# What Is a Generative Summary in AI Search?

> What a generative summary is, what it is made of, why its citations don't always back what it says and how to check whether it cites your store.

URL: https://clione.ai/glossary/generative-summary/
Lang: en

A generative summary is what you read when a search engine answers you with a written paragraph and a row of sources, instead of ten links. It is not the summary of a document that you ask an assistant for, nor the [meta description](/glossary/meta-description/) you write for your [product page](/glossary/product-page/), nor a product's review roundup. It is **the answer a search engine composes in [AI search](/glossary/ai-search/)** from other people's pages, and each service gives it its own name: Google calls it [AI Overviews](/glossary/ai-overviews/). A store should care for two reasons: the shopper can read what they need without opening any site, and **what that text says about your category or your product may come from your page or from someone else's**.

## How it works

The paper that gave [GEO (Generative Engine Optimization)](/glossary/geo/) its name describes the chain behind it. A [generative engine](/glossary/generative-engine/) turns the query into several simpler searches, retrieves a set of sources, summarizes each one with a model and passes those summaries to another model that **writes a single answer backed by the sources**, usually structured text with citations inside.[^geo-paper] The authors add why citations matter: language models tend to make information up.[^geo-paper]

Google tells its version in other words. Its AI features use [retrieval-augmented generation](/glossary/rag/) ([RAG](/glossary/rag/)), which it also calls [_grounding_](/glossary/grounding/): its usual ranking systems pull relevant, up-to-date pages from its index to improve the quality, accuracy and freshness of the answer. And the model may run a set of related queries at once, [_query fan-out_](/glossary/query-fan-out/), to bring back more results.[^google-ai-guide] Bing describes its generative search, introduced in July 2024, as a tailored answer that **understands the query, reviews millions of sources, matches content and generates the results in a new** AI-built format, without removing the usual results from the page.[^bing-generative-search]

### The three parts of a summary

When you look at one, split it into three pieces, because each is earned differently:

| Part                 | What it is                                            | What your store decides                                                                              |
| -------------------- | ----------------------------------------------------- | ---------------------------------------------------------------------------------------------------- |
| The text             | The sentences the model writes                        | Whether your page states the fact the text needs, clearly                                            |
| The citations        | Marks inside the text that point to a specific source | Whether your page really backs the sentence that cites it                                            |
| The supporting links | The list or panel of sources next to the text         | Whether your page can be a source: in Google, indexed and with snippets allowed[^google-ai-features] |

You can show up in the links without the text using anything of yours, and the text can say what your page says while citing another. That is why a summary is not measured with a single question ("am I in it?") but with three.

### Why it gets things wrong

In 2023, three researchers, Nelson F. Liu, Tianyi Zhang and Percy Liang, audited four generative search engines with human evaluators ([Bing Chat](/glossary/copilot/), NeevaAI, [Perplexity](/glossary/perplexity/) and YouChat) and measured whether the citations backed what they said. Their answers were fluent and appeared informative, but **only 51.5% of sentences were fully supported by their citations, and only 74.5% of citations backed the sentence they came with**.[^liu-verifiability] Those are 2023 systems, and the figure does not describe today's, but the method still works for checking any summary.

In May 2024 Google explained the failures of AI Overviews' first days. According to the company, those summaries generally do not make things up the way other language-model products do, because they are integrated with its ranking systems; when they fail it is from **misinterpreting a query, misinterpreting a nuance of language on the web, or not having a lot of good information** available. Its examples were a query about how many rocks to eat and another that ended up recommending glue on pizza, taken from a forum.[^google-aio-may-2024] The lesson for a store is direct: if the only page that explains a fact about your product is a forum or a competitor's product page, the summary works with that.

## In a store

A **made-up example**: someone searches "which manual coffee grinder works for espresso". The summary says three things, each with its citation. This is how you review it, sentence by sentence:

| Summary sentence                                                      | Cited source                 | Does the source say it?                  |
| --------------------------------------------------------------------- | ---------------------------- | ---------------------------------------- |
| "For espresso you need a very fine grind that adjusts in small steps" | A coffee store's guide       | Yes, with the same idea                  |
| "Model X has 38 mm steel burrs"                                       | Another store's product page | No: the page does not give the burr size |
| "It costs about $90"                                                  | A blog comparison            | Yes, but the price is a year old         |

The second row is the opportunity. **If your product page gives the fact the summary asserts and another source does not back it, your page is the better source for that sentence**, and nothing stops the summary from citing it next time. None of these three checks needs tools: a browser and patience.


That a generative summary is a featured snippet with a different look. A featured snippet takes a piece of **a single page** and puts it on top with its link; a generative summary **writes new sentences from several pages** and cites them inside the text.[^geo-paper] That is why being cited does not mean the summary copies your text, or that it says what your page says: the 2023 audit found that one in four citations did not back its sentence.[^liu-verifiability]


## How to check it

1. **Search the question your customer would ask** and separate the text, the citations and the supporting links in a note, with date and time.
2. **Open each citation and read whether it backs the sentence**, fully, partly or not at all: it is the same check the 2023 audit made.[^liu-verifiability]
3. **Repeat it on another day.** In Google, AI Overviews only appears when its systems believe it adds something to classic search, so the same query may bring a summary or not.[^google-ai-features]
4. **Confirm your product pages can be a source in Google**: indexed and with snippets allowed, without a `nosnippet` or a `max-snippet` that trims what can be used.[^google-ai-features]
5. **Open the [generative AI performance report](/glossary/generative-ai-performance-report/) in [Search Console](/glossary/google-search-console/).** It counts how many times links to your site were shown in AI Overviews and in [AI Mode](/glossary/google-ai-mode/), by page, country, device and date, without clicks; it has been available to all sites since August 31, 2026.[^gsc-genai-report] It does not tell you what the text claims: you only see that by reading it.

## Related concept: AI Overviews

Generative summary is the genus; AI Overviews is a species, Google's, which appears inside the regular results page.[^google-ai-features] Bing has its generative search,[^bing-generative-search] and Perplexity and ChatGPT with search answer with summaries of the same kind. Even within Google, AI Overviews and AI Mode **may use different models and techniques**, so their answers and links vary.[^google-ai-features] When someone says "the AI summaries", ask which search engine they mean: each is checked separately.

To go further, the [generative engine](/glossary/generative-engine/) entry explains the system that writes these summaries, and the [AI search reading path](/learn/ai-search/) orders the rest of the terms.