Sage Journals: Discover world-class research

Abstract

Digital search platforms enable real-time observation of relationship distress through behavioral traces. This study tests whether Google Trends predicts official divorce rates in the United States, Germany, the Netherlands, and the United Kingdom from 2009 to 2023. We introduce Digital Behavioral Decoupling, in which online distress signals diverge from legal outcomes as divorce shifts from an institutional procedure to an emotionally mediated digital phenomenon. Methods include unit root tests, cointegration analysis, Granger causality, spectral coherence, and rolling-origin nowcasting. Search queries are grouped into pre-divorce (cognitive distress), during-divorce (procedural action), and post-divorce (emotional recovery) phases. Results show 62.5% of terms lead divorce rates by 1–2 years, yet only 8.3% remain significant after False Discovery Rate correction. The Netherlands demonstrates 100% forecast improvements beyond autoregressive models (DM 2.87–3.43, p < 0.02) across all search terms, from early relationship therapy queries through procedural and post-divorce searches, indicating systematic capture of the entire divorce pathway. Germany shows intermediate results with 33% forecast success beyond autoregressive benchmarks (DM 2.40–2.81) limited to problem-recognition terms, suggesting episodic crisis-driven engagement. The United States and United Kingdom show no forecast gains beyond autoregressive models despite high search volumes, consistent with information saturation in normalized divorce cultures. Lead-lag relationships are frequency-specific, concentrated at 3–5 year periodicities. Findings link family sociology with affective computing and provide a replicable toolkit for tracking relationship dissolution in algorithmically curated information environments.

Keywords

digital behavioral decoupling google trends divorce prediction relationship distress analytics affective computing in family sociology algorithmic diagnostic spirals

Introduction

Divorce is a legally regulated mechanism that allows individuals to end an unsatisfactory marriage while safeguarding personal rights and freedoms (Pešić-Jenaćković, 2018). The process involves recognizing marital dysfunction, evaluating possible responses, and ultimately implementing the decision (Solichah & Istiqomah, 2018). Similarly, it can be viewed as a three-phase transition encompassing pre-divorce decision-making, restructuring during-divorce, and post-divorce adjustment (Salts, 1985).

The perception and stigma of divorce are shaped by legal norms, cultural acceptance, and personal experience, with increasing social tolerance, particularly among younger adults, and reduced stigma in cases of unhappiness or violence (Cooney, 2022). As marriage dissolution becomes more normalized and associated with individual freedom and self-fulfillment, global divorce rates have generally risen, although recent data show stabilization or decline in some regions. High divorce rates are especially evident in Western countries, where estimates indicate that in the United States approximately 39–50% of first marriages end in divorce, and the United Kingdom experienced a sharp rise during the 20th century followed by moderation in the 2010s (Ho, 2025). Societal factors such as economic development, rising social and economic status of women, democratization, and political instability are positively associated with divorce rates, particularly in post-Soviet and transitioning societies (Aghayeeabianeh & Stamatel, 2022). Scientific perspectives also suggest that conflicts rooted in differing reproductive strategies between men and women contribute to divorce across cultures (Mortelmans, 2021).

Current prediction approaches rely on retrospective statistics or late-stage surveys, missing early intervention opportunities for government policies and support services. Although Google Trends is widely used to monitor collective emotions and public health, its application to divorce remains limited despite divorce being a major life transition with significant individual, familial, and societal costs. Questionnaire-based machine-learning models achieve high accuracy but depend on narrow survey samples (Kumar et al., 2023; Castrillón, 2021). Time series forecasts predict regional divorce counts using administrative data but fail to capture real-time behavioral signals preceding divorce (Derwanto, 2022; Yudistira et al., 2025). Moreover, existing studies treat divorce-related searches as uniform and lack systematic cross-national analyses linking search behavior to official statistics. This study addresses this gap by integrating official divorce statistics with Google Trends data and organizing search terms into pre-divorce (cognitive distress), during-divorce (procedural action), and post-divorce (emotional recovery) phases. We examine four countries (United States, United Kingdom, Germany, and the Netherlands) reflecting contrasting welfare regimes and divergent divorce trajectories since 1970 (Esping-Andersen, 1999). In the end of 2021, crude divorce rates were 2.5 per 1,000 in the United States, 1.7 in Germany, 1.5 in the Netherlands, and 1.8 in the United Kingdom, with sharper declines from mid-1990s peaks observed in the European cases (OECD, 2025). These cross-national differences, together with reliable administrative data and sufficient multilingual search volume, provide strong conditions for comparative time series analysis. The study introduces the concept of Digital Behavioral Decoupling, defined as the divergence between online distress signals and real-world divorce outcomes as marital dissolution becomes increasingly digitally mediated, that is, emotionally processed through online searches and information-seeking. Accordingly, we ask: (1) to what extent can Google Trends data predict official divorce rates across national contexts, and (2) under what conditions does digital search behavior decouple from administrative divorce statistics. Hence, the study offers a scalable framework for detecting relational distress in the digital age.

The use of Google Trends as a data source for divorce prediction is supported by growing evidence that couples increasingly seek help online when facing relational difficulties. Research has shown that couples frequently seek external support when facing relational uncertainty, with more than one-third turning to resources such as marital counseling or relationship literature within the first 5 years of marriage, often in response to individual or interpersonal difficulties (Doss et al., 2009). The search for relationship-relevant or distressing information plays a key role in regulating uncertainty and emotional well-being, as explained by the Theory of Motivated Information Management (Afifi et al., 2004). This help-seeking behavior has shifted substantially toward digital platforms. Individuals who seek support digitally report qualitatively different concerns than those in traditional therapy, particularly involving trust deficits, reduced shared time, and parenting strain (Roddy et al., 2019). Digital help-seeking now exceeds conventional services, especially among women experiencing acute relational distress, with nearly two in five turning to online resources (Trillingsgaard et al., 2019). Large-scale analyses further indicate that younger populations use online platforms as spaces for emotional disclosure and problem-solving related to intimacy, communication, and attachment (Entwistle et al., 2021). This trend suggests that relationship distress is increasingly expressed through digital channels, making search engine data a promising source for real-time monitoring of marital strain.

Literature Review

Studies have established that search engines like Google serve as indicators of emotional distress. Google Trends query indices summarizing search term frequency over time and across regions correlate with diverse real-world phenomena, including housing foreclosures, labor market conditions, epidemiological trends and political events. Emerging Google Trends topics can trigger collection and analysis of geolocalized tweets, enabling decision makers to monitor near real-time sentiments and emotions these trends evoke in online populations (D’Avanzo et al., 2017). A study shows that Google Trends reflects population-level emotional states; for instance, searches for “anxiety” significantly correlate with both the UN World Happiness Index and the OECD Life Satisfaction Index. These findings support that Google Trends can serve as a useful indicator for estimating population-level well-being when applied to high-volume search terms, though reliability concerns remain for low-frequency queries (Banerjee, 2018). Emotional search data track established well-being indicators and predict health outcomes such as cardiovascular conditions and depression beyond income and traditional measures. Aggregate Google searches for affect-related terms (e.g., stress, depression, fatigue, and scared) correlate with survey-based well-being, personality, and Twitter (X) language, while uniquely predicting depression rates and circulatory disease mortality at state and metro levels after controlling for demographic and socioeconomic factors (Ford et al., 2018).

Validation during crisis periods, including the COVID-19 pandemic, confirms that search trends can act as sensitive indicators of widespread emotional strain and anxiety, supported by consistent findings across multiple data sources (Boy, 2023). The COVID-19 pandemic provided a natural experiment for observing how external stressors influence divorce-related search behavior. Divorce-related searches initially declined during early lockdown phases, consistent with restrictions on mobility and institutional closures (Brodeur et al., 2021). As stay-at-home conditions persisted, however, a rebound emerged, especially in higher-income nations, suggesting that prolonged confinement intensified relational pressures (Cheng et al., 2025). Parallel analyses confirm this pattern, revealing a temporary 11% decrease in Europe followed by a 13% post-lockdown surge in the United States (Berger et al., 2021). These findings demonstrate that Google Trends data can capture temporal shifts in relationship distress at population scale.

A range of recent studies demonstrates that divorce rate prediction increasingly relies on structured questionnaires or official statistical series, achieving very high accuracy under controlled or narrow data conditions. Several machine-learning studies using the UCI Divorce Predictors dataset show that decision-tree-based models can reach 100% accuracy when coupled with feature selection (Kumar et al., 2023). J48-based decision trees similarly identify a small set of influential variables with perfect classification when more than 90% of the sample is used for training (Castrillón, 2021). Another ensemble-learning study employing Support Vector Machines (SVMs), passive-aggressive classification, and multilayer perceptron reports near-perfect classification performance based on expert-designed questionnaires (Sadiq Fareed et al., 2022). In contrast, time series forecasting studies focus on predicting regional divorce counts using statistical smoothing methods, such as exponential or triple-exponential smoothing, but rely solely on administrative data and do not capture behavioral or attitudinal signals preceding divorce (Derwanto, 2022; Yudistira et al., 2025). Demographic analyses suggest broader social trends, such as declining divorce rates among younger cohorts in the United States, based on large census datasets, but provide limited insight into real-time behavioral indicators or early warning signals (Cohen, 2019).

Despite this growing body of research, no study has systematically compared divorce-related search behavior across multiple countries or examined whether search patterns correspond to distinct phases of the divorce process. Furthermore, existing time series approaches have not tested for bidirectional relationships between search behavior and official statistics, leaving open the question of whether online signals precede, follow, or decouple from administrative divorce outcomes.

