How to fix negative brand sentiment in AI answers

What does “sentiment” mean when an assistant describes your business?

Sentiment here is the slant of the description itself: whether the sentences an assistant writes about you read as favourable, unfavourable, flat, or both at once.

It is a property of one answer, not of your company. Ask for the best options in your category and your name may arrive with a clause attached — though some users report slow support. Ask again with the words reordered and the clause may not appear. The description is written fresh each time, out of training and out of whatever was fetched while answering.

So one answer’s slant is one of four things: positive, negative, neutral, or mixed — praise and criticism in the same breath. A fifth outcome is not a slant at all: the answer may not name you. No description means no sentiment, and counting that as neutral is the first mistake available.

Is this the same as a review score or a star rating?

No, and the difference is what makes it hard to act on.

A star rating is a number computed from inputs you can inspect: reviews, each with an author, a date, and a platform with a written process for disputing one.

An assistant’s description has none of that furniture — no total, no author, no record of which pages contributed, only prose that varies from run to run. So there is nothing to dispute, and no single figure that can be made to go up. A dashboard handing you one has manufactured it.

The honest unit is a count: how many recorded answers described you one way, how many the other, out of how many were actually read.

The hard part: you cannot edit the model

There is no support desk for what an assistant says about your business, and accepting that is cheaper than discovering it later.

Model weights are not editable from outside. You cannot file a correction, you get no case number, and there is no verified-owner route for amending a characterisation the way a directory listing or review platform gives you one.

You cannot change the model. You can change what it reads.

That is the entire available surface, and it is narrower than most advice on this subject admits. An assistant that fetches while answering reads whatever the open web says about you at that moment; a model retrained later reads it again. Every remedy below acts on a page, a thread, a listing or a record. None acts on the model.

How do you find out what is actually being said about you?

Ask the assistants the questions your customers ask, and write down what comes back word for word.

Not what do you think of my company — nobody types that. The questions that matter are buying questions: best tool for this job, is this one any good, alternatives to it, is it worth the money. Ten is enough to start.

Record four things per answer: the exact question, which assistant, the date, and the complete text of the reply — not a summary of it. The phrase matters more than the verdict, because the phrase is what you will later go looking for a source for: reported slow onboarding sends you somewhere specific, negative sends you nowhere.

The manual version of this routine is written out in track brand mentions in AI.

One answer is one observation, not a sentiment

Assistants do not answer the same question the same way twice, so a single reply tells you what happened once.

Sampling, retrieval and the word order of the question all move the output. The same prompt an hour later can name different companies and attach a different clause to yours. A screenshot of something unflattering is not yet a finding.

A pattern is a finding. The same criticism in four of six recorded answers is worth tracing; one in twelve is noise, and chasing it costs more than it returns. Below two measured runs on one question there is nothing that deserves a name — the rule the product further down enforces on itself.

Where does a negative characterisation come from?

Five origins account for most of what people find, and reading the answer carefully usually says which one it is.

1. An old review page. Complaints from a period since fixed, still near the top of the results for your name. Nothing has to be recent to be repeated.
2. A forum thread. One long discussion, strongly worded, often years old, and frequently the most detailed thing written about your product anywhere — and detail is what retrieval rewards.
3. A news item or a public record. A dispute, a filing, an outage. Accurate, permanent, and not something to argue with.
4. A competitor’s comparison page. Phrased as a fair summary, and often the only page describing you in the exact terms a buyer asked about.
5. A plain factual error, repeated. A price that changed, a product discontinued, a market you left. The model is not being unfair; it is quoting stale text.

The last is the most common, the least dramatic, and the one fixable this week.

Correcting the source, when there is a source to correct

When a description traces to a specific page, that page is the job, and the highest-leverage work on this list.

Order matters. A page you own is an edit. A page you do not own falls into three cases: a platform with a documented correction route, such as a business profile or directory listing; a publisher who will consider a factual, specific correction; and everybody else, who will not. Work down that order and stop at the third.

