Why this matters
After this lesson you can explain how an AI assistant can talk about a page published yesterday, and what "grounded" and "cited" really mean. That picture keeps you from believing both extremes: that AI search is magic, and that it is just old search with a new coat of paint.
As in every lesson, we separate what vendors document from what is general knowledge and what is guessing.
A model alone
A large language model is trained on huge amounts of text and learns patterns well enough to produce fluent text. By itself it answers from what it absorbed during training. It has a training cut-off, it does not look things up, and it can state false things confidently. That last failure is often called a hallucination.
Adding retrieval
To answer questions about recent or specific material, products add a lookup step. The general idea is called retrieval-augmented generation, often shortened to RAG:
- The user asks a question.
- The system turns the question into one or more searches and retrieves relevant passages or pages.
- Those passages are placed in the prompt that goes to the model.
- The model writes an answer using them.
- The product may show citations: links or footnotes pointing at the sources.
This is a conceptual pattern. Google's Search Central guidance for generative features says its features use techniques such as retrieval-augmented generation and query fan-out to draw on content from its Search index. That sentence is documented. The internal detail is not.
Grounding
Grounding means tying the answer to retrieved material rather than to the model's memory alone. Google uses the word in product names and in the controls described for Google-Extended: its crawler documentation says the token covers use of content for grounding in Gemini apps as well as for training. Grounding lowers the chance of invention. It does not remove it; models can still misread or compress sources badly.
Query fan-out
Google's documentation on AI features says AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources while building a response. This lets the answer link to a wider range of pages than the one query typed. Practical meaning, as established reasoning: a page can be used for a sub-question the user never typed. Writing a page that fully answers its own topic, with clear subtopics, serves this. Google's guide does not tell you to create a separate page for every variant, and lists mass-creating pages for fan-out queries as scaled content abuse. Do not do it.
Citations
A citation in a generated answer might be a footnote, a card or a "sources" panel. Four things to know:
- It is observational. It tells you what appeared in one answer, for one user, at one time.
- Different products behave differently. Some cite many pages, some few. The behaviour is set by each company and changes.
- A citation is not a quote of accuracy. The answer might describe your page incorrectly.
- Selection is not fully documented. Which retrieved page gets a visible link is decided by product design plus model behaviour. Claims such as "answers cite pages with a table in the first paragraph" are hypotheses.
What is unknown
Be honest about this list; it is long:
- The exact mix of ranking systems and models used in any commercial AI answer.
- How many candidate pages are read versus linked.
- Whether a particular phrasing style increases quotation.
- How results differ between users, countries, languages and days. They do vary, which lesson Measuring AI visibility honestly treats seriously.
A worked mental example
Imagine a fictional tutoring service, "Greenfield Tutors". A parent asks an assistant: "how much homework should a nine-year-old do?" A retrieval-based system may search for homework guidance from schools, education bodies and child-development sources. It reads passages, writes a short answer, and links three pages. If Greenfield Tutors has a clear, well-sourced article on this, hosted on a crawlable page, it is eligible to be among the candidates. Whether it is chosen depends on many things nobody outside the vendor can see: the quality of competing pages, how well the passage matches, and product rules.
What Greenfield controls: access, clarity, accuracy, honest authorship, and not hiding anything. What it does not control: the selection.
Reader question -> searches (possibly several) -> retrieved passages
-> model writes answer using passages -> links shown (maybe)Common mistakes
- Saying "the AI knows our site" when it may only have retrieved one page once.
- Believing a citation proves ranking, or that no citation proves failure.
- Writing text only for the model and forgetting that real readers click through and judge you.
- Assuming one assistant's behaviour applies to another. Documentation is per vendor.
- Ignoring errors about your brand because the answer "has a link".
Exercise
- Ask an AI search product a question that your site answers. Record the exact wording, the date, the product and which sources it links. Repeat on another day and note differences.
- Write one paragraph, in your own words, explaining RAG to a non-technical friend. Use the bakery or tutoring example.
- List three claims about AI source selection you have read and label each documented, established or hypothesis.
Show a sample answer
Step 2 could read: "The assistant first searches, then reads what it found, then writes its answer from that. Links show where it looked." For step 1, expect differences: this variation is normal and is exactly why single screenshots prove little.
Previous: From crawl to ranked results. Next: What Google documents about AI Overviews and AI Mode.
Key points
- A language model on its own writes from patterns learned in training; it does not look up today's web.
- Retrieval-augmented generation fetches sources first and gives them to the model, which is how current pages can be cited.
- Grounding means anchoring the answer in retrieved material; a citation is a link or reference to that material.
- Google documents query fan-out: several related searches may be issued to build one answer.
- How a particular product picks which sources to cite is not fully documented, so claims about it are hypotheses.