Last verified: 22 October 2026
ChatGPT, Perplexity and Gemini choose which brands to cite by turning your question into several narrower web searches, collecting candidate pages, and quoting the few that answer each part most clearly and credibly. Ranking first isn’t enough: only about 38% of pages cited in Google’s AI Overviews rank in the top 10 for the same query (Ahrefs, March 2026).
Key takeaways
- Google’s AI Mode splits a question into subtopics and runs many searches at once, a technique Google calls query fan-out (Google, May 2025).
- Only 37.9% of pages cited in Google’s AI Overviews rank in the top 10 for the same query, down from 76% in July 2025 (Ahrefs, March 2026, 863,000 keyword SERPs).
- Sites that block OAI-SearchBot won’t be shown in ChatGPT search answers, according to OpenAI’s crawler documentation (OpenAI, 2026).
- Adding schema markup made no meaningful difference to AI citations across 1,885 pages, compared with 4,000 control pages (Ahrefs, May 2026).
- Stelrad earned 2,389 AI citations across 405 pages and a 62% share of voice in AI Overviews for radiator questions after Absolute Digital Media rebuilt its advice content (Ahrefs Brand Radar).
When a buyer asks an AI assistant who to trust, they don’t get ten blue links. They get a short answer, a handful of named brands and a few source links. For a regulated brand, the question that matters is no longer only “where do we rank?” but “which page does the answer quote, and is it ours?”
This guide explains the mechanism behind that answer, platform by platform, using published studies and the platforms’ own documentation. It’s written by Absolute Digital Media (ADM), the regulated-sector agency behind the GEO and AI search visibility work shortlisted for Best AI Search Optimisation Campaign at the UK Search Awards 2026. If you’d rather see how your own brand is cited today, you can book an AI visibility audit and we’ll show you.
How do AI assistants decide which sources to cite?
AI assistants decide which sources to cite in four steps: they rewrite the question into one or more search queries, retrieve candidate pages from a search index, rank those pages by how well and how credibly each answers part of the question, and then quote a few of them in the final answer. Citation is the last step of a search process, not a judgement about which brand is best.
The first step is the one most marketers miss. Google describes AI Mode as breaking “your question into subtopics” and issuing a multitude of queries at the same time, which it calls query fan-out (Google, May 2025). Google’s own AI optimisation guide defines fan-out as a set of concurrent, related queries generated by the model to fetch more results (Google Search Central, updated July 2026). Gemini’s grounding tool works the same way for developers: the model decides whether to search, writes its own queries, runs them and attaches citations to the parts of the answer each source supports (Google AI for Developers).
That has a direct consequence. A buyer who asks “which UK lenders are safe to borrow from?” isn’t running one search. The system might look up FCA authorisation, complaint data, representative APR, reviews and recent news separately. Each of those sub-questions can be answered by a different page, and each of those pages is a citation opportunity.
The final step, choosing which pages to quote, is the least documented. OpenAI publishes its crawlers and how to opt in or out, but it doesn’t publish a ranking formula for citations. What we know comes from the platforms’ documentation on retrieval and from large studies of what actually gets cited, which the rest of this guide sets out.
Does ChatGPT use Google or its own index?
ChatGPT doesn’t use Google’s index. OpenAI says ChatGPT search draws on third-party search providers and content from its partners (OpenAI), its help centre names Bing as one of them (OpenAI Help Center), and its own crawler, OAI-SearchBot, gathers pages for search. Sites that opt out of OAI-SearchBot won’t be shown in ChatGPT search answers, though they can still appear as navigational links (OpenAI).
OpenAI runs three separate user agents, and confusing them is one of the most common technical mistakes we see on regulated sites:
| User agent | What it does | What blocking it means |
|---|---|---|
| OAI-SearchBot | Crawls pages so they can appear in ChatGPT search answers | Your pages won’t be shown in ChatGPT search answers |
| GPTBot | Crawls content that may be used to train OpenAI’s models | Your content shouldn’t be used for training; search is unaffected |
| ChatGPT-User | Visits a page when a user’s action asks for it | Not an automatic crawler; doesn’t decide search inclusion |
Source: OpenAI crawler documentation, checked 22 October 2026.
A compliance team that blocks “all OpenAI bots” to keep content out of model training can remove the brand from ChatGPT’s answers by accident. Check robots.txt and any firewall rules for OAI-SearchBot specifically.
