AI Overviews and AI answers: what actually gets cited

Getting cited in AI Overviews, AI Mode, ChatGPT search, Perplexity and Copilot takes the same work as ranking in ordinary search. Google says no special optimization is needed. What raises your odds is a page that can be crawled and indexed, answers the question near the top, cites its sources, and offers something other pages don’t.

That is less exciting than a new acronym, but it is what the primary sources say. This guide walks through what Google, Microsoft and the best-known research paper actually found, what the AI search engines need from your site technically, and which popular tactics you can safely ignore.

What Google says about AI Overviews and AI Mode

Google’s documentation is unusually direct. Its page on AI features and your website says: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” Existing SEO best practices still apply.

The same page sets one hard requirement. To be shown as a supporting link, “a page must be indexed and eligible to be shown in Google Search with a snippet.” If a page can’t appear as a normal result with a snippet, it can’t be a source in an AI answer either.

Google lists the practices that matter:

• Allow crawling in robots.txt.
• Make content easy to find through internal links on your site.
• Put important content in text, not only in images or video.
• Provide a good page experience.
• Make sure structured data matches the visible text on the page.

It also explains why the sources in an AI answer differ from the top ten links. AI Overviews and AI Mode may use “query fan-out”: the model runs several related searches across subtopics and then builds one response from what it found. A page can be cited for a subtopic it answers well, even if it doesn’t rank for the main query.

Why the top ten and the citations no longer match

Fan-out has a measurable effect. Ahrefs found that 37.9% of URLs cited in AI Overviews also appear in the top 10 for the same query, across 863,000 keyword results and 4 million cited URLs. Its earlier study, from July 2025, put that share at about 76%.

The practical reading: a page that answers one narrow question very well can earn citations from many broader queries. That rewards depth on specific subtopics over one thin page per keyword.

Non-commodity content is Google’s main advice

Google’s newer guide, Optimizing your website for generative AI features, adds a useful distinction. Commodity content “is often based on common knowledge, which could originate from anyone.” Non-commodity content “provides unique expert or experienced takes that go beyond common knowledge and the ordinary.”

AI systems can already summarize common knowledge from a hundred pages. They have a reason to cite you when you add something those pages lack:

• First-hand experience. What happened when you actually did the thing.
• Original data. Your own prices, measurements, survey results or counts.
• A point of view. A clear recommendation, with the reasoning behind it.
• Specifics. Names, dates, steps and edge cases that a generic summary skips.

John Mueller’s post on succeeding in Google’s AI experiences gives the same advice first: focus on “making unique, non-commodity content.” It also claims that clicks from AI Overviews “are higher quality, where users are more likely to spend more time on the site.” Google offers no figures for that, so treat it as Google’s claim and measure conversions on your own site rather than raw clicks.

What the GEO study measured, and what re-tests found

The best-known research on this is “GEO: Generative Engine Optimization” by Aggarwal and colleagues, published at KDD 2024. The authors rewrote source pages in different ways and measured how much each page showed up in generated answers. They tested a custom engine built on GPT-3.5 and the live Perplexity engine, across about 10,000 queries.

Change to the page	Effect on visibility in AI answers
Adding quotations from relevant sources	Largest gain, about 41%
Adding statistics	About 33%
Improving fluency	About 29%
Citing sources	About 28%
Making text easier to understand	About 14%
Keyword stuffing	No gain; about 9% worse than the unedited page

The figures are relative gains on the paper’s main measure, which weights how many of a source’s words appear in the answer and how early. The authors sum up the top methods as gaining 30% to 40% on that measure. On Perplexity, adding quotations still gave about 22% and adding statistics up to 37% on a second measure.

Later work has not held up those numbers as a general rule. The GEO gains were measured on pages already handed to the model, so they say nothing about whether a page gets found in the first place. C-SEO Bench, presented at NeurIPS 2025, re-tested conversational SEO methods and found most of them “largely ineffective”, and often harmful to how a document ranks. Traditional SEO, which helps a page get retrieved at all, did much better. A 2026 survey of GEO research reached the same view: the original gains are valid in their setting but don’t establish discoverability or lasting traffic, and rewriting pages to court citations can hurt retrieval.

So don’t rewrite pages for AI, and don’t expect a fixed lift from adding statistics. Source your figures because readers check them and engines quote facts they can verify. Inventing a number to look authoritative is the fastest way to lose trust. The parts that do hold up match what Google and Microsoft say: be relevant to the question, and keyword tricks don’t help.

