# How do AI assistants decide which businesses to recommend?

> They search an index, open a few pages and quote what is clear, current and corroborated elsewhere. You cannot buy the slot, but you can be easy to find and safe to cite.

Short answer: an assistant does not keep a list of favourite businesses. When someone asks for a recommendation it searches an index, opens a handful of pages, reads the plain text and composes an answer from what is clear, current and repeated by more than one source. Nobody can buy or guarantee a place in that answer. What a business can control is whether it is easy to find, easy to read and safe to quote.

## What actually happens when someone asks an assistant for a recommendation?

The assistant turns the question into one or more searches. ChatGPT and Microsoft Copilot lean on Bing's index alongside their own. Google's AI Overviews and AI Mode use Google's index. Claude and Perplexity run their own search crawlers. From the results, the assistant fetches a few pages live, reads them, and writes an answer with links to the pages it relied on.

That means three different kinds of visitor reach a website, and they are controlled separately in robots.txt:

| Visitor | Examples | What it is for |
| --- | --- | --- |
| Training crawler | GPTBot, ClaudeBot | Collects text that may be used to train future models |
| Search crawler | OAI-SearchBot, Claude-SearchBot, PerplexityBot | Builds the index the assistant searches at answer time |
| User-triggered fetcher | ChatGPT-User, Claude-User, Perplexity-User | Opens one page because a person's question needed it |

Blocking a training crawler does not remove a site from answers. Blocking a search crawler does. Many sites block everything with "AI" in the name and then wonder why assistants never mention them.

## Do assistants see what a visitor sees?

Usually not. The crawlers behind ChatGPT, Claude and Perplexity read the HTML a server sends and do not run JavaScript, according to a large log analysis published by Vercel and MERJ. Google's AI features inherit Google's renderer, which does run scripts. The safe rule is simple: anything you want quoted must be in the HTML as delivered, not painted in later by a script. A quick test is to open "view source" and search for a sentence from the page.

## What makes a page likely to be quoted?

Assistants lift short passages that answer a question on their own. The traits that show up repeatedly in citation studies are practical ones:

1. The title is the question a person would ask.
2. The first paragraph answers it completely, without needing the rest.
3. Headings are the follow-up questions, in the order a reader would ask them.
4. Claims are specific, and sourced where they rest on someone else's work.
5. The page shows when it was written and updated, and it is kept current.
6. Steps are numbered and comparisons are in a table, because both extract cleanly.
7. Limits are stated honestly. A page that says when something is the wrong choice reads as more trustworthy, to people and to models.

The best controlled evidence is the GEO study by Aggarwal and colleagues, presented at KDD 2024. In its test setting, adding citations, quotations and statistics raised a page's visibility in generated answers by up to 40 percent, while keyword stuffing did nothing or made things worse. It used a simulated engine rather than a live assistant, so treat the size of the effect as indicative.

## What matters outside your own website?

For "recommend a company that does this" questions, most of the signal is off the site. Assistants corroborate. They look for the same business described the same way in several places: a Google Business Profile and Bing Places listing, review platforms, industry directories, "best of" lists, LinkedIn, forums such as Reddit. Large vendor studies of cited brands find that plain mentions of a brand across the web track citation more closely than backlinks do. Those studies are not peer reviewed, but they agree with each other.

Two practical consequences follow. Use one name, one address, one phone number and one domain everywhere, exactly the same. And earn real mentions and real reviews, because those are what an assistant can check.

## What is proven, and what is only repeated?

| Measure | Evidence today |
| --- | --- |
| Main content in server-delivered HTML | Strong. Vendor documentation and log studies agree |
| Allowing search crawlers in robots.txt | Strong. Documented by OpenAI, Anthropic and Perplexity |
| Being indexed by Google and Bing | Strong. Most assistants search one or both |
| Fresh, answer-first, specific pages | Moderate. Consistent across several studies |
| Brand mentions and genuine reviews elsewhere | Moderate. Large vendor studies, not peer reviewed |
| Structured data such as JSON-LD | Mixed. Helps search engines understand a page; no proven direct effect on answers |
| An llms.txt file | Unproven. Google says it is not needed and no major assistant has confirmed reading it. Cheap to add, not a lever |
| Hidden text or instructions aimed at AI | Harmful. Search engines treat it as spam |

## How do you know whether it is working?

Measure it rather than guess. Server logs show visits from the named crawlers and fetchers. Visitors who arrive from an assistant carry a referrer such as chatgpt.com, perplexity.ai, copilot.microsoft.com or gemini.google.com, and ChatGPT adds utm_source=chatgpt.com to its links. Once a month, ask the assistants the five or ten questions your customers would ask and note whether you appear, who does, and which pages they cite.

## When this is the wrong choice

It is the wrong investment if you expect a shortcut. Answers differ between assistants, between users and from week to week, and a new business with no footprint elsewhere will not be recommended on the strength of its own pages alone. The work is ordinary: be crawlable, be clear, be current, be consistent, and be talked about for real reasons. Anyone promising a guaranteed place in AI answers is selling something that does not exist.

## How we approach it at Ribhu Labs

This site is built the way this page describes. Every page is server rendered and works without scripts, each service opens with a direct answer, articles follow the structure above, and the owner studio counts visits from AI crawlers and referrals from assistants so the effect can be seen rather than assumed. Read how we keep a rich site fast in [briefing a 3D website](/insights/brief-a-3d-website-that-stays-fast), see the [websites and experiences](/solutions/websites) service, or [start a conversation](/project) if you would like the same checks run on your own site.

## Sources

- OpenAI, [Overview of OpenAI crawlers](https://developers.openai.com/api/docs/bots)
- Anthropic, [Does Anthropic crawl data from the web?](https://support.claude.com/en/articles/8896518)
- Perplexity, [Perplexity crawlers](https://docs.perplexity.ai/docs/resources/perplexity-crawlers)
- Google Search Central, [AI features and your website](https://developers.google.com/search/docs/appearance/ai-features)
- Aggarwal et al., [GEO: Generative Engine Optimization](https://arxiv.org/abs/2311.09735), KDD 2024