Inspiration
Short-form videos on TikTok, Instagram Reels, and RedNote have changed how we discover products. A single video can spark interest in something we never knew we wanted.
But these platforms are designed to capture attention, not to help us make informed decisions. Is the product really as good as it looks? Is the creator sharing a genuine recommendation or a sponsored one? And is it actually right for me?
Answering those questions means researching specs, comparing prices, reading reviews, and weighing personal factors like height, budget, and preferences, by hand.
We built WorthIt, an AI research agent for one simple question:
"I saw it. I liked it. But is it worth it?"
What it does
WorthIt turns a product discovered in a short video into a personalized, evidence-backed report.
Users upload a screenshot and tap the product, or paste a TikTok or RedNote link. WorthIt then:
- Identifies the product. It finds the item from the photo itself (Google Lens via Telem), verifies matches against the user's image, and names the brand and model when the content or image search reveals them.
- Researches it. Telem and Tavily gather specifications, prices, and owner reviews.
- Checks for promotional bias. It separates explicitly disclosed sponsorships from unverified promotional signals, and never infers a paid deal without evidence.
- Checks the video's claims. Claims like "doesn't wrinkle" are checked against owner reviews: supported, mixed, contradicted, or no evidence.
- Shows how solid the evidence is. Every finding is cited and labeled with who said it (owner, editorial, brand, creator) and how many independent sources agree. A Report confidence rating (High / Medium / Low) shows how well-backed the report is, not whether to buy.
- Explores on demand. "Dig deeper" options (fit for you, price check, owner reviews, the video's claims) each add a focused, cited section.
- Learns from feedback. Users can open the store page, add the item to a watch list, or say "not interested" in their own words. That feedback changes future research.
Unlike recommendation engines, WorthIt doesn't decide what you discover next. It helps you evaluate what you've already found.
How we built it
WorthIt is a Node.js/Express app with a lightweight, no-build JavaScript frontend. It runs on GPT-4o and a set of independent, testable research skills coordinated by a central orchestrator, with 66 automated tests.
From social media to product identification
The user taps the product, and WorthIt crops just that item. GPT-4o identifies it and extracts any claims the video makes. For links, the post's cover image becomes the photo and its caption helps identification.
Research powered by Telem
Telem is our retrieval backbone. Each research question goes to the provider best suited to it:
- Web search (auto-routed): specifications and product details
- Reddit and Trustpilot: owner experiences
- Amazon: pricing
- TikTok and YouTube: promotional content and sponsorship signals
- RedNote: real people wearing the item, filtered to users of similar height
- Google Lens via Telem's SerpApi provider: finding the product from the image itself
Telem Fetch reads key pages in full, including the photo-matched store page for the item's actual price. Tavily adds complementary evidence. A typical investigation makes 5–6 Telem searches across 5+ providers, plus page reads and an image search. Each Dig deeper adds 2–3 focused Telem calls.
The photo is the source of truth
Results describing a different color or length are filtered out, and image-search matches are visually compared with the original photo (same item vs. similar). Only verified matches become evidence about this item.
Trust built in code
Sources are labeled by perspective and commercial incentive (sells it, sponsored, affiliate, listicle). Every claim must cite evidence. Evidence strength, sponsorship status, and report confidence are calculated by deterministic rules, not by the language model grading itself.
Memory that changes the research
Users explain decisions in their own words. Short-term preferences, like a budget or "too long", update automatically: a repeated reason reinforces the preference, an opposite one replaces it, and they expire after 30 days. Stable traits like height are saved only after a one-tap confirmation.
Memory changes the search queries themselves. After a user says "too long, I'm 5'2", the next search adds garment-length and height-specific Telem queries, and a fit-risk icon appears when an item clearly conflicts with what they've told us.
Independent by design
No affiliate links, and tracking parameters are stripped from store URLs. WorthIt helps people decide; it doesn't sell.
Challenges we ran into
Our biggest challenge was identifying the exact product, especially in fashion.
At first, GPT-4o described the product and we searched by that description. The model recognized type, color, and style, but descriptions weren't distinctive enough: many garments look identical in text but differ in cut, length, or detail. Finding a similar product isn't the same as finding the right one, and a wrong match pollutes the whole report with the wrong prices and reviews.
We fixed this by searching with the image itself through Google Lens via Telem's SerpApi provider, then verifying each match against the user's photo.
The lesson: product identification and product research are two separate problems. Good research depends on getting the identity right first, and even plausible matches need verification before their evidence can be trusted.
Accomplishments that we're proud of
- A working prototype that goes from a screenshot or link to a verified product match, multi-source research, and a cited, personalized report.
- Image-based matching that found the exact item and its store page in our tests, with lookalikes clearly labeled as "similar."
- Trust signals computed in code: who said it, whether they benefit, and how many independent sources agree.
- Interactive "Dig deeper" research instead of a chatbot or a wall of text.
- Memory learned from natural language that changes the next investigation's actual searches.
What we learned
Accurate identification is often harder than the research itself. More information isn't better: focused research, clear evidence, and user-controlled exploration make AI reports more useful. And trust comes from showing who is speaking and why, not from a confident summary.
What's next for WorthIt
- Watch list as decision memory: remember why you were interested, set target prices, and revisit decisions when something changes.
- Ongoing tracking: price changes, availability, and new reviews, without repeating the research.
- Broader identification: more categories and platforms, including content that shows little product information.
- Deeper personalization: recognizing recurring preferences and comparing new finds with items you've already researched.
Our goal isn't another recommendation engine. It's a trusted research companion that works for the shopper, not the seller.
Built With
- express.js
- google-lens
- gpt-4o
- javascript
- node.js
- npm
- openai
- serpapi
- tavily
- telem