What is the Next Big Pulse?
The Next Big Pulse is a weekly reading of how much interest there is in a search, on a scale of 0 to 100 where 100 is its own busiest week. Instead of trusting one company's counter, it blends every source we can read week by week: searches on Google and Bing, reading on Wikipedia and coverage in the news.
Every trend page on Next Big charts it, and its growth over 3 months, 1 year and 2 years and its projection are read off it. It's also read in monthly searches, on Google Ads' scale, so two searches can be compared in size.
Which sources go into it?
One source is one activity, measured by one provider. A search's line blends the ones it has:
- Google web search, from Google Trends: the anchor, corrected for the weeks Google changed how it counts and for bursts of searches from somewhere else.
- YouTube, Google News and Google Shopping searches, also from Google Trends, refreshed weekly for every search we track (Shopping for products).
- Bing searches in the United States, from Microsoft Advertising. Its licence lets us name it, not show it: Bing only moves the line.
- Wikipedia page views, from Wikimedia, when the article is the thing itself, never a broader one.
- News coverage, from GDELT, for searches that are in the news.
- DataForSEO Trends, a separate popularity index built from web pages and anonymous browsing data rather than Google's counter.
- Web mentions, from DataForSEO: how many pages cited the phrase each week, divided by the mentions of the same 24 everyday phrases, because the crawl itself keeps growing.
- Mastodon posts, once a tag has half a year of history.
Google Ads volume isn't a source: it counts the same Google searches again. Yearly counts (Hacker News, Bluesky, The New York Times, The Guardian, research papers) can't shape a weekly line, so they stay as confirmation on each trend's page. A source joins once it has half a year of weeks, few of them empty.
How is it calculated?
Six steps, the same for every search, every night:
- Clean each series. Google's counting changes are corrected. A single week or month several times above both its neighbours is set back to their level, and gaps of up to four weeks are bridged.
- Size each count against its platform. Wikipedia views are divided by all of English Wikipedia's views that week; Bing searches and web mentions by a basket of 24 everyday phrases (weather, maps, recipes, flights and others). Google Trends and GDELT are already shares.
- Take logs, so a doubling counts the same at any size. A started log, log(x + c) with c a twentieth of the series' average, lets small searches move at their real size.
- Weight each source by how steady it is. Its noise is measured around its own trend, and its weight is 1 / (noise² + τ²), the weight a meta-analysis gives each study. τ, now 0.05, keeps one quiet source from taking over.
- Chain the weeks. The line starts at Google's level, and each week it moves by the weighted average change of the sources present that week and the one before, the way a price index chains its basket. A source joins at the line's level, so one arriving or ending never makes it jump.
- Scale to 100 at the line's own peak.
How sure is each week?
Every stretch of the line carries its own uncertainty, from how many sources stand behind it and how noisy they are, and the chart shows it as a band. Each search gets a label too: high with three or more steady sources, low when Google stands alone or the sources are noisy, medium between. Where Wikipedia is in a line, the week after the line's last is read early from its daily views (against the same weekdays a week before, so the weekend dip cancels) and shown as provisional until every source has that week. Tonight: 174 high, 823 medium and 474 low.
Why divide by the platform?
Because platforms grow and shrink on their own. English Wikipedia's views changed -8% over the last year as answers moved into search and chat, and Bing's everyday searches moved too. Without sizing, every article and every Bing search would read as fading interest. Google Trends already reports a share of all searches, so dividing puts every source on the same footing.
How do we know it works?
There's no true count of interest to check against. So every night we hide one source, rebuild the line from the others, and see how well it predicts what the hidden source did over the last 26 weeks: the sensitivity test the OECD handbook on composite indicators asks for. A line that reads interest predicts a source it never saw; a line that copies Google doesn't.
| Method | Typical error | Closer than Google |
|---|---|---|
| Google web search alone | 0.343 | |
| Equal votes, log(1 + x) | 0.293 | 63% |
| Equal votes, started log | 0.290 | 64% |
| Precision weights, τ 0.05, in use | 0.270 | 68% |
| Precision weights, τ 0.1 | 0.272 | 66% |
| Precision weights, τ 0.15 | 0.275 | 65% |
| Precision weights, τ 0.25 | 0.275 | 65% |
| Precision weights, τ 0.4 | 0.276 | 65% |
Typical error is the median gap, in log units, between the predicted and the actual change of 462 hidden sources in 184 searches; closer than Google is the share of them each method predicted better than Google alone did. Some sources are much harder to predict than others (web mentions swing most), so the share is the fairer comparison. The method in use is the one that predicted best, and it only changes for a gain of 2% or more on at least 40 hidden sources.
