How to Get Cited by ChatGPT and Google AI Overviews

Ranking #1 on Google used to be the finish line. Now there's a second race: getting your page picked up as a source inside ChatGPT answers, Google AI Overviews, Perplexity results, and Gemini. Miss that race and you can rank well and still lose the click, because the AI answer already gave the user what they needed, with someone else's brand attached to the sentence that mattered.

This is generative engine optimization, usually shortened to GEO. It overlaps with SEO more than the marketing around it suggests, but the overlap isn't total, and treating the two as identical is how people waste a quarter chasing the wrong fixes.

Why this matters more than it did a year ago

Google has kept widening where AI Overviews show up, and in May 2026 it added two features that make citation status more visible: Preferred Sources, which lets searchers pick the sites they want to see more of and now labels those picks inside AI Overviews and AI Mode, and an expanded Highly Cited label that flags articles other outlets cite most often. Google said people had already selected more than 345,000 unique sources by the time of that announcement (Google). That makes being chosen by your own readers a visibility signal, not just a newsletter-style vanity metric.

That's a meaningful shift. Citation inside an AI answer isn't a side effect of ranking well anymore; it's becoming its own visibility layer, with its own labels and now its own feedback loop in Search Console (Performance report, Search Appearance filter, "AI Overview"). If you've only been tracking blue-link rankings, you're missing half the picture.

None of this replaces classic SEO. A page that isn't indexed, or that Google won't show with a snippet, is never going to get cited, full stop. Google's own documentation is blunt about this: "there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary" beyond meeting the normal technical requirements for Search (Google Search Central). Read our technical SEO checklist first if indexing itself is shaky. GEO is what you layer on top once the basics are solid, not a replacement for them.

How Google AI Overviews actually pick what to cite

Google doesn't publish an exact scoring formula, and anyone who tells you they've reverse-engineered one precisely is overselling a correlation study. But the mechanism Google describes and that independent analysis of AI Overview citations keeps confirming works roughly in two stages.

First, a retrieval stage narrows the open web down using the same relevance and ranking signals that power regular Search. This is why indexing and basic rankability still matter: if you don't make the candidate pool, you can't get cited.

Second, a synthesis stage scores the candidates in that pool on how cleanly a claim can be lifted, attributed to a specific source, and trusted as accurate. This stage rewards things that classic ranking doesn't weigh as heavily: a passage that answers the exact sub-question in one or two self-contained sentences, clear entity attribution (this fact belongs to this organization, written by this person), and signals of real expertise behind the claim.

That second stage explains an odd finding that's shown up across multiple independent citation analyses this year: a meaningful share of AI Overview citations come from pages that don't rank in the top five organic results for the query. Citation selection isn't just a re-ranking of the normal results; it's a separate pass that can reward a well-structured page ranking eighth over a muddled page ranking second.

What AI crawlers actually need from your site

This is the part people get wrong most often, because "AI crawler" isn't one thing. Each major answer engine runs at least two separate bots with different jobs, and you can control them independently in robots.txt.

OpenAI runs three: GPTBot collects content to train future model versions, OAI-SearchBot builds the live index that powers ChatGPT's search and citation features, and ChatGPT-User fetches a page on demand when someone asks ChatGPT to look at a specific URL. These are independent switches. You can block GPTBot to keep your content out of training data while still allowing OAI-SearchBot so you remain eligible to be cited in ChatGPT's search answers, and OpenAI says exactly that: disallowing GPTBot keeps content out of training, while sites opted out of OAI-SearchBot won't show up in ChatGPT search answers at all (OpenAI developer docs).

Perplexity runs a similar split. PerplexityBot is the crawler that builds its search index, and Perplexity recommends allowing it if you want to appear in Perplexity's answers. Perplexity-User is the on-demand fetcher triggered when a person asks Perplexity to visit a specific page, and it generally isn't gated by robots.txt the same way, since a human explicitly asked for that fetch (Perplexity crawler docs).

Google doesn't run a separate "AI Overviews crawler" at all. AI Overviews draw from the standard Googlebot index, so if Googlebot can crawl and index you, you're already eligible on the crawling side; there's no second robots.txt rule to add for Google specifically.

