The Background Check That Never Leaves the Chat Window

2026-08-31

Ninety-three percent of searches inside Google's AI Mode now end without a single click to another website. Zero-click behavior across all of Google search has climbed past 65 percent, and when an AI Overview appears on the results page, that figure jumps to roughly 83 percent. Gartner expects traditional search volume to fall by a quarter by the end of this year as conversational AI systems become the default way people look things up.

Read that as a search-industry statistic and it's interesting. Read it as a description of how due diligence now happens, and it's something else.

For decades, checking out a company, a counterparty, or a person meant clicking through: the corporate registry, the news archive, the court filing, the LinkedIn profile. The paper trail was inconvenient, but it was a trail. Someone could point to exactly which page said what, and when.

That trail is disappearing. A lawyer vetting an expert witness, an underwriter pricing a policy, an investor doing pre-call homework on a founder, a journalist fact-checking a source before publication — increasingly, all of them are typing a name into an AI system and taking the synthesized answer as the finding. No click-through. No second source. The answer that appears in the chat window is, for practical purposes, the due diligence.

This would be a manageable shift if AI systems reliably said the same thing, cited real sources, and held their answers steady over time. They don't. Ask the same question of several AI systems and you'll often get materially different characterizations of the same company or person — not because the underlying facts changed, but because each system was trained on a different slice of the internet, weighted differently, and tuned by a different company with different guardrails. A US-trained system and a Chinese-trained system asked about the same multinational can produce descriptions that a reader would struggle to recognize as being about the same entity.

Layer on the accuracy problem and the picture gets worse. Independent testing of AI search engines has found error rates in cited claims ranging from under 40 percent to over 90 percent, delivered with no visible hedging. The system doesn't flag which parts of its answer it's confident about. It states an invented lawsuit, a fabricated regulatory action, or an outdated leadership fact with the same tone it uses for something it got right. The reader has no way to tell the difference from inside the chat window, and there's no longer a click-through habit that would surface the underlying source and let them catch it.

The result is a background check with no receipt. If an AI system tells an underwriter that a company had an undisclosed safety incident, or tells a diligence team that a founder was named in litigation that never happened, that claim shapes a real decision — a premium, a pass on a deal, a declined introduction. But the claim itself typically isn't saved anywhere. It isn't dated. Nobody wrote down which system said it, in what words, sourced from what. Six months later, when the decision is questioned, there is nothing to point to except someone's memory of what the screen showed.

This is a different problem from the one most reputation and legal teams are set up to handle. Traditional reputation management assumes a fixed, findable artifact: an article, a review, a post. You can screenshot it, archive it, dispute it, get it corrected or removed, and show your work. An AI-generated answer isn't fixed. It can change between two queries typed a minute apart, it rarely lists a URL you can visit, and by the time anyone thinks to check whether it said something damaging, the system may have already produced a different answer to the same question.

None of this means AI-assisted research is going away, and it shouldn't. It's faster and, on plenty of questions, no less reliable than a harried analyst skimming the first page of search results. But treating an AI system's answer as equivalent to a verified source, simply because it arrived instantly and in fluent prose, is a category error that due diligence teams are making at scale without realizing it.

The organizations that will handle this well are the ones that stop treating "what the AI said" as ephemeral. If a synthesized answer is doing the work that a verified document used to do, it deserves the same discipline: know what was said, know where it plausibly came from, and have a timestamped, unaltered record of it before it quietly changes into something else. The zero-click era didn't remove the need for a paper trail. It just moved the trail somewhere nobody thought to keep it.