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

Most teams start asking about AI sentiment a month into monitoring. The brand shows up, share of voice is climbing, yet nothing suggests the models are selling it. The right question now is not how often the brand is named but what gets said about it, and who the AI recommends in the end.

Visibility tells you whether you showed up, not what the AI gave the buyer a reason to choose you for. That is why we shipped Brand perception in GEO Scout on 12 September 2026. Conflict of interest up front: I founded GEO Scout, so where I describe our tool I name its limitations too.

This is a practical guide. Our study of which brands AI praises and criticises by niche is in Sentiment in AI: which brands neural networks praise and criticize. Fixing wrong facts is covered in ChatGPT writes wrong information about my company. Here the focus is measuring sentiment and perception and acting on them.

Sentiment, perception, reputation: what each one measures

Sentiment is the attitude of an answer toward the brand: positive, neutral or negative. One label per answer.

Perception is what that label is made of: the words used to describe the brand, the role it gets, who it is compared with and which caveats are attached.

Brand reputation in AI search is the stable part of perception: attributes and caveats that repeat across answers and providers. One phrase in one answer is noise. The same caveat across four providers is what the models think of you, and every buyer who asks will hear it.

Why positive versus negative tells you very little

Between 16 and 27 September 2026, GEO Scout monitoring collected 5,459 answers that named one of 43 client brands (test and partner projects excluded; 42 of the 43 brands are Russian-language). Sentiment split like this:

• positive — 58.9%;
• neutral — 39.8%;
• negative — 1.3%.

There is an outside benchmark to check this against. Analyze AI, in a study published on 4 August 2026 on B2B prompts across ChatGPT, Perplexity and Google AI Mode, found 58.3% positive, 39.1% neutral and 0.9% negative answers, plus 2.1% mixed. Different language, different market, nearly identical split: once an AI model names a brand, it almost always speaks about it neutrally or well.

So the sentiment chart is green for almost everyone. That does not mean everything is fine.

In the same 5,459 answers, the brand landed at the Secondary option level 382 times (7.0%): the AI named it as weaker than a competitor, framed it negatively, or gave the direct recommendation to another brand. Here is how standard sentiment labelled those answers:

A sentiment metric misses four out of five answers where the brand loses. More than a quarter of them it counts as positive.

The wording explains why. Of the 322 answers where the brand is explicitly named as weaker than a specific competitor, 208 (about two thirds) also credit the brand with a strength in the same answer. The typical pattern: "easy to use and affordable, but falls short of X on speed". There is praise, so the label is positive. The pick went elsewhere, so the sale did too.

The opposite mistake happens too: 22% of neutral answers contain concrete arguments for the brand. The model is not praising it, it is explaining when to choose it — which converts better than an enthusiastic adjective.

Three layers of perception: attributes, role, sources

Instead of one sentiment scale, break perception into three layers. Each answers its own question and leads to its own work.

Attributes: the words used to describe you

An attribute is a short quality the answer assigns to a brand: "reliable", "premium", "for enterprise", "expensive". GEO Scout extracts up to three per answer for the brand and each competitor, only from what the text says — empty praise like "a great option" is dropped. Similar wordings are grouped into attributes with a polarity: strength, objection or neutral fact.

Two scores, both from 0 to 100, answer different questions:

• association score — what the AI connects the brand itself with: the attribute named first in a description of the brand scores 100, each next one 10 less, absent scores 0, averaged over answers that mention the brand;
• market prominence — how early the brand gets named among all brands the AI credited with that attribute, when the user asks about the category rather than about you.

The useful part is where the two disagree. A model can describe your brand as "fast" in nearly every answer about you and still leave you out when someone asks who is fast. In the section summary, that is the Gap card: praised, but not picked.

Since 11 September, descriptions of 49 brands have produced 8,647 extracted qualities. The most common attribute is "reliability" (15 brands across different niches), then "specialisation" (14), "low price" and "functionality" (9 each). The uncomfortable takeaway: "reliable" is not a differentiator, it is the background noise of the category. Differentiation starts with attributes your competitors do not have.

Models rarely state objections outright: 0.2% of brand qualities are negative (18 of 8,647), 0.9% for competitors. The most common objection is "high price", found in the data of 26 out of 63 projects. Negative framing in AI rarely sounds like "bad service". It sounds like "expensive" or "falls short of X".

Role: recommended or ranked below

The second layer is the role the AI gives the brand. In GEO Scout this is a five-level ladder: Benchmark, Primary pick, With arguments, Secondary option, Mention. The labelling rules are covered in Brand presence quality in AI answers; here I only connect it to sentiment.

In our data the split is: Benchmark 5.2%, Primary pick 3.7%, With arguments 55.2%, Secondary option 7.0%, Mention 29.0%. A direct recommendation goes to the brand in roughly 9% of its mentions. Most visibility is "named with arguments" or "just named", and that is where positive tone either turns into a pick or does not.

Two blocks explain why a brand does not climb higher:

• AI caveats about the brand — the exact wording that puts the brand lower: "falls short of X on …", "on par with Y, but …", plus the reason an answer was labelled negative. Each opens the full answer.
• Who gets the recommendation — competitors the AI picks outright where your brand is only mentioned, and what it praises them for in those answers. Clicking a competitor filters the caveats to show how the model justified choosing it, and how often it gave no reason at all.

The step-by-step diagnosis, with three typical scenarios, is in AI names your brand but picks someone else.