How ChatGPT behaves also depends on the mode a user picks. Semrush ran 100 prompts through GPT-5.2 in its minimal-reasoning and high-reasoning modes and found that only 25.6% of cited domains appeared in both answers to the same prompt (Semrush, June 2026). Being cited in one mode doesn’t guarantee being cited in the other, which is why a single screenshot of a ChatGPT answer is a poor measure of visibility.
How do Perplexity, Gemini and Google’s AI Overviews choose sources?
Perplexity, Gemini and Google’s AI features all search the live web before answering, but they draw on different indexes and different favourite sources. Perplexity searches the web in real time and returns numbered citations for each answer (Perplexity Help Center). Google’s AI Overviews and AI Mode use Google’s own ranking and quality systems, plus fan-out.
Google is explicit that its AI features sit on top of ordinary Search. Its guidance says pages must be indexed and eligible to show a snippet to appear in AI features, and that there’s no special markup, text file or rewrite needed for AI (Google Search Central). The same page warns that creating pages mainly to target fan-out queries can breach Google’s scaled content abuse policy.
| Platform | How it retrieves | Most-cited domains (study) | What that tells you |
|---|---|---|---|
| ChatGPT search | Third-party providers (Bing named), partner content, OAI-SearchBot | Wikipedia 7.8% of citations, Reddit 1.8% (Profound, 680m citations, Aug 2024–Jun 2025) | Reference-style pages that define things clearly are favoured |
| Perplexity | Live web search on every question, numbered citations | Reddit 6.6%, YouTube 2.0% (Profound, same study) | Community discussion and video carry weight |
| Google AI Overviews | Google’s index and ranking systems plus fan-out | Reddit 2.2%, YouTube 1.9%, Quora 1.5% (Profound, same study) | A wide spread of sources; forums and video feature heavily |
| Google AI Mode | Google’s index with heavier fan-out | LinkedIn, YouTube, Reddit among the top five (Semrush, 100m+ citations, Jul–Oct 2025) | A more balanced mix of domains than ChatGPT |
| Gemini (API grounding) | Model-written Google searches, citations per sentence | Not published | Each sentence can carry its own source |
Sources: Profound, June 2025, updated August 2025; Semrush, November 2025.
Treat the domain shares as a snapshot, not a law. Promptwatch data reported by Semrush showed Reddit’s share of ChatGPT citations falling from 3.8% to 0.5% in a matter of weeks in August 2026, after ChatGPT changed how it fans out queries (Semrush, August 2026). The platforms change their retrieval often. The principle underneath, that the clearest page for each sub-question gets quoted, has held across every study we’ve read.
Google also changed the shape of its answers this month. On 15 October 2026 it said AI Overviews and AI Mode will show more inline links, higher up in responses, that it is testing a section of official sites at the top of AI answers, and that use of its Preferred Sources feature has grown more than seven times since the start of the year (Google, 15 October 2026). More links inside the answer means more value in being one of the pages it already relies on.
Do you need to rank first on Google to be cited by AI?
You don’t need to rank first on Google to be cited by AI, but you do need to be found by the searches the AI runs. Ahrefs’ March 2026 update, covering 863,000 keyword SERPs and about four million AI Overview URLs, found that 37.9% of cited pages ranked in Google’s top 10 for the same query, 31.2% ranked between 11 and 100, and 31.0% ranked beyond 100 (Ahrefs).
| Organic position of cited page | Share of AI Overview citations (all SERP features) | Share (blue links only) |
|---|---|---|
| Top 10 | 37.9% | 37.1% |
| 11–100 | 31.2% | 26.2% |
| Beyond 100 | 31.0% | 36.7% |
Source: Ahrefs, March 2026, 863,000 keyword SERPs.
A year earlier the same analysis put the top-10 share at 76%. Ahrefs attributes most of the drop to fan-out: Google is taking more of its citations from the results of the sub-queries it generates, rather than from the original results page. Part of the change also reflects better citation detection in Ahrefs’ own tools.
Two practical conclusions follow. First, a page that ranks on page three for your head term can still be the page an AI quotes, if it ranks well for one of the narrower questions underneath. Second, the reverse is true: a brand that ranks first for its head term can be left out of the answer entirely if its page doesn’t answer the specific sub-questions cleanly.
