# 7 AI Overview Signals: Relevance Has the Best Evidence

> Seven commonly named AI Overview signals, graded by evidence. Relevance holds up best; formatting, schema and author claims are much weaker.

URL: https://llmagnet.com/blog/ai-overview-ranking-signals  
Author: Ben Cohen, Founder & CEO at LLMagnet  
Published: 2026-09-08  
Updated: 2026-11-10  
Category: Guides

Google hasn’t published which signals decide AI Overview citations, and the best controlled evidence points to relevance to the question rather than formatting. Of seven commonly named signals, only relevance has strong support, and Google says AI Overviews need no special markup or files.

This post grades each signal by the evidence behind it, and gives one thing to do about each.

## Contents

1. How Often AI Overviews Appear
2. What Google Says
3. Signal 1: Relevance to the Exact Question
4. Signal 2: Question Headers
5. Signal 3: Structured Data
6. Signal 4: Named Author
7. Signal 5: External Citations
8. Signal 6: Freshness
9. Signal 7: Topical Depth
10. What Google Says You Can Skip
11. What Changes by Site Type
12. Measuring Your AI Overview Performance
13. FAQ

## How Often AI Overviews Appear

The 48% figure is real, but it isn’t 48% of all Google searches. [BrightEdge](https://brightedge.com/resources/weekly-ai-search-insights/ai-overviews-one-year-presence-size-citing) reported AI Overviews on about 48% of its tracked keywords in February 2026, up from about 31% a year earlier. That is vendor data on a keyword set BrightEdge chooses and doesn’t size.

Other samples give very different rates:

- [Pew Research](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/) found an AI summary on about 18% of 68,879 Google searches made by 900 US adults in March 2025.
- One audit in the [July 2026 GEO survey](https://arxiv.org/abs/2607.14035) found 13.7% across 55,393 trending queries, rising to 64.7% for queries phrased as questions.
- Another study in that survey found 51.5% on its representative query sample.

Pew’s own breakdown shows how much the mix of queries matters. Short searches rarely triggered a summary, and long or question-style searches often did.

![Share of Google searches with an AI summary: question-word searches 60 percent, 10 words or more 53 percent, noun and verb 36 percent, all searches 18 percent, one or two words 8 percent.](https://llmagnet.com/images/blog/charts/aio-pew-query-type.png)

*Source: Pew Research Center, browsing data from March 2025, published July 2025.*

The rate depends on the queries and the date. For your site, the question is how often they appear on the queries you care about. Export your Search Console queries and split them into questions, long phrases and one or two word terms, then check a sample of each. A site built on how-to questions has more at stake than one built on short product names.

## What Google Says

Google’s [documentation on AI features](https://developers.google.com/search/docs/appearance/ai-features) says there are “no additional requirements to appear in AI Overviews or AI Mode.” A page must be indexed and eligible to show with a snippet. Google also says these features may run several related searches across subtopics, which it calls query fan-out.

Citations also reach beyond the top organic results. One audit in the [GEO survey](https://arxiv.org/abs/2607.14035) found that 53% of domains cited by AI Overviews weren’t in the organic top 10, and 27% weren’t in the top 100.

[Ahrefs’ March 2026 study](https://ahrefs.com/blog/ai-overview-citations-top-10) (vendor data) points the same way.

![Where AI Overview URLs rank for the same query: top 10 37.9 percent, positions 11 to 100 31.2 percent, beyond the top 100 31.0 percent.](https://llmagnet.com/images/blog/charts/aio-ahrefs-top-10.png)

*Source: Ahrefs, March 2026. Vendor data. Counts all results blocks; blue links alone gave 37.1% in the top 10.*

The two studies count different units. Ahrefs counts URLs and the audit counts domains, so don’t subtract one from the other. Both say a top-10 rank isn’t required. Ahrefs’ earlier figure of about 76% isn’t a fair baseline, because its sample, its parsing and the Gemini model all changed between studies. Ahrefs suggests fan-out searches explain the shift, but that’s its inference and not something it measured.

Here is how the seven signals compare before the detail.

