How AI Overviews and ChatGPT Decide Who to Cite

You can hold the #1 spot on Google and still be invisible.

Not invisible in the rankings. Invisible in the answer. When Google’s AI Overview or ChatGPT writes the response your prospect actually reads, it names a handful of sources — and stitches everyone else out of the story. If you’re not one of those sources, your ranking is a participation trophy. The prospect got their answer, formed their impression, maybe even built a shortlist, and never scrolled far enough to know you exist.

This is the shift that matters. For twenty years, search visibility meant position. Now there’s a layer above position — the answer layer — and it has its own selection logic. A citation in that layer is earned at the moment the engine decides your page is the safest thing it can repeat in public. Rankings influence that decision. They don’t make it.

The frustrating part: while this shift happened, an entire cottage industry sprang up selling “AI ranking factors” with suspicious precision. Exact percentage weightings. Minimum statistics per thousand words. Formulas nobody at Google or OpenAI has ever published.

So this post does two jobs. First, it walks through what the engines actually do — drawn from the largest citation studies published so far: millions of AI Overviews, hundreds of thousands of real ChatGPT conversations, and seven months of cross-engine tracking. Second, it draws a hard line between three kinds of claims: what’s verified, what’s reasonably inferred, and what was invented by someone with a service to sell.

By the end you’ll understand the mechanism well enough to stop chasing hacks — and to see why the boring fundamentals suddenly matter more than they have in years.

First, know which ChatGPT you’re talking about

Most confusion about AI citations starts here, so let’s kill it early. ChatGPT answers questions in two fundamentally different ways, and they have almost nothing in common.

Parametric mode: no sources, just patterns

By default, ChatGPT retrieves nothing. It reads nothing live. It generates answers from statistical patterns absorbed during training — a compressed impression of the public web as it existed before the model’s cutoff date. When it answers in this mode, there is no lookup happening. There is no source. There is only pattern completion.

Two consequences follow, and both matter to you.

First: if ChatGPT names your brand in this mode, that’s a mention — no link, no click, and no lever you can pull this quarter. A mention reflects how thoroughly and consistently your brand existed across the web before training. It’s the compound interest on years of presence. Valuable? Absolutely — it shapes what millions of people hear when they ask for recommendations. Controllable in the short term? No.

Second: when this mode produces something that looks like a citation — an author, a journal, a URL — it’s constructing a plausible-sounding reference from memory. That’s why the internet is full of stories about ChatGPT inventing sources that don’t exist. It isn’t lying, exactly. It’s autocompleting what a citation usually looks like. If there’s no clickable link and no visible search activity, treat every “source” as unverified until you’ve checked it yourself.

Browsing mode: real retrieval, real citations

When ChatGPT decides a question needs fresh facts — current events, specific data, anything time-sensitive — it switches behaviour entirely. It runs live web searches, reads actual pages, extracts relevant passages, and composes an answer with numbered, clickable citations pointing to real URLs.

This is the only mode where a citation — a genuine link that can send a visitor to your page — is possible. Which makes it the only mode you can optimise for directly, on a timeline measured in weeks rather than training cycles.

Mentions build awareness. Citations send visitors. They’re different visibility states, driven by different levers, measured differently — and any AI-visibility conversation that doesn’t distinguish them is measuring the wrong thing. Both, however, sit on the same foundation: how your brand shows up across the whole search layer, not on one page you’ve polished for the occasion.

How ChatGPT actually picks its citations

The best public window into ChatGPT’s citation behaviour is Profound’s analysis of roughly 730,000 real ChatGPT conversations that contained web citations — US-based, English-language users, October through December 2025. Not lab prompts. Actual usage. (Source)

Three patterns from that dataset should reshape how you think about being cited.

It searches on the first question, not the tenth

Around 18% of ChatGPT conversations trigger at least one web search — a rate that held steady across all three months studied. But when the search happens is the real story. The opening question of a conversation produced citations 12.6% of the time. By turn ten, that had fallen to 4.5%. By turn twenty, 3%.

The logic is intuitive once you see it. Opening questions need factual grounding: “what is this,” “how does this work,” “what should I do about this.” Follow-ups tend to be clarifications, deeper dives, or requests that lean on what’s already in the conversation — none of which need a fresh search.

The strategic takeaway is sharper than it first appears. The queries worth winning are the ones that start research journeys — the first thing a worried parent, a prospective student, or a business owner types when a problem becomes real. Not the sophisticated long-tail question they ask ten minutes into their research, after they already have their bearings and their shortlist. Map your content to the entry points of the journey, because that’s where the citation events concentrate.

It triangulates — you’re competing for a slot, not the answer

When ChatGPT does search, it doesn’t crown one winner. Conversations with citations averaged about six unique sources, and two-thirds of cited turns drew on one to four sources. The model behaves like a cautious researcher: it pulls several perspectives and synthesises, rather than trusting any single authority.

