Methodology

What this site holds

GINC Trends holds the slow structural forces that bear on national capability: 250 trends, each assigned to the one domain of the National Capability Framework it bears on most. It is one of three bodies of work that fit together.

A trend is not a scenario. A scenario is an event with a date; a trend is what is changing underneath, slowly, and raises or lowers the odds of many events at once. This site says three things about each trend: what it relates to, where it sits among the others, and how much notice it is drawing.

The list

The list was built in three desk passes and then put to a model panel in a second pass, which is the list published now. It began from published trend and risk lists used by investors, governments and international affairs institutes, and from the trends the scenarios themselves depend on. It was then reviewed as a whole: trends too narrow to bear on national capability were cut, near-duplicates were merged, broad trends were broken out into the parts the scenarios turn on, and trends were added where a capability domain was thin. The result was GINC’s own list, which the panel then revised; individual trends no longer carry a source. The lists drawn on are named in the change log.

• Capability domain. Every trend has one primary domain. The domains hold between 27 and 29 trends each.
• Direction. A trend’s name states a direction of change where it can: ‘Entrenched opportunity inequality’, not ‘Inequality’. Single theatres and single nations are left out, on the ground that they are situations and not structural forces.
• Permanence. A trend’s address does not change. A trend merged into another or dropped in a revision is recorded in the change log.

Relations

• Trend by scenario. Every trend is scored against every scenario on a seven-point rubric: not related, very low, low, moderate, high, very high, critical. A trend relates to a scenario when it changes how likely the scenario is, how hard it lands, or is itself sharply changed by it. Each score is the median of three scores given separately by the models of the panel. The scenarios are those of the Scenario Library; this site keeps a copy of each one’s number, name and family, and the matrix shows every score.
• Trend by trend. Similarity between two trends is derived, not scored by hand: 0.30 from shared tags, 0.30 from how alike their scenario profiles are, 0.30 from how alike their names and descriptions are, and 0.10 if they share a capability domain. The result is cut into the same seven levels. Critical similarity is rare by design and marks parent-and-child or sibling trends.
• Sensitivity. A trend’s sensitivity is the sum of its scenario scores as a share of the maximum. Breadth is the number of scenarios it scores High or above. Trends with high sensitivity move the most scenarios.

The network

• Themes. The ten themes are the clusters of the trend network, found by the Louvain method: groups of trends more closely linked to each other than to the rest. The method finds the groups; GINC names them. They were found again, and renamed, after the second pass.
• Network measures. Trends are linked where similarity is Moderate or above. PageRank restarts at each trend in proportion to its sensitivity, so rank flows from scenario-relevant trends along similarity links. Betweenness marks trends that bridge clusters. Eigenvector centrality gathers in the densest cluster and is reported with that caveat.
• Rank. The composite is 0.4 sensitivity, 0.3 PageRank, 0.2 betweenness and 0.1 eigenvector, each scaled so its highest trend is 100.
• Two ways to draw the network. Defined clusters arranges the picture to show the ten themes. Organic gives the layout no themes at all: every trend keeps its one strongest link, the strongest 5, 10, 20 or 50 per cent of all links are added, and the separate pieces this leaves are joined by the strongest link between them. Clusters are then found from those links alone and named after their best-connected trend. At 20 per cent, 215 of the 250 trends sit in an organic cluster dominated by their own theme, which is a check on the themes and not a restatement of them.
• The network picture. The network view is drawn for reading, not for measuring. Each trend is placed using its three strongest links, with the other links inside its theme holding the theme together; the themes are then set further apart than the forces alone would put them. Nearness means likeness. The size of the gaps between themes does not carry meaning, and none of the measures above is taken from the picture.

The similarity, the themes and the measures are computed from the published list and scores by a script in the repository, and can be reproduced from the downloads on the downloads page.

The second pass

The first version of the list was a GINC desk build: the trends and every score were drafted in one hand. In October 2026 the list was put through a second pass, to test it against other judgement before the expert panel sits. Three AI models from three developers were used as a stand-in panel, and a fourth made the final cut to GINC’s brief. The list and the scores on the site now come from that pass. The record, with each decision, is on the downloads page.

How it was run

The same three steps were used for the trends here and for the scenarios at scenarios.ginc.org.

1. Separately. Each model was given the first-pass material and asked to improve it, without sight of the others.
2. Together. Each model was shown all three results and asked to settle on a final version, taking the best of each and not simply returning its own.
3. Final cut. The three final versions were put side by side and a fourth model made one list from them, weighing how many models held each item but deciding on merit, and writing each entry afresh.

