Ask an assistant to recommend a supplier and it names two or three businesses. It does not show a ranked list and it does not explain itself. At The Growth Bully, a Malta performance marketing agency, we get asked constantly how to get into that answer, and the honest starting point is that the mechanism is not a ranking at all. It is a selection, built from whatever the model can corroborate about you across sources it did not have to take your word for.
Once you see the selection for what it is, a lot of the usual advice stops making sense.
How do AI assistants decide which business to recommend?
They look for agreement. An assistant assembles a recommendation from sources that independently say similar things about you: your website, your listings, your reviews, directories, press and other people writing about you. Where those sources agree, the model treats the claim as settled and repeats it. Where only your own site says it, it usually gets dropped.
That is the whole difference between ranking and being recommended. A search engine can rank one page on its own merits. A recommendation is a claim about a business, and a model will not stake a claim on a single interested source. Your own marketing copy is the weakest evidence in the set, because every competitor has written the same thing about themselves.
Why does the same question give two different answers?
Because the assistant builds the answer fresh each time from whatever it retrieves in that moment. Retrieval shifts with phrasing, with location, with the model version and with which pages were crawled recently. Two people asking the same question an hour apart can get different names. Nothing is stored and nothing is owed to you.
So a screenshot of a good answer is not proof of a position, and a bad answer is not a penalty. The thing worth watching is whether you appear at all, across many phrasings, over weeks. We track that for ourselves with a fixed set of prompts rerun on a schedule, because a single check on a single day tells you almost nothing.
Why do reviews decide the best-of answers?
Because a question asking for the best or the top of anything needs a comparison the model cannot make itself. It cannot judge your work. So it borrows the only comparable and countable signal available: how many people reviewed you, how recently, and what they said. Review volume becomes the stand-in for quality.
The effect is blunt and a little unfair. A business with sixty reviews gets named ahead of a better business with four, because only one of them hands the model something to point at. Reviews are also the slowest signal to build, which is why the asking has to start long before you need it. The mechanics of getting them without being awkward about it are a job on their own, and we wrote that process up in our guide to building a review system. Most of those signals are read off your Google Business Profile rather than your website.
Does schema markup make an assistant recommend you?
No. Structured data does not rank you and it does not persuade anyone. It verifies. It lets a machine read your business name, your location, your services and your author attribution without guessing, so the model is confident it is talking about the right company. It removes doubt rather than creating preference.
That distinction matters because schema gets sold as a growth tactic. It is closer to filing paperwork correctly. Leave it out and you can be excluded from an answer you deserved, through nothing worse than ambiguity about who you are. Put it in and you have earned the right to be considered, which is not the same as being chosen.
Why is one excellent page not enough?
Because the assistant checks a claim against the rest of the web before repeating it. A business with one outstanding page and no third-party footprint does not read as modest. It reads as unverifiable. No reviews, no listings, no mentions anywhere else, and the model is left with one interested source and no corroboration, so it names somebody else.
This is the most common pattern we see in Malta. The site is good, the positioning is clear, the work is real, and the business is still absent from every answer, because everything true about it lives in exactly one place. The fix is not better copy. It is getting the same facts stated accurately somewhere you do not own.
Can you control what an AI assistant says about you?
No, and anyone promising otherwise is selling something that does not exist. There is no submission form, no account, no dashboard and no appeal. You can influence the evidence the answer gets built from, and you can make yourself easy to quote accurately. The output itself stays outside your control.
We sell this work and we still say that plainly, because the alternative is a client expecting a rankings report for something that has no rankings. What you can reasonably expect is to move from never appearing, to appearing sometimes, to appearing across a widening set of phrasings. That is the real measure, and it is a slow one. The optimisation checklist that supports it sits in our piece on how businesses show up in AI search answers.
What actually changes whether you get recommended?
Five things, in rough order of effect: facts corroborated across independent sources, review volume and recency, a clearly stated answer to the question being asked, unambiguous structured data, and mentions on properties you do not control. Almost everything else people spend money on is downstream of those five.
- Say the same thing everywhere. One business name, one address, one description, across your site, your profiles and every directory you appear in. Contradictions get resolved by leaving you out.
- Earn reviews continuously rather than in bursts. Recency counts, and a quiet year reads as a business that may have stopped trading.
- Answer the question in the first 40 to 60 words of the page, in the words people actually use. An assistant lifts a complete answer and skips a wind up.
- Publish something only you can publish. First party numbers and clear definitions get quoted because they cannot be sourced anywhere else, which is the whole reason we keep a public benchmark report current.
- Get mentioned off your own property. A directory entry, a client naming you, a published interview. One independent source outweighs ten pages of your own copy.
Notice that four of the five sit outside your website. That is the part most businesses get wrong, because the website is the only bit they can edit on a Tuesday afternoon.
Where does this sit against normal search work?
In the same programme, not in competition with it. The sources an assistant trusts are mostly the sources that already rank, so search visibility and AI visibility are built from the same materials. The difference is emphasis: search rewards the best page, while a recommendation rewards the best corroborated business.
None of this is quick and none of it can be bought as a position. It is the slow accumulation of being a business that is easy to verify, which happens to be exactly what a human buyer checks before they call you. If you want that corroboration layer built alongside the paid and organic work rather than after it, that is what our lead generation and digital marketing programmes are for. Tell us what you want to be recommended for and we will tell you what is missing.