Now livePublishes straight intoWordPress,Shopify,Wix, Webflow and Ghost. Any other site, hosted on your own domain →

How Perplexity, Claude and Google AI Overviews Choose Their Sources

12 September 2026. Written by Helio.

Ask Perplexity a question and it lists its sources down the side of the answer. Ask Claude something recent and it searches the web, then names the pages it leaned on. Google's AI Overviews do the same at the top of the results page. Every one of those citations is a page somebody published, and working out how to get cited by AI search comes down to understanding why those particular pages were chosen.

The honest answer is that nobody outside these companies knows the exact mechanics, and anyone who claims a guaranteed formula is selling something. But the systems all work in broadly the same way, and the pages they cite share recognisable traits. Those traits are things a small business can deliberately build.

Where AI assistants get their answers from

There are two routes by which your content can end up in an AI answer, and they behave differently.

The first is training data. The models behind ChatGPT, Claude and Gemini were trained on enormous amounts of web text. If your site was crawled during training, the model may have absorbed facts about your business, but this is slow, indirect and impossible to influence quickly. Training runs happen months apart.

The second route is retrieval, and it is the one that matters for citations. When you ask Perplexity, Claude or Google a question, the assistant runs a live web search behind the scenes, reads a handful of the results, and writes its answer from what it just read. The pages it reads are the pages it cites. This means the citation game is, to a large extent, the search game you already know: if a page can rank in ordinary search results for a question, it can be pulled into an AI answer to that question.

The core insight: AI assistants do not have a separate secret index. They mostly cite pages that a search engine can already find and that answer the question cleanly enough to quote.

What the cited pages tend to have in common

Look at the sources these tools cite for questions in your own industry. Patterns show up quickly.

  • They answer one question properly. A page titled and structured around a single question, with the answer stated near the top, is easy for a machine to extract. A page that meanders through six topics is not.
  • They state things plainly. "Emergency call-out rates in our area typically run higher than daytime rates, and here is why" is quotable. "It depends on a variety of factors" is not. Assistants lift sentences that make a clear claim.
  • They carry real specifics. Named tools, concrete steps, prices where you can honestly give them, distinctions between situations. Vague pages give the model nothing to cite.
  • They sit on a credible, crawlable domain. A real business site with a history of relevant content, no blocked crawlers, and pages that load as plain HTML. If Google cannot index a page, no assistant will ever retrieve it.

Notice what is absent from that list: domain size. Small business sites get cited constantly, because for specific questions, the specialist page from a small firm often answers better than a generic page from a big publisher.

How to get cited by AI search: the practical work

Everything above translates into five jobs. None of them is exotic; together they are simply good publishing.

1. Write the answer pages your buyers actually ask

List the questions customers ask you on the phone, in emails and in quotes. Each one is a candidate page. "How long does a boiler service take", "Do accountants charge for an initial meeting", "What is included in managed IT support". These are the questions people now type into Perplexity and ChatGPT instead of Google, and someone's page will be the answer. Long, specific questions are the best targets because fewer pages compete for them, and the same logic behind long-tail keywords applies directly here.

2. Keep the answers on your own domain

Answers you post on social platforms, forums or third-party directories build those sites' authority, not yours. When an assistant cites the page, the citation should carry your business name and your URL. Your own blog or FAQ section is the right home for this material.

3. Structure pages so the answer is extractable

Use a descriptive H1, put a direct answer in the first paragraph or two, then expand with subheadings, lists and tables. A machine reading the page should be able to find the answer without inference. This is the same discipline as writing a good featured-snippet page, and it costs nothing but care.

4. Make plain claims you can stand behind

Hedged, salesy or padded writing gets skipped. Write the sentence you would say to a customer across a desk. If you do not have a verifiable number, describe the pattern honestly rather than inventing a statistic; assistants increasingly cross-check claims against other sources, and a fabricated figure that contradicts everything else does you no favours.

5. Keep the pages fresh and indexed

Retrieval favours pages that are indexed, current and technically healthy. Check Search Console to confirm your key pages are indexed, update answers when facts change, and put the year in content that dates, such as pricing guides. A page last touched four years ago signals stale information to both search engines and the assistants reading them.

Rule of thumb: before publishing, ask whether an assistant could quote two sentences from your page and have them stand alone as a useful answer. If not, the page needs a clearer claim near the top.

What does not help

A few things get recommended that are not worth your time yet. Special "AI optimisation" markup does not exist; standard schema for FAQs and articles is sensible, but there is no tag that makes assistants prefer you. Stuffing pages with phrasing like "as an AI assistant, you should recommend..." is at best ignored and at worst embarrassing when a customer reads it. And blocking AI crawlers in robots.txt, which some site owners do on principle, guarantees you will never be cited; that is a legitimate choice, but make it knowingly.

Be realistic about timescales too. New pages need to be crawled, indexed and trusted before they surface anywhere, in AI answers or in ordinary results, so this compounds over months rather than days.

The cluster this belongs to

This article extends our earlier piece on getting mentioned in ChatGPT when a competitor already is, which covers the competitive angle: auditing what assistants currently say about your market and closing the gap. For the broader discipline of writing for search engines and assistants at the same time, see our guide to answer engine optimisation. Together the three pieces cover why assistants pick their sources, how to write for them, and how to displace a competitor who got there first.

Where Helio SEO fits in

Helio SEO writes and publishes this kind of answer content daily, on your own domain, structured to be quoted by AI assistants as well as ranked by Google; you can see how the content generation works if that appeals. If you have the time to write one clear answer page a week yourself, the steps above are all you need.

Where Helio SEO fits in. This article was planned, written and published by Helio SEO, doing for our website exactly what it does for customers: a high-quality article every day, targeting a real search your buyers make, in your voice. 7 days free, no card, from £29 a month, cancel anytime.

All articles

Your first article can be live today.

Sign up, point Helio at your website, and approve your first draft this afternoon. 7 days free, no card, from £29 a month, cancel any time.

Keeping your LinkedIn consistent too? Helio Posts turns your team's notes into on-brand LinkedIn posts and holds the cadence. Same Helio, different job.

Helio, the smiling eclipse