AI Almost Never Says Anything Bad About You

What We Measured

Every error-free answer recorded between 2026-01-25 and 2026-08-28, across 4,703 distinct prompts. When an answer names the brand, the response is classified positive, neutral or negative from the text itself, and the sentences carrying the judgement are kept verbatim.

This post uses the whole corpus rather than the July-onward window our per-engine citation work is restricted to. That window exists because citation capture was broken before the model cutover. Sentiment is read from the answer text, which was never affected, so narrowing the window here would throw away most of the sample to fix a defect this measurement does not have.

The Rate Is Not a Constant

A single pooled percentage invites you to treat it as a property of AI, the way people talk about a spam rate. It is not one. Month by month the negative share of mentions runs from 0.53% in Feb to 7.50% in May, a fourteenfold spread across a period in which nothing about the engines changed that dramatically.

The reason is that negative mentions cluster. Of the 1,395 prompts that ever produced a mention at all, only 104 ever produced a negative one, or 7.5%. A single prompt accounts for 45 of the 299, the five worst for 35.5% between them, and the twenty worst for 62.5%.

Silence is the norm even where there was plenty of opportunity to speak. Among the 347 prompts that produced five or more mentions, 283 of them, or 81.6%, never returned a single negative one.

A clustered rate and a diffuse rate need opposite instruments. A diffuse one belongs on a gauge, where the trend line is the information. A clustered one belongs behind an alert, because the average is quiet in every month where nothing is wrong and the signal arrives as a spike in one place. Monitoring an aggregate sentiment score is close to the worst way to catch a problem shaped like this one.

The Question Decides the Sentiment

The clustering is not random, and it is not mainly about which brand is being discussed. It is about what was asked. Split the same corpus by the kind of prompt behind each answer and the negative rate moves by most of an order of magnitude.

Comparison prompts, the head to head ones, return negative sentiment on 6.63% of mentions. Problem solution prompts, where somebody describes a problem and asks what would fix it, return it on 0.70%, a factor of 9.5 between them.

Those two figures are not like for like, because the two kinds of prompt are not tracked by the same accounts. Comparing only accounts that track both, across 1,566 comparison mentions and 1,377 problem-solution ones, both rates fall a long way, to 1.98% and 0.22%. The gap between them barely moves: 9.0x against 9.5x pooled. Among accounts with at least ten mentions in each, comparison came out higher 7 times and lower 0 times, with 9 showing no difference either way. So the level is a property of who is being tracked; the gap is a property of the question.

It makes sense once stated. A comparison question is a request for drawbacks, so the answer goes and finds some. A problem-solution question is a request for a recommendation, and a recommendation with caveats attached is a worse recommendation. The engine is answering the question it was given.

The consequence for measurement is uncomfortable. Your sentiment score is partly a readout of your own prompt mix. Add comparison prompts and it will fall; add problem-solution prompts and it will rise, with nothing having changed about how AI regards you. Two brands with different tracked sets cannot be compared on it at all.

One category is worth calling out separately. Brand research prompts, where the question names your brand directly, produce the highest neutral share of any category at 31.1%. Asking an engine about you by name mostly returns a flat recital of what you do, which is the least useful thing it can say.

The Engines Disagree About Your Character

To compare engines you have to hold the question constant, so this uses only prompts answered by all three inside the same tracking run: 6,961 answers per engine, same brands, same day.

Perplexity returns negative sentiment on 6.7% of its mentions against 2.6% for ChatGPT, a factor of 2.6 on identical questions. It is also the most reluctant to be enthusiastic, at 67.7% positive against 81.1%.

This is a house style, not a verdict on you. The engines that hedge most are the ones drawing on the widest range of sources, and a wider range means more chance of surfacing a complaint. The same holds for competitor sets: the same brand, asked the same question, comes back described differently depending on what the engine happened to read.

Neutral Costs You More Than Negative

Sentiment is usually reported on its own, as though the label were the outcome. It is more useful next to position, because then it stops being a mood and starts predicting where in the answer you land.

The ordering is not the one the category assumes. Positive mentions average position 2.24. Negative mentions average 2.66. Neutral mentions come last at 3.69. The direction survives outside the matched set too, at 2.17, 2.58 and 3.31 across the full corpus.

Being argued with is a form of being taken seriously. An engine that names a drawback has already decided you belong in the answer and is helping the reader choose between real options. A neutral mention is the other thing: your name appears in a list, nothing is claimed for you, and you sit below the brands the answer had an opinion about.

Neutral is also 17.2% of mentions against negative's 2.31%. It is the common outcome, it costs placement, and almost nobody tracks it as a problem.

What the Criticism Is Actually About

Of the 299 negative answers, 154 carry the sentences that earned the label, 505 in total. Running a keyword pass over them puts 102 sentences, or 20.2%, on price or cost, roughly 2.9x the next theme.

Be careful how much weight that carries. The keyword pass matches only 36.8% of the sentences at all, so it describes a shape rather than a taxonomy, and the themes it does not catch are not evenly distributed. What it does establish is that the leading complaint is commercial rather than reputational. Almost none of this is scandal. It is pricing, support and gaps in the feature list.

The more useful pattern is where the criticism comes from. 97 of the sentences, or 19.2%, name their source out loud: reviews, forums, users on a community site. The engine is not forming an opinion about you. It is relaying one it read, which follows from the community citation share behind these answers.

That changes where the work sits. If a model is passing on a two year old pricing complaint from a review site, no amount of writing on your own domain answers it. The page the engine is reading is the thing that has to change.

What to Do With This

Four things follow, in the order they are worth your time.

Stop watching an aggregate sentiment score. At 2.31%, clustered into a handful of prompts, it spends almost all of its life flat and tells you nothing on the days it moves, because a single reading moves on its own anyway. Set an alert on new negative mentions instead and read the sentence.

Track neutral as the problem it is. It is the label that predicts a worse position, it is far more common than negative, and moving a mention from neutral to positive is a much larger available gain than eliminating criticism that mostly is not there.

Read sentiment per engine, never pooled. A brand can sit at 2.6% negative on one engine and 6.7% on another for the same question, so a blended figure describes neither.

And keep the real risk in proportion. The thing that should worry you is not that AI is rude about you. It is that AI leaves you out of 64.5% of the answers where your buyers are asking, or that it says something confidently wrong about what you do. Both are more common than criticism, and both are fixable.

If you are shopping for a tool to do this, our sentiment monitoring roundup still stands. Just buy it knowing which number in it is going to earn its place.