Agreement is reported throughout as support: the number of the three models whose final version held the item.

The trends

Each model was asked to revise the 250 and to balance them across the nine capability domains, with 24 to 32 trends in each. The three final lists were cut one domain at a time, and two overlaps between domains were then settled by hand.

The panel’s main changes were of three kinds. It renamed trends so that the name states a direction of change: ‘Inequality’ became ‘Entrenched opportunity inequality’. It removed single theatres and single nations, on the ground that they are situations and not structural forces: the Taiwan Strait, Russia and NATO, Middle East instability and China’s slowdown are gone, and their substance sits in trends such as contested maritime access and great-power competition. And it added forces the first pass lacked, most of them in national security, finance and technology: undersea military competition, state cyber intrusion, orbital debris, AI inference efficiency, domestic capital market deepening and agricultural soil degradation among them. The full account is on the trends page.

The domains are now within two trends of each other, from 27 to 29. In the first pass they ran from 25 to 34.

The scores

Every trend was scored against every scenario by each of the three models on the seven-point rubric: 250 trends by 20 scenarios, 5,000 pairs each. The score kept is the median of the three, the middle judgement, which no single model can set.

Most of the agreement is agreement that a trend and a scenario are unrelated: 74 per cent of the final scores are ‘not related’, against 73 per cent in the first pass. The scale is used more sparingly at the top than the desk used it: 0.7 per cent of pairs are Critical, against 2.1 per cent before. The 231 pairs on which two models were far apart are listed in the record and are the first candidates for review.

What to make of it

A panel of models is not a panel of experts, and this pass does not replace a structured expert elicitation. It does three useful things. It shows where the first version rested on one view: the items that changed. It gives each trend a measure of agreement. And it exercises the whole pipeline, from list to network, on material GINC did not write.

Its weaknesses are stated plainly.

• The models share sources. Three models trained on overlapping text agree more readily than three independent experts would. Support of three is weaker evidence than it looks.
• One model could not hold a count. Asked for exactly 250 trends, it returned more or fewer through several corrections. Its list was settled by asking it to choose into numbered places, one domain at a time. Its judgement of which trends to keep is its own; the count was enforced.
• Most scores are noughts. Three-quarters of the trend by scenario pairs are scored ‘not related’. Agreement on those says little; the pairs to review are the 231 on which two models were far apart.

Likelihood and impact

The Outlook page places every trend by two estimates, each out of 100. They are provisional. Each is a weighted average of three readings made separately against the rubric below, each model scoring every trend from its name, capability domain and description without seeing the others: Claude Fable 5.1 counts for half, and a first estimate by one AI model and a reading by DeepSeek for a quarter each. Fable and the first estimate agree closely on the order (a correlation of 0.92 for likelihood and 0.89 for impact); DeepSeek less so (0.79 and 0.72), which is part of why it is kept. Combining the readings breaks most of the ties a single reading leaves, so that 211 trends have a place of their own on the chart. They are a first reading for the picture, not a forecast. They are not used in the ranking, the network or the attention data, and they will be replaced by scores from the model panel and then the structured expert elicitation. The weighted score, each reading and the reason each gave are in the downloads.

Likelihood is the chance that the trend continues as described and is clearly material at national scale by 2035.

Impact is how far the trend would change national capability if it played out: how many nations, across how many domains and how deeply.

Because the list holds structural forces already under way, most trends score high on likelihood: the median is 80, and only five sit below 50. Impact runs from 37 to 92 with a median of 60. Impact is only loosely related to sensitivity to the scenarios (a correlation of 0.26): sensitivity measures how strongly a trend bears on 20 particular shocks, impact how far it changes capability on its own.

GINC Share of Attention

The attention score says whether a trend is gaining notice against its own past. It cannot say whether one trend draws more notice than another. GINC Share of Attention does: each week, each trend’s share of all the public attention the trends in the list draw. The 250 add up to 100%, so a trend’s share can be read against the others’ and against its own past; a share that rises is gaining on the field, whatever happens to the platforms’ traffic as a whole.

It is built source by source, and the two are combined.

• Wikipedia. In each language, a trend’s views (every page that stands for it, as counted for the attention score) as a share of the views of all the trends in that language that week. The languages are combined weighted by the square root of their readers, so that English, read far more than any other, does not decide it alone.
• Google. How many times a month each trend’s search terms are searched, from Google Ads keyword data, as a share of the searches for all the trends. Google Trends cannot give this, since it reports each query against its own busiest week; the Ads counts are absolute. They are monthly, so each month’s count is spread across its weeks by the trend’s Google Trends series in the same place, which says in which weeks it was searched more. The weeks since the last full month take that month’s count, with their own Trends timing. Search terms whose searches are mostly about something else (a service, a brand, another meaning of the words) were reviewed before counting, and replaced or dropped.