Clicks matter too. Ahrefs found AI Overviews are associated with a 58% lower average click-through rate for the top-ranking page, comparing December 2023 with December 2025 across 300,000 keywords (Ahrefs, February 2026). When fewer people click, being named in the answer itself does more of the work.
Which types of pages get cited most often?
The pages AI assistants cite most often are the ones that answer a narrow question in a sentence or two, show their evidence and are easy to lift without rewriting. In practice that means reference pages, comparison tables, ranked lists, forum threads with first-hand answers, and video. Long brand pages that bury the answer under introductions are cited far less.
| Page type | Why AI answers cite it | Typical owners in regulated sectors |
|---|---|---|
| Definitions and reference pages | One-sentence answers to “what is” questions | Wikipedia, regulators, glossaries |
| Comparison tables and ranked lists | Answer “which is best” and “how do they compare” in one place | Affiliates, comparison sites, trade press |
| Forum and community threads | First-hand answers to specific situations | Reddit, Quora, specialist forums |
| Video | Demonstrations and explainers, often summarised by transcript | YouTube creators, brands |
| Official and policy pages | Authoritative facts: licences, rates, terms | Regulators, licensed brands’ own sites |
Sources: Profound (June 2025) and Semrush (November 2025) domain studies; Ahrefs (March 2026) found 18.2% of AI Overview citations outside Google’s top 100 were YouTube URLs.
Notice the gap in the last column. In most regulated categories the reference pages, comparison tables and forum answers are owned by someone other than the brand. The licensed operator, lender or clinic owns the official facts, but often publishes them in a way an AI system can’t easily quote: in PDFs, behind tabs, or spread across long pages without a direct answer.
Some popular tactics don’t show up in the evidence. Ahrefs tracked 1,885 pages that added schema markup between August 2025 and March 2026 against 4,000 pages that didn’t, and found no meaningful increase in citations in AI Mode or ChatGPT, and a small fall in AI Overviews (Ahrefs, May 2026). Google says the same of llms.txt files and rewriting content into chunks for AI: its Search systems don’t use them. Schema is still worth having for rich results. It just isn’t a citation lever on its own.
Why do affiliates and forums get cited ahead of licensed brands?
Affiliates and forums get cited ahead of licensed brands because they answer the buyer’s comparative question directly, on one page, in plain language. A brand’s site answers “what do we offer?”. An affiliate’s page answers “which of these should I choose?”, which is much closer to the question people actually ask an AI assistant.
Our iGaming operator’s guide to AI search visibility sets out why this gap is closable for licensed operators in particular. Google’s quality systems reward first-hand experience and trust, and an operator runs the platform, holds the licence and deals with players every day; no affiliate has that operational knowledge. Google has also narrowed the room for third-party content riding on a host site’s reputation: its site reputation abuse policy took effect on 5 May 2024 (Google Search Central, March 2024) and was tightened in November 2024 to cover third-party content regardless of first-party oversight (Google Search Central, November 2024).
The operator’s advantage is real but not automatic. It only turns into citations when the operator’s own pages state the facts an AI needs, in the form it can lift: licence details, how withdrawals work, what responsible-gambling tools exist, which markets it’s licensed in. We’ve written more on the commercial case in how regulated iGaming operators are cutting affiliate CPA with organic search and LLMs, and on why paying for mentions is the wrong shortcut in Google has declared war on bought AI citations.
The same pattern runs through finance and healthcare. A comparison site answers “which lender approves bad credit?” in a table; the FCA-authorised lender’s own site often says nothing on the question at all. The answer engine quotes the table.
What can regulated brands get recommended for in AI answers?
Regulated brands can get recommended for the factual, checkable parts of a buyer’s question: who is licensed or authorised, what a product does, what it costs, how it works and what protections apply. They can’t control how an AI answer compares them with competitors, and they must not publish anything on their own site that would breach the advertising rules to try to influence it.
