AI Overviews surface negative reviews users never asked for

What happened

Per Search Engine Journal, Google's AI Overviews are now surfacing negative reviews and reputational complaints inside answers to queries that have nothing to do with reviews. A user searches for a company's services, hours, or capabilities. The AI Overview returns a synthesis that includes complaint-site language, lawsuit references, or critical Reddit threads pulled in as "context."

The mechanism is retrieval, not intent. The model decides what counts as relevant background on a brand, then injects it. The user never asked. They got it anyway.

This breaks a long-standing assumption in reputation management: that negative content only hurts when someone goes looking for it. In an AI Overview world, the model goes looking on the user's behalf, and the bar for what gets included is set by the model's notion of "useful context," not the searcher's question.

Why it matters for your brand

For B2B buyers, this is a procurement problem disguised as a search problem. A treasurer at a development bank running a query about a custodian's settlement capabilities does not expect a paragraph about a 2019 enforcement action. They will get one anyway if the model has indexed it and decided it qualifies as material context. The same logic applies to a UN procurement officer querying a logistics provider, or a CFO at an industrial group checking a software vendor before a renewal conversation.

Financial services brands are the most exposed. Regulatory actions, class actions, and consumer complaint databases are heavily indexed, well-structured, and trusted by LLMs as authoritative. A retail banking complaint can surface inside an Overview answering an institutional question, because the model does not always distinguish between the consumer-facing arm and the corporate one. The brand pays for the confusion.

Multilaterals and policy institutions face a different version of the same problem: critical NGO reports, archived investigative journalism, and partisan think-tank commentary all rank as "context" in the model's eyes. A query about a UN agency's mandate can return a synthesis seeded with criticism from a single 2021 op-ed, presented with the same visual weight as the agency's own description of its work. There is no editorial referee.

For content strategy, the implication is direct. Brand-owned content needs to do two jobs at once: answer the literal question, and pre-empt the reputational adjacencies the model is likely to pull in. If your "About" page, your investor relations materials, or your governance documentation does not address the criticisms that exist about you in the open web, the model will fill the gap with whoever does. Silence is not neutrality; it is a vacancy the model fills with the loudest available voice.

Distribution changes too. Earned media on reputable outlets is now a defensive asset, not just an offensive one. A well-sourced Reuters or FT piece that contextualises an old controversy gives the model a counterweight to pull from. Without that counterweight, the model defaults to the complaint site, the Reddit thread, or the plaintiff's law firm landing page, because those are what exists.

The signal in context

The shift here is from "search reputation" to "synthesis reputation." For two decades, brand teams managed Google's first page: push down the bad links, push up the owned and earned ones. That playbook assumed the user would scan a list and exercise judgment. AI Overviews remove the list and the judgment. The model decides what is true about you, blends it into prose, and presents it without source hierarchy visible to the reader. A complaint board and a regulator's official statement can sit in the same sentence.

This connects to a broader pattern across ChatGPT, Perplexity, and Google's AI features: models are increasingly willing to volunteer reputational context unprompted, and they draw from a wider set of sources than classical search results favoured. The brands that will hold visibility in this environment are the ones treating their corpus, owned, earned, and third-party, as training data for how they want to be described, not just as pages to be ranked. The work is no longer SEO. It is building a defensible factual record dense enough that the model has no incentive to wander.