Personalisation and the shared front page

1. What is actually being personalised

On most news sites, personalisation touches a small number of surfaces: the order of a recommendation module, the selection in a “for you” rail, the topics in a newsletter, the timing of a push notification, and occasionally the ordering below a fixed editorial block. It rarely touches the words of an article, the headline or the top of the front page. Being precise about this matters, because readers frequently assume more adaptation than exists and publishers frequently claim more than they have built. A public description of exactly which surfaces adapt, and which are identical for everyone, is a cheap and effective piece of transparency.

2. Segments, not individuals

Genuine one-to-one personalisation is rare outside very large services, because it needs data volumes that most publishers do not have and produces unstable results for infrequent visitors. What actually runs is segmentation: readers are grouped by coarse behaviour — sports-heavy, local-heavy, one-visit-a-month, weekday-morning — and each group receives a variant. Segments are easier to audit than individual models, because a desk can look at what each variant contained yesterday. They also fail in a specific way: a reader with mixed interests gets assigned to their dominant one and then sees that dominance reinforced.

3. What the filter-bubble evidence supports

The strong version of the filter-bubble claim — that algorithmic feeds mechanically seal readers into ideologically homogeneous worlds — is not well supported by the research literature, which repeatedly finds that heavy internet users encounter more varied sources than light users, and that self-selection matters at least as much as ranking. The weaker version is better supported: systems can amplify existing preferences at the margin, reduce incidental exposure to subjects a reader would never choose, and concentrate attention on a small number of items. That weaker effect is the one a publisher can actually measure and mitigate on its own site.

4. Incidental exposure and what a front page is for

A traditional front page performed a civic function that had nothing to do with preference: it put the same small set of items in front of everyone who arrived, including subjects they would not have selected. Personalised surfaces erode that by construction. The fix is not to abandon adaptation but to define an editorial floor — a fixed block that is identical for every reader, before any personalised module appears. Publishing the size and position of that floor turns an internal compromise into a stated commitment, and makes it much harder to shrink quietly under performance pressure.

5. Consent, profiling and the GDPR

Building reader profiles is processing of personal data. That requires a lawful basis, a clear notice describing the logic in intelligible terms, purpose limitation, a retention period and a working route for access, objection and erasure. Where profiling relies on information stored on or read from a device, ePrivacy consent requirements generally apply as well. In Ireland the supervisory authority is the Data Protection Commission. A personalisation feature that cannot be switched off, or whose profile a reader cannot see or delete, is a compliance problem before it is an editorial one — take qualified advice on your specific arrangement.

6. Newsletters as light personalisation

Topic-selected newsletters are personalisation that the reader controls and can inspect. The reader chose the topics, can see what they chose, and can change or cancel it in one action. That transparency makes them a reasonable default for publishers who want relevance without opaque inference, and it produces a cleaner consent record. The cost is coverage: readers under-select, choosing two topics and missing everything else the desk publishes. Editors compensate by keeping a short cross-cutting section in every edition, which restores incidental exposure without overriding the reader’s stated choice.

7. Measuring what personalisation does

Personalisation is usually evaluated on engagement, which guarantees a favourable answer, because the system was optimised for engagement. More informative measures compare the personalised experience with a control: the number of distinct sections seen per reader per week, the share of readers exposed to at least one public-interest item, the overlap between what two random readers were shown, and the concentration of impressions across the catalogue. Report these alongside engagement in the same review. A system that lifts engagement while halving section diversity has made a trade that the newsroom should decide on knowingly.

8. Governing the rules

Personalisation logic tends to be owned by whoever built it, which means editorial consequences are decided in a product backlog. Better arrangements give the rules a named editorial owner, a written specification in plain language, a change log, and a review at a fixed interval that examines outputs rather than intentions. The review should look at what actual readers were shown in a sample week, not at what the configuration says they should have been shown. Drift between the two is the normal state of any deployed system, and finding it is the point of the exercise.

Layers of personalisation

The right-hand column is the part worth publishing. Readers can evaluate a stated floor; they cannot evaluate an assurance that the system is balanced.

Personalisation review

• A public page states exactly which surfaces adapt and which are identical for all.
• An editorial floor is defined, sized and positioned before any adaptive module.
• A control group receives the non-personalised experience for comparison.
• Section diversity and impression concentration are reported next to engagement.
• Readers can view, reset and switch off their profile without creating an account.
• The lawful basis, retention period and objection route are documented.
• The personalisation specification is written in plain language and version-dated.
• A named editor owns the rules and signs off changes.
• The review examines what readers were actually shown in a sample week.

Reviews that examine configuration rather than output miss drift, which is where almost all personalisation problems live.

Questions readers ask