Say you rank first on Google for “emergency plumber Widnes”. You ask Google’s AI Mode or ChatGPT the same question, and it names two other firms and links to a directory. How did it decide?
This guide explains how AI search citations work, using what Google, OpenAI and Perplexity have published and what independent research has measured. It isn’t magic, and it isn’t a ranking. It’s a retrieval process: the tool runs searches, reads some of the pages that come back, writes an answer, then attaches a few links.
This is the mechanism, not a to-do list. For tactics, read how to appear in Google AI Overviews or how to get mentioned in ChatGPT. If “GEO” is new to you, start with what GEO is.
Mentions, Citations and Answers From Memory Are Not the Same Thing
Your competitor gets named in a ChatGPT answer. Nobody gets a link. You can’t work out where the name came from.
That confusion comes from lumping three different things together as “being cited”:
- Citation. A clickable link to a page the tool used, shown next to a sentence or in a sources panel. Conductor’s Patrick Reinhart describes citations as “the links that you see that are either embedded or on the side of the AI response.”
- Mention. Your business is named in the answer, with no link.
- Answer from memory. The model answers from what it learned in training, with no live search. No search, nothing to cite.
The third catches people out. OpenAI’s help centre says ChatGPT decides whether to search the web based on what you ask, unless you choose search yourself. Google’s Gemini API documentation describes the same step: the model “analyzes the prompt and determines if a Google Search can improve the answer.”
So a mention can come from training data or from a page the tool read. A citation only comes from a live search. “Am I cited?” and “am I mentioned?” are different questions with different causes.
If you’re checking your own business, record mentions and citations separately.
How One Question Turns Into Many Searches (Query Fan-Out)
The tool doesn’t search for what you typed. It searches for several things it thinks you meant.
Google says AI Overviews and AI Mode “may use a ‘query fan-out’ technique”, issuing “multiple related searches across subtopics and data sources” to build one response. At Google I/O in May 2025 it said AI Mode issues “a multitude of queries simultaneously”, and Deep Search “can issue hundreds of searches”. AI Mode has been live in the UK since late July 2025.
ChatGPT does the same under a different name. OpenAI’s help centre says ChatGPT search typically rewrites your prompt into one or more targeted queries for its search partners, and may send more specific follow-ups after reading the first results. It can add your general location from your IP address too, so a “restaurants near me” prompt becomes a search for top restaurants in a named city. Gemini’s API docs say “the model automatically generates one or multiple search queries and executes them.”
Here’s a hypothetical for a local trade (no platform publishes the sub-queries for a given prompt). Someone asks about “boiler repair cost Warrington”, and the tool might search separately for:
- typical call-out charges for boiler repairs
- whether the engineer must be Gas Safe registered
- same-day boiler repair availability
- heating engineers serving Warrington
Your page might rank well for the typed phrase and appear in none of those results, while a competitor’s page that answers one sub-question clearly gets picked up. That explains most of the “but I rank first” confusion, and our AI Overviews guide covers what to do about it.
Classic search answers one query. AI search answers a cluster.
Where Each Tool Looks: Google, ChatGPT and Perplexity Compared
Why would you show up in Gemini but not ChatGPT? Because each tool searches a different pool of pages, and visibility in one doesn’t carry over.
| Platform | Where it searches | How it reads pages |
|---|---|---|
| Google AI Overviews and AI Mode | Google’s own index | A page must be “indexed and eligible to be shown in Google Search with a snippet”. Google says there are “no additional requirements”. |
| ChatGPT search | Third-party search partners, sent rewritten queries | OAI-SearchBot is “used to surface websites in search results”. ChatGPT-User may visit a page live when someone asks a question. |
| Perplexity | Its own index, built by PerplexityBot | PerplexityBot “is designed to surface and link websites in search results on Perplexity”. Perplexity-User may visit a page live. |
On ChatGPT’s partners, the best independent evidence is a small Seer Interactive study (February 2025). It found 87% or more of ChatGPT’s citations matched Bing’s top results. The authors call it “directional”: it covered 100 questions, and things may have changed since.
Local answers add a second layer. OpenAI says ChatGPT may share your general location with trusted third-party providers of local information such as nearby business listings. Google’s local AI can draw on Maps place data and reviews, not just web pages: launching Ask Maps in March 2026, Google said Maps “analyzes information from over 300 million places, including reviews from our community of more than 500 million contributors.” Ask Maps started in the US and India, then expanded on 6 August 2026 to over 150 countries and territories in English. Google hasn’t named the UK, so UK availability is unconfirmed. Nor has Google documented how much Business Profile data counts inside AI Overviews.
For crawler settings, see how to get mentioned in ChatGPT.
If you only check one platform, you only know about one index.
Read, Used, Cited: Why Most Pages Never Get the Link
Only 38% of pages cited in Google’s AI Overviews also rank in the top 10 for the same query. That’s from an Ahrefs study published in March 2026, covering 863,000 keyword results pages and 4 million AI Overview URLs. About 31% ranked 11th to 100th, and another 31% ranked beyond the top 100. Ahrefs puts this down to query fan-out.