Data and Method

Data

This study combines Google Trends data with official divorce statistics to enable cross-national comparison of online search behavior and institutional divorce rates. Google Trends was used to collect search terms related to marriage, divorce, and counseling with semantically equivalent translations, normalized to relative search interest over time (Google Trends, 2025). Official divorce rates were expressed as divorces per 1,000 population for comparability. U.S. data were obtained from the CDC report National Marriage and Divorce Rate Trends for 2000–2023 (Centers for Disease Control and Prevention, 2025). German data came from the Marriages, Divorces (Time Series) dataset (German Federal Statistical Office, 2024). Dutch statistics were sourced from Marriage Dissolutions; Due to Divorce and Death (Statistics, 2025). U.K. data were retrieved from Divorces and Dissolutions in England and Wales: 2023 (Office for National Statistics, 2025), and, because separate rates are provided for men and women, the analysis used their average, as the gender difference was minimal (0.2 at most). For Google Trends, A common 15-year period (2009–2023) was selected to ensure temporal alignment and address years with near-zero or inconsistent (sudden changes—0 to high levels). Data is downloaded directly from the website by the authors on October 19, 2025.

Search terms were categorized into three divorce process phases based on semantic content. Two authors independently conducted this categorization, achieving high inter-rater reliability (93.8% agreement; 15 of 16 terms), with disagreements resolved through discussion. Pre-divorce phase queries capture cognitive distress and help-seeking (“marriage problems” and “marriage counseling”), during-divorce terms represent procedural action (“divorce” and “divorce lawyer”), and post-divorce searches address emotional recovery (“after divorce”). Search term identification combined theoretical foundations (Salts, 1985). Large language model consultation (ChatGPT; Claude and Grok) provided culturally appropriate translations; candidate terms were retained only if they returned non-zero search volume in Google Trends. Additional terms tested (Supplementary Material 1) were excluded due to near-zero or inconsistent search volumes with sudden unreasonable changes and recent emergence (i.e., after 2015).

Metrics

This study employs a time series econometric framework to examine the relationship between Google Trends search intensity and official divorce rates over 2009–2023. Our approach integrates unit root diagnostics, cointegration analysis, frequency-domain methods, lag structure identification, and forecast evaluation to address lead-lag dynamics and ensure robustness against spurious correlations in non-stationary time series data (Granger & Newbold, 1974). Given the limited 15-year observation window, we prioritize methods that balance statistical power with finite-sample reliability.

We establish the order of integration for each series using a consensus voting approach across three complementary unit root tests. The Augmented Dickey-Fuller (ADF) test (Dickey & Fuller, 1979) given in equation (1) estimates:

Δ y_{t} = α + β t + γ y_{t - 1} + \sum_{i = 1}^{p} δ_{i} Δ y_{t - i} + ε_{t}

(1)

where

y_{t}

is the time series of interest,

Δ

denotes the first-difference operator,

t

represents a deterministic time trend,

p

is the optimal lag length selected by AIC, and

ε_{t}

is white noise. The null hypothesis

H_{0} : γ = 0

indicates a unit root (non-stationarity).

The Phillips-Perron (PP) test (Phillips & Perron, 1988) provides a non-parametric alternative to serial correlation and heteroskedasticity given by equation (2):

y_{t} = α + β t + ρ y_{t - 1} + u_{t}

(2)

where

u_{t}

may exhibit autocorrelation. The PP test modifies the test statistic using Newey-West heteroskedasticity and autocorrelation consistent (HAC) standard errors.

The Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test (Kwiatkowski et al., 1992) reverses the null hypothesis, testing stationarity against the alternative of a unit root, as given by equation (3):

y_{t} = ξ t + r_{t} + ε_{t}, r_{t} = r_{t - 1} + u_{t}

(3)

where

ξ t

is a deterministic trend,

r_{t}

is a random walk component, and the null hypothesis is

σ_{u}^{2} = 0

(stationarity). A series is classified as I (1) if at least two of the three tests agree on non-stationarity.

For series determined to be I (1), we test for cointegration—the existence of a stable long-run equilibrium relationship (Engle et al., 1987). The Engle-Granger two-step procedure first estimates the cointegrating regression, given in equation (4).

y_{t} = α + β x_{t} + ε_{t}

(4)

where

y_{t}

is the divorce rate,

x_{t}

is the Google Trends index, and

ε_{t}

represents the equilibrium error. The residuals are then tested for stationarity using the ADF test with critical values adjusted for cointegration testing. Cointegration is confirmed if the residuals are I (0) at conventional significance levels.

We supplement this with the Johansen trace test (Johansen, 1988), which estimates a Vector Error Correction Model (VECM) as given by equation (5):

Δ Y_{t} = Π Y_{t - 1} + \sum_{i = 1}^{k - 1} Γ_{i} Δ Y_{t - i} + ε_{t}

(5)

where

Y_{t} = {[y_{t}, x_{t}]}^{'}

is a vector of endogenous variables,

Π = α β^{'}

is the long-run impact matrix with cointegrating vectors

β

Γ_{i}

captures short-run dynamics, and

k = 1

is the lag order. The trace statistic

λ_{trace} = - T \sum_{i = r + 1}^{p} \ln (1 - {\hat{λ}}_{i})

tests the null of at most

r

cointegrating relationships, where

{\hat{λ}}_{i}

are ordered eigenvalues of

Π

When cointegration is detected, we estimate the Error Correction Model (ECM) given in equation (6), to quantify adjustment dynamics:

Δ y_{t} = α + β Δ x_{t} + γ {EC}_{t - 1} + ε_{t}

(6)

where

{EC}_{t - 1} = y_{t - 1} - {\hat{α}}_{LR} - {\hat{β}}_{LR} x_{t - 1}

is the lagged equilibrium error,

β

captures contemporaneous short-run effects of search intensity changes on divorce rate changes, and

γ

(expected negative) measures the speed of adjustment to long-run equilibrium. A significant negative

γ

confirms error correction behavior.

The cross-correlation function (CCF) identifies optimal lead-lag relationships in the time domain, given by equation (7):

ρ_{x y} (k) = \frac{\sum_{t = 1}^{n - ∣ k ∣} (x_{t} - \overset{ˉ}{x}) (y_{t + k} - \overset{ˉ}{y})}{\sqrt{\sum_{t = 1}^{n} {(x_{t} - \overset{ˉ}{x})}^{2} \sum_{t = 1}^{n} {(y_{t} - \overset{ˉ}{y})}^{2}}}

(7)

where

k \in {- L, \dots, 0, \dots, L}

denotes lag in years,

\overset{ˉ}{x}

and

\overset{ˉ}{y}

are sample means,

L = 2

is the maximum lag considered, and negative

k

indicates

x

leads

y

∣ k ∣

periods. We identify the optimal lag

k^{*} = \arg \max_{k} ∣ ρ_{x y} (k) ∣

and report both the lag structure and maximum absolute correlation.

We decompose relationships across frequency bands using spectral coherence using equation (8) (Granger and Newbold, 1974):

C_{x y} (f) = \frac{∣ P_{x y} (f) ∣^{2}}{P_{x x} (f) P_{y y} (f)}

(8)

where

P_{x y} (f)

is the cross-spectral density;

P_{x x} (f)

and

P_{y y} (f)

are auto-spectral densities at frequency

f

. Coherence ranges from

[0, 1]

, with values near 1 indicating strong linear association at frequency

f

(or equivalently, period

T = 1 / f

Phase-based lag is derived from the cross-spectrum phase angle, and it is given in equation (9):

τ (f) = \frac{ϕ_{x y} (f)}{2 π f}

(9)

where

ϕ_{x y} (f) = \arg (P_{x y} (f))

is the phase spectrum. Positive

τ

indicates

x

leads

y

τ

time units at frequency

f

. We compute coherence and phase spectra using Welch’s method with overlapping segments.

To assess temporal stability of lead-lag relationships, we employ rolling window CCF analysis given by equation (10):

ρ_{x y}^{[t, t + w]} (k) = CCF (x_{t : t + w}, y_{t : t + w}, k)

(10)

where

w = 4

years is the window width and

t

indexes window starting positions. We track the evolution of optimal lag

k^{*} (t)

and maximum correlation

∣ ρ_{x y}^{*} ∣ (t)

across windows to identify regime changes or time-varying dynamics.

We implement a rolling-origin nowcasting procedure comparing baseline AR ( $p$ ) forecasts to augmented models incorporating Google Trends. For each evaluation point $t$ in a sliding window scheme:

Baseline model is given by equation (11):

{\hat{y}}_{t + h ∣ t} = α + \sum_{i = 1}^{p} ϕ_{i} y_{t + 1 - i}

(11)

where

h = 1

is the one-year-ahead forecast horizon,

p

is selected by corrected AIC:

{AIC}_{c} = AIC + \frac{2 k (k + 1)}{T - k - 1}

with

k = p + 1

parameters, and

T

is the training sample size. Also in equation (12):

{\hat{y}}_{t + h ∣ t} = y_{t} + {\hat{β}}_{0} + {\hat{β}}_{1} Δ x_{t}

(12)

where

{\hat{β}}_{0}, {\hat{β}}_{1}

are OLS estimates from regressing first-differenced divorce rates

Δ y_{t}

on first-differenced search intensity

Δ x_{t}

in the training window, exploiting the contemporaneous relationship between changes in search behavior and divorce dynamics.