That reframes the competition. You’re not fighting to be the answer. You’re fighting for a slot in a small set of sources the engine trusts enough to blend — and your competitors will routinely appear in that same set, side by side with you, in the same response.

Profound’s co-citation analysis makes this concrete: sources travel in packs. Within any vertical, certain domains are repeatedly cited together — in personal finance, for instance, two well-known advice sites appeared together in 14% of conversations that cited either one. The engine has, in effect, a trusted cluster per topic. Your realistic goal is to join the cluster that already exists around your subject — the institutions, journals, and specialist sites the engine already leans on — so that citing you feels like citing them.

The economy is wide open — and brutally unequal

Here’s the twist in the data. No single domain dominates ChatGPT’s citations: the top ten domains combined capture only about 12% of all citations. The long tail is enormous. Hundreds of thousands of sites get cited.

But the distribution across that long tail is severely skewed — a small set of domains takes a disproportionate share while everyone else splits the scraps. Wikipedia sits at the top as the default knowledge layer, appearing in roughly one in six cited conversations. It wins on breadth and neutrality, and you will not out-Wikipedia Wikipedia.

The winnable position is the source after it. Wikipedia can define the condition; it can’t hold current pricing, local specifics, practitioner nuance, or original data. That next-step layer — deeper, fresher, more specific — is where a focused specialist site earns its slot. Which means the real question was never “how do I get cited?” It’s “who gets cited next to me, and what would make the engine trust my page as much as theirs?”

How Google’s AI Overviews choose sources

Google is more mechanical about the whole thing — and, unusually, has confirmed part of its own mechanism in public documentation. Combine that with the two largest independent datasets and the picture gets reasonably clear.

What an AI Overview actually is

An AI Overview is a composed answer, not an extracted one. Where a featured snippet lifted a passage from a single page, an AI Overview synthesises multiple pages into one summary, attaches a set of source cards, and suggests follow-up questions — all above the traditional results. Surfer’s large-scale study found the average Overview cites around eight sources, and fewer than 1% ship with no sources at all. (Source)

That structure changes the game for publishers: visibility inside the box isn’t about rank alone. It’s about whether the system pulls your page into the synthesis.

Query fan-out: one search becomes many

The confirmed mechanism behind that synthesis is called query fan-out, and it’s described in Google’s own documentation. When a user asks a question, the system expands it into a set of related sub-queries — the definition, the steps, the costs, the exceptions, the comparisons — retrieves candidate pages for each, then composes one answer citing across them.

Sit with the implication, because it’s the single most practical fact in this post. A page that answers only the head question competes in one retrieval. A page that also cleanly covers the adjacent sub-questions competes in several retrievals simultaneously. Every well-structured sub-section is another ticket in the draw. This is why comprehensive, well-organised pages keep beating thin pages that technically outrank them — the fan-out gives coverage more chances to matter than position.

The top 10 still feeds the machine — but the door moved

Now the sizing data, and it comes with a trajectory worth understanding. In July 2025, Ahrefs studied 1.9 million citations across one million AI Overviews and found 76.1% of cited pages ranked in Google’s top 10 for the query. When they re-ran the study in March 2026 — 863,000 SERPs, 4 million cited URLs, more than double the original — that share had fallen to roughly 38%, with the rest split almost evenly between pages ranking 11–100 and pages ranking nowhere in the top 100.

What changed wasn’t the mechanism. It was the weighting. Ahrefs’ read, consistent with Google’s own documentation, is that newer AI Overview models run query fan-out more aggressively — so citations increasingly go to pages that rank well for the sub-queries, not the original keyword.

Read that honestly, from both directions.

Ranking for the exact query is no longer the near-guarantee of eligibility it briefly was — a page can hold #3 for the head term and lose the citation to a page that owns three of the fan-out questions. And the side door for well-structured pages with no head-term ranking is now wider than ever.

But don’t mistake the door moving for the building disappearing. Retrieval still runs entirely through the search index: every cited page is one Google found, crawled, and ranked somewhere across the query’s fan-out universe. The front entrance didn’t close — it moved from “rank top 10 for the keyword” to “rank across the keyword’s whole neighbourhood.” That’s a harder brief, not a softer one, and it makes broad, well-structured organic coverage more valuable, not less. GEO without SEO is still a door with no building behind it — which is exactly why we treat organic foundations as the entry ticket to AI visibility, not the legacy channel AI replaced.

Core sources: why the same pages keep getting cited

One more pattern from Surfer’s dataset deserves attention. For a given query, Google tends to reuse a small cohort of URLs across repeated runs — call them core sources — while a larger group of non-core sources rotates in and out around them.