We also compare with Google Ads search volume, a count of Google searches we never feed in. Across 317 searches, the Pulse came closer to its change than Google Trends did: a typical error of 0.383 against 0.519, and the right direction 72% of the time against 64%.
How did its projection do?
Every week we save each search's Pulse and where its projection says it will be in 8, 13, 26 and 52 weeks, and never change a saved one. As those weeks pass, each projection is scored against what happened, beside Google's own projection. The record started on September 13, 2026 with 1,471 searches; its first scores, 8 weeks ahead (early and noisy), go up here in December 2026, and 13 weeks ahead in January 2027.
When does it disagree with Google?
When Google says a search rose over the year and the Pulse says it didn't, 20 points apart or more, we flag it. Tonight 6 of the 36 searches we publish carry the flag. A search only counts as trending when the Pulse is up over the year too, so a rise Google saw alone isn't called a trend.
How is the projection made?
A model trained on 1,000 keywords chosen without looking at their growth: 500 names from Google's own category tree, and 500 longer searches drawn from Google Ads' ideas for them, spread evenly from 30 to 10,000 searches a month. The library can't teach a projection: every search in it was found because it rose, so its history already knows the ending. The model reads the last 4, 13 and 52 weeks, distance from the peak, noise and last year's pattern, and fits all its horizons together so they make one path; a search with only a year of history gets a model that reads a year. It gives the odds of being higher in 6 months and a range eight in ten outcomes fell in, taken from errors on weeks it never saw by searches like it (as steady, and moving the same way).
| Ahead | Typical error | Right direction | Range held | Old projection |
|---|---|---|---|---|
| 4 weeks | 0.094 | 57% | 85% | 0.086, 60% |
| 8 weeks, in use | 0.119 | 62% | 82% | 0.140, 57% |
| 13 weeks, in use | 0.147 | 63% | 80% | 0.172, 57% |
| 26 weeks, in use | 0.188 | 62% | 78% | 0.218, 60% |
| 39 weeks | 0.214 | 63% | 74% | 0.243, 68% |
| 52 weeks, in use | 0.247 | 66% | 59% | 0.269, 58% |
Scored on the panel's last 26 weeks, never used in fitting; the old projection carried the last half year's slope forward. A horizon uses the new model only while it beats the old one on error without losing on direction. Range held is how often the range, meant to hold eight in ten, held the outcome; where it held under seven in ten, the page shows the middle only, with no odds or range. It's refit every week on 926 panel keywords, with the library added at half weight. Where a horizon or a short line didn't pass its test, it gets no projection rather than a guess.
Does it match what people did?
Sources agreeing with each other isn't proof. So the Pulse is also checked against outcomes it never reads: the Google Ads volume figures above, web traffic for sites and brands, and for products, their place on Amazon's best-seller lists and their reviews a day. Against web traffic (a site's Tranco rank, which averages five traffic and DNS providers) over 277 three-month stretches for 40 sites, the Pulse matched the direction 57% of the time and Google alone 61%: here the Pulse is no better than Google, and we show it anyway. The shelf check starts scoring once products have eight weeks of readings.
Can I use the data?
Yes. Every published search's Pulse, week by week, is open data under CC BY 4.0: download pulse.csv and credit “Next Big Pulse, nextbig.io”. Each search also has a chart you can embed, with the credit in the image:
<a href="https://nextbig.io/trend/ai-agent"><img src="https://nextbig.io/embed/ai-agent.svg" alt="ai agent: Next Big Pulse" width="600" height="300"></a>No source's own numbers are in the file or the charts: Bing's licence lets us name it, not show it.