![Seven AI Overview signals from weakest to strongest evidence: structured data shows no clear effect; topical depth, named author and question headers are untested or weak; freshness depends on intent; external citations are moderate on a benchmark; relevance is strong but indirect.](https://llmagnet.com/images/blog/charts/aio-signal-ladder.png)

*Our grading, based on the sources linked in each section below.*

## Signal 1: Relevance to the Exact Question

**Evidence: strong in controlled tests, indirect for AI Overviews.** The [GEO survey](https://arxiv.org/abs/2607.14035) names relevance as the most reproducible factor. A factorial experiment it reviews, with 252,000 trials across six LLMs, found relevance and position in the model’s context were the main drivers of the first citation.

Placing the answer early on the page is a related but separate idea. [Kevin Indig’s analysis](https://www.searchenginejournal.com/the-science-of-how-ai-pays-attention/561306/) found 44.2% of 18,012 verified citations came from the first 30% of a page, but that was ChatGPT, using vendor data from Gauge.

**Do this:** take your five most important pages and rewrite the first two sentences so each answers the question the page exists for.

The company below is made up for illustration. The question is “Can contractors be paid in another currency?”

| Before | After |
| --- | --- |
| Welcome to Talwen, the modern payroll platform that growing teams love. We make paying people simple, wherever they are. | Yes. Talwen pays contractors in 40 currencies on one pay run, and fixes the exchange rate on the day you approve it. Contractors in sanctioned countries are excluded. |

The first version names no fact a model could use. The second answers in its first word, then gives the number and the limit.

## Signal 2: Question Headers

**Evidence: weak.** We found no study tying question-format headers to AI Overview inclusion. In the experiment above, formatting changes alone had weak effects. Google’s [AI optimization guide](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) says “There’s no requirement to break your content into tiny pieces for AI to better understand it.”

**Do this:** use a question as a header where readers really ask that question. Don’t restructure a page just for this.

## Signal 3: Structured Data

**Evidence: no clear effect in the one controlled test we found.** Google says there is no special schema.org markup you need to add. In a vendor test, [Ahrefs compared 1,885 pages that added JSON-LD](https://ahrefs.com/blog/schema-ai-citations/) with control pages and found AI Overview citations down 4.6% relative to those controls, in a set of pages that were already declining. It can’t say schema caused that. Ahrefs pooled all schema types and tested only pages already heavily cited.

**Do this:** keep your structured data valid and matched to the visible text, for rich results and accuracy. Don’t plan on it to win AI Overview citations.

## Signal 4: Named Author

**Evidence: unproven for AI Overviews.** We couldn’t trace a study showing that AI Overviews prefer named authors, and Google’s guide doesn’t mention authors. It does ask for a “unique point of view,” such as first-hand experience.

**Do this:** put a real name and a short bio on expert content, so readers can check who wrote it. Say what that person has done, not only what they know.

## Signal 5: External Citations

**Evidence: moderate, on a benchmark.** In the [Princeton GEO study](https://arxiv.org/abs/2311.09735) (Aggarwal et al., KDD 2024), Cite Sources gained about 27% on Position-Adjusted Word Count against baseline, on the 1,000-query test split of a 10,000-query benchmark. In the paper’s breakdown by search rank, it helped most at the bottom: +115.1% for sources at rank 5 and -30.3% for those at rank 1. These are benchmark tests, not proof for live engines. The GEO survey says the test is whether the evidence is relevant, verifiable, dated and properly attributed.

**Do this:** link the original source beside each figure on your key pages. Cut any figure you can’t source.

## Signal 6: Freshness

**Evidence: depends on intent, and AI Overviews are the exception in the largest study we found.** [Ahrefs’ analysis of about 17 million citations](https://ahrefs.com/blog/do-ai-assistants-prefer-to-cite-fresh-content/) (vendor data) found AI assistants cited newer pages than organic search on average, but AI Overviews cited content about 16 days older than organic results. The GEO survey reports that recent dates help on time-sensitive queries and not necessarily on stable definitions. Ahrefs also notes that content cited by AI assistants still averages about 2.9 years old, so newness alone doesn’t win.

**Do this:** look at the current results for your query. If most were recently updated, freshness probably matters there. Update the facts and date the change. Don’t change the date alone: Ahrefs points to John Mueller of Google warning against exactly that.

## Signal 7: Topical Depth

**Evidence: untested.** We couldn’t trace a study comparing deep and broad sites in AI Overviews. The reasoning is sound, though: if [query fan-out](https://developers.google.com/search/docs/appearance/ai-features) searches related subtopics, a site that covers those subtopics well has more chances to be retrieved. Google’s guide also asks for content that goes beyond common knowledge.

