AI Search Ranking Factors: How AI Chooses Which Sources to Cite and Recommend

The owner of a bookkeeping firm asks ChatGPT to "recommend a reliable bookkeeping firm for a small business" and sees three competitors in the answer. He asks Perplexity the same question and again does not find his company. The first thought is that someone is buying these spots, or that the AI picks at random. Neither is true. There is a mechanism behind the choice, and it is now understood well enough to influence.

I will explain it the way I would explain it to a colleague without a technical background: where AI gets its knowledge, how it finds and filters pages, why it readily cites some websites and ignores others, and what the AI search ranking factors mean in practice for a company that wants to become a source AI trusts.

What an AI answer is and why the shortlist matters

Let's start with the basics. An AI answer is text an AI assistant writes for a specific question. Not a list of ten links, as in classic search, but a ready-made recommendation: who to choose, what it costs, what to watch out for. Links to the websites the information came from often appear next to it.

People see these answers in two places. The first is right in search: Google AI Overviews and AI Mode on top of Google results, Copilot answers in Bing. The second is chat assistants: ChatGPT, Gemini, Claude, Perplexity, Copilot and others, where a person types a question as if messaging a friend and gets a detailed reply. We compare what this looks like on screen in how AI answers cite sources.

For a business, one thing matters most: an answer typically names only a handful of companies. The user chooses among them and rarely looks further. So the question "how does AI choose sources" is really a question about whether you make the shortlist your customer will choose from.

Two sources of knowledge: training data and real-time search

Every modern AI assistant has two ways of knowing about your company. It is important to tell them apart, because you influence them differently.

Knowledge from training

A large language model (LLM) is trained on a huge body of text collected up to a certain date. The model does not store those texts like a database. It learns relationships between words and concepts. Big brands that thousands of pages talk about are well "known" to the model. A small company with a website and a couple of mentions is probably unknown to it, or remembered inaccurately.

This knowledge has three weak points. It goes stale: the model knows nothing about events after its data cutoff. It cannot be corrected quickly: changes only appear in a future model version. And it is fuzzy on details: the model may confidently quote an old price or confuse companies with similar names.

Real-time search (RAG)

To answer questions about current things like prices, companies, locations and recent events, AI assistants look up information at the moment they answer. This approach is called retrieval-augmented generation, or RAG. Simplified, the process looks like this:

1. Understand the question. The AI decides whether its own knowledge is enough or a search is needed. Questions about companies, prices and choosing a vendor almost always trigger a search.
2. Rewrite it. One question turns into several search queries. A hypothetical example: "recommend a bookkeeping firm for a small LLC" might become "bookkeeping services for small LLC in New York City pricing," "best bookkeeping firms for small business" and "outsourced bookkeeping reviews."
3. Find pages. The queries go to a search index. Google AI Overviews and AI Mode use the Google index, Copilot relies on Bing, and ChatGPT, Claude and Perplexity search the web with their own crawlers (OAI-SearchBot, Claude-SearchBot, PerplexityBot) and search partners.
4. Read and select passages. The AI opens the pages it found, splits the text into passages and keeps the ones that best answer the question.
5. Cross-check and write. Passages from different sources are compared, assembled into a coherent answer, and links to the pages used are added.

Here is the key takeaway. In search mode, a company gets into the answer in two stages: first its page has to be found, then its passage has to be more useful than a competitor's. Classic SEO handles the first stage, generative engine optimization (GEO) handles the second. That is why GEO is built on top of good old SEO.

AI search ranking factors at a glance

There is no official list of AI search ranking factors, and no platform publishes its weights. But the mechanics above make it clear which qualities decide whether a page gets cited. This is the working model we use in audits:

Stage one: the page has to be found and read

The most frustrating reason for missing out on AI answers is purely technical. AI cannot cite what it could not open. These are the typical obstacles we find in audits:

• the website is poorly indexed in Google or Bing because of errors, duplicates or slow loading;
• robots.txt blocks AI search crawlers such as OAI-SearchBot, Claude-SearchBot or PerplexityBot;
• hosting, a CDN or a firewall blocks bots even though robots.txt allows them;
• the main text is loaded by JavaScript, and the crawler sees an empty page;
• important information is buried in PDFs, images or tabs that only open on click.

One nuance worth knowing: blocking training crawlers is a separate decision. GPTBot collects data for OpenAI model training, while OAI-SearchBot powers ChatGPT search, and ChatGPT-User fetches pages when a user asks. The Google-Extended token controls whether your content is used for Gemini training and grounding, and it does not affect Google Search. Mixing these up is a common way to disappear from AI search by accident.

All of this can be checked in a day and fixed in a few days. But as long as the problem exists, no amount of content will help.

Stage two: why AI cites clear pages with facts

Say the page has been found and read. Now the AI decides whether to take a passage from it. Four simple rules apply here.

The passage answers the question directly

Compare two hypothetical passages. The first: "We offer a personalized approach and flexible solutions for any business." The second: "Bookkeeping for an LLC with up to 10 employees starts at $400 per month and includes monthly reports and payroll" (the numbers are made up for illustration). The AI has nothing to take from the first. The second is a ready-made answer to the user's question and is easy to retell.

The passage makes sense out of context

AI does not read the page as a whole but in separate chunks. Phrases like "prices are listed above," "as we mentioned" or "our service" without a name lose their meaning once the passage is cut out. A good block names the company, the service and the terms on its own.

