Build your own video social-listening tool with VideoDB A VideoDB developer guide shows how to build a video social-listening tool that indexes a user-supplied collection of videos to find every moment a brand is spoken, shown on screen, or visible in shot, returning each match with its video and start and end times. VideoDB handles indexing, search and clip generation, while the developer supplies the videos and confirms rights to analyze them; the guide states VideoDB does not pull posts from TikTok, Instagram or X and offers no built-in sentiment analyzer. The guide walks through the same building blocks as the Read the Room social-listening app, which compares how long each brand is on screen across public videos with detected screen time, spoken mentions and on-screen text. Build your own video social-listening tool with VideoDB Most social listening tools read captions and comments. This guide shows how to build one that also hears and sees brand mentions inside videos: index a collection you supply, find every moment a brand is said or shown, and compare brands with a clip behind every number. Who this is for: developers and founders who want to build a focused video-analysis tool, such as a brand tracker for a marketing team or a competitor report for clients. You can follow along with no code in Claude or ChatGPT, or in Python. Most social listening tools read text: captions, comments, hashtags and transcripts. That misses a lot of video. A creator can hold your product for a minute without naming it. A competitor’s logo can sit on screen while the caption talks about something else. A video social-listening tool closes that gap. You give it a set of videos, it finds every moment a brand is said or shown, and it compares brands with a playable clip behind every number. This guide shows how to build one on VideoDB, from a collection of videos you supply to a source-backed comparison. In short: - You bring the videos. VideoDB indexes and searches them. It does not collect posts from social platforms. - Index speech and visuals. A spoken-word index catches what is said. A scene index and the brand and on-screen text analyzers catch what is shown. - Search the whole collection for each brand, and get every match with its video and start and end times. - Compare with evidence. Count moments and seconds per brand, and keep a clip for every row so anyone can check it. What can video social listening find that text-based social listening tools miss? It finds mentions that never reach the text: spoken mentions, products in shot and brand names on screen. Each one comes back with a timestamp, so a person can watch the moment instead of trusting a count. There are three kinds of mention to look for: - Spoken: someone names the brand or product. - On screen as text: the name appears on packaging, a slide, a sign or a caption burned into the video. - In shot: the product is visible, held or used, whether or not anyone names it. Our archival footage page https://videodb.io/archival-footage shows Read the Room, a social listening app built on VideoDB. Per that page, it compares how long each brand is on screen across public videos, with detected screen time, spoken mentions and on-screen text, and then opens the exact moments behind every number. This guide does not reproduce that app. It walks through the same building blocks so you can build your own version. What does VideoDB do, and what is your part? VideoDB handles indexing, search and clips. Collecting the videos, and making sure you are allowed to analyze them, is your part. | | Documented in VideoDB | Your part | |---|---|---| | Getting videos in | Upload from a public URL or a YouTube link. A search youtube tool finds YouTube links to add. The SDK also uploads local files. | Choose the videos and confirm you have the rights to analyze them. VideoDB does not pull posts from TikTok, Instagram or X. | | Speech | Spoken-word index and transcripts with timestamps | Decide which names and phrases count as a mention | | What is shown | Scene index from a prompt you write; brand detection brand names and ocr on-screen text analyzers | Review matches; brand detection returns names, not logo positions | | Finding moments | Search across a whole collection, with video, start and end for each match | Write one query per brand or product | | Evidence | Play any match, or compile matches into one stream | Keep the clip link next to every number you report | | Sentiment and engagement | Not built in. No sentiment analyzer is listed. | Add your own, for example a custom VLM prompt, or engagement data from a source you are allowed to use | Sources: analyzer outputs https://docs.videodb.io/pages/understand/indexing-pipelines/analyzer-outputs , MCP tools https://docs.videodb.io/pages/mcp/tools , collection search https://docs.videodb.io/pages/understand/legacy-indexing-search/collection-search , results, clips and streams https://docs.videodb.io/pages/understand/search-and-retrieval/timestamps-clips-streams . How does the workflow fit together? Five steps take you from a folder of videos to a comparison you can defend: collect, upload, index, search, compare. 