2,202 US adults were asked about AI search. Sixty-five percent use it, fifteen percent trust it a lot, and seventy-six percent say seeing where the information came from matters.

Anthropic described the mechanism behind the watermark it will apply to Claude's text output for EU AI Act compliance. It is not hidden characters or metadata. The pattern is created during generation, by using a watermark key plus the preceding context to bias which word gets selected where the model would otherwise choose at random. Detection requires the key. Anthropic states that light editing probably will not remove it, and that a complete rewrite in which every word is replaced will. Detection is weak on short samples and weak on factual writing, where there are fewer word choices to encode a pattern. The approach is built on Google DeepMind's SynthID-Text method from 2024.

2,202 US adults aged 18 and over, fielded 26 to 28 February 2026. Sixty-five percent had used an AI-powered search tool in the previous six months, and 15% said they trust the information those platforms give them "a lot". Sixty-three percent double-check AI results against other sources they trust. On what would fix it: 72% say platforms should always show where the information came from, 76% say seeing the sources matters, 66% want links as proof, and 69% want the option to leave the AI and visit a trusted site. Fifty-one percent describe AI search as a walled garden. On local specifically, 57% use AI tools to find local businesses at least monthly, 73% say customer ratings and reviews matter in those answers, and 65% say being able to act, such as booking, matters. Gen Z uses it more and trusts it no more: 84% have used it, 60% use it monthly for local business discovery, 72% demand proof of sources.

Judge it honestly. This is Yelp's own commissioned research and Yelp sells the thing the answer implies, which is a trusted third-party source with reviews and links. Morning Consult is a real polling firm and the sample size is stated, which puts it well above most of what this category publishes. It is older than our normal window, from April, and we had not logged it. It measures what people say about trust, not what they do, and stated trust is generally lower than revealed trust. The pattern it describes is a demand for verification, and it is a survey.

First-party platform data across 1,584 Demandbase customer instances and more than 11 billion website visits, from 1 June 2025 to 31 July 2026. Monthly ChatGPT referral visits rose from about 645,000 in June 2025 to 2.6 million in June 2026, up 303%, with a sharp inflection in May 2026 when the figure more than doubled against prior months. ChatGPT is nearly all of it: Perplexity referrals declined, Gemini and Claude were flat. The number that keeps it honest is in the coverage itself: the whole AI referral category is roughly one third of one percent of the traffic Demandbase measures. The same release put connected TV spend up 112% and display up 55% across the same window, on more than $123 million of media across 75,645 campaigns.

Judge it honestly. First-party vendor data, which is a genuine strength on volume and a real weakness on independence. Demandbase sells brand advertising to B2B marketers, and the framing that AI discovery is creating blind spots which brand spend should fill is the conclusion that sells its product. It is enterprise B2B software buying, which is not a coach with a booking page. The release carries no conversion data at all. Coverage quotes conversion figures from Invoca and from Ahrefs which we have not read at source, so do not repeat those.