In short: Consolidating your own naming, category words and stated relationships is necessary and it is not sufficient. Category association is built mostly from what other people have written about you, which means the highest-leverage entity work happens off your own site.
It also does not carry across languages. Measurement on this site found that the same question asked in two languages agrees on 4% to 11% of the domains it cites, against 72% to 74% when a question is compared with its own rerun.
Before doing anything about either, find out where you actually stand. That takes about an hour and no tool.
The Symptom Worth Recognising
There is one pattern that comes up more than any other, and it is the reason this lesson exists.
Ask an assistant about your brand by name and the answer is broadly accurate. Ask it the buying question your category is known by, and you are not mentioned at all. Your competitors are.
What that pattern means: you are resolved as an entity and you are not classified into the category. It reads like a visibility problem and behaves like one, and no amount of work on your own pages will move it, because your pages are never retrieved for a question that does not consider you a candidate.
Recognising this correctly is worth more than any list of tactics, because the two failures it gets confused with have completely different fixes. Failing extraction is fixed on your pages. Failing classification is not.
Why This Is Off-Site Work
A model's sense of which things belong in a category comes from the general written record: coverage, comparisons, documentation, community discussion, lists that other people maintain. Your own description of yourself is one source among thousands, and it is the one source with an obvious interest in the answer.
That is not a complaint about fairness. It is the mechanism working as intended, and it has a practical consequence: the work that moves classification is being written about by other people, in the places where the category is discussed, in terms the category uses.
None of this is new to anyone who has done digital PR or entity work in search. The mechanism changed. The substance largely did not.
It Does Not Transfer Between Languages
If you sell in more than one market, there is a further problem, and it is larger than most reporting assumes.
The study published on this site asked the same eight buyer questions across seven European markets, written natively in each language rather than translated, across four models, with repeats and a second pass days later.
The practical reading is that a single blended visibility score across markets is averaging over environments that barely overlap, and that entity work has to be done per market rather than once in English and translated.
Find Out Where You Stand
Do this before changing anything. It takes about an hour, costs nothing, and gives you the baseline that every later claim of improvement has to be measured against.
Write five questions a buyer would actually type, in the words they would use, with no brand name in them. Write three more that name you directly.
Run all eight in each assistant that matters to your market, three times each, in a fresh session every time. Record what came back and which sources were named. Fresh sessions matter because history changes the answer; three runs matter because the same question does not agree with itself.
Record the sources, not just the answer. Who appeared instead of you, and where that answer came from, tells you which places the category is being read from. That list is the target list for the off-site work, and it is more useful than the answer itself.
The test plan builder on this site will generate the prompt plan for you from your category and markets. It runs nothing and calls no API.
The running is done by you, in the actual assistants, because an API call with a search tool attached is a different product and gives you a different answer dressed up as the same one.
Reading What You Get
Sort the results by which of the three decisions failed.
Not mentioned when named, or described as something you are not: resolution failed. Go back to Making Yourself Resolvable and consolidate.
Accurate when named, absent from category questions: classification failed. That is off-site work, and it is slow.
Present in the category but never the one recommended: association and evidence. What is written about the ones chosen instead, and what do they have that is checkable that you do not?
One more thing to record, because it will save you an argument later: how much the answers differ between your own three runs. Any change you report later that is smaller than that spread is noise, and somebody will eventually ask.
Key Takeaways
- Accurate when named but absent from category questions is a classification failure, and it is not fixed on your own site.
- Category association is built from what other people write, in the places the category is discussed. That is slow work and it is durable once done.
- Retrieval is separated by language. A blended score across markets averages over environments that share almost no sources.
- Measure your baseline before changing anything: five category questions, three brand questions, three fresh runs each, sources recorded.
- Record how much your own runs disagree. That spread is the smallest change you are entitled to call a result.
Check yourself
Before you move on
Not scored, not recorded, and not part of the certificate. Both answers are settled by a sentence in this lesson, and the reasoning appears whichever option you pick.
- 01
Which piece of work does most to move category association?
- 02
You sell in seven European markets. What does the cross-language finding mean for your reporting?