The page has a clear structure

Headings phrased as questions, a short answer at the start of each section, lists, comparison tables and an FAQ block all help the AI find the right passage. Schema markup adds hints about where the organization name, price and address are. More on this in schema markup, FAQ and your company as an entity.

The facts can be verified

A named author with real experience, an update date, company details, an address, licenses and certifications. The more verifiable details a page has, the fewer reasons the AI has to doubt it.

Consistency: why AI cross-checks you against other sources

AI does not take a company's word for it. It treats a company as an entity, a set of facts: name, address, phone number, services, prices, leadership, reviews. It collects those facts from different sources. When the sources agree, the picture is clear. When they contradict each other, problems start.

A hypothetical example. On the website the company is "Ledger Plus Bookkeeping," in its Google Business Profile it is "LedgerPlus," a directory lists an old address and a marketplace still shows last year's prices. The AI sees four versions of one company. There are three possible outcomes, and all of them are bad: it picks the marketplace data and quotes the wrong price, it hedges with "according to some sources," or it avoids mentioning a company it cannot be confident about.

What to align:

• the website and all its sections, including the footer and contact page;
• map listings and business directories;
• industry catalogs, marketplaces and review platforms;
• social media and professional network profiles;
• media coverage and articles by your company's experts.

The other side of consistency is independent confirmation. When industry roundups, reviews and articles write about a company, the AI gets a signal: this company exists, works in this field and is known. We explain how to build this systematically in brand mentions and digital PR for AI search.

Recommendations: why AI names some companies and not others

Now let's put it all together. When a user asks for a recommendation, the AI is essentially looking for companies that at the same time:

• are found for the queries it rewrote the question into;
• match the user's conditions (location, budget, industry, timing), and those conditions are stated explicitly on their pages;
• are confirmed by several independent sources without contradictions;
• are described clearly, so their strengths can be summed up in a sentence or two.

The competitor AI recommends is not necessarily better than you. More often, it is simply described more clearly and is more visible in the sources.

What about ChatGPT ranking factors specifically?

People often search for ChatGPT ranking factors as if ChatGPT had its own secret formula. In practice, when ChatGPT searches the web, it goes through the same stages: its crawler and search partners have to find your page, the page has to be readable, and your passage has to answer the question better than others. The platform-specific part is mostly technical: which crawler needs access and which index the assistant relies on. The rest of the factors are shared across Google, ChatGPT, Perplexity, Gemini, Claude and Copilot.

Why answers change

The same question asked twice can produce different lists of companies. The reasons are clear: the AI writes the answer from scratch each time, the search queries it generates can differ, and the result depends on wording, language, the user's location and how fresh the index is. That is why an honest agency never promises a specific position in an AI answer. What you measure is LLM visibility: the share of tracked prompts where the company appears, checked regularly against the same set of questions.

How to become a source AI cites: what this means for your company

A clear sequence of actions follows from the mechanics. Each step addresses a specific stage of the selection.

• AI visibility audit. Find out what AI assistants say today and uncover technical obstacles at the "found and read" stage.
• Prompt research. Collect the questions customers ask AI, the way keyword research works for SEO. This is the basis for both content and measurement.
• Technical fixes. Open the website to search engine and AI crawlers, fix errors, move text out of scripts, add schema markup and llms.txt.
• Content with facts. Rewrite key pages so every block works as a standalone answer, and create content for your tracked prompts.
• Consistent company data. Align your company information on the website and across all third-party sources.
• Third-party confirmation. Grow brand mentions and publish articles by your experts in external sources.
• AI visibility tracking. Regularly run your tracked prompts and adjust the work based on the results.

On their own, these steps achieve little. A great article on a website blocked for crawlers will not become a source. A perfect website without any third-party mentions loses to a well-known competitor. And conflicting data in directories undermines both. If AI already recommends your competitors, we break down what to do in AI recommends your competitors: what to do.

The result: predictable work instead of a lottery

Once you understand the mechanics, AI search optimization stops looking like magic. AI recommends the businesses it can find, understand and trust. Each of those three conditions turns into concrete tasks: technical, content and reputation. They can be done, verified and repeated when AI platforms change the rules. The effect compounds: the sooner a company becomes a clear, confirmed source, the stronger its position in AI answers.

How AI Search Optimization turns your website into a source for AI

We have been doing SEO since 2002, with 100+ clients around the world and Google and Yandex certifications. We optimize for Google AI Overviews and AI Mode, ChatGPT, Gemini, Claude, Perplexity and Microsoft Copilot, plus other assistants where your market calls for them. Your website and your AI visibility are handled end to end by one team, for businesses from New York City and anywhere else.

The work includes an AI visibility audit, prompt research, technical fixes on the website (AI crawler access, llms.txt, schema markup, speed, errors), new content built for AI answers, brand mentions on third-party sites, articles in external sources that build the authority of your company and its experts, and ongoing AI visibility tracking.

AI evaluates your website as a whole, so we take care of the entire site from day one, from technical fixes to GEO. We guarantee delivery of GEO tasks, error fixes and fast adaptation of your website to changing AI search requirements.

The fee is $290 per month, including the first 10 hours of work. Every additional hour is $45. GEO work is billed time and materials: we invoice only the work completed and delivered that month. No discounts or free months.

Book a call: we will find the stage where AI assistants are losing your company today and what to fix first.