The Algorithm Experiment methodology

We created four personas based on typical 13 year-olds based in the UK (see details below). Each persona was on two platforms each, scrolling up on average 30 minutes a day, across seven days. Individual UK-based VPNs were purchased for each persona and used throughout the experiment.

Gmail accounts were set up for each persona, and during the opening of social media accounts, dates of birth were given that put each persona at just turning 13 within the last month. During the set-up of profiles, interests were listed in line with the pre-defined interests set out in the personas if the platforms asked for them (e.g. TikTok). A list of pre-defined influencers were also followed on day 1 in line with interests.

During each session, researchers took a randomised viewing behaviour pattern to mimic normal viewing behaviour. This means some videos were skipped within 1–2 seconds, others were watched for 10–30 seconds and some random videos were watched nearly in full or briefly re‑watched. This reflects evidence that teenagers’ attention during short‑form video use is fragmented and reactive rather than controlled .

There were limited micro-interactions. To mirror realistic but low‑effort engagement:

• Occasional likes were given to content perceived as entertaining or impressive, or line with interests
• Likes were not strategic, repeated or used to target specific content categories
• Some comments section were read (but no commenting happened)
• No commenting, sharing or direct messaging occurred.

Neutral search terms in line with pre-defined interests were used on day 2, to mimic typical behavior. Provocative search terms were used on day 6 or 7 of the experiment to analyse how this affected the algorithm.

After each session, researchers filled out a predetermined evaluation form to capture the qualitative and quantitative data. Content was flagged as concerning or harmful if it included:

• Sexualised material unsuitable for under‑16s
• Extreme or adult fitness regimes, diets or body transformations, or looksmaxxing (which is defined as a process of maximizing one’s own physical attractiveness through extreme measures, which originated on male incel / manosphere message boards in the 2010s)
• Encouragement of poor mental health behaviours or isolation
• Eating disorder or self‑harm related encouragement
• Graphic violence or aggressive ideology
• Misogynistic, hateful, racist or radicalising narratives
• Explicit drug use or criminal behaviour

The research design was informed by Ofcom’s Children’s Media Literacy (2025), Children and Parents: Media Use and Attitudes (2025), and Children’s Media Lives (2025) reports, which document extensive children / teenage platform use, feed‑driven discovery, and recurring exposure to age‑inappropriate content.