How Do AI Assistants Decide Which Brands to Cite?

AI assistants can only cite what they can reach, retrieve and quote. In practice that means three filters: the engine’s crawler must be allowed to read your pages, your page must show up in the searches the assistant runs behind the scenes, and your content must contain something worth quoting, such as a clear claim, a statistic or a cited source. Most brands that are “invisible” in AI answers are failing at one of those three steps, and each one is something a marketing team can check.

Step 1: Can the AI engine read your site at all?

Every answer engine with live web results uses its own crawler, and each one can be blocked separately in robots.txt, often by accident.

Engine	Crawler that powers answers	What happens if you block it
ChatGPT search	OAI-SearchBot	OpenAI says opted-out sites will not be shown in ChatGPT search answers (apart from navigational links).
Perplexity	PerplexityBot	Perplexity recommends allowing it in robots.txt, plus its published IP ranges in any firewall.
Google AI Overviews and AI Mode	Googlebot	Pages must be indexed and eligible to show with a snippet in Google Search.

The common mistake is mixing up search crawlers and training crawlers. OpenAI’s GPTBot collects content that may be used to train its models. Blocking it is a separate decision from blocking OAI-SearchBot, which is what decides whether you can appear in ChatGPT’s search answers. A blanket rule that blocks every AI crawler also takes you out of AI search.

Check this week:

1. Open your robots.txt and look for rules covering OAI-SearchBot, PerplexityBot and Googlebot.
2. Ask your web or security team whether the CDN or firewall blocks bot traffic by default. Perplexity specifically recommends allow rules that match both user agent and IP.
3. After any change, allow about 24 hours: OpenAI and Perplexity both say robots.txt changes take up to a day to take effect.

Step 2: Do you show up in the searches the assistant runs?

When someone asks an AI assistant a question, the assistant usually doesn’t answer from one search. Google says AI Overviews and AI Mode may use “query fan-out”: they run several related searches across subtopics and data sources, then draw on a wider and more diverse set of pages than a classic results page.

This changes what “ranking” means. A question like “best private bank for expats in Hong Kong” might fan out into searches about fees, minimum balances, app reviews and account opening times. A brand with a clear, specific page for each of those subtopics has more chances to be retrieved than one with a single generic landing page.

It also means classic SEO still matters. Google is explicit that there are no extra requirements for AI features: crawlable pages, internal links and text-based content are the foundation. Google also says you don’t need special AI text files or markup to appear. Proposals like llms.txt may help other AI agents navigate a site, but they aren’t required for Google.

Step 3: Is your content worth quoting?

Once pages are retrieved, the model decides what to use in its answer. The most cited research on this is GEO: Generative Engine Optimization, by researchers from Princeton and IIT Delhi, presented at KDD 2024. They built a benchmark of 10,000 queries across 25 domains and tested nine ways of rewriting content.

What worked:

• Adding quotations gave the largest gain, raising the study’s visibility score from 19.5 to 27.8.
• Adding statistics (25.9) and citing sources (24.9) came next.
• The best methods improved visibility by roughly 30–40%.

What didn’t:

• Keyword stuffing gave little or no improvement, and on Perplexity it performed about 10% worse than the original text.

The most useful finding for challenger brands: lower-ranked pages gained the most. Citing sources improved visibility for fifth-ranked pages by 115%, while top-ranked pages lost about 30%. In AI answers, a smaller brand with better-evidenced content can take share from a market leader.

Two caveats. The study used a research benchmark and 2023–24 engines, which change often, so treat the percentages as directional. The authors also found that the best tactic varies by domain, so test on your own category’s questions instead of copying a checklist.

How do you know whether any of this is working?

Classic SEO reports don’t show how each assistant describes your brand or who it recommends instead, so measurement needs its own process:

1. Write down the 20–50 questions your customers ask before buying, in the languages they use. In Hong Kong that usually means English and Traditional Chinese.
2. Ask each assistant those questions regularly and record whether your brand is mentioned, recommended or cited, and who appears instead.
3. Track changes over time, because answers can vary between runs and shift as engines update.

Doing this by hand works for a handful of prompts. For a larger set, or for engines beyond the Western ones, a monitoring tool saves time. Geoneo, for example, tracks brand mentions and citations across 11 engines, including Asian engines such as DeepSeek, Qwen, Hunyuan and Kimi, and scores visibility across eight dimensions. Whichever route you take, the point is to get a baseline before you change your content, so you can tell what worked.

The short version

• Unblock the right crawlers. OAI-SearchBot, PerplexityBot and Googlebot decide whether you can appear in answers at all; training crawlers are a separate choice.
• Cover the subtopics. Query fan-out rewards specific pages that answer the follow-up questions.
• Make pages quotable. Clear claims, real statistics, quotations and cited sources beat keyword density.
• Measure before and after. Track a fixed set of prompts across engines so you know which changes moved your visibility.