Designing a considerate discovery system for a site with too many articles to browse manually
More articles should mean more choice. In practice, readers can get the opposite experience.
They reopen the homepage and see the same cards. They enter the archive and spin through dozens of titles too fast to inspect. Someone then says, “Fine, show only six,” but if all six miss, the visitor leaves before candidate seven ever appears. Search does not fully rescue the situation because casual visitors often do not have a query yet.
The inventory exists. Search exists. Recommendations exist. Discovery still fails.
The problem is not a lack of content. It is that the site is making the reader pay the full cost of exploration.
0. Thirty-second version: do not shrink the warehouse; make the shopkeeper more helpful
A smaller visible set is useful, but it is only a viewing window.
If the first set misses, immediately offer directions: more first-hand stories, more practical pieces, comparisons, explanations, lighter material, or a completely different topic.
Do not force the reader to invent a fresh query every time.
The useful behavior is:
“Not this? Maybe you mean one of these.”
1. Why a homepage starts to feel stale
Online-news research describes user fatigue: interest drops when the same recommendation is shown repeatedly.[1]
In Bing Now logs, 15 news items were displayed at once. When consecutive visits occurred within two hours, an average of 11 items overlapped. Adding features for prior exposure and interaction improved ranking performance by up to 15% across all users and 34% for heavy users.[1]
The lesson is not “remove old articles.”
A three-month-old evergreen article can be new to a reader. A ten-minute-old article can feel stale after four impressions on the same day.
So separate content freshness from reader-specific freshness. Track exposure count, last exposure, clicks, and meaningful reads. Promote unseen items, gradually demote repeatedly ignored items, and widen recommendations after a click.
Any threshold should be learned from the site’s own exposure-to-response curve.
2. When an archive turns into a waterfall
A long archive is theoretically complete. In practice, a mouse wheel or finger can turn it into a stream of titles.
If one hundred cards exist but only eight are meaningfully inspected, the rest are inventory, not discovered content.
It is also wrong to say that “more choice is always bad.” A 2010 meta-analysis covering 50 experiments and 5,036 participants found an average choice-overload effect near zero, with substantial variation across studies.[2]
The better diagnosis is:
large inventories are not inherently bad; poorly differentiated options are expensive to compare.
A useful archive preserves orientation, makes differences legible, and keeps the next direction visible.
3. “Show six first” is not a complete solution
Small sets are easier to compare and harder to lose in fast scrolling. But they create omission risk.
If the first six are wrong, the session can end before the seventh candidate exists.
Treat six, eight, or ten cards as an observation window, not a magic number.
Place recovery paths next to it: another set, more practical, more personal, deeper on this topic, lighter, or completely different.
A 2026 two-week production experiment on NU.nl found that a BERT-based diversification method that preserved relevance while reducing within-list similarity increased click-through rates and perceived relevance.[3]
Diversity is not random rotation. It is “not too similar, still interesting.”
4. Keep the complete archive, but stop making it the front door
A complete archive still matters for exhaustive browsing, recovering a previously seen article, or deeply exploring a topic.
It simply should not carry the full burden of casual discovery.
Instead of “choose from 127 articles,” the discovery layer can say, “Here are a few promising directions from 127.”
Keep “view all,” precise filters, and date sorting as power-user tools.
Discovery and inventory management do not have to be the same screen.
5. The hidden weakness of search: the user must invent the query
Site search is powerful only when the user can express what they want.
People trust general web search to tolerate rough language. They may not grant the same trust to a smaller site. A casual visitor may not even have a stable intent yet.
A blank box saying “Enter keywords” therefore does little by itself.
Baymard’s large-scale search UX testing found that users often used autocomplete suggestions as a starting point for forming their own queries, not merely as exact suggestions to submit.[4]
The real value is not only typing assistance.
It is query-formulation assistance.
6. Considerate search suggests intents, not only titles
If someone types “work,” suggestions could be:
- I dislike the work itself
- the people are exhausting
- I am deciding whether to change jobs
- I want to work faster
- I want a first-hand story
These are not article titles. They are possible intents.
Even before typing, the surface can offer: How do I do this? Which should I choose? Why does this happen? What happened in practice? Show me something unexpectedly interesting.
After an article has been read, suggestions can use that context: compare it, understand why, see an opposite case.
In Japan, a 2026 meme popularized through a Mikiya Takasu tapioca-video context uses the phrase “Ki ga kiku nee,” roughly “Now that is thoughtful.”[5] The original phrase has competing origin claims, so the important point here is the reaction, not authorship.
A good discovery interface should produce exactly that reaction: the next useful move was already waiting.
7. You do not need an LLM call on every keystroke
Generate discovery metadata during the editorial pipeline.
For each article, store related queries, problem types, nearby topics, contrasting topics, tone tags, and likely next questions.
At runtime, rank a small set using current context, recent exposures, recent clicks, unseen-item bonuses, and list diversity.
Use the expensive model once to help generate structured metadata. Use cheap deterministic ranking on the live site.
That improves latency, cost, and explainability.
8. Do not optimize homepage, archive, and search as separate islands
A fast-rotating homepage, a paginated archive, and smart search can still feel disconnected.
Share the reader’s exploration state across them.
Homepage: show concrete candidates and reduce stale repetition on repeat visits.
Archive: show a manageable comparison window plus several next directions.
Search: work even when the box is empty, then suggest both query strings and intents while typing.
The shared layer is:
reader state + content metadata + exposure history.
9. CTR alone creates a headline-bait factory
If clicks are the only success signal, increasingly aggressive titles will win.
Measure the longer chain:
candidate → impression → click → arrival → meaningful read → next article → return.
For this problem, also measure repeated-exposure CTR, overlap between repeat visits, use of “show another set,” suggestion selection, suggestion editing, zero-result searches, exits after high-speed archive scrolling, continuation after the first set produces no click, and exposure received by previously unshown articles.
The key question is:
When the first six miss, does the reader reach candidate seven?
10. Conclusion: move from “search for it” to “is this what you mean?”
Growing sites eventually hit a warehouse problem.
Show everything and it becomes a stream. Show too little and you miss. Add search and people may not type. Keep recommendations fixed and they go stale. Personalize too hard and the world narrows.
The answer is not a single magic card count.
The inventory can be large. The screen can be small. The next directions can be many. The site can generate half of the query.
Do not merely replace a huge archive with six cards.
Put a thoughtful next move behind the six.
Implementation QC
- Can repeated exposure be tracked?
- Is content freshness separated from reader-specific freshness?
- Are there several recovery paths after a small set misses?
- Is the full archive preserved for power users?
- Can exploration start from an empty search field?
- Do suggestions help with intent, not only strings?
- Can related queries and next questions be created during publishing?
- Is a non-click treated cautiously rather than as an instant dislike?
- Are meaningful reading, next-article behavior, and return visits measured?
- Do previously unshown articles receive exploration slots?
- Are thresholds learned from local data rather than treated as universal laws?
Sources
- Hao Ma, Xueqing Liu, Zhihong Shen (2016), User Fatigue in Online News Recommendation doi.org
- Benjamin Scheibehenne, Rainer Greifeneder, Peter M. Todd (2010), Can There Ever Be Too Many Options? doi.org
- Robin Verachtert, Kim Falk, Christine Bauer (2026), Enhancing Diversity in News Recommendations Increases Click-Through Rates doi.org
- Baymard Institute (2024), Always Copy the Active Autocomplete Suggestion to the Search Field baymard.com
- Public meme circulation indexes and posts checked in September 2026; origin attribution is intentionally left unresolved
