Last week I built a check-up tool, and the first thing I did was take it to myself.
Here is the conclusion first: 8 questions, 8 times, and AI never mentioned me once in its answers.
It is not a low score — it is not being in the pool at all.
First, what the "pool" is
We are used to treating AI search as a smarter search engine, so we put all our effort into "making the page good": load faster, structure more clearly, write more content.
But an AI answer is not picked from the whole web; it is picked from the batch of sources that have already entered the retrieval pool.
What sits in the pool is finite. Platform sites (portals, reposted WeChat accounts, tech communities, the Toutiao family, video sites) occupy most of the slots. If you are not in there, then however beautiful your page is, it will not be seen.
It is like renovating your home beautifully while your home stands on no road that any navigation can find.
What I did
I took the 8 questions people genuinely ask about my kind of work and probed the source pool with them — asking "when AI answers this question, which sources does it have in hand".
The result was this (7 of the 8 questions gave a full distribution):
For each question I kept clickable links to the top sources, so anyone can verify it.
Along the way I also checked my own site's technical readiness: full marks is 100, and we scored 100 (A). Running the same tool against an external tech community gave 72.
Put those two things together and that is the real point: perfect on-site preparation, and still not one appearance in the pool.
Another probe: put the questions straight to AI
The source-pool probe asks "which sources does AI have in hand"; the other method is to take the question directly to AI. The same 8 questions, 7 rounds each, produced 56 real answers, checked one by one for any mention of us.
Both probes agree: a perfect on-site score, 0 hits inside the pool. This should no longer be treated as "the content is not good enough" but as "not being in the pool".
Three observations read out of the real answers:
The accounting is written down here too: only answers genuinely captured are recorded, anything not captured is left blank with the reason stated, and when we are not mentioned the "sentiment" field is left empty — with no object there, positive or negative does not apply.
Three actions, ranked by leverage
1. Take a slot in the pool before optimising your own site
Wherever the head of the pool is, that is where you need a position. Not reposting the same draft everywhere, but turning "one answer to one question" into a single atomic piece of content:
A compendium suits human readers; an atomic answer suits being cited. They are two different products.
2. Entity names must be unified
We stepped into a very typical trap: the same brand written with its full name in some places and its abbreviation in others. It looks like a small thing, but to a machine they are two entities — it cannot assemble information scattered across places into one credible brand.
Now, everywhere I can write it, I use one consistent line: the name, what we do, who we serve, where the boundary is. The boundary goes in too; a statement like "we do not do X" looks as if it weakens credibility, but it is actually a plus, because it is verifiable.
3. Do not treat a "platform" as a channel
This one is counter-intuitive, and it is where effort is most easily wasted.
We are often advised to "publish on WeChat accounts and Zhihu so AI picks it up". But measured in practice: the robots files of WeChat accounts and Zhihu are site-wide Disallow, so AI cannot crawl them.
They are useful to people (people can see them and forward them) but useless for AI crawling. So the right approach is: the same answer needs a crawlable landing page; the platform is only responsible for letting people see it.
One boundary that must be spelled out
We have not treated llms.txt as a weapon.
Many articles online present it as "the standard way to make AI find you", but the fact is: it is not a standard adopted by any mainstream vendor, and mainstream search engines have publicly said it does not affect rankings or AI summaries.
So our tool only checks whether it exists; it does not take part in scoring. Expecting a single file to earn you mass citation will probably end in disappointment.
Just as we do not read "undeclared" as "allowed". A missing rule in robots is neither permission nor prohibition — behaviour is decided by each party. So what we measure is the observed result: we take each crawler's identity and actually crawl once, to see whether the other side hands it over.
Why write all this out
Because the method is recomputable, and hiding it would be pointless.
On "how many repeated asks make a result credible", here is the honest answer: an academic review warns that mechanically stuffing citations into a page can damage retrieval hits, and that the same question must be repeated several times to be stable. So we do not treat a single run as a conclusion.
Finally
If you work on content too, do a minimum-cost check: pick three questions your customers genuinely ask, put them to AI yourself, and see whether you appear among the sources it names.
If you do not, your current optimisation direction probably needs adjusting — it is not that you are not good enough, it is that you are not in the pool.