How AI assistants decide which businesses to mention.

Two places an answer comes from

A large language model knows things in two ways. The first is what it learned in training: a snapshot of the public web, frozen at a cutoff date. If your business was well known and widely written about before that date, the model may name you from memory.

The second is what it looks up right now. For questions about products, prices, local services or anything recent, ChatGPT, Gemini, Perplexity, Copilot, Claude and Google’s AI Mode increasingly run a web search, read a few results and write an answer from them. This is called retrieval, and it’s where most businesses can actually move the needle, because it reflects your site as it is today.

Training decides whether the model has heard of you. Retrieval decides whether it recommends you today.

1. 1SearchRewrites the question into several queries
2. 2ReadFetches a handful of pages
3. 3Cross-checkKeeps what several sources agree on
4. 4AnswerNames businesses and cites sources

Step 1: the assistant searches

A question like “who should I hire for commercial HVAC in Dallas” rarely becomes one search. The assistant breaks it into several: the service, the location, reviews, maybe pricing or certifications. Google calls this query fan-out, and the others work in a similar way.

Those searches run against a web index. Gemini, AI Overviews and AI Mode use Google’s. Copilot uses Bing’s. ChatGPT search draws on third-party search providers and on OpenAI’s own crawler. Perplexity runs its own index alongside others. The practical point: if you don’t rank anywhere for the pieces of the question, you’re unlikely to be read at all. Classic SEO is the entry ticket.

Step 2: it reads a handful of pages

For each search, the assistant opens a small number of results and pulls text out of them. Three things decide whether your page makes it through.

Can the crawler get in?

Each company sends its own crawlers, and each one checks your robots.txt first. Many sites block some of them by accident, often through a security plugin or a CDN bot setting.

Blocking a training crawler and allowing a search crawler is a legitimate choice. Blocking a search crawler means that assistant can’t cite you today.

Is the text in the HTML?

Most AI crawlers don’t run JavaScript the way a browser does. If your services, prices or locations only appear after scripts load, the crawler may see an almost empty page. Server-rendered or static HTML avoids this.

Is the answer easy to lift?

Assistants quote short, self-contained passages. A page that says “We install and service commercial rooftop HVAC units across Dallas–Fort Worth, with 24/7 emergency repair” is easier to use than three paragraphs about your passion for comfort.

Step 3: it looks for agreement

This is the step most businesses miss. An assistant that recommends a business is taking a risk, so it prefers names that several independent sources agree on: your own site, review platforms, industry directories, local news, comparison articles, forums and your Google Business Profile.

If those sources describe you the same way, with the same name, services, location and specialties, the assistant can confirm you. If your site says “commercial HVAC” and your listings say “home heating and plumbing”, it has two weaker signals instead of one strong one.

Consistency across sources does more for AI visibility than any single page on your own site.

What makes a page easy to cite

• A direct answer near the top. State what you do, for whom and where, in one or two plain sentences.
• Specific facts. Service areas, certifications, hours, typical project sizes, what’s included. Numbers with a source and a date.
• Structured data. JSON-LD for Organization or LocalBusiness, Service, Product, FAQPage and Article tells crawlers what each fact is.
• A visible date. “Last updated” helps assistants prefer current information.
• Questions in your buyers’ words. An FAQ written from real sales calls matches the questions people type into assistants.

What doesn’t work

• Paying for a guaranteed placement. No one sells a spot in ChatGPT’s answers. Anyone offering one is selling something else.
• Hidden instructions for AI. Text aimed at manipulating assistants tends to get a page treated as untrustworthy.
• Expecting llms.txt to do it alone. An /llms.txt file is a useful guide to your key pages, and some AI tools read it. It doesn’t replace crawlable pages or outside sources.
• Mass-produced pages. Hundreds of thin, near-identical location pages make it harder for an assistant to trust any of them.

How to measure it

AI visibility moves gradually, so track it month over month against the same set of questions. We look at, in order:

1. Mentions: how often you’re named for the questions your buyers ask.
2. Citations: how often your pages are linked as a source.
3. Share of voice: your mentions compared with your competitors’.
4. Sentiment: whether you’re described accurately and favorably.
5. AI referrals: visits from chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai, grouped as their own channel in GA4.
6. Branded search and sales: the numbers that tell you it’s paying off.

Where to start this week

• Open yoursite.com/robots.txt and check that OAI-SearchBot, PerplexityBot and Googlebot aren’t blocked.
• View your homepage with JavaScript turned off. Is your main service still on the page?
• Write one sentence that says what you do, for whom and where. Put it at the top of your homepage and your About page.
• Compare your name, services and location across your site, Google Business Profile and your top three directories. Fix what doesn’t match.
• Ask ChatGPT, Gemini and Perplexity the three questions your buyers ask most. Write down who gets named.

Written by the AMAGID editorial team. Platform behavior changes often. We review this article when the crawlers or search features it describes change.