When AI recommends a brand, it is almost never weighing criticism against praise. We read the language of 1,144,385 brand mentions inside the web pages AI cited over a 30-day window, and only 3.5% were negative. Sixty percent were flat neutral, name-drops and listings that make no case at all. The evidence base behind AI brand recommendations is not adversarial. It is promotional or indifferent, which changes what your monitoring should actually look for.
What does AI actually read about brands?
Parse labels the sentiment of every brand mention inside the sources AI cites, separately from the AI's own answer. Across 1,144,385 validated mentions covering 108,789 brands and 110,098 domains, the distribution is lopsided toward the harmless. Positive mentions outnumber negative ones by more than ten to one. The single largest bucket is neutral: a brand named in a list, a feature table, or a passing reference, with no opinion attached.
This is the sentiment of the source content (the cited page), not of the AI sentence that quotes it. The gap matters. The story is what the model has to work with before it writes a word.
| Source sentiment toward the brand | Share of mentions | Count |
|---|---|---|
| Neutral | 59.8% | 684,063 |
| Positive | 36.2% | 414,584 |
| Negative | 3.5% | 40,582 |
| Mixed | 0.5% | 5,192 |
| Hedged | 0.04% | 510 |
Read together, 96% of what AI reads about a brand is neutral or positive. The implication is uncomfortable for anyone who assumes AI answers reflect a balanced read of the web: on the input side, balance does not exist. Criticism is the exception, not the counterweight.
- Across 1,144,385 brand mentions in AI-cited sources, only 3.5% carried negative sentiment, while 36.2% were positive and 59.8% were flat neutral.
- Positive source mentions outnumbered negative ones by 10.2 to 1.
- Of 108,812 brands, only 10.5% drew a single negative source; among the 27,051 brands with five or more mentions, 66% drew zero critical sources.
- The deeper a source goes on a brand, the kinder it gets: 44.1% of primary-subject mentions were positive versus 23.8% of passing mentions.
- When AI's sources hedge, 77% of the hedges are "it depends on your use case" or "a rival is better," not safety or risk warnings.
How often is a brand criticized at all?
Rarely, and for most brands, never. Of 108,812 brands in the set, only 11,461 (10.5%) had even one source mention labeled negative. Filter to brands with real coverage, five or more analyzed mentions, and 27,051 qualify; of those, 17,859 (66%) still had zero negative sources. Two-thirds of well-covered brands face an AI evidence layer with no criticism in it whatsoever.
That is not because those brands are flawless. It is because the pages AI reaches for, official sites, listicles, category roundups, and review aggregators, are not built to criticize. The web AI samples skews toward content produced to sell or to summarize, not to scrutinize. For brand teams, the takeaway inverts the usual anxiety. The risk is rarely that AI is repeating something damaging about you. The risk is that your thin slice of negative coverage carries outsized weight precisely because there is so little of it.
Why does deeper coverage read more positively?
Because the sources that go deepest on a brand are usually the ones with a reason to be kind. Parse tags how prominent a brand is in each source: the primary subject, a secondary mention, or a passing reference. Positivity rises with prominence, and neutrality falls.
| Brand prominence in source | Positive | Neutral | Negative |
|---|---|---|---|
| Primary subject | 44.1% | 52.4% | 3.0% |
| Secondary | 37.3% | 58.4% | 3.8% |
| Passing mention | 23.8% | 72.0% | 3.9% |
When a page is mostly about your brand, it is often your own site or a dedicated review, and 44.1% of those mentions read positive. When your brand is a drive-by reference inside content about something else, positivity drops to 23.8% and the mention goes flat neutral 72% of the time. Negativity barely moves across the three tiers (3.0% to 3.9%), so the real swing is between praise and indifference, not between praise and attack. Owning the pages that go deep on you is how positive framing enters the evidence layer in the first place.
If you want to know when AI changes its answer about your brand, start with a free brand check — it takes a minute.
Where does the criticism actually come from?
When a source does criticize a brand, it is disproportionately a community or news page, not a vendor or reference page. Cutting negativity by source type shows a clear hierarchy, against a corpus average of 3.5% negative.
| Source type | Negative | Positive | Neutral |
|---|---|---|---|
| Social (Reddit, forums) | 8.5% | 36.4% | 53.3% |
| News | 4.3% | 36.6% | 58.6% |
| Editorial | 3.7% | 43.2% | 52.5% |
| SaaS / vendor | 3.6% | 36.5% | 59.5% |
| Encyclopedia | 2.6% | 5.0% | 92.4% |
| Ecommerce | 2.3% | 31.2% | 66.3% |
| Government | 2.0% | 12.0% | 85.8% |
Social sources criticize at 8.5%, roughly 2.4 times the corpus average, and Reddit alone accounts for 9.7% of every negative brand mention in the data. Reference and institutional sources (encyclopedia, government, documentation) almost never criticize, but they almost never praise either; they sit at 85% to 92% neutral. If you want to know where your AI reputation is most exposed to a real knock, it is the community layer, which is exactly where you have the least control and the most need to monitor. We have written before on negative Reddit threads and AI visibility and on which subreddits get cited most by ChatGPT; this data quantifies why those threads punch above their weight. They also punch for longer than most sources: Reddit citations are among the most durable in AI answers, the class least likely to appear once and vanish, so a critical thread that enters the evidence layer tends to keep getting cited.