That second point is easy to overlook. The Advertising Standards Authority (ASA) has regulated marketing claims on companies’ own websites and social media pages since 1 March 2011 (ASA). A page written to be quoted by an AI is still a marketing communication. If it makes a claim the CAP Code wouldn’t allow in an ad, it’s a breach whether a human or a machine reads it first.
| Sector | What your own pages can state clearly | What they must never say |
|---|---|---|
| iGaming | Licence numbers and markets, payment and withdrawal terms, safer-gambling tools, how products work | Anything implying a likely win or profit, content that appeals to under-18s, bonus language that misleads |
| Financial services and crypto | FCA authorisation and register details, how a product works, representative examples where required, complaint routes | Anything that reads as a financial promotion without the right approval, or as personal advice |
| Healthcare | Registration (CQC, GPhC), who delivers treatment, process and aftercare, how to book | Clinical outcome claims, before-and-after promises, promotion of prescription-only medicines |
| Regulated ecommerce | Age verification, delivery and returns terms, product specifications | Health or performance claims you can’t substantiate |
| Adult | Age verification, privacy, product information quoted from platform policies | Explicit content or claims that breach platform policies |
Sources: ASA online remit; UK Gambling Commission LCCP 1.1.2; FCA FG24/1.
The ASA’s ruling of 14 October 2026 against HRY Games, trading as AI Scorecast, shows where the line sits. A Facebook ad for its football-results app claimed to predict outcomes with 97% accuracy; the ASA upheld the complaint and found the ad implied people were very likely to profit from betting decisions, breaching CAP Code rules 1.3 and 3.1 (ASA, 14 October 2026). “AI” in a product name doesn’t change the rules.
Gambling operators also carry responsibility for their partners. Licence condition 1.1.2 makes licensees responsible for the actions of third parties they contract with, including marketing affiliates (UK Gambling Commission). If an affiliate page is the one AI answers quote about your brand, it’s worth knowing exactly what it says.
Marketing guidance, not legal or financial advice. Rules and platform documentation checked 22 October 2026. Check sector rules with your compliance team before publishing.
How does a brand become a cited source?
A brand becomes a cited source by finding the questions AI assistants are answering in its category, identifying which pages they quote today, and publishing a clearer, better-evidenced answer on its own domain for each one. The work is closer to editorial planning than to traditional link building, and it should start with measurement.
- Measure before you change anything. Record which prompts in your category mention your brand, which competitors appear and which URLs are cited, across ChatGPT, Perplexity, Gemini and Google’s AI features. Without a baseline you can’t tell whether anything worked.
- Map the sub-questions. For each important buyer question, list the narrower questions an AI is likely to search for: licensing, costs, process, comparisons, risks. These are your fan-out targets, and each needs a direct answer somewhere on your site.
- Answer first, on your own domain. Open each page or section with a one-to-two-sentence answer that names the subject, then give the detail. Put the facts AI systems need (licence numbers, rates, timings) in HTML text and tables, not images or PDFs.
- Date and source every claim. A figure with a named source and date is easier to trust and easier to quote than an adjective. Link the primary source, not a blog that summarised it.
- Make proof checkable. Licences, registrations, published case studies and independent awards give an AI system something to verify. Keep them on crawlable pages.
- Keep crawlers in. Allow OAI-SearchBot and Googlebot, make sure key pages are indexed and snippet-eligible, and don’t hide answers behind tabs, accordions loaded by JavaScript or login walls.
- Fix the sources that already speak for you. Review the affiliate, comparison and directory pages AI answers quote about your brand. Correct outdated facts and, for gambling operators, make sure partners meet the same standards you do.
- Measure citations, not just clicks. Track citations and share of voice monthly alongside conversions. Referral data from AI tools is unreliable: from 9 October 2026 many ChatGPT links arrived without their tracking tag, and LLM Pulse measured a 46% fall in ChatGPT sessions in GA4 over three days while total traffic stayed flat (LLM Pulse, October 2026).
The order matters. Brands that start by producing content at volume often create pages aimed at fan-out queries, which Google’s guidance specifically warns against, without knowing which questions they were losing in the first place.
What does this look like for a real brand?
For Stelrad, the UK radiator manufacturer, it meant turning a brand that ranked first in Google into one that AI answers quote. Stelrad’s site already held position one for “radiator” and “radiators” in UK search, but AI answers to radiator questions were drawing on retailer buying guides, with the manufacturer mentioned as a product rather than cited as the authority (ADM case study).
ADM followed the order above. Citations and share of voice were baselined first. The advice hub, not the category pages, became the citation source: guides on radiator types, heat output and column versus panel radiators were rebuilt as answers, each opening with a direct answer. Thin, duplicated and legacy pages were merged or removed, and the link building and digital PR programmes were pointed at the same guides so the brand’s claims were corroborated off-site.
| Measure (Ahrefs Brand Radar) | Stelrad |
|---|---|
| AI citations, all platforms | 2,389 across 405 pages |
| Google AI Mode citations | 557 across 231 pages |
| Share of voice in AI Overviews, radiator questions | 62% (next retailer 30%) |
| Manufacturer share of voice in ChatGPT | 19% (next manufacturer 10%) |
Source: Stelrad GEO case study, Absolute Digital Media, figures from Ahrefs Brand Radar.
