Why AI cites some pages and ignores others: how LLMs select sources

You ask the same question in ChatGPT and Perplexity. Each cites different pages. One company appears in every answer; another never does, despite excellent SEO and regular publishing. Why? It is neither chance nor black magic. Language models apply specific, comprehensible criteria when selecting sources. Once you know them, you can plan work that genuinely changes your visibility in AI.

According to a report from The Digital Bloom, just 11% of websites are cited by both ChatGPT and Perplexity. The rest appear in one model, in both occasionally, or not at all. That is not accidental. It follows from different model architectures and different criteria for assessing content credibility.

Two citation modes: with and without the web

Before the mechanism, a key distinction: there are two fundamentally different scenarios in which AI cites content. Call them Case L (knowledge only) and Case L+O (knowledge plus online retrieval).

• Case L - the model answers purely from training knowledge, such as ChatGPT without browsing or Claude without tools. It cites brands and facts that appeared repeatedly in the training data. It has no access to current pages.
• Case L+O - the model fetches current web results before generating an answer: Perplexity by default, ChatGPT Search, Google AI Overviews. It can cite a page published yesterday, provided it meets the selection criteria.
• The practical difference: in Case L, the history of mentions matters - how often you were written about before the training cut-off. In Case L+O, current page quality and crawler accessibility matter.

Seven decision stages: how an LLM chooses what to cite

In models with web access (Case L+O), source selection passes through several assessment layers. Each can disqualify a page - or promote it. Here is a simplified but faithful description of the process:

1. 1Technical accessibility - can the crawler reach the page? robots.txt, WAF blocking and load time are checked. A page blocked to AI crawlers (GPTBot, PerplexityBot, OAI-SearchBot) is not considered at all, however good the content.
2. 2Topical relevance - does the page address what the user asked? The model assesses semantic fit between content and query. Pages that are too general or thematically diffuse rank lower on relevance.
3. 3Freshness - the publication and last-modified dates matter, particularly for fast-moving topics. A 2022 article about AI tools has a far weaker chance than a 2025 one.
4. 4Source authority - assessed through signals similar to conventional PageRank, but with more weight on mentions and citations by other resources. Wikipedia and major publications carry a large premium.
5. 5Structural content quality - is the text organised in a way the model can process? Headings, lists, short paragraphs, figures and definitions in the first paragraph are all structural quality signals.
6. 6Distinctiveness of data and perspective - AI prefers content with original data, case studies and expert opinion it cannot reconstruct from general knowledge. If your article repeats what others wrote, it adds nothing to the answer.
7. 7Fit with the query context - the final filter: does citing this passage actually answer the user's question better? The model weighs whether adding the quote improves the answer or clutters it.

A weighting matrix: what really decides citation

Not all factors carry equal weight. From available research and empirical testing, an approximate matrix of each signal's influence on the probability of citation:

Structured data and citations: what the numbers say

Schema.org can help search systems identify the author, the organisation and the material type, but there is no credible basis for attributing a universal uplift in citability to it. Treat implementation as a layer that organises data, not as a standalone success factor.

Structured data describes entities and material types, but it does not reveal the full source-selection mechanism of an AI system. The most important information should remain available in the readable text of the page, and the schema must match the visible content.

Perplexity cites differently from ChatGPT

Perplexity and ChatGPT Search are both L+O models, but they take different approaches to source selection. Understanding the differences helps target optimisation:

• Perplexity always shows sources as links, which is a strong incentive to cite pages with readable URLs, short titles and a clear meta description. A page whose meta description does not explain precisely what it contains loses on relevance ranking.
• ChatGPT Search draws heavily on conventional SEO signals. Domain authority and backlinks matter more here than in Perplexity.
• Google AI Overviews favours Google's own index. Standard SEO carries over most strongly here - a good organic position substantially improves the chance of appearing in an AI Overview.
• Claude with web access applies very cautious selection, favouring sources with high editorial credibility and avoiding pages with low author transparency.

Common mistakes that exclude pages from citation

Analysing B2B sites that publish good content yet do not appear in AI answers, the same problems recur:

• Blocking AI crawlers through Cloudflare or an aggressive WAF - GPTBot and PerplexityBot get treated as scrapers and blocked. Check robots.txt: there should be an explicit Allow for those crawlers.
• Content loaded through JavaScript with no server rendering - AI systems crawl HTML and do not render JS. Content in a React single-page app without server-side rendering is invisible to them.
• No named author with a bio and profile links - models assess E-E-A-T. An article with no signed author and no credentials has lower credibility.
• No publication or update dates - a page without a date is hard to assess for freshness. That is particularly damaging on technology and market topics.
• Very large HTML files (over 500KB) or slow response times (over 3 seconds) - AI crawlers have resource limits. Slow loading means being skipped.
• A canonical pointing to a different page - the model sees content at one URL but the canonical sends it elsewhere. It may skip the content as a duplicate.

How to work out why your page is not cited

A systematic diagnosis has three layers:

1. 1Technical layer: check robots.txt for GPTBot, PerplexityBot and ClaudeBot. Use a crawler such as Screaming Frog to audit JS rendering. Measure Core Web Vitals - time to first byte should be under 800ms.
2. 2Content layer: ask 15 to 20 category questions in ChatGPT, Perplexity and Gemini. Note which pages are cited instead of yours. Compare their structure with yours - what do they have that you do not?
3. 3Authority layer: check in Ahrefs or Semrush how many external pages mention competitors versus you. Set up alerts on cited competitors' names - where are they written about and you are not?

An action plan: what to do first

If you want to be cited systematically by AI, prioritise by effort-to-effect ratio:

• Week 1 - technical foundations: unblock AI crawlers in robots.txt, enable server rendering for content, and optimise load time to under two seconds
• Weeks 2-3 - structured data: implement schema.org Article on posts, FAQPage on pages with an FAQ, and Organization globally. Add an FAQ section to the end of every article.
• Month 2 - content quality: rebuild your top five articles by adding proprietary data and statistics, clear H2 and H3 headings, definitions in the first paragraph, and original perspectives
• Month 3 onwards - external authority: a campaign for mentions in industry media, activity on LinkedIn and Reddit, and guest articles on high-authority publications