WeChat's AI rewrites searches, so your video can match words nobody typed
Tencent says the change lifted views of over 10 seconds by 2.593% in a search handling over 100 million queries a day. Brands now need to say clearly what each video shows.
The short version
- Tencent has put a small AI model into WeChat's short-video search that quietly rewrites vague searches based on what each person seems to want.
- In a live test, views over 10 seconds rose 2.593% and people searched again 2.350% less often. The system handles over 100 million searches a day.
- Your video may now be found through words the shopper never typed, so keywords alone won't explain your search results.
What Tencent built
People type very little into a video search box. Someone who enters "crepe cake" could want a recipe, or could want to watch someone eat one. WeChat's search now tries to settle that before it looks for videos.
A paper from Tencent's Weixin Group and East China Normal University, updated on October 30, describes the system, called WeFair. It reads the search along with what the person just searched for and watched. It then adds a clearer version of the search, like "crepe cake eating videos". If the original is already clear, it leaves it alone. The original words stay in use for finding and ranking results, so the rewrite widens the search and doesn't replace it.
Tencent says this is live in WeChat's short-video search, which handles over 100 million searches a day. It compared the new system with the old one on real traffic. Views of more than 10 seconds rose 2.593%. The rate at which people gave up and searched again fell 2.350%.
What is new here
Search systems are usually trained on what people watched afterward. That has a flaw. People may watch something entertaining that has nothing to do with what they came for, so the AI learns to steer searches toward whatever holds attention. In the paper's crepe cake example, an older approach sends the person to recipes when they wanted eating videos.
WeFair puts more trust in a clearer signal. When a person doesn't like the results and types a better search, that second search shows what they meant. Tencent used those real corrections as its strictest training material. Viewing behavior still plays a part, but with less weight in the final stage.
The model is also small, which matters because search has to answer almost instantly. That brings this kind of personal rewriting within reach for a system of this size.
What it means for your brand
This research covers WeChat alone. It tells you nothing about what other platforms do inside their own search. But the problem it solves exists on any short-video platform, so it matters if your brand in Indonesia sells through video search.
If this approach spreads, 3 things follow:
- 2 people typing the same words may see different results, based on what they watched just before. A single report of where you rank can't capture that.
- Your video can match a search the person never typed. If your caption says clearly what the video shows, a rewritten search has more to match.
- Purpose matters more. A video that teaches a recipe and a video that shows someone enjoying the dish serve different searchers, even with the same keyword.
Tencent reported views and repeat searches. It didn't report any effect on brands or sales.
What to do next
- Sort your priority keywords by what the searcher wants, such as learning or buying. Make sure you have video for each one you care about.
- Say plainly what each video is in the caption and the spoken words. "How to make" and "taste test" send clearer signals than a product name alone.
- Watch which searches bring in viewers in your platform analytics, and look for words you never targeted.
- Check your search results from several accounts with different viewing histories, instead of relying on one person's phone.
- If you sell in China, ask your WeChat partners whether they see changes in where search traffic comes from.