The two shares are combined by a geometric mean, each with half the weight, and scaled so the trends add up to 100%. A trend is large only where both sources say so: neither a burst of reading nor a heavily searched term can make it large alone. A trend one source does not read at all is given a tenth of that source’s median share, so it is kept small rather than dropped.

By place

The same share is read for regions and nations, from the Wikipedia languages mostly read in each and the Google countries it is made of, their searches added together: Japanese reading and searching in Japan for Japan; the languages of Europe other than English, and ten European countries’ searches, for Europe; and so on. Wikipedia is used only where a place’s languages are its own. English, read across North America, Britain, India and Oceania alike, cannot be split among them, so the United States, the United Kingdom, North America and Oceania rest on Google alone. Where a place’s languages reach fewer than half the trends (Africa, India, Indonesia, Türkiye), Google stands alone there too. Russia rests on Russian reading alone, since Google’s figures for Russia cannot be bought, and China cannot be read at all: Wikipedia and Google are both blocked there. Each trend’s page lists every place and says what it is read from.

On the home page the tiles can be shown for Asia Pacific, Europe, the Americas and Africa. There, Top is the place’s share of attention, and Rising, Declining and Volatile are read from the place’s attention index, which follows Google searching in its countries and Wikipedia reading in its languages, each trend against its own past; Asia Pacific and the Americas average two regional indices each.

The share is shown on the home page (Top), on each trend’s page and on the Outlook, and the weekly figures are in the downloads. Nothing in the ranking, the network or the attention score uses it.

Trend attention

Each trend carries a weekly score of public attention over three years. It is built from two open sources: what people read on Wikipedia and what they search for on Google. It is published on the Index and on each trend’s own page.

The score measures attention, which is not the same as the strength of a force. A trend can deepen for years with little notice, and a single news event can lift attention for a fortnight. The score is offered as a proxy and a prompt: a trend whose attention is climbing is one to look at again, and one that draws less notice each year while its scenario scores stay high may be under-watched.

What is counted

• Wikipedia pages. Each trend is mapped to between four and eight English Wikipedia pages, each with a weight that says how central it is to the trend. The weights sum to one and no page carries more than 0.6. Pages are chosen to be specific: a trend about advanced chips counts the page on export controls on semiconductors, not the page on a country. Each page is followed to its counterpart in the other languages through Wikipedia’s own language links, so the count covers 28 languages: English, Chinese, Spanish, Arabic, French, Russian, Portuguese, German, Japanese, Hindi, Indonesian and Korean, and sixteen more added so that regions can be read through them, among them Italian, Turkish, Persian, Hebrew, Vietnamese, Thai, Bengali, Polish and Ukrainian.
• Google search terms. Each trend has two to five search terms, short and plain, which are put to Google Trends as alternatives. One query returns interest in any of them, worldwide.

The mapping was drafted one capability domain at a time and every page title was checked against Wikipedia. The pages and terms counted for a trend are listed on its page and in the downloads.

The basket

Both sources move for reasons that have nothing to do with any trend. Reading of the typical Wikipedia page in this set fell by more than a third over the three years, far faster than Wikipedia’s readership as a whole, which is consistent with readers taking answers from search summaries and AI assistants instead. Google’s index moves in steps that unrelated queries share: in the queries harvested so far, interest in subjects as far apart as fusion, gene editing and space debris rises together in the middle of 2025 and falls together a year later, which points to a change in how searches are counted and not to the subjects. Left alone, the first would make every trend appear to fade and the second would make every trend appear to surge.

So each source is set against its own basket: the typical trend that week, taken as the median of the whole set. Dividing by the basket removes what is common to all and leaves what is particular to each. The score therefore shows whether a trend is gaining or losing attention against the rest of the list. A rise in attention to every trend at once would not show, and that is the price of the correction. The baskets are published with the data.