Here's the practical table worth pinning to a wall:

Check your robots.txt for accidental blanket rules. A lot of sites picked up a User-agent: GPTBot / Disallow: / line in 2023 out of training-data caution and never revisited it, which also quietly blocks nothing for OAI-SearchBot (it's a separate rule) but can create confusion if copied sloppily across bots. Audit the file directly rather than assuming last year's decision still matches this year's goal.

The content patterns that actually get lifted

Once you're crawlable and indexed, the difference between getting cited and getting skipped comes down to how extractable your content is. A few patterns show up consistently in what gets pulled.

Answer the question in the first sentence or two of the section, not the fourth paragraph. Models weight the direct statement near the heading more heavily than context that arrives after a long wind-up. If your section starts with "When thinking about X, it's helpful to consider several factors," you've buried the actual answer under throat-clearing that a model has to work around.

Use headings that match real questions, written in plain language. "What is a canonical tag" and "How long should a title tag be" are chunkable. "Let's talk titles" is not. This also happens to be good classic SEO, since it mirrors how people phrase searches.

Make facts self-contained. A statistic, a definition, or a step should still make sense if someone lifts just that sentence and drops it into a different context, because that's exactly what extraction does to it. If a number only means something after three paragraphs of setup, it won't survive being pulled out.

Combine text with original tables, images, or video where it's genuinely useful. Independent analysis of AI Overview citations has repeatedly found that pages mixing prose with original visual or tabular elements get selected noticeably more often than text-only pages covering the same ground. This tracks with how synthesis works: a table is already structured data a model can quote directly.

Keep structured data accurate, not just present. Article, FAQ, and HowTo schema don't force a citation, but invalid or mismatched schema (markup describing something the visible page doesn't actually say) works against you, since Google's structured data guidelines explicitly require markup to reflect what's visible on the page (Google structured data general guidelines). One caveat worth knowing: Google retired FAQ rich results from regular search results in mid-2026, along with the related Search Console reporting (Search Engine Journal). FAQ schema can still help a model parse your page's structure, but don't expect the old FAQ snippet in blue-link results anymore, and don't build a page around FAQ rich results as the goal.

Publish something that isn't already the fifth rewrite of the same five facts. Original data, a first-hand account, a direct quote from someone with standing to make the claim: these are more citable than another summary of what's already published, because corroboration matters here too. If three independent, credible sources already say the same thing in roughly the same way, a model treats that convergence as more trustworthy, but being the only source saying something unusual (without evidence) makes you less likely to get picked up, not more.

A worked example of the rewrite

Here's what the "answer first" fix actually looks like on the page, not just as advice. A typical under-performing section opens like this:

"When thinking about page speed, there are a lot of factors to consider, and it's a topic that comes up constantly in technical audits. Google has talked about this for years, and the tooling has changed a lot over that time, so it's worth understanding where things stand today before making changes."

That's four sentences before any actual information shows up. A model trying to extract an answer to "what counts as good page speed" has nothing to grab here; it's all framing. The fixed version puts the answer first and folds the framing into supporting sentences afterward:

"Good page speed, by Google's current thresholds, means an LCP under 2.5 seconds, an INP under 200 milliseconds, and a CLS under 0.1 on real-user field data. Lab scores from Lighthouse are a useful diagnostic, but Search Console's Core Web Vitals report, which uses actual visitor data, is what determines whether a page passes."

Same topic, same level of technical accuracy, but the second version gives both a human skimmer and an extraction model a complete, quotable answer in the first two sentences. Everything else on the page can now build on that without repeating it.

