Sentiment: what engines say about you, not just whether they name you
Most AI visibility work counts appearances. A prompt is run, the answer is scanned for the brand name, and the result is recorded as a yes or a no. That measurement captures half of what happened. The answer also said something about the brand, and what it said decides whether the appearance was worth having. A brand described as the expensive option with slow support has been named. It has also been argued against.
Presence and characterization move independently
Presence asks whether the brand appears in the answer. Sentiment asks how the brand is characterized once it does appear. These are two separate measurements, and they do not track each other. A brand can appear in most answers for its category and carry the same caveat in every one of them. A brand can appear rarely and be described accurately each time it does.
Ranked against each other, high presence with a recurring negative clause is the worse position. Absence means the reader learns nothing about you. Presence with a qualification means the reader learns a specific reason not to choose you, delivered by a system they treat as neutral. Work that optimizes only for appearance rate can make that position worse by making it more frequent.
The characterization is assembled from sources
An answer engine composes its answer from content it retrieved. It holds no opinion about your company. When a criticism recurs across reviews, forum threads, comparison articles and support discussions, it recurs in the retrieved set, and a clause built on it recurs in the answers. When the same criticism appears once, in one obscure place, it usually does not survive into the answer.
That makes an unfavorable characterization a retrieval problem before it is a brand problem. The useful question is not what the model thinks of you. It is which pages the model is reading. Community discussion is a large part of that set, and how it gets there is covered in Reddit, forums and UGC in AI answers.
Measuring characterization without fooling yourself
Three rules separate a usable sentiment measurement from a number that feels informative and is not.
- Record the verbatim answer text, not a machine label. The sentence is the evidence. A label is a summary of evidence you have already discarded, and you cannot go back to it later to ask what was actually said.
- Classify recurring claims, not mood. A 1 to 5 sentiment score compresses away the part you can act on. "Negative, 2.1" names nothing. "Four of eleven answers said onboarding takes months" names a claim, which can be checked, corrected or fixed.
- Sample across engines, languages and phrasings. The reasons are the same ones that apply to any AI measurement. Analysis of variance in LLM brand responses found that the language of the query accounts for 26.5 to 32.0 percent of the variance in what comes back, while brand identity accounts for 1.5 percent. A characterization seen once, in one language, on one engine, is not a finding.
The full sampling requirements are in measuring AI search visibility honestly, and what to record run by run is in the metrics worth keeping.
Four kinds of negative claim
Once you classify claims rather than score mood, the claims sort into four types, and each has a different fix.
| Type | What it is | Where the fix sits |
|---|---|---|
| Accurate criticism | The claim is true | The product or the policy |
| Outdated criticism | True once, not true now | The sources still saying it |
| Competitor framing | A rival's positioning repeated as fact | Your own published substance |
| Misattribution | Wrong company, or a detail that exists nowhere | Entity clarity and correction |
Accurate criticism
No content strategy fixes a true complaint. If buyers consistently report that setup is slow, a page asserting that setup is fast contradicts a large body of retrieved material and loses. The record changes when the underlying thing changes, and only then. Treating accurate criticism as a visibility problem is the most common way to spend a year without moving anything.
Outdated criticism
This is the most common type and the most tractable. Two actions matter. Make the current facts easy to retrieve by stating them plainly, with dates, on a page written to be quoted, as described in the grounding page. Then ask the sources still carrying the old claim to update: the review, the comparison article, the thread. A publisher who corrects a two year old statement changes the retrieved set directly, which is the only thing that changes the answer.
Competitor framing
Comparison content written by a competitor states your limits in their terms, and engines retrieve it because it answers comparison questions directly. The counter is your own comparison substance, published with the real trade-offs stated, including the ones that do not favor you. Material that claims no trade-offs is not retrieved as an answer to a trade-off question. The approach is set out in writing comparison pages.
Misattribution
A claim that belongs to a different company, or a detail present in no source at all, is a different failure with a different remedy. It is covered in when an AI answer states something false about your company.
What nobody can promise
No publisher can edit an engine's characterization. There is no field for submitting a preferred description, no setting that adjusts tone, and no access to the sentence a model is about to compose. Each answer is assembled at request time from whatever was retrieved at that moment, which means the same brand can be described two ways in two consecutive runs.
For the same reason, no vendor can promise sentiment improvement. Improvement is not something anyone can apply to a model's output. What is available is slower and indirect: fix what is fairly criticized, correct what is stale at the source, publish substance where the framing is contested, and sample answers over months to see whether the recurring clause recurs less often. That is the whole mechanism, and anything sold as more than that is selling access nobody has.