--- title: "Do Google AI Overviews downplay negative content? | Underneath" description: "In one audit of 11,000 Google questions, AI Overviews drew less from negatively toned sources. Brand reviews were not tested, and criticism still appears." canonical: "https://underneath.agency/resources/do-ai-overviews-downplay-negative-content" published: 2026-10-07 updated: 2026-10-08 publisher: "Underneath (https://underneath.agency/agent)" entity: "https://underneath.agency/.well-known/entity.json" --- Guide · AI search # Do Google AI Overviews downplay negative content? In one large audit, yes: Google’s AI Overviews drew noticeably less of their text from negatively toned sources than from neutral or positive ones. That audit used general questions, not brand or review searches, and AI answers about brands still contain plenty of criticism. The realistic risk is not that complaints vanish, but that they end up in the citation list, further down the answer, or left out. ## The short version 1. In an audit of 11,000 real Google questions, AI Overview text drew 13.8 points less than expected from the negatively toned pages it cited (University of Illinois researchers). 2. The same summaries drew 22.1 points less from the forums and social sites they cited, which is where many customer complaints live. 3. When an AI Overview claim is unsupported, it is about 2.6 times more likely to state something no cited page mentions than to contradict a page (Washington University researchers, 98,020 claims). 4. In our test of 79 brands, AI answers raised at least one problem 99.7% of the time, but Gemini and Google AI Mode put a negative point in the first paragraph only 1.3% of the time. ## Do AI Overviews give negative sources less weight? In one large audit, yes: negatively toned sources were under-used in the summary text. [Huang and colleagues at the University of Illinois](https://arxiv.org/abs/2603.16138) ran 11,000 real search questions, drawn from a public collection of Google queries, through Google and ChatGPT search. Only 57.8% of the questions triggered an AI Overview, the AI summary at the top of Google’s results. For a sample of answers, the researchers split each summary into single facts and checked which cited pages supported each one. That shows how much of each cited page made it into the summary. Compared with an even split, the summaries used cited sources in these ways: | Kind of cited source | Use in the AI Overview text | |---|---| | Negative tone | 13.8 points less than average | | Positive tone | 1.1 points more than average | | Social media and forums | 22.1 points less than average | | Wikipedia | 5.4 points more than average | ChatGPT search showed no reliable link between a source’s tone and how much it was used. The authors concluded that “Google AIO’s synthesis implicitly filters against negative content.” ## Does that mean Google hides bad reviews about your brand? Not as far as the evidence goes, because the audit did not test brand or review searches. The questions were general information questions from a collection first gathered around 2018. They were run from one location, Urbana, Illinois, at one point in time. The tone of each source was scored by automated tools, not people, and the study shows a pattern, not its cause. The authors note that an under-used source may simply be lower quality, repetitive or less relevant. A negative page could be under-used for any of those reasons. AI Overviews are not uniformly upbeat either. Compared with ChatGPT answering without search, their wording used 44% fewer positive emotion words but only 25% fewer negative ones. On contested debate questions, [Grossman and colleagues](https://arxiv.org/abs/2604.27790) found that 33.4% of AI Overviews opened with a plain yes or no, against 5.6% of Gemini answers. ## Where do complaints and reviews end up in AI Overviews? Often in the citation list rather than the summary, because forums are cited but little used. In the Illinois audit, AI Overviews cited Reddit and Quora but drew 22.1 points less content from social and forum sources than from others. Reddit was among the most under-used sources of all. [Xu and colleagues at Washington University in St. Louis](https://arxiv.org/abs/2605.14021) tracked 55,393 trending searches over 40 days. AI Overviews cited user-generated platforms less often than Google’s own first page did, in all 19 topic categories. In science searches, 56.22% of first-page links came from those platforms, against 15.33% of AI Overview citations. The same team checked 98,020 individual claims against the pages the AI Overviews cited. 11.0% were [not supported by the cited pages](https://underneath.agency/resources/ai-answers-unsupported-claims), and 4.1% were contradicted by them. Unsupported claims were about 2.6 times more likely to state something no cited page mentioned than to contradict a page. ## Do other AI engines treat negative information the same way? Partly: some also under-use negative sources, and several place criticism below the opening. In the Illinois team’s check on Perplexity, negatively toned sources were under-used by 10.1 points, close to the AI Overview pattern. Our [“Is this brand legit?” study](https://underneath.agency/research/is-it-legit-ai-reputation-study) asked ChatGPT, Gemini, Perplexity and Google AI Mode about 79 brands on 26 September 2026. Criticism was everywhere: 99.7% of complete answers made at least one negative claim, and 35.5% of all claims were negative. Placement differed sharply by engine: | Engine | Answers with a negative claim in the first paragraph | |---|---| | Perplexity | 31.6% | | Gemini | 1.3% | | Google AI Mode | 1.3% | | All four engines | 10.9% | [Kumar](https://arxiv.org/abs/2606.20065), a co-founder of an AI visibility company, reports from tracked brands that 0.0% of brand, question and engine combinations were consistently negative. Tone flipped between positive and negative across runs in 45.5% of them, so negativity came and went rather than settling. ## What does an AI assistant weigh when reviews are mixed? In one controlled test, star ratings mattered most and replies to reviews barely registered. [Baig and colleagues](https://arxiv.org/abs/2606.16344) asked 12 AI models to pick among fictional hotels whose ratings, reviews, prices and other details were set at random. A 4.7-star hotel was recommended 31.6 points more often than a 3.9-star one. A visible management response to reviews had no detectable effect, at +0.1 points. This was a simulated choice among five listed hotels, and two of the authors are affiliated with a travel-sector company. Where machine and customer priorities part ways is weighed in [AI versus human reputation priorities](https://underneath.agency/resources/ai-vs-human-reputation-priorities). Review sites carry most of the criticism in real answers. In our brand study, claims backed only by review or complaint platforms were negative 56.5% of the time. Claims backed only by the brand’s own website were negative 6.4% of the time. ## What should you do about it? Do not count on AI summaries to hide criticism; fix what customers complain about and check where it appears. 1. Search your brand with words such as “reviews”, “complaints” and “legit” in Google and in the main assistants. Note whether criticism appears, and whether it is in the opening or further down. 2. Treat review platforms as the source of your AI reputation. Fix the recurring complaints there, such as billing and support problems. 3. Protect your ratings. In the hotel test, ratings moved recommendations far more than replies to reviews did. 4. Publish a clear, factual page about who you are and how you handle complaints, so assistants have accurate material for that part of the answer. 5. Check more than once. AI answers vary from run to run, so one result is not your reputation. For help building this into regular monitoring, see our [generative engine optimization service](https://underneath.agency/services/generative-engine-optimization). ## What does the research not tell us yet? Nobody has tested whether AI Overviews downplay negative content about brands or products specifically. - The Illinois audit used general questions from an older collection, one location and automated tone scoring. - It shows a pattern, not a cause; negative pages may be under-used for reasons other than tone. - Our brand study covered Google AI Mode, not AI Overviews, with one answer per engine and brand. - The finding that no tracked brand was consistently described negatively comes from a vendor’s own customer data. - Whether moving criticism below the opening changes what buyers decide has not been measured. ## Frequently asked questions ### Will Google’s AI Overview show negative reviews of my company? It may cite them but use little of what they say. In one audit, AI Overviews drew 13.8 points less from negatively toned sources and 22.1 points less from forums, though brand searches were not tested. ### Does Google’s AI favor positive content? Only slightly, in the one audit that measured it. Positively toned sources were used 1.1 points more than average, while negatively toned ones were used 13.8 points less. ### Why does an AI Overview cite Reddit but not repeat what Reddit says? AI Overviews tend to list forum pages without drawing much from them. In the Illinois audit, social and forum sources were cited but under-used by 22.1 points in the summary text. Being cited is not the same as being used; see [pages cited for claims they do not make](https://underneath.agency/resources/ai-citing-pages-for-claims-they-do-not-make). ### Should I respond to negative reviews to improve AI recommendations? Respond for your customers’ sake, but do not expect replies alone to move AI picks. In a test of 12 AI models, management responses had no detectable effect, while a higher star rating added 31.6 points. ## Sources - Michelle Huang, Agam Goyal, Koustuv Saha and Eshwar Chandrasekharan (2026), [Answer Bubbles: Information Exposure in AI-Mediated Search](https://arxiv.org/abs/2603.16138), arXiv:2603.16138. - Haofei Xu, Umar Iqbal and Jacob M. Montgomery (2026), [Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact](https://arxiv.org/abs/2605.14021), arXiv:2605.14021. - Riley Grossman and colleagues (2026), [How Generative AI Disrupts Search: An Empirical Study of Google Search, Gemini, and AI Overviews](https://arxiv.org/abs/2604.27790), arXiv:2604.27790. - Pratyush Kumar (2026), [Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines](https://arxiv.org/abs/2606.20065), arXiv:2606.20065. - Mirza Samad Ahmed Baig, Syeda Anshrah Gillani and Asher Ali (2026), [Whose hotel does the AI recommend? An algorithm audit of reputation signals in LLM-assisted hotel selection](https://arxiv.org/abs/2606.16344), arXiv:2606.16344. - Underneath (2026), [“Is this brand legit?” How AI assistants build a reputation](https://underneath.agency/research/is-it-legit-ai-reputation-study) --- This is the Markdown twin of https://underneath.agency/resources/do-ai-overviews-downplay-negative-content. The HTML page is canonical. Publisher: Underneath, https://underneath.agency/agent. Site index: https://underneath.agency/llms.txt.