How assistants choose which businesses to recommend.
A three-name answer looks effortless. Underneath it is retrieval, source weighting and a confidence check. Understanding those three steps tells you where to work.
Step one: the assistant retrieves
When the question has a local or commercial intent, most assistants do not answer from memory alone. They run searches — often several — and pull back pages: directories, review sites, news, and business websites. What comes back depends on how well each page matches the question, but also on how clearly the page states what it is about. A page that says, in plain text, “emergency plumber, Sabadell, 24 hours, call-out fee €60” is retrievable for a dozen different phrasings. A page that says “we care about your home” is retrievable for none.
Step two: the assistant weighs sources
Retrieved pages are not equal. Assistants prefer sources that are consistent with each other, that look authoritative, and that are specific. Three sources saying the same thing about a business make the assistant confident. Three sources saying different things make it hedge — or pick the one it trusts most, which is frequently a large directory rather than your own site.
This is why a single stale directory entry can do so much damage: it is often the most “authoritative-looking” source the assistant has about you, and if your own website does not contradict it clearly, the stale fact wins.
Step three: the assistant checks its own confidence
Assistants are tuned to avoid saying things they cannot support. Before naming a business, the model effectively asks: do I know what this is, where it is, and what it does — and would I be embarrassed to be wrong? Businesses with strong entity clarity pass that check. Businesses that are ambiguous — same name as another company, no address, vague services — fail it silently. The assistant does not say “I'm not sure about X”; it just names Y and Z instead.
What this means for the three or four names
- The businesses named are those the assistant can describe confidently, not necessarily the best ones.
- The order is influenced by how specifically each business matches the question, so a page per service beats one generic page.
- The sentence about each business is assembled from whichever sources were retrieved, so errors in those sources appear verbatim.
The four assistants are not identical
In our reports the four assistants frequently disagree. Perplexity leans hardest on live retrieval and cites sources openly, so it is quickest to reflect a fix — and quickest to repeat a stale directory. ChatGPT and Gemini blend retrieval with what they already know and tend to favour businesses with broad web presence. Claude is more likely to hedge or ask a clarifying question when its picture of a business is thin. A business can be a default answer in one and absent in another, which is why we measure all four rather than treating one as representative.
Where to work
In order: make your own website unambiguous and structured; make the two or three sources the assistants actually rely on agree with it; then publish pages that answer specific buyer questions. A Visiqn report tells you which sources those are, so you do not spend a month updating forty directories when two of them matter.