Core sources share traits: tight semantic alignment between the page’s title and the query’s actual phrasing, answer-first formatting, and generally stronger organic positions. Once a page becomes a core source, it captures repeated Overview visibility even as rankings wobble day to day.

The rotation among non-core sources is the opening. Overviews are volatile — the cited set shifts between runs — and every rotation is a fresh audition. A page that keeps proving it’s the cleanest answer to a specific sub-question graduates from occasional walk-on to recurring cast. That’s the realistic path: not gaming your way into the box once, but becoming the page the system keeps reaching for.

Every engine has a personality — and they don’t agree

Everything so far covers the two engines in the title. But your prospects don’t confine themselves to two engines, and the engines don’t share a rulebook.

Conductor tracked citation behaviour across seven AI engines, seven query-intent categories, and seven months — September 2025 through March 2026, 1,056 data points — and found that every engine has what they call an editorial identity: a persistent preference for certain source types that holds month after month. (Source)

Three observations worth more than the table itself.

Even sibling products disagree. ChatGPT and ChatGPT Search share a company and a name and behave like different publications. Google’s three engines — Overviews, AI Mode, Gemini — share infrastructure and diverge on source preference for the same query. There is no such thing as optimising for “AI” as one channel.

Claude is the outlier that proves the point. It skips the social and encyclopedic layer entirely and goes straight to primary, institutional sources. For anyone in a trust-critical field — a hospital, a college, a law practice — that’s quietly excellent news: the careful, formal, source-heavy content you’re obliged to produce anyway is exactly what at least one major engine treats as citation-grade. The content style that feels unfashionable is a competitive asset.

Format coverage now matters alongside topic coverage. If everything you publish is long-form text, you’re contesting citations on some engines and forfeiting others entirely. That doesn’t mean panic-launching a video channel this month. It means knowing which engines your actual audience uses, and matching the format those engines demonstrably trust.

The part nobody says out loud: half the “ranking factors” are folklore

Now the uncomfortable section — the reason this post exists. The AI-visibility space has a data problem: verified research, plausible inference, and outright invention circulate side by side, formatted identically, and almost nobody labels which is which. So let’s label it.

Verified, from primary research or official documentation:

• Retrieval runs through search indexes. Query fan-out is confirmed by Google’s own docs. Search visibility feeds AI visibility — that’s structural, not opinion.
• Position on the page matters. A study of 1.2 million ChatGPT answers, reported by Search Engine Land, found 44% of citations were pulled from the first third of a page. Bury your answer under a long wind-up and you’ve roughly halved your odds before quality even enters the picture.
• Freshness correlates hard. Ahrefs’ study of ChatGPT’s most-cited pages found 89.7% had been updated within the year, and 60.5% were published within the previous two years.
• The citation pool is top-heavy. The same Ahrefs study concluded that about two-thirds of ChatGPT’s thousand most-cited pages are effectively out of reach for ordinary site operators — locked up by major publishers and institutions. The realistic play for everyone else is niche depth, not head-on competition with household names.

Inferred, plausibly, from correlation across independent datasets: authority signals, entity consistency, structured formatting, and third-party presence all correlate with citation frequency in study after study. Acting on them is reasonable. Calling them proven mechanisms is not — correlation in observational data is a hint, not a law.

Invented: the precise citation formulas doing the rounds. The tidy percentage splits between authority, content quality, and trust. The “minimum statistics per thousand words” rules traded in marketing forums like settled science. No AI company has published a citation algorithm. Every exact weighting you’ve encountered was reverse-engineered from the outside by someone observing outputs and guessing at the machine — and, more often than not, someone with a tool or a retainer attached to the conclusion.

Most “AI ranking factors” you’ve read were reverse-engineered by people selling you the fix.

Strip the folklore away and the mental model that best fits the verified data is this: citation is a risk decision. The engine is about to make a claim in public, attach your name to it, and let millions of people check its work. The question it’s effectively asking is not “what’s the best page on this topic?” It’s “what’s the safest thing I can repeat without being wrong?”

That single reframe explains the patterns the formulas can’t. It explains why Wikipedia dominates — not because its prose is the finest available, but because it’s the lowest-risk thing to reference. It explains why engines triangulate across several sources instead of trusting one. It explains why unambiguous, verifiable, consistently-described entities keep beating better-written pages with fuzzy identities. A low-risk page has four properties: it’s unmistakable about what and who it is, its claims can be checked, its structure lets the relevant chunk lift out cleanly, and everything the rest of the web says about it agrees. That’s not a hack. That’s a standard — and it happens to be the same standard a careful human researcher would apply.

So what do you actually do? The five moves that survive scrutiny

Everything below follows from the verified mechanics — no folklore required. None of it is exotic. All of it compounds.