When AI recommends you "with a but," what shape does the but take?
A reluctant recommendation is rare, 3.4% of all mentions, but when it appears it follows a small set of patterns. Parse classifies the reluctance, and the shape is more useful than the rate. Of the reluctant recommendations:
- Conditional (42.9%): good, if your use case fits. The recommendation survives, scoped to a segment.
- Inferior comparison (33.8%): fine, but a named rival is better. You are present and losing the head-to-head (how AI picks a winner, axis by axis).
- Risk warning (9.8%): an explicit caution about a downside or limitation.
- Fallback (7.2%): a second choice if the first option does not work out.
- Budget only (4.9%): recommended mainly because it is cheap.
Together, "it depends" and "a rival is better" account for 77% of all hedged recommendations. That reframes the work. Most reluctance is not a reputation crisis to defend against; it is a positioning problem to fix. If AI keeps recommending you conditionally, the fix is content that resolves the condition (clarifying who you are right for). If it keeps recommending a rival over you, the fix is comparison evidence, not crisis management.
What does this mean for AI brand monitoring?
It means the metric most teams reach for, sentiment, is close to a constant for most brands, and a near-constant is a poor early-warning signal. Parse tracks AI visibility across ChatGPT and Google AI, covering a public index of 603,456 brands and more than 57 million citations, and the source-language layer is consistent across that index: criticism is scarce and most coverage is characterless. The better questions to monitor are which neutral mentions you can convert into positive, distinctive ones (the actual words AI uses to describe brands), and whether your rare negative sources are spreading. Watch the community layer for new criticism, watch the inferior-comparison hedges for which rival is winning the head-to-head, and audit the sources behind your brand to see whether the pages that go deepest on you are ones you influence. Treat the thin negative slice as high-signal, because in an evidence base that is 96% non-critical, the 3.5% that pushes back is the part that moves.
Methodology and caveats
This study covers 1,144,385 brand mentions with a validated sentiment label, drawn from the source content AI cited across a 30-day window in mid-2026 (108,789 brands, 110,098 domains). Sentiment is labeled by Parse's source-language model on the cited page text, so it measures what the source says about the brand, not what the AI's answer says. The two layers diverge, and the answer-side gap is its own study.
The figures describe the sources behind AI answers, not the AI's own wording. A brand can be framed positively by an AI sentence that cites a neutral source; this data is the input layer, measured separately. Brand mentions here are weighted toward whatever AI chose to cite, so high neutrality partly reflects listings and roundups, which are heavy in AI citations. Sentiment is model-labeled at scale, and "negative" includes inferior-comparison mentions where a competitor is praised over the brand. One window is a point in time, not a trend.
Numbers come from Parse's first-party AI-monitoring dataset. You can run the same source-sentiment cut for your own brands and competitors inside Parse Sources and Brands.
Does AI say negative things about brands?
Rarely. In Parse's analysis of 1,144,385 brand mentions inside AI-cited sources, only 3.5% carried negative sentiment, against 36.2% positive and 59.8% neutral. Positive mentions outnumbered negative ones by more than ten to one, and two-thirds of brands with substantial coverage drew zero critical sources at all.
What is AI brand sentiment and how is it measured?
AI brand sentiment can mean two things: how the AI's own answer frames a brand, and how the sources it cites frame the brand. This study measures the second, the source-content layer, labeling each brand mention in cited pages as positive, neutral, negative, mixed, or hedged. It is the evidence AI reads before it writes, which is why it predicts how answers form.
Where does AI find negative information about a brand?
Mostly community and news sources. Social pages such as Reddit and forums carried negative sentiment 8.5% of the time, roughly 2.4 times the overall average, and Reddit alone accounted for 9.7% of all negative brand mentions. Reference and institutional sources (encyclopedias, government, documentation) almost never criticize brands, but they almost never praise them either.
Why does AI recommend my brand with a caveat?
Reluctant recommendations are only 3.4% of mentions, but 77% of them are either conditional ("good if your use case fits") or an inferior comparison ("a named rival is better"). Only about 10% are explicit risk warnings. Most caveats are positioning problems, who you are right for or who beats you, rather than reputation damage.
Should I monitor AI sentiment for my brand?
Monitor it, but do not expect it to swing. For most brands sentiment is near-constant and non-critical, so it is a weak early-warning signal. The higher-value signals are new negative sources in the community layer, the rival showing up in your inferior-comparison hedges, and whether the pages that cover you most deeply are ones you influence.