AI content tagging for UGC: how it works, shown on sample posts
How AI tagging links UGC to your products. Tags on sample posts, where it misses, and cost.
A customer photo arrives with a caption, a handle and nothing else. Nobody has said which product is in it. Until someone does, it cannot sell anything. Tagging is that step, done for every post.
In this article
A UGC library grows faster than anyone can label it. An untagged post cannot be found, filtered or bought from. Tagging fixes that, one post at a time, without a person doing each one.
What is AI content tagging for UGC?
Tagging attaches structured labels to a photo or video. There are two kinds. Product tags say which item from your catalogue is in the post. Descriptive labels say what else is there: colour, setting, activity.
A post stops being "image 4471". It becomes "linen overshirt, beige, outdoor, worn". Product tags make it shoppable. Labels make it findable, which is the idea behind natural-language UGC search.
How Idukki's tagger works
From post to product tag
- 01
Ingest
Photos use the image. Videos use the thumbnail.
- 02
Shortlist
Over 150 products, the caption and labels pick the 60 closest products.
- 03
Match
One Gemini vision call returns the catalogue products it can see, with confidence.
- 04
Route
At 0.75 or above it tags. From 0.5 to 0.75 it waits for review.
- It only picks from your catalogue. Any product id the model invents is thrown away.
- Low matches are dropped. Below 0.5 confidence, nothing is tagged.
- You set the bands. Thresholds move between 0.3 and 0.98 on the Recommendations page.
- Video tags can be timed. In the hotspot editor, a tag can appear at a set second.
Annotated examples: what the tagger adds
The table below is illustrative. It shows the tags a tagger would aim to add to typical posts. It is not output from a named store.
| Post type | Product tag | Colour | Scene | What to check |
|---|---|---|---|---|
| Outfit photo on a street | Two-piece set, sunglasses, shoulder bag | Cream, black, tan | Outdoor, city, daylight | Bag may be a lookalike |
| Skincare clip at a desk | Serum bottle | White, pale blue | Indoor, desk, home | Thumbnail may hide the label |
| Watch on a wrist | Watch | Silver, brown strap | Close-up, indoor | Strap variant is easy to miss |
| Sofa in a living room | Sofa | Grey | Living room, daylight | Cushions may not be yours |
| Unboxing on a table | Product plus accessories | Packaging colours | Indoor, table | Box contents are not products |
| Kitchen recipe clip | Pan | Black | Kitchen, cooking | Product may appear after the thumbnail |
| Gym selfie in a mirror | Leggings, top | Navy, white | Gym, mirror | Mirror flips printed text |
| Flat-lay of three items | All three products | Mixed | Overhead, studio-like | Each item needs its own tag |
Where AI tagging misses
- Small or hidden products. A ring half in shadow may not be seen.
- Lookalike variants. Two shades of one jacket can look the same in warm light.
- Late appearances in video. The tagger reads the thumbnail, so later products need a manual tag.
- Items you do not sell. It only matches your catalogue, so other items get no product tag.
- Mirror shots. Printed text and logos read backwards.
- Busy flat-lays. Several products in one frame each need checking.
This is why the middle confidence band waits for a person. A wrong tag on a live gallery costs more than a missing one.
Manual tagging vs AI tagging
| Dimension | Manual tagging | AI tagging |
|---|---|---|
| Speed | Limited by a person's time | Runs on each post as it arrives |
| Consistency | Varies between people and days | The same rules every time |
| Small details | A person spots context | Can miss small or hidden items |
| Catalogue link | Typed or picked by hand | Matched to a real product id |
| Where people fit | Every post | The review band only |
What AI tagging costs
Model cost depends on tokens per image and the model's rates. Our calculator defaults to one Gemini vision call per image. It assumes 2,100 input tokens and 200 output tokens.
| Rates | Per 1,000 images | Per 5,000 images |
|---|---|---|
| Introductory, to 31 Dec 2026 ($0.75 in, $3.75 out per 1M tokens) | $2.33 | $11.63 |
| Standard, from 1 Jan 2027 ($1.50 in, $7.50 out per 1M tokens) | $4.65 | $23.25 |
These figures cover model calls only. Change the inputs for your own library in the AI tagging cost calculator. On Idukki, product tagging is listed on every plan, including Free, on the pricing page.
