How AI Decides Who to Recommend

When an assistant answers "who's a good family accountant in Halifax?", it isn't consulting a ranking. There's no list at the top of which you could sit. Something stranger happens. In the seconds after the question, the assistant assembles an answer from what it already knows and what it can quickly verify. The businesses it names are the ones that survive that assembly.

Understanding the machine matters, because everything in this cluster follows from it.

Assistants are consensus machines. Before naming a business, an assistant cross-references what the web says about it. The business's own pages. Its Google profile. Its reviews. Directories. The odd mention. Where those sources agree, the business is a safe answer. Where they conflict, or barely exist, it's a risk. Assistants are tuned above all to avoid confident errors. They quietly skip risks.

The two mechanisms

Every answer gets built from two very different sources. And the difference decides what you can influence, and how fast.

What the model already knows. Assistants carry a compressed memory of the public web from their training. If your practice has existed publicly for years, some trace of you is probably already in there. Pages. Mentions. Reviews. This layer changes slowly, on the timescale of model updates. You influence it the way you influence reputation: by existing consistently in public, for a long time.

What it looks up right now. For current, local, factual questions — exactly the "who should I call" kind — assistants search the live web and read what they find, in real time. This layer responds to your work within weeks. Fix your pages. Strengthen your profile. Accumulate reviews. The next lookup meets the improved record. Nearly everything actionable in this cluster acts here.

What gets weighed, in practice

The verifiable pattern across assistants is consistent with what they can actually read.

Your own pages. Can the assistant retrieve them? Do they state plainly what you do, for whom, where, and at what cost? A page that renders empty without JavaScript, or buries its facts in brochure prose, contributes nothing. (Making your site legible to AI is the checklist.)

The review record. Volume, recency, and above all detail. "Excellent service" proves almost nothing. "Handled our cross-border filing after we moved from the US" gives the assistant exactly the evidence it needs to match you to a real question.

The independent record. Your profile, directories, association listings, mentions. This is where consensus actually gets computed — one consistent story across sources, or noise. (What AI reads about you elsewhere covers making the story agree.)

Citable substance. A practice that has published real answers to real questions gives the assistant something to quote. Being quoted is the strongest form of being named. (That asset gets built in content that compounds.)

Notice what's missing here. Tricks. There's no keyword density for a consensus machine. No schema hack that outweighs a thin review record. The system rewards being verifiably, consistently, specifically good. That's either bad news or excellent news, depending on whether you are.

Why consistency beats cleverness

One detail deserves its own paragraph, because it inverts the old SEO instinct. In a consensus system, a contradiction costs more than a gap. An old address on a directory. Two different practice names. Services claimed on your site but absent everywhere else. Each mismatch gives a cautious machine one more reason to prefer a competitor whose record simply agrees with itself. Before you add anything new to your public record, make what already exists tell one story.

Questions practices actually ask

Do assistants just repeat whoever ranks first on Google? Search results are one input, and there's real overlap. But assistants routinely name businesses from outside the top results when the record supports them, and skip top-ranked ones whose substance is thin. The overlap is the shared foundations, not the decision itself.

Can I pay to be recommended? Not currently. Any vendor implying otherwise is selling smoke. Ad placements inside assistants, where they exist, are clearly labelled — the organic answer is assembled, never sold.

Why does my competitor get named and not me? Run the same comparison the machine runs. Their pages state facts yours only imply. Their reviews name services yours don't. Their record agrees with itself. The DIY check turns this comparison into a monthly habit.

Does this differ between ChatGPT, Claude, and the rest? In details, yes. In mechanism, no. They're all consensus machines running the same two layers — which is why the work in this cluster transfers across all of them, rather than chasing any one.