geo fundamentals
How do AI assistants pick the businesses they recommend?
Last updated
How does an AI assistant choose who to name?
Two mechanisms produce every recommendation: what the model already “knows” from training (model memory) and what it looks up on the live web at answer time (grounded retrieval). Most consumer assistants blend both, then generate a short answer naming only the two or three options best supported by that evidence.
Nothing about this is a ranking. The process is probabilistic: run the same prompt twice and the shortlist can change. What is stable — and optimizable — is the evidence pool each answer draws from.
What is the difference between model memory and grounded answers?
Model memory is baked in at training time and changes only when a new model ships; grounded answers are assembled from live web fetches and can change within days. They fail differently, and they are fixed differently — which is why they should never be measured as one number.
| Model memory | Grounded retrieval | |
|---|---|---|
| Source | Training data (web snapshots, licensed corpora) | Live search + page fetches at answer time |
| Freshness | Months old — frozen at the training cutoff | Days old or better |
| You influence it via | Consistent public facts: directories, Wikipedia-class sources, stable entity data | Crawlable pages, review platforms, quotable answer-shaped content |
| Speed of change | Per model release | Weeks — as soon as sources are re-crawled |
| Blocked by | Training-bot blocks (GPTBot, ClaudeBot…) | Retrieval-bot blocks (OAI-SearchBot, PerplexityBot…) |
Where do grounded answers get their information?
From a small, repeated set of source types: search-index results, review platforms, industry directories, editorial “best of” roundups, community threads, and — when it is machine-readable — your own website. Which platforms dominate varies sharply by vertical.
- Review platforms — ratings and review language are quoted as the justification for recommendations (Google reviews, Yelp, G2, Healthgrades…)
- Vertical directories — license and credential records assistants use to verify a business is real and in good standing
- Editorial roundups — “best X in Y” articles often get near-verbatim reuse
- Community threads — Reddit and neighborhood forums supply “what locals actually say” evidence
- Your website — prices, services, hours, and FAQs, if retrievable as plain HTML with valid schema.org structured data
Every industry guide on this wiki lists the specific platforms assistants cite in that vertical.
Which crawlers need access to your site?
Six retrieval agents matter most for live answers, and six training crawlers shape future model memory. Blocking the first group removes you from grounded answers now; blocking the second only limits future training presence — a legitimate choice, but know which lever you are pulling.
retrieval — live answers
- OAI-SearchBotOpenAI — ChatGPT search results
- ChatGPT-UserOpenAI — Pages ChatGPT opens for a user mid-conversation
- PerplexityBotPerplexity — Perplexity’s answer index
- Perplexity-UserPerplexity — Live page visits during answers
- Claude-UserAnthropic — Pages Claude fetches when users search
- Claude-SearchBotAnthropic — Claude’s search indexing
training — future memory
- GPTBotOpenAI
- ClaudeBotAnthropic
- Google-ExtendedGoogle (Gemini training)
- CCBotCommon Crawl (feeds many models)
- Applebot-ExtendedApple
- meta-externalagentMeta
What makes content quotable to an LLM?
Content that already looks like an answer: a question heading, a direct one-to-two sentence response, then detail — plus lists, tables, and statistics with linked sources. The original GEO study measured up to 40% visibility gains from adding citations, quotations, and statistics alone.
Equally important is what breaks quotability: content rendered only by JavaScript, PDFs and images holding key facts, vague claims with no numbers, and structured data that fails to parse. Validate yours at validator.schema.org.
Frequently asked questions
Do AI assistants have a ranking algorithm for businesses?
Not in the search-engine sense. There is no indexed leaderboard — each answer is generated fresh by weighing whatever the model remembers and retrieves. That is why the same question can name different businesses on different runs.
Why does ChatGPT recommend my competitor but not me?
Usually one of three reasons: your site blocks or hides content from retrieval bots, your entity is inconsistently described across the web, or your competitor simply has more presence in the sources the answer was assembled from — reviews, directories, roundups.
Does blocking GPTBot remove me from ChatGPT answers?
Blocking GPTBot only affects future model training. Live grounded answers come through OAI-SearchBot and ChatGPT-User — blocking those removes your pages from real-time answers. The two decisions are independent and worth making deliberately.
Can I pay to be recommended by an AI assistant?
No major assistant sells recommendation placement in organic answers today. Influence runs through the retrievable evidence about your business — which is good news: it is earnable, and the playbook is what this wiki documents.
How fresh is the information assistants use?
Grounded answers can reflect pages crawled within days. Model memory is only as fresh as the last training cutoff, often many months old — which is why measuring the two separately matters when you evaluate your visibility.
What single fix most improves the odds of being recommended?
Access. If AI retrieval bots cannot fetch your site — blocked robots.txt, JavaScript-only content — assistants literally cannot quote you, and every other optimization is moot. Verify access first with a free audit.
see where you stand
Is AI already recommending your business?
Run the free audit to score any page against the 19 GEO checks this wiki teaches — no account, no API keys. Then probe real ChatGPT, Claude, Gemini, and Perplexity answers with your own keys to measure your actual mention rate.