A step-by-step audit you can actually run this week

1. Confirm indexing and snippet eligibility first. Check Search Console's Page Indexing report for the URL. If it's not indexed, or it's set to noindex, nothing downstream matters yet.
2. Check robots.txt against the current AI crawler list. Confirm GPTBot, OAI-SearchBot, PerplexityBot, and Googlebot are allowed if you want citation eligibility across engines. Decide deliberately, not by inheriting a 2023 blanket block.
3. Rewrite the first two sentences of every H2 and H3 section. State the answer, then support it. Read each section's opening out loud; if the real point doesn't land until sentence three, move it up.
4. Convert vague headings into real questions. Pull actual queries from Search Console's query report or from "People also ask" for your target terms, and use that phrasing.
5. Isolate your key facts. For every statistic or claim, check whether it reads cleanly on its own. If it needs three sentences of setup to make sense, rewrite it as a standalone sentence with the context folded in.
6. Add one genuinely original element per page. A comparison table built from your own research, an annotated screenshot, a short original dataset: something that isn't a rehash.
7. Validate structured data against the visible page, not just against the schema.org spec. Use the Schema Markup Validator for syntax, then manually confirm every field matches what a reader actually sees.
8. Track citations in Search Console. Go to Performance, Search Appearance, and filter for "AI Overview" to see which queries and pages are already getting pulled into AI Overviews. There's no click data in that view yet, only impressions, so treat it as a visibility signal, not a traffic report.

What doesn't move the needle

Keyword-stuffing headings for AI doesn't work, because models match meaning, not strings. An FAQ section that just repeats the main content in question form adds nothing either; if it's not a distinct, extractable answer, it's dead weight on the page. And chasing every answer engine with equal effort is usually a mistake: Google AI Overviews reach the widest audience for most sites, so get that right first, then extend the same structural fixes to ChatGPT and Perplexity rather than building separate content for each.

If you want a deeper look at a single engine, we've got dedicated pieces on getting cited by Claude and getting cited in Perplexity.

The honest tradeoff on traffic

AI Overviews can answer a query completely enough that the user never clicks through, even when your page is the cited source. That's a structural change in how informational search works, not a problem you optimize your way out of entirely. Preferred Sources and Highly Cited labels add a layer of brand visibility on top of that even when the click doesn't happen: being named as the source in front of millions of AI Overview impressions is worth something, even at a zero-click rate, if the goal is brand recall and trust rather than direct traffic.

The practical response isn't to abandon informational content. It's to be honest about which pages are traffic plays and which are visibility plays. Comparison pages, tool pages, and anything with transactional intent still drive clicks even inside an AI-heavy results page, because the AI answer usually can't fully replace the decision those pages help someone make. Purely informational explainer content is increasingly a brand-visibility play: worth doing well, worth optimizing for citation, but budgeted and measured differently than a page you expect to convert. For more on how the overall traffic picture is shifting, see our breakdown of how Google keeps expanding AI Overviews.

Being cited as a source and being recommended as a product are related but different. For "best X for Y" prompts, read our SEO tactics to get your website recommended by ChatGPT; for product listings, see the ChatGPT merchant setup guide.

If your pages are built from generic divs, the content is harder to extract than it needs to be. See our guide to semantic HTML for SEO and AI answers for the elements worth fixing first.

FAQ

Do I need separate content for ChatGPT, Google AI Overviews, and Perplexity?

No. The structural changes that help one (direct answers, real headings, self-contained facts) help all of them, since they're all doing some version of retrieval plus extraction. The only engine-specific work is in robots.txt, where you decide which crawler gets access.

Does adding FAQ schema guarantee I'll get cited?

No. FAQ schema can help a model parse structure, but it doesn't force a citation, and Google removed FAQ rich results from standard search results in 2026, along with the related reporting. Treat FAQ sections as genuinely useful content for readers and models, not as a shortcut to a guaranteed result.

Can I block AI training crawlers without losing citation eligibility?

Yes, for OpenAI specifically. Blocking GPTBot keeps your content out of model training while leaving OAI-SearchBot free to crawl for ChatGPT's search and citation features, since they're separate crawlers controlled independently in robots.txt.

How do I know if I'm already getting cited in Google AI Overviews?

Check Search Console's Performance report, filter by Search Appearance, and select "AI Overview." It shows impressions for queries where your page was cited, broken out by page, country, and device, though it doesn't yet report clicks separately from blue-link traffic.

Is GEO just a rebrand of SEO?

Mostly no, though they share a foundation. Both reward crawlability, indexing, and real expertise. GEO adds a specific emphasis on passage-level extractability and multi-engine crawler management that classic SEO never had to think about, because classic SEO only had to satisfy one ranking algorithm's retrieval step, not a second synthesis step on top of it.

If you're building out a broader content plan around this shift, our GEO for beginners piece is a good next stop, and the free AI SEO tools roundup covers ways to check citation status without paying for an enterprise platform.