Can I check nothing was changed later?
Yes. Each Monday we write a file of every published search's Pulse and its projection as they stood that week, and never write that week again. Its SHA-256 fingerprint is listed here, so a copy you keep can be checked against it: any later edit would change the fingerprint.
- Week of October 12, 2026, 36 searches:
7181e967c93558d56218c1a7c71eee1b0a12842cb102af48ffbcc75f20aa3f67
What changed, and when?
Every change to how the Pulse is made is listed here, and every saved week records the version that made it.
- 2.3, October 13, 2026: Honest newest weeks, one projection path, a long-tail panel. Confidence is graded on the last 13 weeks, and the newest weeks that rest on fewer sources (a monthly source not in yet) are named. Bing is read as soon as a month closes. The projection's horizons are fitted together, each range comes from lines like it, a zigzag the season doesn't explain is straightened, and lines with one year of history get a one-year model. The test panel gained 500 long-tail keywords. Published searches get a provisional week from Google's daily figures too. Why: Most lines read low because their last few weeks waited for Bing's month. A projection that fell at six months and recovered at a year had no reason to. Head terms alone said little about the library's long tail.
- 2.2, October 13, 2026: Projection v2, a test panel, volume and outcome checks. The projection became a model trained on 500 keywords chosen without looking at growth, giving odds and a range. It replaced the damped trend at each horizon only after beating it on weeks it never saw. The Pulse is also read in monthly searches, and checked against Amazon's best-seller shelf. A provisional week and Wikipedia in eight more languages were added. Why: The damped trend did worse than predicting no change, and the library alone can't test a projection: everything in it was found because it rose.
- 2.1, October 13, 2026: More sources, confidence and a track record. Added Google's YouTube, News and Shopping lines, DataForSEO Trends and web mentions for every search, within a fixed budget. Each stretch of the line now has a band and each search a confidence label. Every week's Pulse and projection is saved, never changed. Why: Most lines rested on Google and Bing alone.
- 2.0, October 13, 2026: Next Big Pulse. Wikipedia views and Bing searches are now divided by their platform's size. Each source is weighted by how steady it is, logs are started at a twentieth of the average, and a nightly check hides each source in turn and tunes the weights. Why: Platforms grow and shrink on their own; equal votes let a noisy source count as much as a steady one.
- 1.0, October 13, 2026: The interest index. Google web search, Bing, Wikipedia, Google's other search lines and the news are blended into one weekly line, each source chained in from the week it joins. Why: Every chart showed Google alone.
What doesn't it do?
It doesn't count people: 100 is a search's own busiest week, and its monthly searches are an estimate on Google's scale, not a count across every engine. A projection is odds, not a promise. And for a new or niche search the line may rest on two sources: each trend's page names the ones in its line and says how sure it is.
Further reading
The method draws on these, and the full list of sources is on the methodology page:
- OECD and the European Commission's Joint Research Centre, Handbook on Constructing Composite Indicators: Methodology and User Guide (2008): how a composite is built, weighted and tested for sensitivity.
- R. DerSimonian and N. Laird, Controlled Clinical Trials 7(3), Meta-analysis in clinical trials (1986): random-effects weights, 1 / (noise² + τ²).
- F. R. Hampel, Journal of the American Statistical Association 69, The influence curve and its role in robust estimation (1974): the median absolute deviation, and setting single outliers aside.
- F. Mosteller and J. W. Tukey, Data Analysis and Regression (1977): the started log, log(x + c).
- IMF, ILO, OECD, Eurostat, UNECE and the World Bank, Consumer Price Index Manual: Concepts and Methods (2020): chain-linking an index whose basket changes.
- H. Choi and H. Varian, Economic Record 88, Predicting the Present with Google Trends (2012): search interest as a reading of what people do.
- M. Mestyán, T. Yasseri and J. Kertész, PLoS ONE 8(8), Early prediction of movie box office success based on Wikipedia activity big data (2013): Wikipedia views as a reading of interest.