How AI assistants decide which websites to cite

An assistant can only quote what it can reach. When a question calls for current or specific information, ChatGPT, Claude, Gemini, Copilot or Perplexity will usually run a web search, open a few promising pages and build the answer from the passages that tackle the question head-on. The goal is to be one of those pages: easy to fetch, easy to parse, and plainly the best answer in the pile.

None of these providers hands out the full recipe, and their systems change often. The broad mechanics are well documented, though, and they reward the same qualities that have always made a website useful. Here's what we know, what helps, and what we'd happily skip.

Two kinds of knowledge

An assistant draws on two reservoirs. The first is what the model absorbed during training: a vast snapshot of the web and other texts, frozen at a cut-off date. It shapes what the assistant "knows" about your brand, rarely comes with a link, and you have no way to update it.

The second is retrieval at answer time. When the question calls for it, the assistant queries a search index (its provider's own, a partner's, or both), fetches the pages that look relevant and reads them. That's where citations come from, and it's the part you can influence today.

What happens during retrieval

1. The question becomes searches. The assistant rewrites the request into one or more search queries, often sharper than what the person actually typed.
2. Candidates are retrieved. Pages come from a search index, so classic discoverability still counts: crawlable pages, a sitemap, sensible titles and links.
3. Pages are fetched and read. Some assistants fetch the live page, others work from a cached copy. Either way, the system only has the text it can extract.
4. Passages are selected. The model hunts for the parts of each page that answer the question, compares sources and writes a synthesis.
5. Sources are cited. The pages whose passages shaped the answer are the ones most likely to get the link.

What tends to make a page citable

• A direct answer near the top. An opening paragraph that answers the question in plain words is a gift to anyone quoting.
• Wording that matches real questions. Headings and FAQ entries phrased the way people actually ask help the page match the query.
• Self-contained passages. A paragraph that still makes sense once lifted out of the page is far more usable than one leaning on what came before.
• Specifics. Named criteria, dates, prices and steps, as long as they're true and you can stand behind them.
• Freshness and ownership. A visible date, a named author or organization, and facts that line up with what the rest of the web says about you.

Let the right crawlers in

AI providers document the user agents they use, and most of them keep crawling for model training separate from fetching pages for search and live answers. OpenAI, for example, documents distinct crawlers for training (GPTBot) and for search (OAI-SearchBot). Check your robots.txt, your firewall and your CDN's bot settings: plenty of sites block these agents by accident. Then make the call on purpose. Blocking training crawlers while allowing search and user-initiated fetches is a reasonable middle ground for many organizations.

Put the content in the HTML

Many AI crawlers read the HTML your server sends and never run JavaScript. If your text only shows up after a script runs, they may find an empty page. Server-side rendering or pre-rendering fixes that. This one is close to our hearts, since we build a lot of creative, motion-rich websites: the experience can be as ambitious as you like, as long as the words live in the markup. Our guide to the front-end stack for creative websites shows how to have both.

Lead with the answer

Give each important question its own clear answer. Open with the answer, then explain. Write headings that read like the questions people ask. Add a short FAQ where it genuinely helps. One topic per page will outperform a page that tries to cover everything.

Make the structure explicit

Semantic HTML (real headings, lists and tables) helps every parser. On top of that, three additions make your site easier for machines to understand:

• structured data that states who you are and what each page is;
• an llms.txt file that summarizes your site for AI agents;
• Markdown versions of your pages that agents can read without the layout.

Earn authority signals

Assistants compare sources. When your website, your profiles and third-party mentions all tell the same story, your version is easier to trust. Reviews, press coverage, industry recognition and links from reputable sites all help, just as they do in search. And they're very hard to fake for long.

A new name for familiar work

"Generative engine optimization" (GEO) is the label that stuck. A 2023 research paper popularized it, and today it sits alongside "AI search optimization" and "answer engine optimization" as shorthand for any effort to show up in AI-generated answers. In practice, it overlaps heavily with good SEO and good content strategy: crawlable pages, clear answers, credible sources. New acronym, familiar homework.

What we'd skip

• Promises of guaranteed placement. Nobody controls which sources an assistant cites. Be wary of anyone who claims otherwise.
• Content written only for machines. Keyword-stuffed pages and hidden text read badly for people, and systems keep getting better at discounting them.
• Showing bots something different from people. Machine-readable versions should mirror what visitors see. Slipping the robots a sales pitch misses the whole point.

How to measure it

There's no complete dashboard yet. Ask the main assistants the questions your clients ask, regularly, and note who gets cited. Watch referral traffic from assistant domains in your analytics. Check your server logs for AI user agents to confirm they can reach your pages. Then follow the trend over several months. Next Friday's numbers won't tell you much.

1. List ten questions your prospects ask before they contact you, in their own words.
2. Put them to two or three assistants and note which sources get cited. Are you there? Who is, and why?
3. Check your answers. For each question, does a page on your site answer it in its first paragraph?
4. Check what machines see. Load a key page with JavaScript disabled, or fetch it with a command-line tool, and make sure the text is there.
5. Fix the basics first, then look at structured data, llms.txt and Markdown versions.

If a redesign is on the horizon, write these requirements into the brief from day one: they cost far less to build in than to bolt on later. Curious what it looks like on a real site? Here's how we made locomotive.ca ready for AI assistants, with our own website as the test bench.