How AI Assistants Actually Decide Which Businesses to Recommend

AI assistants don't pick businesses the way Google ranks pages. There's no single index, no ten blue links, no crawl-and-rank loop you can game with backlinks. When someone asks ChatGPT or Gemini "who's the best plumber near me" or "what CRM should a small agency use," the model assembles an answer from three separate systems, each with its own rules. Most businesses optimize for one of them, ignore the other two, and then wonder why they never get named.

This matters more every quarter. 51% of buyers now start product research inside an AI chatbot, up from 29% a year earlier, according to G2. That's not a future trend. That's already where the first impression happens, before anyone touches a search bar.

So here's how the decision actually works, why your homepage almost never gets cited, and what to do about it.

The three layers behind every AI recommendation

Every recommendation an AI makes comes from some mix of three layers. Understanding which layer you're trying to win changes everything about your strategy.

Layer 1: Training memory

This is what the model already "knows" from its training data. When an AI confidently names a business without searching anything, it's pulling from this baked-in memory. The catch: training data is frozen at a cutoff date, skewed toward whatever appeared most often across the open web, and impossible to edit directly. You can't submit a form to update it.

What dominates this layer is repetition. Brands mentioned constantly across forums, articles, and discussion threads get encoded as "known." Over 25% of ChatGPT's US citations trace back to just two sources, Wikipedia and Reddit, per 5W Research. The model learned the web's consensus, and that consensus lives in a handful of high-trust, high-volume places.

Layer 2: Live retrieval

When the model isn't sure, or the question is time-sensitive or local, it searches the live web and reads what comes back. This is retrieval-augmented generation, and it's where most of your near-term opportunity sits. The pages it pulls in this moment shape the answer in real time.

The good news is you can influence this layer directly. Yext analyzed 6.8 million AI citations and found 86% of the sources AI cites to recommend businesses are ones you can influence. Directories, review platforms, third-party roundups, your own structured pages. These are all in play.

Layer 3: Entity understanding

This is the model's grasp of who you are as a distinct thing in the world. Not a URL, an entity. Is "Acme Plumbing" a real business with a consistent name, location, service area, and category across the web? Or is it a fuzzy collection of conflicting listings the model can't confidently resolve?

Entity clarity is the quiet gatekeeper. If the AI can't pin down what you are, it won't risk recommending you. This is why a small business with sloppy, inconsistent listings gets skipped even when its actual service is excellent.

Why your homepage rarely gets cited

Here's the uncomfortable part. You probably spent years and real money making your homepage rank on Google. In the AI world, that work barely transfers.

Only about 12% of AI-cited URLs overlap with Google's top 10 results, according to Ahrefs. And the overlap is shrinking fast. AI Overview citations coming from Google's own top 10 fell from 76% to 38% in under a year. The two systems are diverging, not converging. Winning Google no longer means winning AI.

Your homepage loses for specific, fixable reasons. It's a marketing page, written to persuade a human who's already on your site, not to answer a neutral question a stranger asked an AI. It rarely contains the structured, factual, citable content a model wants to quote. And AI assistants prefer third-party validation over self-promotion, because a roundup or review reads as more trustworthy than your own "we're the best" copy.

The signals that actually predict AI visibility look nothing like classic SEO. Ahrefs studied 75,000 brands and found YouTube mentions and branded web mentions predict AI visibility far more than Domain Rating. Being talked about across the web beats having a high domain score. The model is tracking your reputation footprint, not your link equity.

What gets recommended instead

If not your homepage, then what? The model reaches for sources that read as objective and that it has seen validated repeatedly.

Source type	Why AI favors it	Your level of control
Wikipedia and large reference sites	Treated as authoritative consensus	Low, but possible over time
Reddit and forums	Read as real, unscripted opinion	Medium, through genuine presence
Directories and review platforms	Structured, comparable, third-party	High
Third-party roundups and "best of" lists	Pre-digested recommendations	Medium to high, through outreach
Your own structured FAQ and data pages	Directly quotable facts	High
Your homepage	Self-promotional, hard to quote	High but low payoff

The pattern is clear. The model wants facts it can lift and attribute to something that looks neutral. The more your information lives in those formats, the more often you get named.

The local recommendation squeeze

Local is where the stakes get brutal. AI names only 3 to 5 businesses in a typical local answer. There's no page two. You're either in the short list or you're invisible.

The numbers expose how tight that funnel is. Roughly 1.2% of locations get recommended by ChatGPT, compared to 35.9% that appear in Google's local 3-pack. AI is dramatically more selective. The businesses that make the cut have clean entity data, consistent listings everywhere, and a visible trail of third-party validation the model can lean on.

What actually moves AI visibility

Now the useful part. Princeton researchers (KDD 2024) ran causal experiments on what lifts AI visibility, and the results are concrete. Adding statistics to your content lifted visibility around 32%. Adding expert quotations lifted it around 41%. Adding citations to authoritative sources lifted it around 30%. Treat the exact magnitudes as directional, but the direction is unambiguous: AI rewards content that reads like a well-sourced reference, not a sales page.

The most striking finding is who benefited most. Lower-ranked pages gained the most, up to 115%. If you're not the established leader in your space, this is good news. The playbook that wins AI citations favors the underdog who structures content well over the incumbent who coasts on brand.

What that means in practice:

• Write factual, structured content that answers real questions directly, with numbers, named experts, and links to authoritative sources baked in.
• Get listed accurately and consistently everywhere that matters, so your entity is unambiguous across the web.
• Earn third-party mentions and roundup placements, because the model trusts those more than anything you say about yourself.
• Build a real reputation footprint, including video and branded mentions, since those predict visibility better than domain scores.

One thing to skip: llms.txt. It has roughly zero correlation with citations, and no major AI provider honors it. Don't waste a sprint on it.

Where to start

You can't fix what you can't see. The first move is finding out how AI assistants currently perceive your business across all three layers, what they already know, what they retrieve, and whether your entity is even legible.

Run the free AI visibility audit. It shows you exactly where you stand and which gaps are costing you recommendations right now.

If you want the full playbook to act on what the audit surfaces, the $39 GEO Action Kit walks you through the specific fixes, templates, and checklists step by step.

The shift is already here. Half of buyers start in a chatbot, the citation rules have nothing to do with Google, and the underdog who structures content well can leapfrog the incumbent. The businesses that move now get named. The ones that wait stay invisible.

FAQ

How do AI assistants decide which businesses to recommend? They combine three layers: training memory (what the model already learned from the web), live retrieval (pages it searches and reads in real time), and entity understanding (how clearly it can identify your business). A recommendation usually draws on all three.

Why doesn't my homepage get cited by AI? Homepages are written to persuade humans, not to answer neutral questions, and AI prefers third-party validation over self-promotion. Only about 12% of AI-cited URLs overlap with Google's top 10, so ranking on Google no longer guarantees AI visibility.

Does ranking high on Google help with AI recommendations? Less than you'd expect. AI Overview citations from Google's top 10 fell from 76% to 38% in under a year, per Ahrefs. The two systems are diverging.

What actually increases AI visibility? Princeton research found adding statistics (32%), expert quotations (41%), and citations to authoritative sources (~30%) causally lift AI visibility, with lower-ranked pages gaining the most. Consistent listings and third-party mentions matter more than domain authority.

Does llms.txt help with AI citations? No. llms.txt has roughly zero correlation with citations, and no major AI provider currently honors it.

How many businesses does AI recommend in a local answer? Usually only 3 to 5. About 1.2% of locations get recommended by ChatGPT, versus 35.9% appearing in Google's local 3-pack, so AI is far more selective than traditional local search.