Guide
How Do AI Engines Choose Their Sources?
Most engines document how their crawlers work; almost none publish how they rank sources. This page keeps what is known apart from what is not.
AI Visibility Turkey is the Turkish-market home of Zen Media, Optimum7 and ecommerce.com.
Three steps: crawl, retrieve, generate and cite
Whichever engine you study, a source gets into an answer by passing three gates.
Crawling and indexing. The engine’s bot must reach your page and read it. For Google the page must be indexed and eligible to show with a snippet; for OpenAI, OAI-SearchBot access; for Perplexity, PerplexityBot access.
Retrieval. When a question arrives, the engine picks relevant pages from its index or from live searches. Per Google, AI Mode does this with query fan-out: it splits the question into subtopics and runs many related searches together.
Generation and citation. The model writes the answer and shows some sources. Which ones are shown is largely the engine’s decision; xAI, for example, says inline citations are added by the model itself.
Source: Google Search Central: AI features ↗; OpenAI bots ↗; Perplexity bots ↗; xAI: citations ↗
The access gate: if bots cannot fetch you, you cannot be cited
This is the most commonly missed problem, and it is written in the engines' own documentation.
Indexed and snippet-eligible
Per Google, AI Overviews and AI Mode have no additional technical requirements; the page must be indexed and eligible to show with a snippet. Appearing is never guaranteed.
OAI-SearchBot access
OpenAI documents that OAI-SearchBot surfaces sites in ChatGPT search. Sites that opt out do not appear in ChatGPT search answers (they can still show as navigational links). Changes take about 24 hours to take effect.
Claude-SearchBot and Claude-User
Anthropic documents that blocking these two bots may reduce visibility in Claude search answers. ClaudeBot is for training data; that is a separate decision.
PerplexityBot
Perplexity documents that PerplexityBot surfaces and links sites in Perplexity search and says it is not used to crawl for foundation-model training. Perplexity-User fetches on a user’s request and generally ignores robots.txt.
Source: Google ↗; OpenAI ↗; Anthropic ↗; Perplexity ↗
AI crawlers and robots.txt guide → · Engine-by-engine details →
What the evidence says gets cited
Because engines do not publish ranking criteria, we look at outside studies. These are vendor studies and they have limits.
Vendor research Ahrefs analyzed about 78.6 million AI responses (AI Overviews, ChatGPT and Perplexity) for June 2025. The table shows the share of each engine’s most-cited sources held by these platforms in that engine’s top 10.
Scroll the table sideways.
| Platform | ChatGPT | Perplexity | AI Overviews |
|---|---|---|---|
| Wikipedia | 16.3% | 12.5% | 8.4% |
| YouTube | Not in top 10 | 16.1% | 9.5% |
| Not in top 10 | Not in top 10 | 7.4% (AI Overviews top 10 only) | |
| Quora | Not in top 10 | Not in top 10 | 3.6% (AI Overviews top 10 only) |
Source: Ahrefs: most cited domains in AI assistants
How to read these numbers
This is one vendor’s own study, based on a single month and top-10 lists only. Other studies report very different shares because they count differently, so we do not present any single percentage as a universal fact. The conclusion we can safely draw: third-party platforms (Wikipedia, Reddit, YouTube, news sites) are heavily cited, and overlap with Google’s organic top 10 is partial and varies by engine.
The practical meaning: fixing your own site is often not enough; you also need to be mentioned on the third-party sources engines already turn to. Which sites do AI answers cite? →
Content format: the engines' own advice
No ranking formula is published, but Google and Microsoft give written advice on what content is usable.
Google’s advice
- Unique, non-commodity content
- Good page experience and indexable pages
- Structured data that matches visible text
- Current Merchant Center and Business Profile data
- Measure the value of visits, not only clicks
Microsoft Advertising’s advice (8 October 2025)
- Align title, H1 and meta description
- Use descriptive H2 and H3 headings
- Write a question followed by a short, direct answer
- Use lists and tables, and add JSON-LD
- Do not hide key answers in tabs or accordions
- Do not keep information only in PDFs or images
- Write self-contained, snippable sentences
Source: Microsoft Advertising ↗
On the academic side, the GEO paper (KDD 2024) reports that citing sources, adding quotations and adding statistics raised visibility, while keyword stuffing did not perform well. It was tested on a research engine and the GEO-bench dataset, not on live ChatGPT or Google; we do not promise the same lift.
Source: arXiv 2311.09735 ↗
Unknowns: where the knowledge ends
We treat this as the most important part of the page. We do not present what we do not know as if we did.
Not documented
Not documented
- ChatGPT and Perplexity ranking criteria. Only crawler documentation is published; how sources are selected or ranked is not explained. Perplexity does not describe it in its documentation.
- DeepSeek. We found no primary documentation on how it picks sources. We go no further than saying general web-grounded principles apply.
- Whether Bing’s index feeds Copilot and ChatGPT. Claims that it does come only from third-party blogs; they are not confirmed in primary documentation, so we do not state them as fact.
- Turkish-language citation studies. We found no study of which sources are cited for Turkish queries. There is also no evidence that engines prefer one language when citing.
Source: Perplexity documentation ↗
What this means for our work
There is no ranking formula, but three levers are clear. We shape our work around them.
Be fetchable
Bots must be able to fetch your site: robots.txt policy, CDN and WAF rules, server-rendered HTML, indexing. Which bots you allow is a business decision.
Be clear and citable
Question-led headings, short direct answers, tables, original data and sourced facts. Structured data must match visible text.
Be mentioned elsewhere
Being named on the publications, directories and communities engines already cite. This is digital PR work; its effect per brand is not proven, it is a strategy.
As a group, this is how we divide the work: on digital PR, the published work of Zen Media; on technical and content execution, Optimum7's ecommerce development and digital marketing experience. Our services: technical crawlability, content and entity optimization, digital PR and citation building.
Which bots can see your site, and which engines name you?
Send us your website and topics and we will start with measurement. We reply by email.