Google has said this publicly since about 2013 and reaffirmed it in 2017 and again recently in the context of AI search. In 2007 the company's figure was 20 to 25 percent, so the number has moved in the direction you would expect as the index matured.

It is self-reported, unaudited, and has not budged in thirteen years, which is a long time for a measured quantity to sit on a round number. Google has never published what counts as a query (normalization, spelling variants, voice, Lens) and the denominator is not disclosed. Howell's riff about 52 characters raised to the thirtieth power is fine arithmetic and does not bear on whether the 15 percent is measured or curated.

The order of magnitude is consistent with the other number Google has given out, which is that roughly 7 percent of daily searches are health-related. Against current search volume that arrives at several hundred million, so Howell's figure is internally coherent.

It does not reconcile with Google's own headline claim, made by a prior health executive in 2019 and repeated ever since, of more than one billion health questions a day. Both figures are still in circulation and they cannot both be right under the same definition. Neither has a published methodology, and "related to health in some way" is doing an unknown amount of work.

The study exists and the result is real. In the Med-PaLM 2 human evaluation, 15 physicians from the United States, the United Kingdom and India judged 1,066 consumer medical questions with answers blinded and order randomized. Model answers were preferred on eight of nine axes, including better alignment with medical consensus (72.9 percent) and lower likelihood of harm. Only "inclusion of irrelevant content" favored the physicians, and that difference was not significant.

Howell describes it as error rates. The study measured preference. Those are adjacent, not identical, and the authors flag three reasons the gap flatters the model. Physicians answered consumer questions cold, with no clinical scenario and no communication context. Each produced exactly one answer. Model answers were more than twice as long (median 794 characters against 337.5), which the authors themselves name as a likely driver of preference. This is a writing exercise, not care.

Google did build citations into AI Overviews from the start, and Howell is right that this was a decision made up front rather than bolted on. In May 2026 Google added five more link surfaces.

The number. Google has declined to release AI Overview click data for two years, Search Console still does not separate AI Overview clicks from ordinary search clicks, and publishers cannot A/B test the feature. The independent evidence on the volume question runs hard the other way. Pew examined 68,879 searches by 900 US adults in March 2025 and found users clicked a traditional result in 8 percent of visits when a summary was present against 15 percent when it was not, clicked a link inside the summary in 1 percent of visits, and ended the session entirely in 26 percent of visits against 16 percent. A pre-registered randomized experiment on 1,065 Chrome users in early 2026 found that switching the summaries off raised organic clicks 38 percent, dropped zero-click searches from 72 percent to 54 percent, and left satisfaction scores nearly unchanged.

One caveat in Howell's favor, and it matters. Fewer clicks does not logically refute greater variety among the clicks that remain. Diversity is a distribution claim and nobody outside Google has the data to test it. The problem is that this is true by construction: the claim is unfalsifiable from outside, was offered in answer to a question about whether people click at all, and sat next to two other assertions ("people like it", "people ask more questions") whose closest independent test found no satisfaction difference.

Supported, and any hospitalist reading this already believed it. Of 22,889 patients who read an ambulatory note, 21.1 percent reported a mistake and 42.3 percent of those judged it serious. Among the very serious ones, diagnosis (27.5 percent) and medical history (23.9 percent) led. Howell's example (a note advising continued care of the right toe on post-op day three from a below-knee amputation) is the copy-forward failure mode exactly.

That study measures patient-perceived mistakes in notes patients chose to read, not an adjudicated error rate across all documentation, so it is the right direction and the wrong instrument for a baseline. The larger issue is rhetorical. Howell answers a question about model safety by moving the comparator to human error, which is a fair and necessary point that also ends the conversation. Human and model errors are not distributed alike and do not fail alike at scale. One bad note harms one patient. One bad summary pattern reaches several hundred million questions a day.

Google does publish seriously in health, including in Nature, JAMA and Lancet, and the volume is plausibly in that range. Pointing an interviewer at the public list is the right instinct.

The page Howell names states no total. It mixes blog posts with peer-reviewed citations behind "Load More" buttons, so the figure cannot be checked from the source he offered. The comparative half of the claim is weaker still, attributed only to unnamed "groups that aggregate the impact of companies" on the literature. That is the claim doing the persuasive work and it is the one with no citation.