Other studies measure it differently. BrightEdge (September 2025) found 54.5% of AI Overview citations overlapped with organic rankings at any position, but only 16.7% came from the top 10. The numbers disagree. The conclusion doesn’t: ranking first for your phrase guarantees nothing.
To see why, split how AI search citations work into three filters:
- Fetched. The tool pulls candidate pages from the results of its many searches.
- Used. The model reads some of them and writes its answer from what it found.
- Cited. Links are attached to particular sentences. Google’s Gemini API shows this plainly: each citation ties a specific stretch of the answer text, marked by start and end positions, to a source URL.
So a page can be fetched, can even shape the wording, and still get no link. The link goes to whichever source the tool matched to that sentence.
Google also says its AI features show “a wider and more diverse set of helpful links” than a classic search. That’s why an unexpected directory or forum can take the link instead of you.
A citation is a sentence-level credit, not a position. You earn it one claim at a time.
What the Research Says Makes a Source More Likely to Be Used
You’ve probably heard that “GEO boosts visibility by 40%”. It comes from a real study, and what it tested is more useful, and more modest, than the way it gets quoted.
The paper is “GEO: Generative Engine Optimization” by Aggarwal and colleagues from Princeton and other institutions, published at KDD 2024. Its abstract says GEO “can boost visibility by up to 40% in generative engine responses.”
The researchers rewrote source pages nine different ways and measured how much of a generated answer drew on each source. The test engine was GPT-3.5 writing answers from the top five Google results, and they also tested live on Perplexity.
What helped:
- adding quotations (about 41% relative improvement)
- adding statistics (about 33%)
- citing sources (about 28%)
Keyword stuffing scored below the unedited baseline.
The interesting part for small firms is who gained most. A source in fifth position that added citations gained 115.1%, while the top-ranked source lost 30.3% under the same method.
Now the caveats. This was a lab setting with an older model, and visibility meant share of the answer, not clicks or enquiries. The authors say the methods may need to change as engines evolve. It shows what gets used, not a guaranteed route to being recommended. For the practical side, see what GEO is.
If someone quotes you the 40% figure, ask them which setup it came from.
Why the Same Question Gets a Different Answer Tomorrow
There’s less than a 1 in 100 chance of getting the same list of brands twice from the same AI prompt. SparkToro and Gumshoe.ai found this in late 2025, when 600 volunteers ran 12 prompts through ChatGPT, Claude and Google’s AI 2,961 times. Getting the same list in the same order was closer to 1 in 1,000.
The reason follows from everything above. The fan-out queries, the pages fetched and the answer written can all change from run to run. The GEO researchers sampled five answers per query for exactly this reason.
Some things do hold. SparkToro found that “visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric”, and narrower markets gave more consistent answers. Rand Fishkin’s verdict on the alternative: “any tool that gives a ‘ranking position in AI’ is full of baloney.”
So never treat one screenshot as your score. Track the share of runs you appear in, with the same prompts, repeated over time. Our guide to AI search visibility for local businesses explains how to run that check. SparkToro do say they need bigger samples, and newer models may behave differently.
A Google rank is a position. An AI visibility figure is a percentage.
Citations Can Be Wrong, and That Matters for Your Business
A link next to a sentence feels like proof. It isn’t.
In March 2025, Columbia’s Tow Center for Digital Journalism asked eight AI search tools to identify the source of quoted news excerpts, across 1,600 queries. They were wrong more than 60% of the time, from 37% (Perplexity) to 94% (Grok 3). They often cited syndicated copies instead of originals, made up URLs and rarely admitted uncertainty.
That tested news, not local business details, but the lesson carries over. A citation shows where the tool says it got something, and a directory’s copy of your details can be cited instead of your own site.
The tool can only repeat what it finds. If your prices, hours or services differ across the web, it has more chances to repeat the wrong version.
If you’d rather someone checked how AI tools describe your business, that’s our GEO work. Tell us what you do and where, and we’ll give you a straight answer on whether we can help. If we can’t, we’ll say so.
FAQ
How does ChatGPT choose which sources to cite?
First it decides whether to search. If it does, it rewrites your prompt into one or more queries for its search partners, reads the results and links the pages it used for particular points. A small Seer Interactive study found most of its citations matched Bing’s top results, though the authors call that directional.
Do I need to rank in Google’s top 10 to be cited in AI Overviews?
No. Ahrefs found only 38% of pages cited in AI Overviews rank in the top 10 for the same query, largely because of query fan-out. To get cited by AI search features like AI Overviews, your page must still be indexed and eligible to show in Google Search with a snippet. Google says there are no additional requirements.
Why was my business in the AI answer yesterday but not today?
AI answers are generated fresh each time. SparkToro found less than a 1 in 100 chance of the same list of brands appearing twice for one prompt. Measure how often you appear across repeated runs, not once.
Can I see AI Overview traffic in Search Console?
Yes, but it isn’t separated out. Google says sites appearing in AI Overviews and AI Mode are included in the Search Console Performance report, under the “Web” search type.