How does an AI assistant decide who to recommend?

Section 01

The four layers of evidence

Think of an assistant as a very fast, very literal researcher with no local knowledge and a fear of being wrong. Asked for the best physio in Croydon, it doesn't have opinions. It goes looking for evidence, and it prefers evidence that agrees with itself. There are four layers of that evidence, and you can influence all four.

Layer 1 — Memory (what it absorbed in training)

Old, slow, and largely out of your hands. Businesses with long-standing mentions in news, Wikipedia, big directories and forums are baked in. You can't fix this month. You can start feeding it now.

Layer 2 — Live retrieval (what it fetches mid-answer)

This is your fastest win. When an assistant searches while answering, it can find your page today. That means crawlability, clear headings, real prices and dates. Fix it this week, benefit this month.

Layer 3 — Licensed and structured feeds

Some data arrives by contract, not crawl. Yelp licenses review data to OpenAI and Perplexity; Tripadvisor feeds Perplexity for travel; Google's AI Overviews, Gemini and AI Mode read Google Business Profile directly. That's why a neglected profile on a platform you personally dislike still costs you money.

Layer 4 — Consensus (does the internet agree?)

The tie-breaker. If four independent sources say the same thing about you, confidence goes up and you get named. If your address differs across three sites, the model quietly drops you rather than risk being wrong. Consistency beats cleverness.

Section 02

Where do AI citations actually come from?

Peec AI mapped 30 million cited sources in July 2026 and the top ten across major platforms were Reddit, YouTube, LinkedIn, Wikipedia, Forbes, G2, Yelp, Facebook, Medium and TechRadar. Engines differ: ChatGPT leans editorial, Google's AI surfaces lean more on YouTube, Facebook and Yelp.

• 32%

  Of citations in one early-2026 test of local queries in Google's AI Mode came from Yelp.

  (Four Horse Digital local citation test, early 2026)

• 30%

  Came from Reddit threads on the same test.

  (Four Horse Digital local citation test, early 2026)

• 16%

  The business's own website supplied only around 16% of citations.

  (Four Horse Digital local citation test, early 2026)

Read that again. Roughly five out of six citations came from somewhere you don't own. You have spent years polishing the one-sixth. This guide is mostly about the rest.

And don't build your whole plan on one platform: Semrush analysed 230,000 prompts and over 100 million citations and watched Reddit fall from roughly 60% of ChatGPT answers to about 10% after a 2025 update. Channels move. Being credible in several places is the only hedge.

The split by surface matters here too: only 11% of the URLs AI assistants cite rank in Google or Bing's top 10 for the same query, while 76% of Google AI Overview citations come from pages ranking in Google's top 10 (Ahrefs, 15,000 prompts, August 2025). AI Overviews follow the search results; AI assistants do not.