Methodology / v2.0
Google Trends measures queries surging relative to their own normal level. It is a view of accelerating attention, not an audited list of every search made on Google. Everything below follows from taking that distinction seriously.
Once each morning we request the United States trending window from the endpoint Google's own Trending Now page calls, covering the previous seven days. Each story arrives with an approximate volume, a growth figure, a start and end time, a category, and the related queries Google clustered into it. One request reconstructs a complete calendar day, so nothing depends on catching a moment.
Stories are assigned to the United States Central Time day in which they began trending, and ranked by the highest approximate volume bucket Google reported. Where the leading stories would otherwise all sit in one category, the next distinct category is promoted, because a ranking that reads as a single sport tells a reader less than the day actually contained.
For each ranked story we retrieve current news coverage for that query and keep the headlines and their publishers. Those headlines are the only permitted basis for any statement about a real-world event. Where coverage does not explain a surge, the edition says so rather than filling the gap.
Two readings are written. Each story gets what the subject is, what the coverage establishes, and what the search behaviour itself reveals, which is the part a news site does not provide. The day then gets an interpretation: how concentrated attention was, what mechanism produced that shape, and what has duration rather than having spiked once.
Before anything publishes, every figure in the prose is matched against the computed set, every copied query against the observations, and every causal claim against the cited headlines by an independent reviewer that only sees the evidence. Findings are sent back for revision, then for line-by-line repair. A claim the evidence cannot carry is cut rather than softened.
An edition publishes only when nothing blocking survives review. If a story cannot be described from available coverage, it is dropped and the omission recorded. If too little of the day survives, no edition publishes that day. The post announcing an edition is sent only after that edition is confirmed live, and only if it passed review, so nothing is announced that a reader cannot then read.
Language discipline
We say trending searches because the source identifies queries experiencing unusual growth. We never describe these numbers as exact totals, as market share, or as a census of Google searches. A category share on this site is a share of the combined volume of the stories that were trending, which is a small and unusual slice of all searching.
Volume labels such as 100K+ are Google’s approximate buckets and are lower bounds. Several stories often share one bucket, which means the ranking among them is not resolved by the data, and each edition says so where it applies.
Growth is capped by the source. A reported 1000% means at least 1000%, so it marks a ceiling rather than measuring a rate. Category labels come from Google and a few of them are unreliable, so they are treated as a hint rather than a fact.
Every edition carries the reading’s own caveat: what would have to be true for that day’s interpretation to be wrong. That line is written as part of the analysis, not added afterwards.
Read Google’s Trending Now documentation ↗No. They are the searches growing fastest relative to their own baseline. A query millions of people run every day at a steady rate does not trend, because trending measures acceleration rather than size.
Because a story was dropped. If a subject cannot be described from the coverage available without asserting something the coverage does not support, it is removed rather than guessed at, and the omission is recorded with the edition.
The reasoning is, and the checking is done by a second model that sees only the observations and the cited headlines. Nothing numeric is generated: every figure on the site is computed from the collected data, and any figure appearing in the prose that is not in that computed set blocks publication. The post announcing each edition is generated with the edition and checked by the same review, then sent only after the edition is confirmed live.
It splits the difference between American coasts, so a story that broke in the evening lands in the day readers associate it with.
Yes, with attribution. Each edition is also published as structured data on its page, including a Dataset description of that day’s observations.