AI Search Visibility: Getting Cited by ChatGPT and AI Overviews

Something changed in search and it has not stopped changing.

For twenty years, the goal of a good page was to earn a click. Now a growing share of queries are answered on the results page, by an assistant summarising sources it retrieved and cited. The click does not happen. The work still does.

Getting cited in those answers is now a distinct discipline, and it is won or lost on structure and clarity rather than on keyword density.

How the retrieval actually works

The mechanics differ between systems, but the shape is consistent.

A query goes to a retrieval layer, which selects a set of documents — usually by semantic relevance, often with a recency preference, sometimes restricted by crawl permissions. A synthesis layer then reads those documents and composes an answer, citing the ones it used.

Three things follow from that, and they are the whole basis of the discipline.

If you are not retrieved, you cannot be cited. No amount of quality helps if the crawler cannot reach the page, if the robots directives exclude it, or if the page is not about the question.

The pages that get retrieved are the ones that clearly answer a question. Retrieval favours documents with an explicit, self-contained answer to the thing being asked. A page that is about the topic is less likely to be cited than a page that answers the question.

The pages that get cited are the ones that make quoting easy. A clear claim, attributable to the page, on a page that a language model can parse without guessing at the structure.

What actually to do

Write answer-first

This is the single biggest difference between content written for a human skimming and content written to be retrieved.

Put the answer in the first paragraph. Then support it. The pattern is inverted from traditional long-form writing, where the answer came last.

A paragraph like this gets cited:

Core Web Vitals are three metrics — LCP, INP and CLS — that measure a page's loading speed, responsiveness and visual stability. Google considers them when ranking results, and poor scores correlate with lower engagement and fewer conversions.

The same paragraph written as "In this comprehensive guide we will explore everything you need to know about Core Web Vitals, from the history of page performance measurement through to advanced optimisation techniques..." gets skipped. The model has to do work to extract the answer from it, and when a cleaner source exists, it will not.

Ask the question your customer asks

This matters more in AI search than anywhere else, because the model is matching a natural-language question.

"How long does a logo design take?" beats "logo design services" as a heading. "How much does a website redesign cost?" beats "website redesign pricing". The phrasing should match how a real person phrases the problem, including the slightly awkward way they would say it.

One clear question per section, answered in the first two sentences of that section.

Make claims attributable and specific

Models cite things that look like facts with a source. Vague claims get skipped in favour of specific ones.

Weak: "Redesigns can significantly improve conversion rates."

Strong: "A redesign that rebuilds the two highest-traffic journeys typically lifts conversion on those journeys by 15–30% in our engagements."

Both are true in spirit. Only one is quotable. Numbers, named methods, specific timeframes and clear attributions give a model something to point at.

Use structured data that reflects the actual page

Schema markup does not directly cause citations, but it removes ambiguity about what a page is.

• Organization on the homepage, with real addresses, contacts and a link to the real profile

• LocalBusiness for each location, with genuine coordinates and hours

• Service on service pages, describing what the service actually is

• FAQPage where the page genuinely contains questions and answers

• Article and BlogPosting for posts, with accurate dates and authors

• BreadcrumbList on every nested page

The rule: the markup must describe what is visibly on the page. Markup that claims something the page does not say is a liability in AI search, not an asset.

Keep the site technically reachable

This is not new, but it became critical.

• Server-side rendering, or a fully static build. A page that requires JavaScript to render is often not retrieved.

• Clean semantic HTML with real headings, lists and tables. A model parses structure far better than it parses visual layout.

• Sensible internal linking with descriptive anchor text.

• Fast responses. A slow page frequently times out during retrieval.

• No noscript fallbacks that hide the content.

• Indexability: one canonical URL per page, no duplicate parameter URLs, correct robots.txt.

Build the entity

AI systems resolve entities, and they resolve them from many sources at once. A company with a consistent presence across its own site, its profiles, its directory listings and its press mentions is easier to identify as one thing.

Consistency of name, description, contact details and links is doing more work here than it ever did. If your business is called "Acme Plumbing" on the website and "Acme Plumbing Co." on the directory and "Acme PlUMBING" on a social profile, you are making the entity harder to resolve every single time.

The measurement problem

There is no equivalent of Search Console for AI citations, at least not one with the rigour of a ranking report. Most tools that claim to track this are estimating.

What we do instead is simple and reasonably reliable:

1. A fixed question set. Twenty to forty real questions your customers ask, written in the words they use. Not keywords — questions.

1. A fixed set of assistants. Two or three platforms, checked on a schedule.

1. A manual monthly review. Ask each question, record whether you are cited, whether you are mentioned without a link, and what the answer says about you.

1. Accuracy as a first-class metric. Being cited accurately is worth more than being cited often. Errors get reinforced.

This takes an hour a month. It is more honest than a dashboard that models something it cannot observe.

What this changes about content strategy

Not as much as the headlines suggest, which is reassuring.

The content that performs in AI retrieval is the same content that performs in organic search: genuinely useful, clearly organised, specific, well-maintained and structured. The difference is the shape — answer-first, question-headed, quotable.

The content that stops working is the same content that was always going to stop working: thin, keyword-stuffed, padded to hit a word count, and written for a crawler rather than a person.

There is one genuine strategic shift. Writing for a human who will click means optimising the first impression. Writing for a retrieval system means optimising the whole document for extractability. That favours depth over punchiness, and it favours genuinely comprehensive coverage of a topic rather than a single sharp angle repeated across ten posts.

A practical sequence

1. Fix the technical foundations. Rendered HTML, fast responses, clean schema, correct robots directives. Nothing else works without this.

1. Find your real questions. Mine support tickets, sales calls, search console queries, reviews and on-site search. This is the highest-value input to the whole effort.

1. Restructure your strongest pages to answer-first, with one question per section.

1. Add the specifics. Numbers, named methods, clear attributions, real dates and authors.

1. Make the entity consistent everywhere you control.

1. Measure with a fixed question set, monthly, by hand.

Expect the effects to compound rather than arrive. AI retrieval favours pages that have been stable, specific and maintained for a while, because the model is more confident citing a source it has seen repeatedly than one it encountered once.

---

If you want to know where your own content and structure stand against this, our SEO and growth practice covers AI search readiness as part of the technical work — and if the underlying problem is that the content does not exist yet, that is a web design brief.