What this is not is a takedown campaign. A legitimate criticism is not an error, and an accurate page that happens to be unflattering is not a correction target. Pushing on one is how a small complaint becomes a larger story — which the assistants then read.

Answering the objection on a page of your own

If a criticism keeps recurring, publish the page that answers it directly, in the words the criticism uses.

The mechanism is dull and it is the only one on offer. An assistant that fetches while answering retrieves pages matching the question in front of it, and a page that states an objection and answers it matches that question better than one which never mentions it.

So name the thing: the setup time, the pricing confusion, the limitation that is genuinely real — what is true, what changed, and when. Say the unflattering part out loud. A page that concedes a limitation and explains the trade-off is usable evidence; one that only praises you reads as marketing and weighs nothing against the thread that prompted the question.

Getting other sources to agree with you

One page saying the right thing is weaker than three independent pages saying it, and that asymmetry is the whole of this step.

A description assembled from many documents moves when many documents move. The practical list is short: the factual pages on your own site, your profile on whichever platforms your category uses, any directory that accepts submissions, and whatever independent coverage can honestly be earned.

The flip side is unglamorous. Where one fact is written four ways in four places — a changed tier, a dropped feature, a new location — the inconsistency becomes the description.

Fixing the plain factual errors on pages you control

A large share of unflattering descriptions is not criticism at all. It is stale fact, faithfully repeated.

A discontinued product still listed. A support policy that changed two years ago. A minimum contract you no longer require. An assistant reading your site has no reliable way to know which page is current, and a model that absorbed the old version keeps reproducing it until something newer displaces it.

The audit is mechanical: list the facts a buyer asks about — what it costs to begin, what is included, how long it takes, who it is not for — check each against every page of yours that states it, and delete the contradictions. The cheapest item in the guide, because every page involved belongs to you.

What does not work, and gets sold anyway

Three remedies are sold for this problem and none of them can work.

What brand sentiment tracking in an LLM can honestly report

A tool can tell you how the answers it collected described you, and over how many runs. Anything beyond that is invented.

Here is how that works in this product, so the limits are visible rather than implied. When a scan runs, every stored answer naming the brand is read by a classifier, and the label it returns — positive, neutral, negative or mixed — is written onto that answer then and there, stamped with the classifier’s name and version. Later views read the stored label, and a cell labelled by two different classifiers says so.

Three outcomes are not labels, and keeping them apart is the point. An answer that never names the brand is absence: no description exists, so no sentiment exists, and it is not neutral. An answer with no text to read — a failed run, an engine not asked, text past its retention window — is not measured. A run where the text was there and the classifier was not is a failure, carrying its reason, and the loudest thing in its cell. None of the three is folded into neutral, and none is a zero.

There is no score in it. The output is a set of counts — this many runs positive, this many negative — printed beside the number of runs actually measured. No mean, no net figure.

And it refuses to publish a distribution from fewer than two measured runs. Each plan asks a question once per scan, so one scan cannot reach that floor alone; the screen pools the labels its recent scans already stored, matched on the exact question and engine, and every cell states how many scans it drew from. Nothing is padded. Sentiment reporting is on the paid plans.

A routine you can keep

Quarterly beats monthly here, because what is measured moves at the speed of published pages rather than of days.

1. Re-ask the same questions, word for word. A reworded question is a different question, and its answers do not belong in the old column.
2. Count, do not judge. How many answers carried the criticism, out of how many came back. Write both numbers; the second is the one people drop.
3. Trace anything new to a source. If none is found, log the exact phrase and wait for the next round.
4. Check whether last quarter’s fix is visible. Two rounds with no movement in the wording means the evidence you changed was not the evidence being used.

A pair of counts per quarter is a record. Three records is a trend, and a trend is the only thing that tells you whether any of this worked.

FAQs