Getting the most out of Discover search

What the search is actually doing

When you type a query into Discover, Eden embeds your text as a 3072-dimensional vector using Google's Gemini embedding model. Every post in the index is already embedded the same way. The search ranks posts by how close their vectors are to yours in that shared semantic space.

Two consequences worth knowing:

The post-side vector isn't built from the caption alone. For every post Eden ingests, we run a separate model (Reka) across the image or video frames and generate a short description of what's actually in the picture. That description, plus the post's title and body and a small set of AI-generated tags, all get embedded together. So a TikTok of someone making coffee with no caption is still findable by typing "morning coffee routine" because Reka described the coffee.

The match is on meaning, not words. "How small accounts price their first product" pulls posts about pricing pages, launch tweets, and pricing screenshots even when the posts never use the word "small". The vector picks up on intent.

What works well

The pattern that gets the best results: a short descriptive phrase, three to eight words, that describes the kind of post you want to study. Not the topic word, the specific intersection.

Some queries that pull strong feeds:

• morning routine ideas for early risers
• how solopreneurs price their first offer
• before and after kitchen renovation
• negative space minimalist photography
• B2B SaaS retention tactics
• cluttered home office desk setup
• controversial hot take about AI
• behind the scenes studio tour

Notice what these have in common. They name a topic, but they also imply a format, an angle, or a vibe. The more of those signals you bake in, the more the vector ranking earns its keep.

A few other patterns that work:

• Domain language is fine even if posts don't repeat it. "B2B SaaS retention metrics" finds posts that talk about churn, cohorts, MRR, NDR, expansion revenue. The model knows those concepts cluster together.
• Visual descriptions work. Typing what you'd want a photo of will surface photos like that, because Reka described every post's pixels. Try minimal walnut desk with single monitor or golden hour street photography in Tokyo.
• Tone and format work. personal essay about burning out will favor long-form Substack posts and reflective LinkedIn writeups over rage-bait tweets, because the embedding picks up on register, not just topic.

What doesn't work well

A few things to skip or rephrase.

Single common words. Typing design or ai or growth doesn't help the ranker, and you'll get a noisy top result set ordered by mostly other factors. Reach for a pillar chip or a follower-range filter instead. If you do want to search, expand: design portfolio sites for senior product designers is night-and-day different from design.

Exact handles or hashtags. @levelsio or #buildinpublic are substring territory, not semantic. The Creators tab is the right place for handles. For hashtag-style discovery, the underlying concept is usually what you want anyway. Try the concept directly: building in public revenue updates.

Numbers and dollar amounts. Embeddings de-emphasize numerics. $10k MRR will pull posts about MRR but the dollar figure won't filter. Use the Audience follower-range filter if you want a creator-size cut, and the outlier-minimum filter if you want a performance cut.

Boolean operators. There's no AND, no OR, no NOT. Eden treats design AND ai NOT crypto as a six-word phrase. If you want exclusion, just leave the concept out. If you want intersection, write the intersection: ai tools for designers.

Very short queries. One or two characters get filtered out by the debounce. Even three or four character single-word queries don't carry enough signal to rank well. Below five or six characters you're usually better off browsing.

How search composes with the other filters

Search runs on top of the filter set, not instead of it. The order things apply:

1. Platform toggles cut the candidate pool to the platforms you've enabled.
2. Pillar chips cut it to the topic.
3. Follower range and outlier minimum cut it by audience size and performance multiple.
4. Then the vector ranker reorders what's left by closeness to your query.

This composes really well. A query like productivity systems for ADHD founders inside the Productivity pillar, on YouTube only, with a 5x outlier minimum, will give you something close to a personalized research feed for that exact problem space. None of the four filters alone gets you there.

The flip side: a noisy filter set drowns out the search. If you've toggled off five of the six platforms and set the outlier minimum to 20x, you might be left with a candidate pool of fifteen posts before the ranker even runs. The query is just choosing the best fifteen. Loosen filters before assuming the search is broken.

A workflow that uses search well

A workflow that works for me:

1. Pin the pillar you write in.
2. Pick the platform you're studying that day.
3. Search for the specific kind of post you want to learn from, not the topic. "How a 20K substack writer onboards new subscribers" beats "substack growth" every time.
4. Skim the top ten or so. Save anything that hooks you to a board.
5. When you sit down to draft, open the board next to your draft.

The search is most useful as the last filter you apply, not the first. The chip and platform filters narrow the pool to the right neighborhood; the search picks out the houses worth visiting.

Where to go next

• For what each filter does and how they compose, read filtering Discover by platform, pillar, and outlier.
• For what the outlier number means and how to read it, read reading the outlier multiplier.