Do suicide-related internet searches track suicide deaths, and what do search pages show?

What went wrong with Google Flu Trends

Google Flu Trends (GFT) estimated flu levels from what people typed into Google. In the 2011–2012 season it overshot the actual US flu level by more than 50%. From 21 August 2011 to 1 September 2013 it reported too much flu in 100 of 108 weeks. The authors discuss possible causes, such as changes to Google search itself, but do not test them (Lazer et al., Science 2014).

Search data still helped when paired with official figures. GFT combined with delayed figures from the US Centers for Disease Control (CDC) had a lower error than either source alone. The error measure here is mean absolute error (MAE), the average size of the miss: 0.232 combined, 0.311 for the CDC figures alone, 0.486 for GFT alone. This is about flu, not mental health; it shows search data were useful only next to official figures (Lazer et al. 2014).

Do suicide-related searches track suicide deaths?

A study of the USA, Germany, Austria and Switzerland compared suicide-related search volumes with suicide rates. Significant links appeared in 8.34% of tests, against 5% expected by chance alone, so roughly what chance would produce. The data were monthly national series (USA and Switzerland 2004–2010, Germany and Austria 2004–2012); they say nothing about individuals or other countries (Tran et al., PLoS One 2017).

The same authors argue that forecasting suicide from searches assumes two things: that searchers are mostly people with suicidal thoughts, and that searching is strongly linked to suicidal behaviour. They argue both are likely not met. This is an argument from their results and earlier work, not a direct measurement (Tran et al. 2017).

An earlier US study (CDC statistics, 2004–2007) found that Google searches for suicide-related terms went down as the general suicide rate went up, but went up with suicide and self-injury among people aged 15–25. This rests on only 3–4 yearly data points, shows no cause, and the direction differs by age (McCarthy, J Affect Disord 2010).

In Ireland (2004–2015), a model using one search term forecast monthly suicide counts with a mean absolute error of 4.14, against 6.87–10.35 for comparison models without search data. A check was also made on published UK data. This is one small country, and a forecast of counts, not of who is at risk (Barros et al., IJERPH 2019).

What does the search page show?

In May 2007 researchers in the UK ran 12 searches about suicide methods on the 4 most popular UK search engines and looked at the top ten results. Of 480 results, 62 (13%) were prevention or support sites and 45 (9%) were sites encouraging suicide. In the same sample, 21% of support or prevention sites and 55% of academic or policy sites also contained information about methods. This is a snapshot of what was listed in 2007; what people clicked was not measured, and the presence of information is not proof of harm (Biddle et al., BMJ 2008).

When the same search was repeated in 2014, formal help sites fell from 13% to 6.5% of results, while blogs and discussion forums rose from 3% to 18.5%. Searches by people looking for help were not studied, so the overall balance of helpful and harmful sites cannot be judged (Biddle et al., J Affect Disord 2016).

Other studies used simulated users: automated accounts that run searches as a person would. In one experiment, previous search history did not change Google results, but whether a helpline box appeared was decided by the algorithm and then stayed the same for that user for the rest of the observation. The published abstract gives no display rates (Haim, Arendt & Scherr, Health Commun 2017).

In 137,937 simulated Google searches in German in April 2019, the helpline box appeared moderately often in Germany and Switzerland, never in Austria and Belgium, and never for searches about celebrities who died by suicide. This is a snapshot; Google's behaviour changes, and the abstract gives no exact rates (Arendt et al., Soc Sci Med 2020).

In 215,999 simulated US searches, the helpline box was shown often for harmful suicide queries but substantially less often when drug-related words were added. The abstract gives no exact rates, and the effect on deaths was not measured (Haim et al., Drug Alcohol Depend 2021).

How carefully are search-data studies done?

Google Trends is a free Google tool that shows how often a term was searched over time. A systematic review of 70 health studies using it found that 24% were on mental health or substance use, and only 7% described their search strategy fully enough for someone else to repeat it. The review is about how methods are reported, not about whether Trends measures illness (Nuti et al., PLoS One 2014).

A second review of Google Trends health papers (2006–2016; 109 included, 104 analysed) found that 39.4% used correlations and only 8.7% used the data for forecasting. It describes the methods used; it does not test whether they are valid (Mavragani et al., JMIR 2018).

How to cite this page

Cite one finding together with its limit, as given next to it on this page, and the original study it comes from. Give the date of this page, 2026-10-02, because search engines change and the studies are snapshots.