1. Collect videos you can use. For example: your own brand’s videos, creator videos you sponsored, or public videos that your team has the rights to analyze. Keep a list of where each one came from. 2. Upload them into one collection per project , such as “Headphones, Q4”. A collection keeps the set together so you can search it as one library. 3. Index each video. Run a spoken-word index for speech and a scene index with a prompt that asks for brands, products and on-screen text. For more structured results, run the brand detection and ocr analyzers too. 4. Search the collection once per brand, for what is said and for what is shown. 5. Compare. Group the matches by brand and video, count moments and seconds, and attach a clip to each row. Try it with no code in Claude or ChatGPT The quickest way to test the idea is to ask an assistant with the VideoDB plugin added. See how to add VideoDB https://videodb.io/agents to Claude or ChatGPT. Replace the parts in brackets. These prompts use the documented connector tools; we have not published a test run of them. Create a VideoDB collection called " Project name ". Upload these videos into it: link 1 , link 2 , link 3 . Index every video in " Project name " for speech and for scenes. In the scene index, describe any brands, products and on-screen text you see. Search " Project name " for every moment where Brand A is said or shown. Then do the same for Brand B . Group the results by video, with start and end times. Make a table comparing Brand A and Brand B : moments per video, total seconds, and whether each was spoken, on screen or in shot. Then compile the Brand A moments into one clip I can play. What you get: a list of moments per brand with video and timestamps, a comparison table, and a link that plays the matching moments back to back. Build it in Python When the workflow is worth keeping, move it into code so it runs the same way every time. Install the SDK with pip install videodb and get an API key from the VideoDB console https://console.videodb.io . This example follows the calls in the VideoDB docs https://docs.videodb.io/pages/understand/legacy-indexing-search/collection-search and the Python SDK README https://github.com/video-db/videodb-python . Method names can change between SDK versions, so check them there before you build on it. python import time import videodb from collections import defaultdict from videodb import IndexType, SceneExtractionType from videodb.exceptions import InvalidRequestError conn = videodb.connect api key="YOUR API KEY" 1. One collection per project coll = conn.create collection name="Headphones Q4", description="Brand mentions in videos we are allowed to analyze", is public=False, 2. Upload the videos you collected public URLs or YouTube links sources = "https://www.youtube.com/watch?v=VIDEO ID 1", "https://www.youtube.com/watch?v=VIDEO ID 2" videos = coll.upload url=url for url in sources 3. Index speech and visuals for video in videos: video.index spoken words waits until the transcript index is ready scene index id = video.index scenes extraction type=SceneExtractionType.shot based, prompt="List any brand names, logos, products and on-screen text visible in this scene.", index scenes returns while the scene index is still processing, so wait for it while True: status = next i "status" for i in video.list scene index if i "scene index id" == scene index id if status == "done": break if status == "failed": raise RuntimeError f"Scene indexing failed for {video.id}" time.sleep 10 def search or none run search : Legacy search raises an error when nothing matches; treat that as zero moments try: return run search except InvalidRequestError as e: if "No results found" in str e or "Query scene index failed" in str e : return None raise 4 and 5. Search per brand, then total moments and seconds per video brands = "Brand A", "Brand B" report = {} for brand in brands: shown = search or none lambda: coll.search query=f"{brand} logo or product visible", index type=IndexType.scene said = search or none lambda: coll.legacy search brand spoken-word search, per the collection search docs shown shots = shown.get shots if shown else said shots = said.get shots if said else per video = defaultdict lambda: {"moments": 0, "seconds": 0.0} for shot in shown shots + said shots: row = per video shot.video id row "moments" += 1 row "seconds" += shot.end - shot.start report brand = {"per video": dict per video , one playable stream of the visual matches, or