How answer engines choose which sources to quote

When someone asks ChatGPT or Perplexity a question, the model does not rank ten blue links. It retrieves a handful of documents, pulls specific passages out of them, and assembles an answer — usually citing a few of the sources it used. Whether you are one of them comes down to a different set of properties than classic SEO optimises for.

This guide covers what those properties are, what the published evidence supports, and — just as important — what nobody can honestly promise.

What nobody can promise you

Start here, because the category is full of claims that cannot be tested. No tool, plugin or agency can tell you whether an answer engine will cite your page. Citation depends on the exact prompt, the other documents retrieved alongside yours, and ranking behaviour inside systems nobody outside those companies can observe. It also varies between runs of the same prompt.

What can be established is the opposite and more useful thing: the reasons an engine would skip you. Those are observable, fixable, and the rest of this guide is about them.

1. Reachability: can the engine fetch you at all?

Before content quality matters at all, a crawler has to be allowed in. This is where the most common and most self-defeating mistake lives, because the major vendors run two different kinds of crawler under similar names.

Training crawlers collect content used to train models. GPTBot (OpenAI), ClaudeBot (Anthropic) and Google-Extended (Google) are in this group. Blocking them keeps your content out of training data. Whether you want that is an editorial and commercial judgement, and there is a reasonable case either way.

Retrieval crawlers fetch pages to answer live questions. OAI-SearchBot, Claude-SearchBot and PerplexityBot are in this group. Blocking these removes you from the answers themselves.

The failure mode is a site owner who reads about AI scraping, adds a robots.txt rule intending to opt out of training, and catches the retrieval crawler in the same block. The site then does everything else right and remains invisible, because it made itself unfetchable. If you check one thing after reading this, check that.

2. Extractability: is there a passage worth lifting?

An answer engine does not quote a page. It quotes a passage. That distinction has a practical consequence most writing advice misses: a page can contain exactly the right answer and never be used, because the answer only makes sense after three paragraphs of context.

The single most effective structural change most pages can make is an opening that directly answers the question implied by the title, short enough to stand on its own. Not a summary of what the article will cover — the actual answer, in a form someone could paste into a reply without editing it.

Two related habits help for the same reason:

• Question-shaped headings. People prompt in questions. A heading reading "How do I verify hreflang is working?" matches the shape of a prompt far more closely than "Verification".
• Self-contained sections. Avoid "as described above" and pronouns that reach back across headings. Each section should survive being read alone, because that is how it will be read.

3. Attribution: is the claim traceable?

This is where the published research is most useful. Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024), tested a range of content interventions against generative engines and measured their effect on visibility. Among the interventions tested, citing sources, adding quotations and adding statistics produced the largest gains.

That result is intuitive once you consider the position the engine is in. It is assembling an answer it will be held to, from documents of unknown reliability. An unattributed assertion on an unfamiliar domain is the weakest possible evidence. The same assertion, traced to a primary source with a date, is something it can repeat without exposure.

Practically: cite the source you already used. Most factual claims in most articles came from somewhere — a spec, a vendor's documentation, a study. Linking it costs a sentence and converts an assertion into evidence.

4. Specificity: is there anything concrete to quote?

Vague copy is unquotable copy. "Significantly faster" gives an engine nothing to lift; "cuts median response time from 800ms to 210ms" is a passage that can be dropped into an answer verbatim.

This is the same finding as above from a different angle — the statistics result in the KDD paper — and it is the easiest to act on, because in most cases the number already exists and simply was not written down.

5. Machine-readable structure

Two mechanisms let a site describe itself without an engine having to infer everything from prose.

Structured data (schema.org JSON-LD) states what a page is, who wrote it, and when it last changed. A dateModified matters more than it used to: an engine choosing between two sources on a fast-moving topic has an obvious reason to prefer the one that can prove it is current.

llms.txt is a proposed convention — a plain-text file at the site root listing what the site covers, with links and one-line descriptions. It is not a standard and no engine is obliged to read it. It is also nearly free to publish, and it gives a model a map of the site without crawling every page. Treat it as cheap insurance rather than a lever.

What to do, in order

1. Check your robots.txt for retrieval crawlers. Everything else is wasted while you are unfetchable.
2. Rewrite your best pages' opening paragraphs into self-contained answers. Highest return per hour of any item here.
3. Add the citations you already have — the sources you used but did not link.
4. Add the numbers you already know.
5. Emit dateModified and keep it honest.
6. Publish an llms.txt. Cheap, quick, no downside.

You can check most of this on any URL with the AI Visibility Checker, which reports each of these signals for a page and lists which AI crawlers the site blocks.

A note on how to read any of this

The evidence base for generative-engine optimisation is thin compared to two decades of search-engine research — one widely cited paper and a great deal of vendor assertion. The recommendations above are the ones that either have published support or follow directly from how retrieval works.

Be sceptical of anyone, including us, who tells you more than that. If a claim about AI visibility cannot be traced to a measurement or a mechanism, it is a guess wearing a confident tone.