Getting cited by ChatGPT, Perplexity, or Google’s AI Overviews depends less on keywords or backlinks than on how well a machine can parse the page.

When LLMs retrieve and score information they do so based on individual passages rather than whole pages, so getting mentioned relies on markup structure a crawler can read, not necessarily on how the page looks to a human visitor.

Here’s how an AI’s retrieval pipeline works (as far as we can tell at this point), what a crawlability audit for these systems involves, and why so many of the fixes turn out to be the same ones that make a page accessible for users with a screen reader.

Being cited by an AI is a parsing problem

An Ahrefs analysis found that pages later cited in AI Overviews already carried more than 100 citations back in February 2025, months before any schema markup touched them. 

The data shows those pages were already well structured enough to be machine-readable. Underneath every AI-generated answer sits the same pipeline: 

  1. a query gets embedded, then
  2. matched against chunked passages of indexed content, and 
  3. the highest-scoring chunks get synthesized into a response with citations attached.

Ranking first in Google doesn’t guarantee a citation in an AI-generated answer, because the system is scoring one passage’s relevance inside a chunk and not the authority of the whole page. That chunk fails an AI crawler the same way it fails someone using a screen reader. Content rendered only through client-side JavaScript, meaning conveyed by color or layout instead of markup, or an image with no text equivalent all disappear to a non-human reader parsing the underlying DOM rather than the rendered page.

How each AI sources citations

The main AI engines source citations through different methods, so fixing crawlability for one doesn’t fix it for the rest.

ChatGPT’s browsing and retrieval

Current GPT-5-class models decide, on the fly, whether a question needs a live fetch or can be answered from training data. OpenAI’s documentation describes the model applying reasoning automatically when a response would benefit from it, which includes triggering a search. In its default, non-browsing mode, it cites nothing, because it isn’t retrieving anything.

Perplexity works differently. It treats almost every query as a live search rather than leaning on memorized training data, which is why its answers come with numbered citations attached by default.

Google AI overviews’ dependence on the existing index

AI Overviews pull from Google’s existing web index and knowledge graph, the same crawling and quality pipeline behind organic rankings. Google Search Central confirms Overviews draw from that same index, meaning a strong organic ranking is still the foundation (but not a guarantee) of citation potential there.

Answer-first paragraphs

Put the direct claim in the first sentence, then support it. A 2026 study tracked which source gets cited first across a controlled dataset built to isolate individual content factors, and answer-first paragraphs consistently won out over scene-setting SEO intros.

Embedding-based retrieval weighs the opening of a chunk more heavily than what follows, so the sentence carrying the answer needs to sit at the front. A legacy SEO intro (the “in today’s fast-paced world” throat-clearing before the useful sentence) pushes that sentence past the point most retrieval systems bother pulling into a single chunk.

Two rewrites do most of the work. 

  1. Phrase section headings as the query someone would type into ChatGPT or Perplexity, since that maps cleanly onto how query embeddings get matched against document embeddings. 
  2. Write the key fact as a standalone sentence, one that reads correctly with zero surrounding context, because that’s the sentence a system will use in its answers.

Most retrieval-augmented pipelines chunk text in windows of roughly 200 to 500 tokens (around 150-400 words). Keep the claim and its immediate support inside that span.

AI corrections and “hallucinations”

A wrong fact on a webpage is easy enough to fix once it’s noticed, but hallucinations don’t work that way. Those happen when a model fills the gaps between retrieved passages with parametric memory (its existing background training), so the answer sounds fully sourced while secretly blending or inventing the attribution.

The only real defense is watching for it before it spreads. That means running a fixed set of prompts about the brand across ChatGPT, Gemini, and your choice of other LLMs on a recurring schedule, then logging what each one says. Treat it the way a communications team treats social listening. A wrong answer caught in week one is far cheaper than one that’s been circulating for a quarter.

When something is wrong, the fixes are all indirect. Update whatever source page is most likely being retrieved, since that’s the raw material the next answer draws from. Use each platform’s feedback mechanism, like ChatGPT’s response feedback or Google’s “About this result,” even though neither guarantees a fast fix. 

Be prepared for corrections to propagate slowly, dragged down by caching and training-data lag on the model side. 

Screen readers and AI crawlers

A screen reader user gets nothing when an image carries meaning and the alt attribute is blank. An AI system summarizing that page hits the same wall: no alt text, no adjacent data table, and so nothing to lift into an answer. 

Here are some things to check and fixes to implement to make your pages maximally readable for both LLMs and screen readers:

  1. Crawler access: confirm GPTBot, ClaudeBot, PerplexityBot, and Google-Extended reach the page and get full rendered HTML.
  2. Semantic structure: check the heading hierarchy and whether DOM content order matches what a sighted visitor sees on screen.
  3. Machine-readable metadata: validate that schema.org markup is present and renders without errors.
  4. Text equivalents: verify alt text and data tables exist for anything conveying information visually.
  5. Answer-shaped prose: confirm a direct, extractable claim sits near the top of the section it belongs to.

The W3 catalog of accessibility evaluation tools already tests most of what an AI-citation audit needs, because the underlying question is the same one: Can a non-human reader get through this page? 

Getting cited starts there, not with schema or prompt strategy. Confirm that any non-human reader, assistive software or retrieval model, can actually parse the page, and everything else in this framework has somewhere to land.

References

  1. CJS2030: The initiative on AI | Columbia Journalism School. (2025a). In Columbia.edu. https://journalism.columbia.edu/CJS2030/AI 
  2. From Googlebot to GPTBot: Who’s crawling your site in 2025. (2025b). In The Cloudflare Blog. https://blog.cloudflare.com/from-googlebot-to-gptbot-whos-crawling-your-site-in-2025/ 
  3. Initiative (WAI), W. C. W. A. (2024). Web accessibility evaluation tools list. In Web Accessibility Initiative (WAI). https://www.w3.org/WAI/test-evaluate/tools/list/ 
  4. Linehan, L. (2026). We tracked 1,885 pages adding schema. AI citations barely moved. In SEO Blog by Ahrefs. https://ahrefs.com/blog/schema-ai-citations/ 
  5. OpenAI. (2025). Introducing GPT-5. In Openai.com. https://openai.com/index/introducing-gpt-5/ 
  6. Popsight Team. (2026, April 28). AI visibility tools compared: WHO publishes their methodology? Popsight. https://popsight.ai/blog/ai-visibility-tools-compared/ 
  7. Pru, M. (2026, August 8). Google search central’s AI overviews guidance: How to get cited. Stackmatix. Stackmatix.Com. https://www.stackmatix.com/blog/google-search-central-ai-overviews-guidance 
  8. Research guides: Digital accessibility skills: Using headings. (2026). Springshare. https://guides.lib.purdue.edu/c.php?g=1471250&p=10950446 
  9. Vishwakarma, R., Gurugram, S., Kumar, I., Dubai, S., Uae, R., & Jamidar. (2026). What gets cited: Competitive GEO in AI answer engines. 5. https://arxiv.org/pdf/2605.25517