Super Search: Finding Your Best UGC in Seconds with Natural Language

Natural-language UGC search lets you describe the content you want in plain English ("close-up of the watch on a wrist", "outdoor shots of the tote bag, no faces") and get the matching customer photos and videos back in seconds. Instead of relying on manual tags or filenames, an AI model reads what is actually in each clip (objects, scenes, products, sentiment) so you can query a library of thousands the way you would ask a colleague.

Idukki calls this Super Search. It is the third pillar of the platform: type what you are looking for, filter in real time across connected social and review sources, and pull the result straight into a shoppable gallery. The folder-scroll above is the problem it removes.

Why manual tagging dies at scale

Manual tagging works fine for the first few hundred pieces of content. Someone watches each clip, writes "beach", "blue dress", "smiling", and moves on. The trouble is that customer content does not arrive in batches of a few hundred. A live campaign with a branded hashtag can pull thousands of clips in a week, and every one of them needs a human eye before it is usable.

The deeper problem is that tags are guesses about what you will want later. The person tagging in March cannot know that in June you will need "the tote bag held in one hand, outdoors". If that tag was never written, the clip is invisible to search even though it exists. You end up re-watching the library every time the brief changes, which is exactly the 9pm scroll. See our note on AI content tagging for UGC for why model-generated tags hold up where manual ones break.

• Tagging cost scales linearly with volume: 10x the content is 10x the human hours.
• Tags freeze one interpretation; future briefs ask questions the tags cannot answer.
• Inconsistent vocabulary ("jumper" vs "sweater" vs "knit") fragments your own library.
• Video is worse than photos: a 30-second clip can contain ten taggable moments.

How natural-language UGC search actually works

Behind the search box, every piece of content is run through a vision-language model when it is ingested. The model produces a rich description and a numeric representation (an embedding) of what the clip contains: products, scenes, colours, the presence of faces, the general mood. Your typed query gets turned into the same kind of representation, and the system returns the clips whose meaning sits closest to your words.

That is why you can search for "autumn colours, no logos visible" and get sensible results even though nobody ever wrote those exact words on a clip. The match is on meaning, not on a string. Connected sources (Instagram, TikTok, YouTube, plus review platforms) all feed the same index, so one query searches everything at once instead of forcing you to hop between tabs.

From a query to a live, shoppable gallery

Finding the clips is only half the job. The point of UGC is to sell, so Super Search hands its results straight into the same tagging and display tools the rest of Idukki uses. You confirm the shortlist, drop product tags onto the moments where a product appears, and the row goes out as a shoppable gallery with one-click checkout. The merchandiser from the lede goes from nine clips to a published, on-brand row before midnight.

On mobile, where most of this content is watched, the tagged clip becomes a tap-to-shop surface: a hotspot on the product opens a card with the name, price and an add-to-cart, without leaving the video. That is the bridge from "we found good content" to "the content is doing work on the storefront".

Real searches that find money

The abstract version ("query your library in natural language") undersells it. The value shows up in specific, repeatable searches that map directly to a merchandising or marketing need. These are the kinds of queries operators actually run before a launch or an ad refresh.

• "Best clips of the Airlift overcoat in daylight" to refresh a hero PDP gallery.
• "Outdoor lifestyle shots, no faces" to source legally simpler content for paid ads.
• "Unboxing reactions of the jute tote" to feed a post-purchase email flow.
• "Five-star review videos mentioning fit" to pair social proof with sizing copy.
• "Watch on a wrist, close-up" to fill a detail row on a single product page.

Each of those used to be a folder dive. As a search, it is a sentence and a few seconds, repeatable every time the catalogue or the campaign changes. The clips you already paid to collect stop being a graveyard and start being inventory you can pull from on demand. For the downstream economics, our piece on how to measure UGC ROI covers what that retrieval speed is worth.

Pairing search with auto-curation

Search is the manual lever: you ask, you get answers. Auto-curation is the standing version of the same engine, where rules and a quality score keep good content flowing into galleries without anyone typing a query at all. The two work as a pair. You search when you have a specific brief, and you let auto-curation handle the steady state so the always-on galleries never go stale.

A sensible setup runs auto-curation across your live galleries for freshness, then uses Super Search for the targeted pulls (a launch, an ad set, a seasonal edit). Both lean on the same indexed understanding of your content, and both feed the same rights workflow so nothing publishes without consent. Our walkthrough on AI auto-curation of UGC goes deep on the standing side of this.