How AI Assistants Choose Which Businesses to Recommend

The single most useful frame for understanding why AI assistants recommend one business and not another is this: a generative engine is composing an answer from sources it trusts, and it is using whatever real-time tools its operator has plugged in to ground that answer.

Different products plug in different tools. ChatGPT uses web browsing for recent information; Gemini is tightly coupled to Google's search index; Claude uses its own retrieval; Perplexity is an answer engine first and an LLM second. The exact ranking formulas are proprietary and change. But across all of them, there is a consistent short list of signals you can actually influence.

1. Whether your site can be read at all

The most boring answer and the one we see fail most often. Each AI vendor publishes a named crawler with a user-agent string. If your robots.txt blocks GPTBot, Google-Extended (Gemini's grounding feed), ClaudeBot, or PerplexityBot, you have removed yourself from the sources the model could draw on. This is the single highest-leverage check on the list.

2. Entity consistency across the web

Models are trying to resolve a single entity (your business) out of many separate mentions. If your business name, address, phone, and core services appear with slight variations across your own site, Google Business Profile, Yelp, state licensing pages, and BBB, the model's confidence drops. In the worst case it splits you into multiple weak entities, neither of which gets cited.

The fix is unglamorous: pick one canonical version of every fact and propagate it.

3. Structured data that names the things you do

Schema.org markup — Organization, LocalBusiness, Service, FAQPage — is not a ranking factor in the SEO sense. For generative engines it is a labeling system. Markup makes the model's job trivial: you have already told it 'this is a service, here is its name, here is its area, here is its provider.' Markup also feeds Google's AI Overviews and Gemini directly.

4. Citation frequency in trusted sources

Models triangulate. If a claim about your business — service area, certifications, response time — only appears on your own site, it is fragile. If the same claim appears on your site, an industry directory, a review platform, and a news article, it is durable.

This is why third-party platforms still matter, and why pruning bad citations is as important as building good ones. The version of your business on the open web is the average of how it is described everywhere, not just on your homepage.

5. Citation-ready content shape

Generative engines prefer to quote discrete, well-formed sentences. Marketing prose that requires the model to summarize is harder to use than a sentence that already states the fact.

Compare: 'For decades, our team has been dedicated to serving the homeowners of the greater metro area with the highest-quality water mitigation services available' vs 'We provide IICRC-certified water mitigation in Hennepin County, with 60-minute emergency response 24/7.' The second sentence is the one a model can drop into an answer verbatim.

6. Real reviews from real customers

Models trust review platforms because they have their own anti-fraud systems and a broad sample. A small business with a thin but real review presence on Google and a few specialized directories beats a business with paid placements and no organic reviews.

What you cannot influence (and shouldn't worry about)

The model's training data is locked in for any given snapshot. You cannot retroactively get into a model trained before your business existed, and you cannot tune the weights. What you control is the snapshot the next training and the live retrieval will see — your site, your structured data, your citations across the web, and your reviews from this point forward. That is what AI visibility work actually moves.