We evaluate forecast accuracy using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) using equation (13):

MAE = \frac{1}{n} \sum_{t = 1}^{n} ∣ e_{i} ∣, RMSE = \sqrt{\frac{1}{n} \sum_{t = 1}^{n} e_{i}^{2}}

(13)

where

e_{i} = y_{i} - {\hat{y}}_{i}

are forecast errors and

n

is the number of out-of-sample evaluations. Lower MAE and RMSE values indicate superior predictive performance. The relative improvement

Δ MAE = {MAE}_{baseline} - {MAE}_{augmented}

quantifies the informational value of incorporating Google Trends data.

Methodology and Experimental Setup

This study employs a comprehensive time series econometric framework to examine the relationship between online search behavior and official divorce rates over the period 2009–2023. The analytical pipeline consists of seven stages. The overview of the proposed methodology is given in Figure 1 and explained in detail below.

(1) Data Collection: Google Trends data on marriage-related searches and official divorce statistics from national agencies (CDC, German Federal Statistical Office, Statistics Netherlands, and UK ONS) were compiled and standardized across countries.

(2) Stationarity Testing: Series integration orders were identified through combined ADF, PP, and KPSS tests.

(3) Cointegration Analysis: Long-run relationships between search intensity and divorce rates were examined using Engle-Granger, Johansen, and Gregory-Hansen tests.

(4) Causality Testing: Directional effects were evaluated via Δ-Granger and VAR-level Granger tests with AIC-based lag selection and integration order adjustment.

(5) Frequency and Lag Analysis: Cross-spectral coherence, phase spectra, and cross-correlation functions (including rolling windows) captured frequency-specific and temporal lead-lag structures.

(6) Forecasting Evaluation: Rolling-origin nowcasting compared baseline autoregressive models with search-augmented versions using MAE, RMSE, and Diebold-Mariano tests for forecast improvement.

(7) Statistical Inference: Benjamini-Hochberg False Discovery Rate correction (α = 0.05) was applied to control for multiple testing.

Figure 1.

Method overview

All statistical analyses were performed using Python 3.12.12. Data manipulation and preparation utilized Pandas, a high-performance library for structured data operations (McKinney, 2010). Numerical computations employed NumPy (Harris et al., 2020). Statistical testing, including unit root tests, cointegration analysis, and probability distributions, leveraged SciPy (Virtanen et al., 2020). Data visualization relied on matplotlib (Hunter, 2007). For publication-quality figures, Seaborn was employed (Waskom, 2021).

This econometric framework characterizes the dynamic link between online search behavior and divorce trends by controlling for spurious correlations through integration, cointegration, and multiple testing corrections. Combining time-, frequency-, and forecast-domain analyses enables evaluation of both contemporaneous and predictive relationships. All results are reported with FDR-adjusted significance levels.

Results and Discussion

Descriptive Results

Tables 1 and 2 present statistics on Google Trends and official divorce rates, revealing substantial cross-country differences in search intensity and variability. Cultural norms and institutional frameworks shape marital information-seeking. Lower-divorce countries such as Germany and the Netherlands show more differentiated search behavior, whereas high-divorce contexts like the United Kingdom display less intensive engagement.

Table 1.

Google Trends: Descriptive Statistics (2009–2023). Search Terms Related to Relationship Problems and Divorce Showed Substantial Cross-Country Variation, With Some Terms Exhibiting Up to 4-Fold Differences Between Minimum and Maximum Values

Country	Google Trends: Descriptive Statistics
Country	Search Term	Mean	Median	Std.	Min.	Max.
Germany	Beziehungsprobleme (relationship problems)	27.41	21.33	13.92	13.75	57.67
Germany	Eheberatung (marriage counseling)	30.49	29.50	7.06	20.58	43.92
Germany	Eheprobleme (marriage problems)	24.49	22.83	6.04	15.25	33.75
Germany	Scheidung (divorce)	42.63	43.75	4.64	33.25	49.08
Germany	Scheidungsanwalt (divorce lawyer)	67.04	70.42	17.32	35.42	88.75
Germany	Nach der scheidung (after divorce)	53.98	54.67	7.37	38.67	70.50
Netherlands	Echtscheiding (divorce)	31.00	29.92	11.08	16.50	48.67
Netherlands	Echtscheidingsadvocaat (divorce lawyer)	53.90	53.75	9.54	39.25	72.54
Netherlands	Huwelijksproblemen (marriage problems)	38.10	34.21	15.62	22.51	76.18
Netherlands	Na de scheiding (after divorce)	49.11	48.42	4.52	41.67	60.63
Netherlands	Relatieproblemen (relationship problems)	25.47	19.75	14.73	9.50	55.25
Netherlands	Relatietherapie (relationship therapy)*	75.66	75.42	4.85	69.17	88.33
United Kingdom	Divorce	69.81	71.08	7.73	49.50	82.42
United Kingdom	Divorce lawyer	26.03	21.83	17.54	14.08	83.17
United Kingdom	Life after divorce*	14.21	13.83	3.03	10.17	22.00
United Kingdom	Marriage counseling	57.67	67.08	21.85	18.33	80.75
United Kingdom	Marriage problems	24.24	23.58	6.66	14.50	33.08
United Kingdom	Relationship problems	28.96	27.42	8.41	17.42	40.17
United States	After divorce	60.59	59.42	7.49	46.17	76.25
United States	Divorce	54.83	53.00	7.88	45.58	70.33
United States	Divorce lawyer	38.56	38.67	4.66	32.17	46.17
United States	Marriage counseling	80.87	81.83	7.31	67.42	90.58
United States	Marriage problems	35.52	32.75	12.26	19.50	55.17
United States	Relationship problems	53.46	54.17	6.87	42.92	63.00

Note. During data compilation, search volumes for “marriage therapy” and “marriage counseling” in the Netherlands, and “after divorce” in the United Kingdom, were too low (zero or near-zero) for values to be returned by Google Trends. Consequently, “relationship therapy” was used for the Netherlands and “life after divorce” for the United Kingdom. Google Trends monthly search interest indices (0–100 scale, 100 = peak) were aggregated to yearly means. Months with zero values were included in calculations, as zero might reflect low rather than absent search activity. Searches were conducted in each country’s local language(s). Period: 2009–2023 (n = 15 years per term). N = 24 country-term pairs across 4 countries. Std. = standard deviation; Min. = minimum; Max. = maximum.

Table 2.

Official Divorce Rates: Descriptive Statistics (2009–2023). Crude Divorce Rates Per 1,000 Population; Higher Values Indicate More Divorces Relative To Population Size. UK Divorce Rates are Calculated Per 1,000 Married Population (ONS Methodology), While Other Countries Report Per 1,000 Total Population. Cross-National Rate Comparisons Should be Interpreted With This Difference in Mind. Period: 2009–2023. Std. = Standard Deviation; Min. = Minimum; Max. = Maximum

Country	Official Divorce Rates: Descriptive Statistics
Country	Mean	Median	Std.	Min.	Max.
Germany (German Federal Statistical Office $—$ Destatis)	1.95	2.00	0.26	1.50	2.30
Netherlands (Statistics Netherlands $—$ CBS)	1.81	1.91	0.25	1.33	2.10
United Kingdom (Office for National Statistics)	9.02	9.00	1.01	6.70	10.30
United States (Centers for Disease Control and Prevention—CDC)	2.99	3.00	0.45	2.30	3.60

Search volatility reflects cultural differences in responses to relationship distress. Germany shows high variability for “Beziehungsprobleme” (relationship problems) (SD = 13.92) and “Scheidungsanwalt” (divorce lawyer) (SD = 17.32), indicating episodic, crisis-driven attention consistent with its low and stable divorce rate, where divorce remains a reactive outcome rather than a continuous public concern. In contrast, the Netherlands demonstrates highly stable and intense searches for “relatietherapie” (relationship therapy) (M = 75.66, SD = 4.85), reflecting a routinized, institutionalized approach to early help-seeking, alongside similarly stable post-divorce adjustment searches for “na de scheiding” (after divorce) (SD = 4.52). This pattern suggests that marital dissolution is often mitigated through preemptive intervention rather than crisis response. The UK presents a paradox: despite the highest divorce rate (M = 9.02), search intensity for “divorce lawyer” remains low (M = 26.03). High averages for “divorce” (M = 69.81) and “marriage counseling” (M = 57.67) confirm continued public salience, but the high variability of “marriage counseling” (SD = 21.85 vs. “divorce” SD = 7.73) indicates fluctuating, non-institutionalized help-seeking. The U.S. shows consistently strong engagement with “marriage counseling” (M = 80.87, SD = 7.31) and “after divorce” (M = 60.59, SD = 7.49), reflecting an institutionalized therapeutic culture embedded in its professionalized mental health framework.

Across the four countries, the overlay figures (Figures 2 –5) illustrate clear cross-national differences in how online search behavior corresponds with official divorce dynamics. As shown in Figure 2, Germany’s search patterns move largely in line with divorce rates but reflect a reactive, legally oriented approach. Figure 3 shows that the Netherlands demonstrates stable and preventive engagement, with consistent interest in therapy and mediation closely matching gradual declines in divorce rates. Figures 2, 4, and 5 reveal partial decoupling in Germany, the United Kingdom, and the United States, where divorce-related searches remain prominent or increasing despite declining official rates.

Figure 2.

Germany $—$ Z-score overlay: Trends terms and divorce

Figure 3.

Netherlands—Z-score overlay: Trends terms and divorce

Figure 4.