How the Wikipedia index is built

1. Weekly human page views are taken for every page in every language over 156 weeks, Sunday to Saturday. Wikimedia removes spiders and automated traffic.
2. Each page is set against its own average, so a page read a million times a week counts no more than one read a thousand times.
3. Each page is then divided by the basket for its language: the median page of the set that week. Where a language has fewer than 60 usable pages, all reading in that language is used instead.
4. A page’s week is capped at eight times its average, so that one page in one week cannot carry a trend.
5. A page created inside the window counts only from its first week of reading, and only if that leaves 26 weeks. Otherwise a new page would look like attention rising from nothing. Where an English page was renamed, its earlier readers are recovered by counting its redirects.
6. A page is used in a language only if it averages 20 readers a week there. A language is reported for a trend only if the pages read from the first week carry at least 0.3 of the weight.
7. Within a language the pages are combined by their weights. Across languages the indices are combined by the square root of each language’s readers of the trend, so that English, which holds most of the reading, leads the result without drowning the rest.

The index is stored for each language as well as for all languages together.

How the Google index is built

Google reports search interest as a relative index: 100 is the busiest week of the window for that query, and every other week is a share of it. The series is worldwide and weekly. It is set against its own three-year average and then divided by the basket, the median trend that week.

Google also reports how interest is spread across countries over the window, as a share of each country’s own searching, with the highest place set to 100. That spread is stored for every country Google reports, including those it marks as low volume, and shown for the nations GINC covers, rescaled so the highest nation is 100. The search terms are English, so the spread favours countries that search in English. It shows where a trend’s English-language attention sits; it is not a measure of how much a nation cares.

Google turns away anyone who asks its Trends site for more than a few queries an hour, and its official interface is open only by application. The figures are therefore bought from a commercial data service that supplies them query by query; it is named in the data files. A test of the same queries asked both ways gave the same series.

The score

The attention score is the plain average of the two indices, rescaled so that the trend’s three-year average is 100. Where one source has nothing to report, the score rests on the other and says so.

• Now is the average of the latest 13 weeks.

• Change over a year compares the latest 13 weeks with the same 13 weeks a year earlier, which removes the season. It is put in words: surging above 40 per cent, rising above 12, steady within 12 either way, easing below that and fading below a fall of 30.

• Change over three years compares the latest 52 weeks with the first 52.

• Agreement is the correlation between the Wikipedia and Google lines for the trend. Where it is low the two sources are reading different things, and the score should be treated with care.

• Heatmap periods. The network can be coloured by the change over 90 days, 180 days, one, two or three years. Each compares the latest 13 weeks with the 13 weeks that ended that long before; three years reaches back to the first 13 weeks of the record. Colour is placed by ratio, so a doubling and a halving are equally strong, and runs to full strength at the ninth decile of the changes in view.

The line drawn on the site is a trailing four-week average. The weekly figures are in the downloads.

Attention by place

The site also shows attention to each trend inside each region and each member of the G20. Two sources are used and averaged where both exist, as for the world.

Google searching, country by country. Each trend’s weekly search interest is taken inside each of 48 countries: every G20 member except Russia, and enough others to stand for each region. In a country that searches mostly in another language, the query adds the names of the trend’s main Wikipedia pages in that language to the English terms, so that it is not left to English alone.

1. A trend is read in a country only if it was searched for in at least half of the 156 weeks. Weeks with too little searching to report count as nought.
2. Each series is set against its own average and divided by the country’s own basket, the median trend there that week. A country with fewer than 60 trends read uses the world’s basket.
3. A region’s Google line combines its countries in the set, each weighted by the square root of its internet users.

Wikipedia reading, by language. No open source counts what each country reads about each subject: Wikimedia’s figures by country are published only for pages read about ninety times a day in a country, which leaves out most of the pages counted here. So a place is read through the languages that are mostly read there.

1. Wikimedia publishes, month by month, how many page views of each language edition come from each country. Twelve months are averaged, to September 2026.
2. A language counts for the region that holds the largest share of its readers, if that share is at least 40 per cent, and for a G20 member on the same test. German stands for Germany, Japanese for Japan, Portuguese for Brazil.
3. A member that shares its language has no Wikipedia line of its own. English is read in the United States, the United Kingdom, Canada, Australia and South Africa, and Spanish in Mexico and Argentina; only the United States, which holds the largest share of English, is given the English line, marked weak. The others rest on Google.
4. Where readers’ locations cannot be trusted, a language is placed by where it is spoken and marked as such. Most Russian, Turkish, Persian and Vietnamese reading is counted in other countries, which points to networks abroad, and Arabic is spread across two regions.
5. A place’s Wikipedia line combines its languages, each weighted by the square root of the readers it brings.

The place’s score is the average of its Google line and its Wikipedia line, rescaled so its three-year average is 100. Its change is measured over the same periods as the world score. Every place is listed on the Index with the countries and languages it is read from and the number of trends that can be followed there.