1. Keep the SEO foundation — it’s the infrastructure now. Every AI Overview citation is a page ranked somewhere across the query’s fan-out universe, and ChatGPT’s browsing mode retrieves through a search index too. Rankings stopped being the trophy and became the plumbing. Practically: technical health stays non-negotiable (an engine can’t cite what it can’t crawl), intent coverage across the full journey matters more than a few hero keywords, and earned authority still opens the retrieval door. Nothing about the AI layer made any of this optional. It made it prerequisite.

2. Answer in the first three sentences — of the page, and of every section. The extraction data is unambiguous: engines lift from the top. So give every section a liftable unit — a direct, self-contained answer of one to three sentences that would make sense quoted on its own, before you expand into nuance. Test it by reading only a section’s opening lines: if a stranger couldn’t take the answer away from those alone, restructure. This single habit serves human skimmers, featured snippets, and AI extraction at once, and it costs nothing but discipline.

3. Structure for extraction. Engines cite chunks, not essays. Make the chunks obvious: headings written as the questions people actually ask, sections that hold one idea each, tables for anything numeric or comparative, a genuine FAQ built from genuine questions, and schema markup that mirrors the visible text — FAQ, HowTo, Organization — rather than decorating the page with markup for things it doesn’t say. Remember the fan-out: every cleanly-bounded sub-section is a separate candidate in a separate retrieval. Structure isn’t presentation. Structure is entries in the draw.

4. Make every number traceable. This one is close to home for us, and the engines have made it a visibility issue rather than just an ethics issue. A claim the engine can verify against a primary source is a claim it can safely repeat; a naked statistic with no source is a risk it will route around. Link every borrowed figure to where it actually comes from — the study, not the blog post that quoted the blog post that quoted the study. And if a number can’t be traced, cut it. The page survives. Your credibility with both readers and machines is worth more than one impressive-sounding stat.

5. Keep pages current and your entity consistent everywhere. Freshness is one of the strongest verified correlations, so put your flagship pages on a real update cadence — refreshed data, dated timestamps, visible evidence the page is alive. Then widen the lens: engines cross-check you against the whole web. Your name, services, locations, and claims should match across your site, directories, review platforms, professional profiles, and everywhere else you appear. Every inconsistency raises the engine’s risk of citing you wrongly, and it resolves that risk by citing someone clearer instead. Ambiguity kills citations quietly.

Notice what’s not on the list: a trick. The engines are converging on the standard good marketing always claimed to hold — be clear, be verifiable, be the source others would point to. The teams that treated content as a credibility asset for the past few years are already ahead. The teams that treated it as a volume game are finding out why.

You don’t need an AI hack. You need one system where SEO earns the citation and GEO shapes it. That’s how we run search for clinics, institutions, and firms — every channel accountable to enquiries and booked appointments, every published number traceable to a source. It’s also why we’ll never sell you GEO on its own: the data above shows exactly why that wouldn’t work.

Frequently Asked Questions

Do I need to rank #1 to get cited in AI Overviews?

No — and increasingly you don’t even need to rank for the exact query. Ahrefs found top-10 overlap fell from 76.1% (July 2025, 1.9M citations) to roughly 38% (March 2026, 4M cited URLs) as query fan-out grew more aggressive. But cited pages still overwhelmingly rank somewhere across the query’s related searches — visibility moved from single-keyword position to topic-wide coverage, not away from search rankings.

Why does ChatGPT sometimes invent sources?

Because in its default mode it isn’t retrieving anything — it generates text, including reference-shaped text, from training patterns. That can produce convincing citations to papers and pages that don’t exist. Real citations only appear when ChatGPT actively searches the web, and those come with clickable links. No link, no search log — treat the “source” as unverified until you’ve checked it yourself.

Is getting mentioned by ChatGPT the same as being cited?

No, and conflating them wrecks your measurement. A mention comes from training memory: your brand appears in the answer, with no link, shaped by years of prior web presence. A citation happens only in browsing mode and links to your actual page, driven by current authority, structure, and freshness. Mentions build awareness on a long timeline; citations send visitors on a short one. Track them separately, work them separately.

Is GEO different from SEO, or just rebranded SEO?

It’s an extension, not a replacement — and definitely not a standalone product. AI engines retrieve through search indexes, so the SEO foundation still decides most of who’s even eligible for a citation. GEO is the layer on top: structuring, phrasing, and evidencing content so an engine can lift it cleanly and credit you. Anyone selling GEO without the SEO underneath is selling you a door with no building behind it — every citation still routes through pages the search index found, crawled, and ranked.

Can a new website get cited?

Yes — on specific, fresh, or under-served questions. The data shows engines citing pages with no meaningful rankings when the structure was clean and the answer fit a sub-query precisely, and newer engines reward recency heavily. Sustained citation across a whole topic still requires authority that takes time to earn. So start narrow: pick questions the big players answer generically, answer them completely and verifiably, and expand outward from what works.