United Kingdom—Z-score overlay: Trends terms and divorce

Figure 5.

United States—Z-score overlay: Trends terms and divorce

Stationary and Cointegration Analysis

Unit root tests revealed that 75% (18/24) of Google Trends series were integrated of order one [I (1)], indicating non-stationary behavior typical of online search intensity data, while the remaining 25% exhibited stationarity [I (0)]. Conversely, official divorce rates displayed the opposite pattern, with 75% (18/24) classified as I (0) and 25% as I (1), reflecting the relatively stable institutional nature of divorce statistics. This asymmetry suggests that search behavior exhibits greater volatility and trending characteristics compared to formal demographic outcomes.

Among the six country-term pairs where both series were I (1), Engle-Granger tests identified long-run cointegration in 33.3% (2/6) of cases, specifically for Netherlands’ “huwelijksproblemen (relationship problems)” (EG p = 0.0213) and “relatietherapie (relationship therapy)” (EG p < 0.001). However, after FDR correction, none remained statistically significant, highlighting the limited evidence for robust long-run equilibrium relationships when accounting for multiple testing. Complete results are presented in Table 3.

Table 3.

Stationarity and Cointegration Analysis Results. Unit Root and Cointegration Test Results by Country and Search Term (2009–2023). Integration order I (0) Indicates Stationary Series; I (1) Indicates Non-Stationary Series Requiring Differencing. Cointegration tests (EG and Johansen) Applicable Only When Both Series are I (1), Indicating Potential Long-Run Equilibrium Relationships. Significant p-Values (<0.05) Suggest Cointegration, Though None Survived FDR Correction. Gregory-Hansen Test Identifies Structural Break Years, Concentrated Between 2012 and 2020

Country	Term	Trends Integration	Divorce Integration	EG Cointegrated	EG p-Value	Johansen Cointegrated	GH Break Year
Germany	Eheprobleme (marriage problems)	I (1)	I (0)	—	—	—	2019
Germany	Scheidungsanwalt (divorce lawyer)	I (1)	I (0)	—	—	—	2019
Germany	Beziehungsprobleme (relationship problems)	I (1)	I (0)	—	—	—	2018
Germany	Scheidung (divorce)	I (0)	I (0)	—	—	—	2015
Germany	Nach der scheidung (after divorce)	I (1)	I (0)	—	—	—	2013
Germany	Eheberatung (marriage counseling)	I (0)	I (0)	—	—	—	2016
Netherlands	Echtscheiding (divorce)	I (1)	I (1)	No	0.637	No	2017
Netherlands	Huwelijksproblemen (marriage problems)	I (1)	I (1)	Yes	0.021	No	2014
Netherlands	Relatieproblemen (relationship problems)	I (1)	I (1)	No	0.218	Yes	2017
Netherlands	Relatietherapie (relationship therapy)	I (1)	I (1)	Yes	<0.001	No	2014
Netherlands	Echtscheidingsadvocaat (divorce lawyer)	I (1)	I (1)	No	0.185	No	2017
Netherlands	Na de scheiding (after divorce)	I (1)	I (1)	No	0.951	No	2014
United States	Marriage problems	I (1)	I (0)	—	—	—	2016
United States	Marriage counseling	I (0)	I (0)	—	—	—	2016
United States	Divorce lawyer	I (1)	I (0)	—	—	—	2019
United States	Divorce	I (1)	I (0)	—	—	—	2020
United States	Relationship problems	I (1)	I (0)	—	—	—	2016
United States	After divorce	I (0)	I (0)	—	—	—	2015
United Kingdom	Divorce	I (0)	I (0)	—	—	—	2016
United Kingdom	Marriage problems	I (1)	I (0)	—	—	—	2012
United Kingdom	Relationship problems	I (1)	I (0)	—	—	—	2014
United Kingdom	Life after divorce	I (0)	I (0)	—	—	—	2016
United Kingdom	Divorce lawyer	I (1)	I (0)	—	—	—	2014
United Kingdom	Marriage counseling	I (1)	I (0)	—	—	—	2016

Note. EG = Engle-Granger cointegration test; GH = Gregory-Hansen structural break test; indicates test not applicable (cointegration tests require both series to be I(1)). Bold p-values indicate significance at α = 0.05 before FDR correction. After FDR adjustment, no cointegration results remained statistically significant.

Gregory-Hansen structural break tests detected potential level shifts across all 24 pairs, with break years concentrated between 2012 and 2020 (Table 3), suggesting that the relationship between search intensity and divorce rates may be subject to regime changes or exogenous shocks. The clustering around 2014–2016 coincides with increased smartphone penetration and changes in Google search algorithms, while later (2019–2020) may reflect pandemic-related behavioral shifts. The Netherlands showed earlier rises (2014+) in “huwelijksproblemen” (marriage problems) and “relatietherapie” (relationship therapy) searches, indicating an early move toward proactive relationship management.

Causality and Dynamic Relationships

Granger causality tests in first-difference form (Δ-Granger) revealed statistically significant predictive relationships in only 8.3% (2/24) of country-term pairs after FDR correction, specifically for United States’ “marriage counseling” (FDR-adjusted p = 0.001) and United Kingdom’s “divorce lawyer” (FDR-adjusted p < 0.001). VAR-level Granger tests yielded no statistically significant results after multiple testing correction, indicating limited evidence for systematic directional causality from search intensity to divorce rates. The low incidence of significant Granger causality suggests that while search behavior and divorce rates may co-move, the predictive power of online searches remains weak in most contexts when subjected to rigorous statistical testing.

Cross-correlation function (CCF) analysis provided more nuanced insights into lead-lag structures as given in Table 4. In 62.5% (15/24) of pairs, optimal negative lags were identified, indicating that Google Trends series preceded divorce rates by one to 2 years. Contemporaneous relationships (zero lag) were observed in 20.8% (5/24) of cases, while positive lags occurred in 16.7% (4/24). The average maximum absolute cross-correlation was 0.781, with Germany and the United States exhibiting the strongest synchronization (0.885 and 0.887, respectively), while the Netherlands displayed more moderate correlation (0.617).

Table 4.

Causality Testing and Lead-Lag Structure. Cross-Correlation, Granger Causality, and Spectral Coherence Results (2009–2023). Optimal Lag Indicates Temporal Relationship: Negative Values Suggest Search Behavior Precedes Divorce (Anticipatory), Positive Values Suggest Divorce Precedes Search (Reactive). Higher |CCF| Values Indicate Stronger Correlation; Coherence >0.4 Indicates Frequency-Domain Co-Movement. Granger Tests Assess Predictive Causality After FDR Correction. Majority of Pairs (62.5%) Showed Negative Optimal Lags, Suggesting Anticipatory Search Patterns

Country	Term	Optimal CCF Lag	Max \|CCF\|	Mean Coherence	Granger Δ p-adj	Granger Δ Sig	VAR-G p-adj	VAR-G Sig
Germany	Eheprobleme (marriage problems)	1	0.972	0.434	0.253	No	0.571	No
Germany	Scheidungsanwalt (divorce lawyer)	−2	0.964	0.3	0.76	No	0.891	No
Germany	Beziehungsprobleme (relationship problems)	0	0.897	0.344	0.365	No	0.891	No
Germany	Scheidung (divorce)	−1	0.879	0.255	0.092	No	0.094	No
Germany	Nach der scheidung (after divorce)	0	0.638	0.131	0.615	No	0.993	No
Germany	Eheberatung (marriage counseling)	−2	0.961	0.269	0.092	No	0.212	No
Netherlands	Echtscheiding (divorce)	−2	0.896	0.349	0.151	No	0.091	No
Netherlands	Huwelijksproblemen (marriage problems)	−2	0.61	0.207	0.31	No	0.696	No
Netherlands	Relatieproblemen (relationship problems)	−2	0.77	0.199	0.795	No	0.091	No
Netherlands	Relatietherapie (relationship therapy)	−2	0.225	0.26	0.253	No	0.696	No
Netherlands	Echtscheidingsadvocaat (divorce lawyer)	−2	0.579	0.444	0.521	No	0.696	No
Netherlands	Na de scheiding (after divorce)	−1	0.624	0.343	0.188	No	0.696	No
United States	Marriage problems	−1	0.961	0.306	0.521	No	0.258	No
United States	Marriage counseling	0	0.782	0.454	0.001	Yes	0.086	No
United States	Divorce lawyer	0	0.851	0.103	0.521	No	0.993	No
United States	Divorce	0	0.908	0.268	0.76	No	0.917	No
United States	Relationship problems	−2	0.975	0.583	0.063	No	0.091	No
United States	After divorce	−2	0.842	0.367	0.253	No	0.086	No
United Kingdom	Divorce	−2	0.762	0.356	0.067	No	0.091	No
United Kingdom	Marriage problems	2	0.637	0.056	0.546	No	0.993	No
United Kingdom	Relationship problems	2	0.767	0.139	0.615	No	0.854	No
United Kingdom	Life after divorce	−1	0.666	0.473	0.521	No	0.258	No
United Kingdom	Divorce lawyer	2	0.816	0.678	<0.001	Yes	0.189	No
United Kingdom	Marriage counseling	−1	0.752	0.043	0.615	No	0.199	No

Note. Negative lag indicates search precedes divorce; Positive lag indicates divorce precedes search; p-adj = FDR-adjusted p-value; Bold |CCF| > 0.900 or coherence >0.4. Only 2 pairs (12.5%) retained Granger significance after FDR correction.

