own rules. Take part openly.

Mass-produced pages to drown out criticism. Google calls generating many pages mainly to manipulate rankings scaled content abuse.

Chasing removal of accurate criticism, or pressuring reviewers. It rarely works and can create a worse story.

Promising a result. No one controls what an answer says. A provider that guarantees a change is promising something it does not own.

When to hand it to a service

You can run this yourself when you have a handful of prompts, a few sources and control of the pages involved. Consider help when false or defamatory content repeats across many pages, when rule-breaking reviews or posts need filings on several platforms, when personal data is involved, or when you lack the hours. Our comparison of AI reputation management companies lists questions to ask. Ask any provider for a dated before-and-after panel with the number of runs, which sources they will work on, and what happens if a removal fails.

Reputation Resolutions has done source-level reputation work since 2013. We are A+ BBB accredited and a Google Partner, with clients in 40+ countries. Our AI reputation management work starts from a free AI visibility audit that records what each engine says, with the sources it cites. Removal of qualifying content is results-based, billed only after it is confirmed. We do not promise that an AI answer will change, and we will tell you when a complaint is accurate and will stay. For ongoing sampling, see AI reputation monitoring.

Frequently asked questions

What is brand sentiment in AI search?+

Brand sentiment in AI search is the tone, framing and selection of facts an AI answer uses when it describes your brand, including whether it recommends you, lists you neutrally or cautions the reader. It is separate from accuracy, since an answer can be accurate and unfavorable. Measure it by running a fixed set of prompts repeatedly and coding each answer against a written rubric.

How do AI search engines determine brand sentiment?+

No provider we checked publishes a separate sentiment component. When search is on, the engine retrieves pages and summarizes them, so the tone of the answer follows the tone of the pages it chose. When search is off, the answer comes from training data. The provider pages we read describe sourcing, but none says how one source is weighed against another.

How do I improve brand sentiment in AI generative search engines?+

Work on the sources, not the chatbot. Correct inaccurate pages, earn accurate independent coverage, reply to and legitimately earn reviews, publish clear pages about the complaints that keep appearing, and use platform channels for content that breaks the rules. Then re-run the same prompt panel to see what moved. No one can guarantee that an answer will change.

Can I delete or edit a negative AI answer about my brand?+

Not directly. We know of no control that edits an answer for everyone. You can report personal-data problems through routes such as OpenAI's removal request or Gemini's report option, which are reviewed case by case. For the rest, you change the sources the engine reads, and the answer may follow as engines re-read the web.

How do I track brand sentiment in AI answers, and do I need a tool?+

You can start with a spreadsheet, a clean browser session and the rubric above. Tools help when the panel grows beyond what you will finish by hand. Whatever you use, ask how it defines sentiment, how many runs sit behind each number, and whether you can read the underlying answers.

How many times should I run each prompt?+

There is no validated number. A SparkToro and Gumshoe study ran prompts 60 to 100 times and found lists almost never repeated, which is not practical by hand. We suggest at least five runs per evaluative prompt per engine as a minimum, and we say that is a practical choice, not a proven threshold. Look for patterns across prompts and rounds.

What does HubSpot's AI Search Grader measure?+

HubSpot's page describes a free, one-time check of how ChatGPT, Perplexity and Gemini represent a brand, scored out of 100 across five dimensions, with sentiment worth up to 40 points. The weights are published but the scoring method is not, and the page calls the result a snapshot, not continuous tracking. Use it as a first look, then run your own repeated panel.

Why does AI say negative things about my brand when most of my reviews are positive?+

Selection. An answer may surface the most detailed or most repeated complaints, or lean on a forum thread or complaint board that ranks well. Check the cited sources. If you cannot see citations, the claim may come from training data, which you cannot inspect. A competitor control also helps, because one 2024 study found models favored global brands over local ones.

Can removing negative reviews fix my AI sentiment?+

Only when the reviews break a platform's rules, such as fake engagement or personal information, and the platform agrees. Honest criticism generally stays, and removing a few reviews may not change the answer if the same complaint appears elsewhere. Replying well and earning genuine new reviews is the dependable route.

How long does it take for AI sentiment to improve?+

No provider publishes a schedule, and the vendor timelines we found cited no verifiable source. Answers that use live search can reflect new pages sooner than answers drawn from training data. Re-run your panel monthly, and judge by trends in the components over several rounds.

Prefer we handle it?

See what each AI engine says about your brand today

Our free AI visibility audit records what ChatGPT, Gemini, Perplexity and Google's AI features say, with the sources they cite, so you can see which complaints come from which pages. We have done reputation work since 2013, and we do not promise that an answer will change.

Explore AI reputation management

Ready to See What AI Finds About Your Reputation?

Schedule your free, confidential consultation and we'll show you exactly where you stand across search engines, review sites, and AI platforms.

$0
Upfront Cost
Within 30 Days
Typical Removal