What good tags are for
- Internal search: find "before and after, kitchen, daylight" in seconds.
- Shoppable galleries: a product tag makes the post tappable.
- Filters: galleries can show only the colour or scene a visitor picked.
- Machine-readable evidence: tags tell an AI agent what each post depicts.
That last use ties into structured data and schema for product UGC. It also feeds visual search and shop the look.
FAQs
Does AI tagging work on video?
Yes, from the video's thumbnail. For a product that appears later, add a tag with a timestamp in the hotspot editor.
Can the AI tag a product I do not sell?
No. It only returns products from your synced catalogue. Anything else is discarded.
How accurate is AI tagging?
It depends on your catalogue and photos. That is why matches between 0.5 and 0.75 wait for a person by default.
How much does it cost to tag a library?
About $2.33 per 1,000 images in model cost, at the calculator defaults, until 2027.
Sources & notes
- 1Idukki, AI tagging feature page · Shortlist size, confidence bands, thumbnail use on video.
- 2Google Cloud, generative AI pricing · Gemini 3.6 Flash rates. Checked 24 September 2026.
- 3Idukki, AI tagging cost calculator · Token defaults used for the cost table.
Continue reading
8 pieces in this clusterThese long-form pieces on the Idukki blog link back to this article, go deeper on the cluster.
- Strategy
Why we made video testimonials a free feature, not a new product
Idukki already had the collection link, moderation queue and gallery engine, so video testimonials shipped as a free feature on every plan, not a new bill.
- Agentic commerce
What is agentic commerce? How AI shopping agents change product discovery
Agentic commerce is shopping where an AI agent does the searching, comparing and shortlisting. Agents read facts, reviews and customer evidence, then decide.
- AI search
How to make your product content readable by AI shopping agents
AI agents shortlist products by reading facts and evidence, not design. Fix three surfaces: structured data, plain-text claims, and a reachable review corpus.
- AI search
Visual search and "shop the look": turning a customer photo into a buyable cart
Visual search and shop-the-look turn a photo, often a real customer's, into a path straight to a buyable product. How it works and where it earns its place.
- Strategy
AI in the UGC loop, part 3, moderation: the layer most teams skip
Brand-safe is not the same as profanity-filtered: the gap is the competitor logo in the unboxing clip. Run a three-tier queue with human-review SLAs.
- Strategy
AI in the UGC loop, part 2, tagging: a content dump becomes a catalogue
AI tagging turns a UGC dump into a shoppable catalogue: the model pre-fills metadata, a human confirms. Tag inside 72 hours and UGC works as a merchandising lever.
- Strategy
AI in the UGC loop, part 1, ingestion: stop chasing creators
Every brand has a creator-chaser and the role does not scale. The 2026 constraint is discoverability, not creator supply; AI turns sourcing into an inbound stream.
- Strategy
A kitchen table in Egham, why I built Idukki
Day job: SAP architect on UK government software. Night job: founder of a UGC platform. The Venn diagram of those two communities is roughly one person.
More from Rohin Aggarwal
- AI search
Answer Engine Optimisation: the 2026 playbook for ecommerce
AEO gets you cited inside ChatGPT, Claude, Perplexity and Gemini, not just ranked. Covers JSON-LD, llms.txt, conversational PDPs, evidence and citation tracking.
- AI search
The 42-signal AEO scorecard: rate any PDP for citation-readiness in 10 minutes
Score any PDP for AI-citation readiness in 10 minutes: 42 signals, 7 categories, each scored 0/1/2. The median DTC PDP sits at ~38% of max. Includes a CSV download.
- AI search
Schema.org for ecommerce in the agentic era: 14 schemas every storefront should emit, ranked by AI-engine pickup
Ship six mandatory schemas first: Product, Offer, Review, AggregateRating, FAQPage, BreadcrumbList. Then four high-ROI shapes, ranked by AI-engine pickup.