What Microsoft says for Bing and Copilot

Microsoft published its own guidance on AI search answers in October 2025. It is more tactical than Google’s and fits the same pattern:

• Keep the title, the H1 and the meta description consistent with each other.
• Answer questions with “one- to two-sentence responses that directly address a question.”
• Use specific H2 and H3 headings, not vague labels.
• Use numbered steps and comparison tables for how-to content and comparisons.
• Avoid long blocks of text.
• Don’t hide key answers in tabs or expandable menus, since AI systems may not render them.
• Don’t put core information only in PDFs or images.

Each of these makes a page easier for a person to scan. Making it easier for a machine to quote is a side effect.

What ChatGPT search and Perplexity need from your site

OpenAI and Perplexity publish less about ranking, but they are clear about access. Each uses a dedicated crawler for search, separate from any training crawler.

OpenAI says that “OAI-SearchBot is for search,” and that sites opted out of OAI-SearchBot “will not be shown in ChatGPT search answers.” Its settings are independent, so you can allow OAI-SearchBot while blocking GPTBot, the crawler used for model training.

Perplexity says its crawler, PerplexityBot, is designed to surface and link websites in its search results, and is not used to crawl content for AI foundation models.

So the first check is simple. Open your robots.txt and confirm you aren’t blocking the search crawlers you want to be cited by. Some sites blocked every AI user agent when these crawlers first appeared, and never revisited the decision.

Tactics you can safely skip

Google’s optimization guide names several popular ideas and dismisses them:

1. llms.txt and AI-only files. Google says creating llms.txt files “will neither harm nor help your site’s visibility or rankings in Google Search, as Google Search ignores them.” Our piece on whether llms.txt is worth adding covers where it does help.
2. Chunking. “There’s no requirement to break your content into tiny pieces for AI to better understand it.” Google also says there is no ideal page length.
3. Rewriting for AI. “You don’t need to write in a specific way just for generative AI search.”
4. A page for every query variant. Creating separate content for each way people might search, mainly to manipulate rankings or AI answers, violates Google’s scaled content abuse policy.
5. Buying or chasing mentions. Google says seeking inauthentic mentions “isn’t as helpful as it might seem.”
6. Structured data as a ticket in. It “isn’t required for generative AI search,” though it still helps with rich results.

Google adds that no third-party tool has access to its internal ranking or AI systems. Be skeptical of any tool that claims to know your “AI visibility score” from the inside.

A checklist for each article

Put the sources together and you get a short list you can apply to every page. It works for search and AI answers at once.

1. Confirm access. The page is indexable, eligible for a snippet, and not blocked for Googlebot, Bingbot, OAI-SearchBot or PerplexityBot.
2. Answer first. Open with a direct answer of 40 to 60 words that names the topic. That range is our convention, not a published rule; Microsoft’s guidance is one or two sentences.
3. Use descriptive headings. Each H2 says what the section answers. Use a question only when it matches what people actually search.
4. Pick the right format. A table for comparisons, numbered steps for procedures, and prose for reasoning.
5. Source every figure. Link each statistic and quotation to its primary source, inline. Our guide to fact-checking AI-written drafts shows a claim-by-claim method.
6. Add something new. Include at least one thing a summary of the top results wouldn’t have: your data, your experience or your recommendation.
7. Link it in. Link to the new page from at least one existing page, and link out from it to related pages. These internal linking rules give concrete numbers.
8. Keep markup honest. Structured data describes what is visible on the page and nothing else.
9. Measure what you can. Watch queries and pages in Search Console and Bing Webmaster Tools, and judge AI traffic by conversions rather than clicks.

How Wavefront applies this research

Wavefront writes articles to this checklist, because the research above is what its pipeline is built on. Each draft opens with a short answer, cites a source for every figure, and links to pages from the site’s own sitemap. An independent fact check reviews each claim against the evidence before a person sees the draft. It generates nothing written only for machines.

The article writer uses the workspace’s own facts to supply the non-commodity part, and it never invents experience. The strategy planner chooses subjects from real search data once you connect Google Search Console. The performance view then shows how published pages do in search. It can’t promise citations, and neither can anyone else: Google and Microsoft describe these practices as improving your odds, not guaranteeing a place in the answer, and the research agrees no rewrite does.

Sources

1. Google Search Central: AI features and your website
2. Google: Optimizing your website for generative AI features on Google Search
3. Google Search Central Blog: Top ways to ensure your content performs well in Google's AI experiences on Search (May 21, 2025)
4. Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024)
5. Puerto et al., C-SEO Bench: Does Conversational SEO Work? (NeurIPS 2025)
6. Martinez, Optimizing Visibility in Generative Engines: A Critical Survey of GEO (2026)
7. Microsoft Advertising: Optimizing your content for inclusion in AI search answers (October 8, 2025)
8. Ahrefs: Update: 38% of AI Overview citations pull from the top 10 (March 2, 2026)
9. OpenAI: Overview of OpenAI crawlers
10. Perplexity: Perplexity crawlers