Online information-seeking can precede formal divorce outcomes, though the strength and direction vary across countries and search terms, reflecting differences in institutional responsiveness, help-seeking norms, and divorce social visibility. Germany’s “Eheprobleme (Marriage Problems)” shows a positive lag (+1), indicating retrospective searching after divorce. In contrast, the Netherlands exhibits consistent negative lags, with five of six showing −2 year leads, signaling systematic early planning behavior.

The gap between high CCF values and weak Granger significance reflects methodological differences: CCF identifies correlated lead-lag structure, while Granger applies stricter tests accounting for autoregressive dynamics and multiple testing. After FDR correction, most relationships fail to demonstrate robust predictive causality despite clear temporal patterning.

Figures 6 –8 illustrate key cases. Figure 6 shows U.S. “marriage counseling” with consistently positive phase lags across 2 to 6 year cycles, indicating divorce rates precede counseling searches despite high search intensity (M = 80.87). Figure 7 shows U.S. “relationship problems” reversing from positive lag at short cycles (+1.0 year) to negative lag at longer cycles (−0.7 years), demonstrating that predictive direction depends on time scale, reconciling its high CCF (0.975) with positive median phase lag (+0.846).

Figure 6.

Phase-derived lag: “marriage counseling”—United States. Positive lag indicates divorce leads search; negative lag indicates search leads divorce. Dashed line represents synchrony (zero lag). X-axis: cycle periodicity in years; Y-axis: temporal lag (τ) in years. Source: CDC

Figure 7.

Phase-derived lag: “relationship problems”—United States. Positive lag indicates divorce leads search; negative lag indicates search leads divorce. Dashed line represents synchrony (zero lag). X-axis: cycle periodicity in years; Y-axis: temporal lag (τ) in years. Source: CDC

Figure 8.

Phase-derived lag: “divorce lawyer”—United Kingdom. Positive lag indicates divorce leads search; negative lag indicates search leads divorce. Dashed line represents synchrony (zero lag). X-axis: cycle periodicity in years; Y-axis: temporal lag (τ) in years. Source: ONS

Figure 8 presents the most complex pattern for UK “divorce lawyer.” Despite the highest coherence in the dataset (0.678, Table 5), its phase shifts from +1.0 year at short cycles to −0.7 years at intermediate frequencies, reflecting competing dynamics: reactive post-divorce searches in short-term fluctuations, anticipatory planning at medium-term cycles, and long-term co-evolution. These frequency-domain results explain why aggregate CCF metrics (Table 4) and time-domain Granger tests often diverge from spectral findings. Lead-lag relationships vary by time scale, with strongest predictive value at intermediate periodicities of 3−5 years, consistent with the Netherlands’ strong DM performance (Table 6).

Table 5.

Spectral Coherence and Phase Spectrum Analysis. Mean Coherence Measures Strength Of Relationship Across Frequency Bands (0–1 Scale). Only 25% of Pairs Exceeded High-Coherence Threshold (>0.4). Phase Lag Indicates Temporal Precedence: Positive Values Mean Divorce Rate Leads Search Behavior (Reactive). Negative Values Mean Search Leads Divorce (Anticipatory). Synchronous Relationships (|lag| < 0.1 years) Suggest Contemporaneous Dynamics. Results Reveal Heterogeneous Lead-Lag Structures Across Countries and Search Terms

Country	Term	Mean Coherence	High Coh. (>0.4)	Median Phase Lag (yrs)	Phase Direction
Germany	Eheprobleme (marriage problems)	0.434	Yes	0.707	Divorce leads
Germany	Scheidungsanwalt (divorce lawyer)	0.3	No	0	Synchronous
Germany	Beziehungsprobleme (relationship problems)	0.344	No	0.025	Divorce leads
Germany	Scheidung (divorce)	0.255	No	0.459	Divorce leads
Germany	Nach der scheidung (after divorce)	0.131	No	0.797	Divorce leads
Germany	Eheberatung (marriage counseling)	0.269	No	−0.061	Search leads
Netherlands	Echtscheiding (divorce)	0.349	No	−0.293	Search leads
Netherlands	Huwelijksproblemen (marriage problems)	0.207	No	0.989	Divorce leads
Netherlands	Relatieproblemen (relationship problems)	0.199	No	−0.407	Search leads
Netherlands	Relatietherapie (relationship therapy)	0.26	No	−1.147	Search leads
Netherlands	Echtscheidingsadvocaat (divorce lawyer)	0.444	Yes	0.666	Divorce leads
Netherlands	Na de scheiding (after divorce)	0.343	No	−0.060	Search leads
United States	Marriage problems	0.306	No	0.189	Divorce leads
United States	Marriage counseling	0.454	Yes	0.307	Divorce leads
United States	Divorce lawyer	0.103	No	−0.967	Search leads
United States	Divorce	0.268	No	0.025	Divorce leads
United States	Relationship problems	0.583	Yes	0.846	Divorce leads
United States	After divorce	0.367	No	0.524	Divorce leads
United Kingdom	Divorce	0.356	No	0.145	Divorce leads
United Kingdom	Marriage problems	0.056	No	0.568	Divorce leads
United Kingdom	Relationship problems	0.139	No	1	Divorce leads
United Kingdom	Life after divorce	0.473	Yes	1	Divorce leads
United Kingdom	Divorce lawyer	0.678	Yes	−0.388	Search leads
United Kingdom	Marriage counseling	0.043	No	−0.423	Search leads

Note. Phase lag derived from cross-spectrum. Positive phase indicates divorce leads at given frequency. Negative indicates search leads. Only 6 pairs (25%) exceeded high-coherence threshold.

Table 6.

Forecasting Performance: Diebold-Mariano Test Results. Positive DM Statistic Indicates Augmented Model (Incorporating Google Trends Changes) Outperforms Baseline AR Model. Higher DM Values Suggest Greater Forecast Improvement. After FDR Correction, 8 of 24 Pairs (33.3%) Showed Statistically Significant Improvements, Indicating Google Trends Data Provides Incremental Predictive Value for Divorce Rates in Select Country-Term Combinations, Particularly for Relationship Problem Searches

Country	Term	DM Statistic	DM p-Value	DM p-adj (FDR)	DM Sig	Nowcast n
Germany	Eheprobleme (marriage problems)	2.811	0.005	0.017	Yes	11
Germany	Scheidungsanwalt (divorce lawyer)	1.408	0.159	0.348	No	11
Germany	Beziehungsprobleme (relationship problems)	2.396	0.017	0.05	Yes	11
Germany	Scheidung (divorce)	1.729	0.084	0.224	No	11
Germany	Nach der scheidung (after divorce)	1.428	0.153	0.348	No	11
Germany	Eheberatung (marriage counseling)	1.134	0.257	0.382	No	11
Netherlands	Echtscheiding (divorce)	2.971	0.003	0.014	Yes	11
Netherlands	Huwelijksproblemen (marriage problems)	3.198	0.001	0.014	Yes	11
Netherlands	Relatieproblemen (relationship problems)	3.425	0.001	0.014	Yes	11
Netherlands	Relatietherapie (relationship therapy)	2.966	0.003	0.014	Yes	11
Netherlands	Echtscheidingsadvocaat (divorce lawyer)	2.873	0.004	0.016	Yes	11
Netherlands	Na de scheiding (after divorce)	3.083	0.002	0.014	Yes	11
United States	Marriage problems	1.074	0.283	0.382	No	11
United States	Marriage counseling	1.18	0.238	0.382	No	11
United States	Divorce lawyer	1.145	0.252	0.382	No	11
United States	Divorce	1.065	0.287	0.382	No	11
United States	Relationship problems	1.145	0.252	0.382	No	11
United States	After divorce	1.171	0.242	0.382	No	11
United Kingdom	Divorce	0.512	0.608	0.697	No	11
United Kingdom	Marriage problems	0.438	0.661	0.697	No	11
United Kingdom	Relationship problems	0.296	0.767	0.767	No	11
United Kingdom	Life after divorce	0.429	0.668	0.697	No	11
United Kingdom	Divorce lawyer	0.47	0.638	0.697	No	11
United Kingdom	Marriage counseling	0.583	0.56	0.697	No	11

Note. Positive DM statistic indicates augmented model outperforms baseline AR; p-adj = FDR-adjusted p-value at α = 0.05; the Netherlands achieved 100% (6/6) significant forecast improvements; US and UK had 0% despite high correlations.

Frequency and Spectral Insights

Spectral coherence analysis revealed moderate average coherence of 0.307, with only 25% (6/24) of pairs exceeding the high-coherence threshold of 0.4 (Table 5). This indicates that strong frequency-specific co-movement between search intensity and divorce rates is the exception, suggesting that the two processes operate on partially independent cyclical patterns. High coherence concentrated in proactive terms like “marriage counseling” and “relationship therapy,” implying preventive information-seeking shows stronger synchronization with divorce outcomes than reactive searches.

Phase spectrum analysis produced an average median phase lag of 0.188 years. Positive phase lags, where search activity leads divorce rates, appeared in 62.5% of pairs (15/24), supporting CCF evidence of temporal precedence. However, Table 5 shows nuanced patterns: although CCF optimal lags indicate time-domain precedence, phase results often reveal divorce leading search within high-coherence pairs at dominant frequencies, reflecting methodological differences between overall temporal alignment (CCF) and frequency-specific lead-lag identification.

Negative phase lags were observed where divorce is highly normalized, notably UK “divorce lawyer” with strong coherence (0.678) but negative phase (−0.388), suggesting reactive or post-divorce searching rather than anticipatory planning at certain frequencies. Conversely, UK “life after divorce” showed high coherence (0.473) and strong positive phase (+1.0 year), indicating divorce cycles precede post-divorce searches.