None if the brand was never shown "evidence clip": shown.compile if shown shots else None} What you get: for each brand, moments and seconds per video, plus a stream URL that plays the visual matches back to back None if the brand was never shown . Each underlying match carries its video id , start , end and matched text , so every number in your report can link to the moment behind it. A few choices to make as you build it out: - Scene search finds what your prompt describes. If the prompt does not ask for brands and on-screen text, the scene index will not reliably mention them. - Count carefully. A brand that is said and shown in the same moment appears in both searches. Decide whether that is one mention or two, and remove overlaps if you count it once. - Seconds are scene-level. A match covers its whole scene or shot, so screen time is an estimate at that granularity. - For structured brand and text results, run the brand detection and ocr analyzers with video.understand and index their output. See running analyzers https://docs.videodb.io/pages/understand/indexing-pipelines/understanding-artifacts and create an index https://docs.videodb.io/pages/understand/indexing-pipelines/create-an-index . For a worked example of searching across many recordings, see how I built call intelligence with VideoDB https://videodb.io/blog/how-i-built-call-intelligence-with-videodb . For more on bringing YouTube videos in from code, see giving an agent a YouTube video it can search https://videodb.io/blog/ai-agent-watch-youtube . What does a useful report look like? A useful report puts a number and the evidence in the same row. Here is the shape to aim for; the column names are suggestions and the values come from your own run. | Brand | Video | Moments | Seconds | Type | Evidence | |---|---|---|---|---|---| | Brand A | Video 1 | from your run | from your run | spoken / on screen / in shot | clip link | | Brand B | Video 1 | from your run | from your run | spoken / on screen / in shot | clip link | Two habits keep it trustworthy: - Every number links to clips. A reviewer should be able to click any count and watch the moments behind it. - Report the sample. Say how many videos were analyzed, where they came from and when. A comparison is only as fair as the set of videos behind it. What are the limits? Know these before you put the tool in front of a client. - No automatic collection. VideoDB does not monitor or pull posts from TikTok, Instagram, X or other platforms. You supply the videos. - Only videos you can reach and are allowed to use. Private videos, and content whose terms do not allow analysis, are out of scope. - No built-in sentiment or engagement data. Add sentiment with your own analyzer or prompt, and bring engagement data from a source you are allowed to use. - Detection is not perfect. Small, blurred or partly hidden logos can be missed, and similar names can match by mistake. Spot-check clips before reporting. - Brand detection returns names, not positions. If you need where a logo sits in the frame, you need another approach. - Processing uses credits. Uploading, indexing and search use VideoDB credits, so start with a small collection. FAQ What are social listening tools? Social listening tools track what people say about a brand, product or topic in public content, so teams can follow mentions, compare competitors and spot trends. Many work mainly from text such as captions, comments and hashtags. A video social-listening tool also looks inside the video: what is said and what is shown. Can VideoDB collect TikTok, Instagram or X videos for me? No. VideoDB has no tool that collects or monitors posts on social platforms. You bring the videos: your own exports, partner or creator content you are allowed to use, or public video links. The connector’s upload tool takes public URLs and YouTube links, and a YouTube search tool helps you find YouTube links to add. Can VideoDB detect logos in a video? VideoDB has a brand detection analyzer that returns the brand names it recognizes in each scene, and an OCR analyzer that reads on-screen text. Brand detection returns names, not the position of the logo in the frame. Check results against the clips before you rely on them. Do I need to write code? No. You can run the same workflow by asking Claude or ChatGPT with the VideoDB plugin added. The Python SDK is there when you want to turn the workflow into your own app or a scheduled job. Is VideoDB free to try? New VideoDB accounts get $20 of free credits, and no credit card is needed. Uploading, indexing and search use those credits. Start with one collection Pick three to five videos you are allowed to analyze, two brands you care about, and run the four prompts above. If the moments it finds are the ones you would have flagged by hand, move the workflow into code.