What this leaves unread should be plain:

• Places are not compared. Each line sets a trend against its own past in one place. Nothing here says Japan attends to a trend more than Germany does.
• Mainland China is barely read. Wikipedia is blocked there, and Google is too, so the Google line reflects only the few who reach it. Chinese-language Wikipedia reading comes mostly from Taiwan, Singapore and Chinese speakers abroad and is counted under the Indo-Pacific.
• Russia has no Google line. The data service does not serve Russia or Iran. Russia rests on Russian-language Wikipedia reading, placed by where the language is spoken.
• Small and less connected countries are thin. Many trends are not searched for often enough there to follow, and Africa and South Asia are read from fewer trends than Europe or North America.
• Translation is by title. The local-language part of a query is the name of a Wikipedia page, which is the term for the concept and not always what people type.

What it cannot say

• It does not compare the size of trends. Both sources are relative. A score of 150 means half as much attention again as this trend’s own average. It does not mean more attention than a trend at 120.
• It is relative to the list. Because each source is set against the basket, a trend is rising only in the sense that it is gaining on the typical trend among the 250.
• It follows the words chosen. A trend is only as well measured as its pages and search terms. Some trends have no Wikipedia page of their own and rest on pages for their parts; these are noisier.
• It leans to English and to the connected. The world score is asked for in English search terms and led by English reading. Wikipedia reading and Google searching are thin where other platforms dominate or where access is limited.
• Attention is not strength. Nothing in the trend scores, the network or the scenario tests uses the attention score. It sits beside them.

The run so far

The harvest was last run on 16 October 2026 and covers the 156 weeks from 15 October 2023 to the week of 4 October 2026, one week on from the first run of 10 October.

• Wikipedia. 1,666 English pages were followed into 21,854 page-and-language pairs across 28 languages, holding 0.98 billion page views, 52 per cent of them in English. All 250 trends have an index. The median trend is measured in fourteen languages; six are measured in English only. 120 English pages begin inside the window, and their redirects give 70 of them an earlier start. The English basket averaged 116 in the first year and 71 in the third.
• Google, worldwide. All 250 trends have a series. The basket averaged 47 in the first year, 72 in the second and 167 in the third, which is the step described above and the reason it is taken out.
• How far the two agree. The median correlation between a trend’s Wikipedia line and its Google line is 0.32. It is 0.5 or better for 94 trends and below zero for 61. Where the two disagree the score is an average of two different readings, and the trend’s page shows the figure so that it can be weighed.
• Google, by country. 12,000 queries were asked, 250 trends in 48 countries. 6,244 were searched for in at least half the weeks and are read. The United States reads 248 of the 250 trends, India 235, the United Kingdom 234, Canada and Japan 220 each and Germany 215. At the other end are Qatar with 14, Uzbekistan and Ethiopia with 25 each and Morocco with 28. China has 68, under the caution above.
• By place. Every region but China can be followed on more than half the trends: from 250 in North America and the Indo-Pacific and 249 in Europe, to 214 in the Gulf and Middle East and 140 in Africa. Both sources are present for most trends in North America, Europe, the Indo-Pacific and Latin America; South Asia and Africa rest mostly on Google.
• The picture. Against the list, 15 trends are surging and 55 rising over the last year, 132 are steady, 42 are easing and 6 fading. The largest rises are in automated AI research, contested maritime access, ransomware and strategic stockpiles. The largest falls are in stablecoins, tariffs and regional trade agreements, each of which had its peak of attention earlier in the window.

What a refresh changes

Moving the window on by a week re-reads every week, not only the new one, and the two sources behave differently.

• Wikipedia counts do not change. Page views for weeks already harvested came back the same: 36 of the 21,854 page series differed at all, and by too little to measure in the total. A week’s reading is final once published.
• Google’s figures are re-scaled each time. Google sets each series against its own busiest week in the window it is asked for, and draws on a sample of searches, so asking again moves the whole series up or down by as much as 12 per cent and each week by about 2 per cent. That is not a change in searching; it is the cost of asking afresh.
• The published score moves little. Because each trend is set against its own average and against the basket, the score for past weeks shifted by about 1 per cent. The change over a year moved by 2 points for the median trend and by up to 9 points for one in twenty. Of the seven trends rising and falling most over each period, five to seven stayed in place; 26 trends crossed into a neighbouring band over a year, most of them sitting near a threshold.
• By place moves more. Country searching is thinner, so a place’s change over a year moved by 3 points for the median trend and by 37 points or more for one in twenty. A single week’s change in a small place should be read with that in mind.