How AI Assistants Choose Which Businesses to Recommend
AI assistants choose businesses by blending reviews, directory listings, entity facts, and citable web content — with different source mixes per platform and no public ranking formula. Local owners influence outcomes by fixing verifiable signals and measuring mention rates across engines, not by chasing a single SEO metric.
Nobody has the algorithm — but patterns are visible
OpenAI, Google, Anthropic, and Perplexity do not publish "local business ranking factors" for their assistants. Internally, systems blend training knowledge, retrieval, browsing, and safety filters in ways that change between releases.
Still, when you sample hundreds of buyer-intent prompts, patterns emerge. Local businesses that appear repeatedly tend to share signal classes — not a secret trick.
This article explains those classes and what you can actually control.
The decision in one sentence
When a user asks "Who's the best plumber near me?", the model estimates which businesses plausibly satisfy the intent given public evidence — then names zero to three options in fluent prose.
That is closer to evidence aggregation than keyword matching.
Signal classes models appear to weight
1. Reviews and social proof density
Aggregate star rating, review count, recency, and text themes matter. A 4.8 with 400 reviews mentioning "emergency" and "fast" beats a 5.0 with twelve generic reviews when the prompt stresses urgency.
Models often quote counts and paraphrase sentiment. Review gating destroys trust — platforms prohibit it, and skewed distributions look unnatural.
Guide: Google reviews the right way.
2. Directory and map entity graphs
Google Business Profile, Apple Business Connect, Bing Places, Yelp, and vertical directories (Healthgrades, Avvo, Angi) feed entity resolution. NAP consistency — matching name, address, phone — reduces ambiguity about which business is which.
Under-claimed Apple listings are a common blind spot: Apple Business Connect guide.
3. Query–theme alignment
The user's phrasing selects which review themes and service labels matter. "Cosmetic veneers" matches different evidence than "family dentist accepting kids."
Service lists on GBP, schema areaServed, and FAQ content sharpen alignment.
4. Citable, grounded facts
Perplexity and Claude especially favor pages with traceable claims. A data study citing public datasets outperforms adjectives. Thin "best in town" copy does not ground well.
5. Third-party mentions
Press from real milestones, merit-based awards, and authoritative listicles add corroboration. Pay-to-play badges add little durable signal.
6. Recency and operational signals
Fresh GBP posts, recent reviews, and accurate hours suggest the business is active. Stale profiles lose to competitors with current data.
Platform differences matter
Engines do not share one corpus. Independent analyses of AI visibility sampling report low cross-platform domain overlap — on the order of ~11% in some market samples — meaning Gemini and ChatGPT may name different winners on the same prompt class.
| Platform | Observable tendencies |
|---|---|
| ChatGPT | Blends training + optional browsing; sensitive to widely discussed brands |
| Gemini | Google ecosystem heavy — GBP, reviews, Maps adjacency |
| Claude | Strong grounding on provided/cited pages |
| Perplexity | Explicit URLs; rewards citable statistics |
| Grok | Variable; growing local use in tech-forward demos |
| AI Overviews | Google SERP synthesis; local pack adjacency |
Strategy implication: measure all major platforms, not one screenshot from ChatGPT.
Read: 11% platform overlap.
What models do poorly (and how to respond)
Hallucinated facts
AI may state wrong hours, phone numbers, or services you do not offer. Usually this traces to stale or conflicting listings, not malice.
Response: trace sources, fix NAP/schema, queue changes through an approval workflow. Guide: AI reputation repair.
Popularity bias
Famous brands surface on vague prompts. Niche local winners need stronger local evidence density — reviews, directories, entity clarity — to compete on specific intent queries.
Training cutoff vs live retrieval
A business launched last year may be invisible in memory-heavy modes but appear when browsing retrieves current directories. You cannot control which mode a user hits — cover both with listing and web presence.
What you cannot control
- Model version and retrieval toggles
- Regional product differences
- Safety refusals on certain query types
- Competitor actions
- Guaranteed placement — any vendor promising this is not credible
What you can control
- Review velocity and ethical solicitation
- Listing accuracy across ecosystems
- Entity profile — llms.txt, JSON-LD, factual About copy
- GBP completeness and freshness (Dominance-tier autopilot helps)
- Citable content — studies, FAQs, sourced stats
- Monthly mention rate tracking and competitor comparison
Operational playbook
Week 1: Free six-platform scan — establish baseline share of AI voice.
Weeks 2–4: Fix NAP drift, claim Apple BC, align schema. Approve review response drafts.
Days 30–90: Build review themes matching buyer prompts; publish entity-dense FAQ content.
Ongoing: Monthly rescan; Super Pixel attribution on AI traffic; accuracy repair when wrong facts persist.
AIrecommend.ai maps each step to Growth Engine modules with client approval on outbound work.
Programs: Growth $4,997/mo · Dominance $9,999/mo — pricing.
Related services
Understanding selection logic is not academic. It is how you stop losing calls to a competitor who simply looked more real to the model.