An assistant that answers "what is the best X for Y" is not picking a favourite. It is running three separate steps, and a brand can fail at any one of them for completely different reasons. Knowing which step you are failing at is the whole diagnosis.
Step one: becoming a candidate
Before the model writes anything, something fetches a handful of documents to read — a search index, a crawl, or a cached retrieval set. Typically a few dozen pages, out of millions. If your page is not in that handful, nothing else in this article applies to you.
What puts you there is unglamorous: pages a crawler can reach, other sites referencing you, and text that plainly contains the words of the question. This is the step most companies fail, and the step they least want to hear about, because the fix is slow.
Step two: being selected from the candidates
Now the system has thirty documents and room to use maybe four. The selection leans on things you can influence directly:
- How directly a passage answers. A paragraph that states the answer beats a page that eventually gets there.
- Agreement with other sources. A claim repeated across independent sites is treated as settled; a claim that only you make is treated as a claim.
- Specificity. A sentence with a number, a date or a named limit survives selection; a sentence of adjectives does not.
- Freshness, where the question implies it. For "best tool in 2026", an undated page is a weak candidate.
Step three: surviving the sentence
Finally the model writes prose. Here brand names get dropped for a mundane reason: the fact being stated did not need a name. "Most tools in this category charge monthly" needs no source. "Magneth publishes your site across hundreds of partner sites" cannot be said without naming who does it.
So the brands that survive synthesis are the ones attached to something only they do, at a stated price, with a stated limit. Generic positioning is the thing that gets edited out — not by malice, but because a name adds no information.
The three steps, side by side
| Step | What decides it | What you do about it | How fast it moves |
|---|---|---|---|
| Candidate set | Crawlability and off-site references | Be published and referenced elsewhere | Weeks to months |
| Selection | Directness, corroboration, specificity | Rewrite the page's opening paragraph | Days |
| Synthesis | Whether a name carries information | Attach a number to your claim | Immediate, once indexed |
What does not influence the choice
Advertising budget does not buy a mention in the organic part of an answer — paid placements, where they exist, are labelled separately. Page volume does not help: forty thin pages saying the same thing is one source, repeated. And nothing described as a hidden instruction to the model works; production assistants do not follow text buried in a page.
The limit worth saying out loud
These systems are not deterministic. Ask the same question twice and you can get two different sets of brands. That is not a malfunction and it is not something you can tune away — it means a single check tells you almost nothing. Ask the same question ten times across a month and count how often you appear. The rate is the measurement; one answer is an anecdote.