Rolling window CCF analysis revealed temporal instability, with optimal lags shifting across 4-year windows, underscoring the non-stationary and context-dependent nature of search and divorce dynamics shaped by changing digital use, economic conditions, and policy environments.

Forecasting and Predictive Performance

Rolling-origin nowcasting compared baseline autoregressive (AR) models with augmented specifications incorporating first-differenced Google Trends data. Diebold-Mariano tests with FDR correction identified significant forecast improvements in 33.3% (8/24) of country-term pairs (Table 6), concentrated in Germany and especially the Netherlands, with no gains observed for the United Kingdom or United States.

Geographic heterogeneity in predictive performance indicates that the value of Google Trends for nowcasting divorce rates depends on institutional and cultural contexts. The Netherlands shows consistent forecasting gains with 100% success rate across all terms (DM = 2.87–3.43, all p-adj <0.02), indicating online searches capture the full divorce decision process.

Germany’s partial success indicates problem-recognition searches hold predictive value, while legal and counseling terms do not add information beyond autoregressive dynamics, aligning with episodic, crisis-driven search behavior observed in Germany.

In the highly digitized, high-divorce contexts of the United States and United Kingdom, widespread access to divorce information and normalization of marital dissolution weaken the marginal predictive value of searches. Despite strong correlations (U.S. average |CCF| = 0.887) and high coherence for specific terms (UK “divorce lawyer” coherence = 0.678), search data did not improve forecasts beyond autoregressive models, suggesting information saturation where baseline models already capture underlying dynamics. Overall, DM tests confirm Google Trends adds meaningful predictive value primarily in information-sensitive settings.

Robustness, Limitations, and Theoretical Implications

To address the multiple testing problem in analyzing 24 country-term pairs across four hypothesis test families (Engle-Granger, Δ-Granger, VAR-Granger, Diebold-Mariano), Benjamini-Hochberg False Discovery Rate (FDR) correction was applied at α = 0.05. This adjustment substantially reduced statistically significant results, with only 10 out of 96 tests (10.4%) retaining significance (Table 7). This conservative approach ensures that reported findings are robust to inflated Type I error risk, though it may reduce statistical power given the limited 15-year observation window.

Table 7.

Summary of Statistical Significance After FDR Correction (q = 0.05). Success rate Indicates Proportion of Tests Remaining Significant After FDR Adjustment. Diebold-Mariano Tests Showed Highest Robustness (33.3%), While Cointegration and VAR-Granger Tests Showed Complete Attenuation (0%)

Test Family	Total Tests	Sig. (Unadjusted)	Sig. (FDR-Adjusted)	Success Rate
Engle-Granger cointegration	6	2 (33.3%)	0 (0%)	0%
Δ-Granger causality	24	8 (33.3%)	2 (8.3%)	8.30%
VAR-Granger causality	24	9 (37.5%)	0 (0%)	0%
Diebold-Mariano forecast	24	10 (41.7%)	8 (33.3%)	33.30%
Overall	96	29 (30.2%)	10 (10.4%)	10.40%

Note. Success rate = proportion retaining significance after FDR correction; DM tests showed highest robustness (33.3% retained), while VAR-Granger showed complete attenuation (0% retained).

Several limitations apply. First, the 5-year window limits statistical power of cointegration tests, likely contributing to significance loss after FDR correction (Table 7, row 1). Second, annual data constrains temporal resolution and may obscure higher-frequency dynamics, as median phase lags below 1 year (Table 5) indicate sub-annual effects. Third, Google Trends provides relative rather than absolute search volumes, preventing direct cross-country magnitude comparisons; differences such as Netherlands “relatietherapie” (M = 75.66) versus UK “life after divorce” (M = 14.21) reflect within-country normalization. Fourth, the analysis does not control for confounders such as economic recessions, legislative changes, or COVID-19. Gregory-Hansen breaks clustered in 2014 to 2016 and 2019 to 2020 (Table 3) suggest such exogenous factors may influence both search behavior and divorce rates.

Despite these constraints, the findings support digital behavioral indicators as useful proxies for institutional processes. Evidence of search activity leading divorce rates in 62.5% of cases (Table 4) and strong forecasting gains in selective contexts, notably the Netherlands with 100% DM success rate (Table 6), indicate that online searches capture latent decision dynamics before formal legal action. However, weak overall Granger causality (8.3% retaining significance) and substantial cross-country heterogeneity confirm that digital trace data complement rather than replace traditional demographic indicators, with predictive value shaped by cultural norms and institutional responsiveness.

Lead-lag patterns can be interpreted through theories of intention formation and help-seeking. The Theory of Planned Behavior argues that actions follow intentions shaped by attitudes, norms, and perceived control (Ajzen, 1991). Online searches represent a low-cost step that externalizes intention before formal action, explaining why search intensity often precedes divorce outcomes, consistent with the predominance of negative lags (Table 4). Google Trends therefore functions as a nowcasting indicator that can enhance short-term forecasts when paired with autoregressive baselines (Choi & Varian, 2012). However, digital indicators require calibration because algorithmic and behavioral shifts can destabilize measurement, and models integrating search data with official series and regular recalibration perform more reliably (Lazer et al., 2014). Digital traces increasingly surface ahead of official releases, anticipating real-world indicators by capturing shifts in information-seeking behavior linked to life events (Clements & Horwood, 2024; Golder & Macy, 2014; Mangono et al., 2021).

Query content matters. Solution-oriented searches such as “marriage counseling” expose users to therapeutic information that may enhance hope and self-efficacy, potentially reversing early divorce intentions. This aligns with the weak predictive relationship observed for U.S. “Marriage counseling” despite very high search intensity (CCF = 0.782, zero lag, Table 4; M = 80.87, Table 1), and with its positive phase lag indicating effects that operate mainly at the individual rather than aggregate level.

In contrast, problem-focused queries such as “marriage problems” and “relationship problems” may trigger diagnostic spirals driven by confirmation bias (Nickerson, 1998). Filter-bubble effects further reinforce this pattern (Allen, 2012). Engagement-optimizing personalization amplifies users’ initial frames (Golbeck, 2020; Zanathy, 2021). Evidence of algorithmic echo chambers confirms that this escalation from tentative concern to conviction is common. Several studies confirm algorithmic eco chambers that transform tentative concern into conviction. (Gombar & Boban, 2025; Jiang et al., 2019; Schoenebeck et al., 2025). These mechanisms likely explain why U.S. “Relationship problems” shows the strongest predictive relationship in the dataset (CCF = 0.975, Granger significant, Table 4), suggesting that searching itself forms part of the causal pathway toward divorce consideration.

Additional limitations warrant acknowledgment. Although term selection was theory driven, alternative terminology and evolving cultural usage may not be captured. Data reliability is also a concern, as Google Trends downloads may vary across days, particularly for low-volume terms (Behnen et al., 2020; Eichenauer et al., 2022; Franzén, 2023; Gummer & Oehrlein, 2023, 2025; Hölzl et al., 2025; Mavragani & Ochoa, 2019; Nuti et al., 2014). While we did not use multi-download averaging, reliance on annual aggregates and high-volume terms reduces this risk compared to monthly analyses of low-frequency queries. The consistency of results across four countries and multiple methods suggests substantive patterns rather than download artifacts. Nevertheless, future studies should apply multi-day sampling, especially for monthly data or rare queries. Future research should also integrate additional covariates (unemployment rates, legal reforms, and social policy changes) to isolate the independent informational contribution of online search behavior.

The observed non-stationarity in the search-divorce relationship reflects not only changing social behavior but also known measurement instabilities in Google Trends. Search indices can be distorted by sampling noise and frequency inconsistency, so variation in model performance may partly reflect data artifacts rather than true behavioral change (Eichenauer et al., 2022). Algorithmic adjustments can also alter search visibility and have previously caused misinterpretation of search-based predictors (Lazer et al., 2014). Looking ahead, increasing use of large language models may reduce the informational value of search data as users seek private guidance without generating explicit queries (del Rio-Chanona et al., 2024; Zhou & Li, 2024). As a result, LLM adoption becomes a new contextual factor that may gradually weaken the predictive capacity of search-derived measures.

Conclusion

This study examined whether Google Trends search intensity predicts official divorce rates across four Western nations. Using a comprehensive econometric framework, it found that digital distress signals are context-dependent rather than universally predictive. Although 62.5% of search terms preceded divorce rates by one to 2 years, only 8.3% remained Granger significant after false discovery correction. Forecasting gains were highly country-specific: the Netherlands showed consistent improvements (DM = 2.87–3.43, p < 0.02) across all search terms, indicating systematic capture of the entire divorce pathway. Germany demonstrated intermediate results with 33% forecast success (DM = 2.40–2.81) limited to problem-recognition terms, suggesting episodic crisis-driven engagement. The United States and United Kingdom showed no gains despite high search volumes, consistent with information saturation in normalized divorce cultures. Digital Behavioral Decoupling captures this divergence between online distress expression and institutional outcomes. Spectral analysis showed predictive value concentrated at three to 5 year periodicities, and algorithmic feedback loops indicated that problem-focused searches may amplify negativity while solution-oriented queries promote corrective engagement. Overall, Google Trends can serve as an early warning indicator only in contexts characterized by preventive, institutionally supported help-seeking. Future research should extend temporal coverage, integrate policy and macroeconomic moderators, and combine computational analysis with qualitative inquiry to map cultural mechanisms underlying digital distress signals.