**Do this:** for your main topic, list the follow-up questions a buyer would ask, and write the pages you’re missing. Write them for readers. Google’s guide says that creating pages for fan-out or query variations mainly to manipulate rankings or AI responses violates its scaled content abuse policy.

## What Google Says You Can Skip

Google’s [guide](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) lists several tactics as unnecessary for its AI features:

- **New machine-readable files.** You don’t need “new machine readable files, AI text files, markup, or Markdown” to appear in Google Search, and that includes llms.txt.
- **Chunking.** You don’t need to split content into small pieces.
- **Length targets.** Google says there is no ideal page length.
- **Inauthentic mentions.** Google names these among the hacks to avoid.

## What Changes by Site Type

The signals apply everywhere, but the first job differs.

- **Ecommerce.** Google’s [AI features page](https://developers.google.com/search/docs/appearance/ai-features) says to check that Merchant Center information is up to date. Do that before you touch copy.
- **Local businesses.** The same page names Business Profile. Keep hours, address and services current there.
- **B2B.** If your buyers’ research questions are long, that is where Pew found summaries most often. Start with pages that answer them directly.
- **Small sites.** With most cited URLs outside the top 10 in the Ahrefs data, you aren’t locked out by rank alone, though that doesn’t show a weak page can win. Give each page one clear question to answer.
- **Large sites.** Fix templates, not single pages. Group the pages in the AI performance report by template and find which templates never appear.

One mistake is worth checking for on every site. Google lists `nosnippet`, `data-nosnippet`, `max-snippet` and `noindex` as ways to limit what appears, and a page must be eligible for a snippet to appear at all. A leftover `nosnippet` tag from an old project can keep a good page out, so check your key templates for one.

## Measuring Your AI Overview Performance

Search Console’s [generative AI performance reports](https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports) launched in June 2026 and reached all websites by August 31, 2026. They show impressions in AI Overviews, AI Mode and Discover’s AI features, by page, country, device and date. Google’s announcement lists no click metric. Google says [traffic from AI features](https://developers.google.com/search/docs/appearance/ai-features) is included in the standard Performance report, under the Web search type.

Expect few clicks. [Pew](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/) found users clicked a traditional result on 8% of visits to pages with an AI summary, against 15% without one, and clicked a link inside the summary on about 1%. Sessions also ended on the results page more often, at 26% against 16%. That data is from March 2025, so judge AI Overviews on impressions and on the clicks you do get, and don’t assume a citation brings a visit.

Open the report and list your top 20 pages by AI impressions, then compare that list with the pages you most want to appear. LLMagnet measures AI visibility across seven engines and has an agent that prepares fixes for your approval. For a view beyond Google, start with the [free LLMagnet audit](https://ai-visibility.llmagnet.com).

## FAQ

### How often do Google AI Overviews appear?

It depends on the sample. BrightEdge reported about 48% of its tracked keywords in February 2026, while Pew found about 18% of real searches by US adults in March 2025.

### What are the seven commonly named AI Overview signals?

Relevance to the question, question headers, structured data, a named author, external citations, freshness and topical depth. Only relevance has strong controlled evidence, and even that isn’t specific to AI Overviews.

### Do AI Overviews only cite pages in the top 10?

No. Ahrefs found 37.9% of AI Overview URLs ranked in the top 10 for the same query, across 4 million URLs in March 2026. One audit in the GEO survey found 53% of cited domains were outside the organic top 10.

### Do I need special markup to appear in AI Overviews?

No. Google says there are no additional requirements. A page must be indexed and eligible to show with a snippet, and no special schema.org markup or AI text file is needed.

### Does schema markup increase AI Overview citations?

It hasn’t been shown to. An Ahrefs test of 1,885 pages found no major uplift on any platform, but it pooled all schema types and covered only pages already heavily cited.

### Does freshness help in AI Overviews?

Not uniformly. Ahrefs found AI Overviews cited content about 16 days older than organic results on average, unlike other AI assistants. Recent dates appear to matter mainly on time-sensitive queries.

### How can I see whether my site appears in AI Overviews?

Use the generative AI performance reports in Search Console, available to all websites since August 31, 2026. They report impressions by page, country, device and date, and Google’s announcement lists no click metric.