Sentence pattern evolution
With a fixed few phrases 「recommendations」 and 「is there any」 searched to the end, the needs you can dig up quickly run dry. This page is the result of actually running 「letting search terms evolve on their own」 for two rounds, including where it failed.
How it runs
Two-round actual run results (2026-11-09)
| Phrase pattern | Return | Buying requests | Pain points | Notes |
|---|---|---|---|---|
| Making spreadsheets repeatedly every day (seed) | 18 | 0 | 13 | Most pain points, but on manual review only 3 are real |
| Looking for an invoicing tool (mutation) | 19 | 3 | 1 | Most buying requests, but on manual review 0 are looking for tools |
| Looking for something to organize meeting minutes (mutation) | 18 | 1 | 2 | Changing the scenario produces small output |
| Where can I find something to organize meeting minutes（mutation） | 14 | 0 | 3 | Rewording brought no incremental gain |
| Are there any AI tool recommendations（seed） | 15 | 1 | 0 | — |
| Looking for a recommendation for an AI tool（seed） | 19 | 0 | 0 | All collections and tutorials |
| Now it all relies on manual work（seed） | 20 | 0 | 0 | The wording is too broad，and I can't find specific scenarios by searching |
| "is there an AI that"（seed） | 17 | 0 | 0 | The rule judged 0，but manual review found 6 out of 10 are specific scenarios——it's the classifier's fault |
| "wish there was a tool"（seed） | 18 | 1 | 0 | — |
Then I manually read through the 39 items，and found that the judgment was wrong
The reason is not complicated：The structure of advertorials is exactly 「first talk about pain points，then promote the product」，and they inherently look like needs。The rule lets it through when it sees pain point and scenario words，and is exactly fooled by advertorials。
So the mechanism needs to be changed in three places
One、the judgment stage must use a model to read the full text，and cannot use a word list。A word list can never keep up with how scenarios are phrased，and it will be fooled by advertorials。The model only needs to answer one question：Is this author 「asking」，or 「offering」？Anything that is offering is not counted at all。
Two、The focus of evolution is not sentence patterns，but scenarios。In actual testing with 「asking for an X」，replacing X with making invoices yielded 3 results、replacing it with changing the background of product images yielded 0 results；whereas 「doing spreadsheets repeatedly every day」 one result is worth ten of others。The difference in scenarios is far greater than the difference in phrasings。So what should change each round is 「what to ask」，not 「how to ask」。
Three、platforms need to be cultivated separately。On X，「is there an AI that」manual review shows 60% are genuinely asking；on Xiaohongshu，the same sentence pattern，search results are all clickbait。It's not a problem of words，but that people in these two places do different things——one is asking questions publicly，the other is looking at others' product recommendations。
There are also two upper limits，which must be acknowledged in advance
Phrasings will run out。In Chinese expressing 「I want a tool that can X」 there are only a few dozen ways to write it，after running for one or two weeks you will have found all the useful ones，after that the incremental gain brought by new phrasings is close to zero。When this day comes，you should stop changing phrasings，and only change scenarios。
Only needs that have been spoken aloud can be uncovered。People who cannot say what they want will never be reached by this mechanism。Don't expect it to discover 「unknown needs」，its upper limit is to fully cover 「needs that have been spoken」、and fully hear all phrasings。
Where this data comes from
This page is not a product feature，but lays out “which information can actually be obtained” for you to see。6 items，all from new products on HN on 2026-11-08，with the official website、pricing page、discussion section visited one by one。
Actual acquisition rate for each information item
| Information item | Obtained out of the 6 | Status | How it was obtained / why it could not be obtained |
|---|---|---|---|
| Whether the official website can be opened | 5 / 6 | Hard facts | Direct request。domdb.com returned 403，it was blocked by anti-scraping |
| What it is（in one sentence） | 5 / 6 | Hard facts | Page title + description tag，basically reliable |
| Whether there is a pricing page | 5 / 5 | Hard facts | Check whether the page contains a pricing link |
| Specific price tiers | 3 / 5 | Requires a human to read | Prices can be scraped as text，but which number corresponds to which plan must be read by a human，machines easily mismatch them |
| Whether it accepts payment | Cannot determine | Will misjudge | Only 0 payment codes were detected on the homepage, but 3 of them have complete pricing pages—they are clearly collecting money, just the payment code is not on the homepage. So this item cannot automatically be judged as “not collecting money” |
| Real user reviews | 2 / 5 | Hard facts | HN discussion section. The other 3 have 0 comments; no one is discussing the new products at all |
| Discussion heat | 5 / 5 | Hard facts | HN score and comment count |
| Monthly revenue (MRR) | 0 / 6 | Not available | There is no payment provider verification channel. Any “estimate” is a guess; do not write it |
| How many users | 0 / 6 | Not available | These are web/desktop products; there is no store install count available to check |
| Is it being discussed on Chinese social media? | 0 / 3 | Effectively none | Searching the project name on Xiaohongshu returns 16–20 results, but 0 of them actually mention it in the body. Overseas new products effectively do not exist in the Chinese-speaking circle |
Three conclusions
1. “What problem does it solve” can be used as a tag, but cannot be used as a basis for judgment.A statement written by the official website itself is reliable; user reviews only exist for projects that have been discussed, and only 2 of the 6 have them. Using 2 comments to infer that “the demand is valid” is the mistake I made in the previous version.
2. Pricing structure is hard data; revenue is not.The plans and unit prices of the three projects were all fully captured, so we can calculate “how much one customer is worth.” But how much was sold can never be seen from the outside. Therefore the “revenue” field on the card is left blank, with “not found” written in it.
3. Chinese-language sources cannot keep up with overseas new products.This is not because the sources are bad, but because the use case is wrong. Xiaohongshu is suitable for seeing “what Chinese people are willing to pay for,” not for seeing “whether anyone is using something newly released overseas.” The two need different sources and cannot be mixed into one database.
So what can this database answer?
It cannot answer: how much money it makes, how many active users it has, and whether this demand is truly valid.
For these last three things, you can only rely on your own observation and judgment. That is the reason the interface does not draw conclusions.