Sources: where the wording comes from

The third layer is the pages behind the wording. Clicking any attribute opens the domains and URLs cited in answers where the AI credited it to your brand, a competitor or any brand in the market. The same panel exists for negative answers about the brand.

The trap: a page cited next to a brand mention does not necessarily talk about the brand. So GEO Scout opens the page and looks for the brand name on it. Confirmed pages come first, pages verified not to mention the brand are excluded from the brand view, and the "brand is on the page" badge expands the matching text fragments.

A note on mechanics. In the same study, Analyze AI compared the sentiment of cited pages with the sentiment of the answers citing them across 985 page–answer pairs and got a correlation of 0.00 (Spearman, p = 0.983). The authors flag the sample as partial — 270 answers, B2B only — but the result is telling: the tone of a single page does not carry over. What carries over is wording and facts the model meets across many sources. So the work is not making pages "more positive" but targeting one attribute or caveat across every page that sustains it. How models pick what to cite: Cited sources in AI.

How to audit sentiment and perception manually

A spreadsheet and one working day are enough for a first picture.

1. Write 10–15 unbranded prompts the way a buyer asks: "which tool should I pick for …", "alternatives to X". The buying decision happens here, not in "tell me about Brand". Add 3–5 comparison prompts: "Brand vs Competitor", "downsides of Brand".
2. Pick the AI providers your audience uses. For most Western markets: ChatGPT, Gemini or Google AI Mode, Perplexity and Copilot. If you sell in Russia or the CIS, add Alice AI, GigaChat and Yandex Search with Alice — perception there often differs from ChatGPT.
3. Run every prompt 2–3 times in a fresh chat. Answers vary between runs, and a single sample misleads.
4. Label each answer using the columns below.
5. Roll it up: which attributes appear three or more times, which caveats show up on two or more providers, who takes the recommendation most often.

The limit is obvious: 15 prompts × 5 providers × 2 runs is 150 answers, and nobody repeats that weekly. Without repetition you cannot see whether perception changes after your edits.

How to monitor perception over time

For a recurring report I recommend a short set of numbers instead of one sentiment score:

• strong share — mentions at Benchmark, Primary pick and With arguments;
• weak share — mentions at Secondary option;
• top 3 attributes with association score, plus the biggest gap;
• objections and repeated caveats with answer counts;
• rivals — who takes the recommendation, and for what;
• negative share — as a sanity check, not the headline.

Track it weekly and per provider: a blended average hides one provider recommending the brand while another keeps it as a secondary option.

In GEO Scout this set lives in the Brand perception section (full walkthrough in the knowledge base guide): a summary of how AI describes the brand, the presence ladder by provider and against competitors with its trend, Who gets the recommendation and AI caveats about the brand, a brand shape radar, a heat map comparing brands by attribute, What justifies picking competitors, and a sources panel behind any attribute or negative answer. The "Explain the perception" and "Fix AI's reservations" buttons start a guided analysis in Agent mode, and over MCP the same data comes from the get_brand_perception, get_presence_quality and get_perception_sources tools.

There are no separate probe queries behind this: attributes, role and caveats are extracted from the same monitoring answers visibility is calculated from. The upside: it works on every plan, including free, and describes perception on exactly the prompts that matter to you. The downside: nobody asks the model "what is this brand known for" separately, so perception is bounded by your prompt set.

A second limitation: we have no automated fact-checking of AI claims against a fact base you define. Facts are checked against the full answer texts.

What matters most for Russian-speaking markets is coverage. GEO Scout monitors 12 AI providers: ChatGPT, Claude, DeepSeek, Gemini, Google AI Mode, Google AI Overview, Grok, Perplexity, Yandex (Search with Alice), Alice AI, GigaChat and Microsoft Copilot. 11 are captured from the live user-facing interface; Claude comes through its official API. Alice AI and GigaChat shape how millions of Russian-speaking buyers see a brand, and Western monitoring tools do not see them at all.

You can start free: 5 prompts across 6 AI providers (ChatGPT, Gemini, DeepSeek, Perplexity, GigaChat, Alice AI) with a refresh every 7 days — for perception, a pointer rather than statistics. Daily monitoring is on paid plans from $24/mo (pricing). Whatever the plan, periods with few mentions produce sharp jumps, so read weekly and monthly views and check the answer count in the tooltip.

What to do about negative framing and caveats

The main rule: work on a specific phrase, not on "sentiment" in general.

Three things that almost never work:

• One positive article. The tone of a single page does not carry into the answer. You need the same wording across many pages, not one glowing review.
• Arguing in generic terms. A caveat like "falls short of X on speed" is answered by a page with numbers and test conditions, not a paragraph about "high quality of service".
• Ignoring reviews. Objections like "pushy upselling" or "complex interface" come from reviews and discussions. How to work with those platforms: review platforms for AI.

To check the effect, note the date a page was published or edited, then after two or three monitoring runs compare the weak share, the repeat count of the caveat and the attribute gap, per provider. Providers with live search pick up changes once updated pages are retrieved; answers from model memory change only with a model update, so the blended chart may move later than you expect.

The bottom line

Brand sentiment in AI answers is a useful sanity check and a poor target. It is almost always positive and misses most answers where your brand loses, because praise and a pick for the competitor sit comfortably in the same paragraph.

Measure perception in three layers instead: the attributes that describe the brand, the role it gets and who takes the recommendation, and the pages the wording comes from. Work on one repeated caveat and every source that sustains it, then judge the result weekly and per provider — by the weak and strong shares, and by whether the phrase that started it all is still being said.