Supplemental Material

Supplemental Material - From Search to Separation: Digital Behavioral Decoupling and the Predictive Power of Google Trends for Divorce Outcomes Across Four Western Nations

Supplemental Material for From Search to Separation: Digital Behavioral Decoupling and the Predictive Power of Google Trends for Divorce Outcomes Across Four Western Nations by Emre Can Kuran and Umut Kuran in Social Science Computer Review.

Footnotes

Acknowledgments

The authors thank the national statistical agencies - the U.S. Centers for Disease Control and Prevention, German Federal Statistical Office (Destatis), Statistics Netherlands (CBS), and the UK Office for National Statistics—for providing open access to divorce statistics. We also acknowledge Google for maintaining the publicly accessible Google Trends platform. No writing or editing assistance was received from third parties in the preparation of this article.

ORCID iDs

Emre Can Kuran

Umut Kuran

Ethical Considerations

Ethical approval was not required for this study. The research involves secondary analysis of publicly available aggregated data from Google Trends and official divorce statistics published by national statistical agencies (CDC, German Federal Statistical Office, Statistics Netherlands, and UK Office for National Statistics). No human participants were directly involved, no individual-level data were collected or analyzed, and all data sources are publicly accessible and anonymized at the population level. The study complies with institutional research ethics guidelines for the use of publicly available secondary data.

Consent to Participate

Not applicable. This study did not involve human participants.

Consent for Publication

Not applicable. This study did not involve individual participant data, images, or personally identifiable information.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

The data that support the findings of this study are openly available from the following sources: Google Trends data at ; United States divorce statistics from the Centers for Disease Control and Prevention (CDC); German data from the Federal Statistical Office (Destatis); Dutch data from Statistics Netherlands (CBS); and United Kingdom data from the Office for National Statistics (ONS). The analytical dataset compiled for this study and replication code are available from the corresponding author upon reasonable request.

Supplemental Material

Supplemental Material for this article is available online.

Author Biographies

Emre Can Kuran was born in Izmir, Turkey, in 1997. He received his B.S. and M.Sc. degrees in Computer Engineering from Harran University, Turkey, in 2020. He is currently a Ph.D. student in Computer Engineering at Bandırma Onyedi Eylül University, Turkey. His research interests include artificial intelligence, image processing, computer vision, data science, machine learning, and software development.

Umut Kuran was born in Sanlıurfa, Turkey, in 1982. He received the B.S. degree in mathe matics from Harran University, Turkey, in 2005, the M.Sc. degree in mathematics from Celal Bayar University, Turkey, in 2007, and in computer science from Rutgers, The State University of New Jersey, USA, in 2013. He is currently working as a full time lecturer in Computer Engineering, Harran University, Turkey. His research interests include artificial intelligence, image processing, statistics, neural networks, fuzzy logic, numerical computation and cryptology.

References

Afifi

W. A.

Dillow

M. R.

Morse

(2004). Examining predictors and consequences of information seeking in close relationships. Personal Relationships, 11(4), 429–449. https://doi.org/10.1111/j.1475-6811.2004.00091.x

Aghayeeabianeh

Stamatel

J. P.

(2022). The effects of sociopolitical factors on divorce rates after political upheavals in post-soviet countries. Comparative Sociology, 21(3), 300–319. https://doi.org/10.1163/15691330-bja10055

Ajzen

(1991). The theory of planned behavior. Organizational Behavior and Human Decision Processes, 50(2), 179–211. https://doi.org/10.1016/0749-5978(91)90020-T

Allen

(2012). The filter bubble: What the internet is hiding from you. Policy Perspectives, 19, 131. https://doi.org/10.4079/pp.v19i0.10431

Banerjee

(2018). How does the world google the internet, anxiety, and happiness? Cyberpsychology, Behavior, and Social Networking, 21(9), 569–574. https://doi.org/10.1089/cyber.2018.0206

Behnen

Kessler

Kruse

Gómez

J. M.

Schoenmakers

Zerr

(2020). Experimental evaluation of scale, and patterns of systematic inconsistencies in google trends data. Communications in Computer and Information Science, 1323, 374–384. https://doi.org/10.1007/978-3-030-65965-3_25

Berger

L. M.

Ferrari

Leturcq

Panico

Solaz

(2021). COVID-19 lockdowns and demographically-relevant Google Trends: A cross-national analysis. PLOS ONE, 16(3), Article e0248072. https://doi.org/10.1371/journal.pone.0248072

Boy

(2023). Google TrendsTM dynamics for the guidance of open-source intelligence: Augmentation of Social-Media and Survey surveillance of population mental health. 2023 IEEE international symposium on technology and society (ISTAS). IEEE (pp. 1–9). https://doi.org/10.1109/ISTAS57930.2023.10305931

Brodeur

Clark

A. E.

Fleche

Powdthavee

(2021). COVID-19, lockdowns and well-being: Evidence from google trends. Journal of Public Economics, 193, Article 104346. https://doi.org/10.1016/j.jpubeco.2020.104346

10.

Castrillón

O. D.

(2021). Predicción del divorcio por medio de técnicas inteligentes. Información Tecnológica, 32(5), 111–120. https://doi.org/10.4067/S0718-07642021000500111

11.

Centers for Disease Control and Prevention . (2025). Marriage and divorce. Centers for Disease Control and Prevention. Retrieved October 19, 2025, from. https://www.cdc.gov/nchs/fastats/marriage-divorce.htm

12.

Cheng

L.-H.

Lin

Y.-H.

Lin

S.-H.

Hsia

P.-H.

Chen

I.-M.

(2025). The COVID-19 pandemic and internet searches for divorce with stay-at-home behavior as a mediator: A multinational longitudinal study. Journal of Social and Personal Relationships, 42(8), 2212–2230. https://doi.org/10.1177/02654075251339268

13.

Choi

Varian

(2012). Predicting the present with Google trends. Economic Record, 88(s1), 2–9. https://doi.org/10.1111/j.1475-4932.2012.00809.x

14.

Clements

Horwood

(2024). Shadow footprints and the provision of digital behavioral data. Journal of Digital Social Research, 6(3), 187–205. https://doi.org/10.33621/jdsr.v6i3.33352

15.

Cohen

P. N.

(2019). The coming divorce decline. Socius: Sociological Research for a Dynamic World, 5, 2378023119873497. https://doi.org/10.1177/2378023119873497

16.

Cooney

T. M.

(2022). Introduction to special issue on “Divorce and the Life Course.”. Social Sciences, 11(5), 202. https://doi.org/10.3390/socsci11050202

17.

D’Avanzo

Pilato

Lytras

(2017). Using Twitter sentiment and emotions analysis of Google Trends for decisions making. Program, 51(3), 322–350. https://doi.org/10.1108/PROG-02-2016-0015

18.

Del Rio-Chanona

R. M.

Laurentsyeva

Wachs

(2024). Large language models reduce public knowledge sharing on online Q&A platforms. PNAS Nexus, 3(9), 400. https://doi.org/10.1093/pnasnexus/pgae400

19.

Derwanto

(2022). Implementation of triple exponential smoothing method in the number of divorce rate in North Sumatra. ZERO: Jurnal Sains, Matematika Dan Terapan, 6(1), 9. https://doi.org/10.30829/zero.v6i1.12452

20.

Dickey

D. A.

Fuller

W. A.

(1979). Distribution of the estimators for autoregressive time series with a unit root. Journal of the American Statistical Association, 74(366), 427–431. https://doi.org/10.1080/01621459.1979.10482531

21.

Doss

B. D.

Rhoades

G. K.

Stanley

S. M.

Markman

H. J.

(2009). Marital therapy, retreats, and books: The who, what, when, and why of relationship help‐seeking. Journal of Marital and Family Therapy, 35(1), 18–29. https://doi.org/10.1111/j.1752-0606.2008.00093.x

22.

Eichenauer

V. Z.

Indergand

Martínez

I. Z.

Sax

(2022). Obtaining consistent time series from google trends. Economic Inquiry, 60(2), 694–705. https://doi.org/10.1111/ecin.13049

23.

Engle

R. F.

Granger

C. W. J.

Grangeri

C. W. J.

(1987). Co-integration and error correction: Representation, estimation, and testing. Source: Econometrica Econometrica, 55(2), 81–111.

24.

Entwistle

Horn

A. B.

Meier

Boyd

R. L.

(2021). Dirty laundry: The nature and substance of seeking relationship help from strangers online. Journal of Social and Personal Relationships, 38(12), 3472–3496. https://doi.org/10.1177/02654075211046635

25.

Esping-Andersen

(1999). Social foundations of postindustrial economies. Oxford University PressOxford. https://doi.org/10.1093/0198742002.001.0001

26.

Ford

M. T.

Jebb

A. T.

Tay

Diener

(2018). Internet searches for affect‐related terms: An indicator of subjective well‐being and predictor of health outcomes across US states and metro areas. Applied Psychology: Health and Well-Being, 10(1), 3–29. https://doi.org/10.1111/aphw.12123

27.

Franzén

(2023). Big data, big problems: Why scientists should refrain from using google Trends. Acta Sociologica, 66(3), 343–347. https://doi.org/10.1177/00016993221151118

28.

German Federal Statistical Office . (2024). Marriages, divorces (time series). German Federal Statistical Office. https://www.destatis.de/EN/Themes/Society-Environment/Population/Marriages-Divorces-Life-Partnerships/Tables/lrbev06.html

29.