The programme is shortlisted for Best AI Search Optimisation Campaign at the UK Search Awards 2026, with winners announced on 25 November. Stelrad isn’t a regulated brand, but the problem it solved, being out-quoted by third parties on its own subject, is the one licensed operators, lenders and clinics face every day. Our sector pages for GEO in iGaming, financial services and healthcare explain how the same method adapts to each set of rules.
Find out which pages AI answers cite in your sector
Absolute Digital Media’s AI visibility audit asks ChatGPT, Perplexity, Gemini and Google’s AI Mode the questions your buyers ask, records which pages they cite for you and up to three competitors, and gives you a prioritised list of fixes that stay inside your sector’s rules. It’s built for regulated brands in iGaming, finance, crypto, healthcare, adult and regulated ecommerce.
Sources
- Ahrefs – Update: 38% of AI Overview citations pull from the top 10, March 2026
- Ahrefs – Update: AI Overviews reduce clicks by 58%, February 2026
- Ahrefs – We tracked 1,885 pages adding schema. AI citations barely moved, May 2026
- Google – AI Mode in Google Search: updates from Google I/O 2025
- Google – AI Mode and AI Overviews updates to explore the web, 15 October 2026
- Google Search Central – Optimising for generative AI features on Google Search, updated July 2026
- Google AI for Developers – Grounding with Google Search
- Google Search Central – March 2024 core update and new spam policies
- Google Search Central – Updating our site reputation abuse policy, November 2024
- OpenAI – Overview of OpenAI crawlers
- OpenAI – Introducing ChatGPT search
- OpenAI Help Center – ChatGPT search for Enterprise and Edu
- Perplexity Help Center – How does Perplexity work?
- Profound – AI platform citation patterns, June 2025 (updated August 2025)
- Semrush – The most-cited domains in AI: a 3-month study, November 2025
- Semrush – Reddit’s citations in ChatGPT fall from 3.8% to 0.5%, August 2026
- Semrush – Only 25% of cited sources overlap between ChatGPT’s reasoning modes, June 2026
- LLM Pulse – ChatGPT referral traffic fell by almost half on 9 October 2026
- ASA – Ruling on HRY Games t/a AI Scorecast, 14 October 2026
- ASA – New online remit enhances consumer protection
- UK Gambling Commission – LCCP condition 1.1.2, responsibility for third parties
- Financial Conduct Authority – FG24/1 finalised guidance on financial promotions on social media
Frequently asked questions
How does ChatGPT decide what to recommend?
ChatGPT decides what to recommend by rewriting your question into searches, retrieving pages through third-party search providers and its OAI-SearchBot crawler, and citing the pages that answer each part most clearly. OpenAI doesn’t publish its ranking formula, but studies show reference-style pages such as Wikipedia are cited most often (Profound, 2025).
Does ChatGPT use Google?
ChatGPT doesn’t use Google’s index for its search answers. It uses third-party search providers such as Bing, partner content and OpenAI’s own crawler, OAI-SearchBot. Sites that block OAI-SearchBot won’t be shown in ChatGPT search answers, according to OpenAI’s crawler documentation, even if they rank well in Google.
Why do Reddit threads get cited so often?
Reddit threads get cited often because they answer specific, real-world questions in plain language, often from first-hand experience. Reddit was among the most-cited domains on Perplexity and Google’s AI features in Profound’s 2025 study, though its share of ChatGPT citations fell sharply in August 2026 (Semrush), so it varies by platform.
Can a casino get recommended by ChatGPT?
A licensed casino can be cited by ChatGPT for factual questions such as licensing, payment methods, withdrawal times and safer-gambling tools, if its own pages state those facts clearly. How assistants handle gambling questions is up to each platform, and the operator’s pages must still follow the CAP Code and UK Gambling Commission rules.
How do I get my brand into Perplexity answers?
To get your brand into Perplexity answers, publish clear, dated, sourced answers to the specific questions buyers ask, on crawlable pages of your own site, and make sure trusted third parties describe you accurately. Perplexity searches the web live for every question, so recently updated pages and active community discussion both count.