Golbeck

(2020). Optimizing for engagement can be harmful. There are alternatives. IEEE Intelligent Systems, 35(4), 117–118. https://doi.org/10.1109/MIS.2020.3005577

30.

Golder

S. A.

Macy

M. W.

(2014). Digital footprints: Opportunities and challenges for online social research. Annual Review of Sociology, 40(1), 129–152. https://doi.org/10.1146/annurev-soc-071913-043145

31.

Gombar

Boban

(2025). Research on the impact of algorithmic echo chambers on perceptions and attitudes of social network users in a digital Society. 2025 MIPRO 48th ICT and Electronics Convention, 259(9), 1026–1033. https://doi.org/10.1109/MIPRO65660.2025.11131918

32.

Google Trends (2025). https://trends.google.com/trends/explore?geo=US&hl=en

33.

Granger

C. W. J.

Newbold

(1974). Spurious regressions in econometrics. Journal of Econometrics, 2(2), 111–120. https://doi.org/10.1016/0304-4076(74)90034-7

34.

Gummer

Oehrlein

A.-S.

(2023). Using google trends data to learn more about survey participation. Social Science Computer Review, 41(6), 1968–1985. https://doi.org/10.1177/08944393221129179

35.

Gummer

Oehrlein

A.-S.

(2025). Using google trends data to study high-frequency search terms: Evidence for a reliability-frequency continuum. Social Science Computer Review, 43(4), 814–826. https://doi.org/10.1177/08944393241279421

36.

Harris

C. R.

Millman

K. J.

van der Walt

S. J.

Gommers

Virtanen

Cournapeau

Wieser

Taylor

Berg

Smith

N. J.

Kern

Picus

Hoyer

van Kerkwijk

M. H.

Brett

Haldane

del Río

J. F.

Wiebe

Peterson

Oliphant

T. E.

(2020). Array programming with NumPy. Nature, 585(7825), 357–362. https://doi.org/10.1038/s41586-020-2649-2

37.

W.-C.

(2025). Explaining the prevalence of marital conflict: Conceptual bifurcation and sociological explanations. Frontiers in Sociology, 10, 1490385. https://doi.org/10.3389/fsoc.2025.1490385

38.

Hölzl

Keusch

Sajons

(2025). The (Mis)use of google trends data in the social sciences - A systematic review, critique, and recommendations. Social Science Research, 126, Article 103099. https://doi.org/10.1016/j.ssresearch.2024.103099

39.

Hunter

J. D.

(2007). Matplotlib: A 2D graphics environment. Computing in Science & Engineering, 9(3), 90–95. https://doi.org/10.1109/MCSE.2007.55

40.

Jiang

Chiappa

Lattimore

György

Kohli

(2019). Degenerate feedback loops in recommender systems. Proceedings of the 2019 AAAI/ACM conference on AI (pp. 383–390). Ethics, and Society. https://doi.org/10.1145/3306618.3314288

41.

Johansen

(1988). Statistical analysis of cointegration vectors. Journal of Economic Dynamics and Control, 12(2–3), 231–254. https://doi.org/10.1016/0165-1889(88)90041-3

42.

Kumar

Saha

Mondal

Mitra

Chaudhuri

A. K.

(2023). Prediction of the probability of divorce using machine learning algorithms. International Journal of Engineering Technology and Management Sciences, 7(2), 195–209. https://doi.org/10.46647/ijetms.2023.v07i02.024

43.

Kwiatkowski

Phillips

P. C. B.

Schmidt

Shin

(1992). Testing the null hypothesis of stationarity against the alternative of a unit root. How sure are we that economic time series have a unit root? Journal of Econometrics, 54(1–3), 159–178. https://doi.org/10.1016/0304-4076(92)90104-Y

44.

Lazer

Kennedy

King

Vespignani

(2014). The parable of google flu: Traps in big data analysis. In Science, 343(6176), 1203–1205. https://doi.org/10.1126/science.1248506

45.

Mangono

Smittenaar

Caplan

Huang

V. S.

Sutermaster

Kemp

Sgaier

S. K.

(2021). Information-seeking patterns during the COVID-19 pandemic across the United States: Longitudinal analysis of google trends data. Journal of Medical Internet Research, 23(5), Article e22933. https://doi.org/10.2196/22933

46.

Mavragani

Ochoa

(2019). Google trends in infodemiology and infoveillance: Methodology framework. JMIR Public Health and Surveillance, 5(2), Article e13439. https://doi.org/10.2196/13439

47.

McKinney

(2010). Data structures for statistical computing in python. In: Proceedings of the 9th python in science conference, Austin, Texas, June 28–30, 2010. https://doi.org/10.25080/majora-92bf1922-00a

48.

Mortelmans

(2021). Causes and consequences of family dissolution in Europe and post-divorce families. In Research handbook on the sociology of the family. Edward Elgar Publishing. https://doi.org/10.4337/9781788975544.00024

49.

Nickerson

R. S.

(1998). Confirmation bias: A ubiquitous phenomenon in many guises. Review of General Psychology, 2(2), 175–220. https://doi.org/10.1037/1089-2680.2.2.175

50.

Nuti

S. V.

Wayda

Ranasinghe

Wang

Dreyer

R. P.

Chen

S. I.

Murugiah

(2014). The use of google trends in health care research: A systematic review. PLoS ONE, 9(10), Article e109583. https://doi.org/10.1371/journal.pone.0109583

51.

OECD . (2025). SF3.1: Marriage and divorce rates. Retrieved November 22, 2025, from OECD. https://webfs.oecd.org/els-com/Family_Database/SF_3_1_Marriage_and_divorce_rates.pdf

52.

Office for National Statistics . (2025). Divorces and dissolutions in England and Wales: 2023. Retrieved October 19, 2025, from Office for National Statistics. https://www.ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/divorce/datasets/divorcesinenglandandwales

53.

Pešić-Jenaćković

(2018). Divorce: Historical-legal, demographic and sociological aspects. Socioloski Godisnjak, 13, 41–58. https://doi.org/10.5937/SocGod1813041P

54.

Phillips

P. C. B.

Perron

(1988). Testing for a unit root in time series regression. Biometrika, 75(2), 335–346. https://doi.org/10.1093/biomet/75.2.335

55.

Roddy

M. K.

Rothman

Cicila

L. N.

Doss

B. D.

(2019). Why do couples seek relationship help online? Description and comparison to in‐person interventions. Journal of Marital and Family Therapy, 45(3), 369–379. https://doi.org/10.1111/jmft.12329

56.

Sadiq Fareed

M. M.

Raza

Zhao

Tariq

Younas

Ahmed

Ullah

Jillani

S. F.

Abbas

Aslam

(2022). Predicting divorce prospect using ensemble learning: Support vector machine, linear model, and neural network. Computational Intelligence and Neuroscience, 2022,(1) 1–15. https://doi.org/10.1155/2022/3687598

57.

Salts

C. J.

(1985). Divorce stage theory and therapy: Therapeutic implications throughout the divorcing process. Journal of Psychotherapy & The Family, 1(3), 13–23. https://doi.org/10.1300/J287v01n03_03

58.

Schoenebeck

Goray

Gunal

(2025). Algorithmic self-diagnosis from targeted ads. Proceedings of the ACM on Human-Computer Interaction, 9(2), 1–19. https://doi.org/10.1145/3710926

59.

Solichah

Istiqomah

(2018). Process of decision-making divorces on women who are undergoing pregnancy in the first year of marriage. International Conference on Sustainable Health Promotion, 81–87. https://doi.org/10.29080/icoshpro.v1i0.19

60.

Statistics

Netherlands.

(2025). Marriage dissolutions; due to divorce and death. Retrieved October 19, 2025, from Statistics Netherlands. https://www.cbs.nl/en-gb/figures/detail/37425eng

61.

Trillingsgaard

Sørensen

T. K.

Fentz

H. N.

(2019). Relationship help‐seeking in a Danish population sample. Journal of Marital and Family Therapy, 45(3), 380–394. https://doi.org/10.1111/jmft.12347

62.

Virtanen

Gommers

Oliphant

T. E.

Haberland

Reddy

Cournapeau

Burovski

Peterson

Weckesser

Bright

van der Walt

S. J.

Brett

Wilson

Millman

K. J.

Mayorov

Nelson

A. R. J.

Jones

Kern

Larson

Vázquez-Baeza

(2020). SciPy 1.0: Fundamental algorithms for scientific computing in python. Nature Methods, 17(3), 261–272. https://doi.org/10.1038/s41592-019-0686-2

63.

Waskom

(2021). Seaborn: Statistical data visualization. Journal of Open Source Software, 6(60), 3021. https://doi.org/10.21105/joss.03021

64.

Yudistira

Romlah

Yulianto

Faisol

Mardianto

M. F. F.

(2025). Prediction of divorce data in pamekasan district based on comparison of exponential smoothing and moving average. Tensor: Pure and Applied Mathematics Journal, 5(2), 67–78. https://doi.org/10.30598/tensorvol5iss2pp67-78

65.

Zanathy

(2021). Burst the bubble: How to defend freedom of expression from algorithmic personalization. International Journal of Scientific and Research Publications (IJSRP), 11(7), 95–99. https://doi.org/10.29322/IJSRP.11.07.2021.p11513

66.

Zhou

(2024). Understanding user switch of information seeking: From search engines to generative AI. Journal of Librarianship and Information Science, 09610006241244800. https://doi.org/10.1177/09610006241244800

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.01 MB