You’re ranking on page one, and your traffic is still dropping, because Google answered the question itself in an AI Overview before anyone reached your listing. I explained why that happens in a separate post on what answer engine optimization actually covers, and the generative-engine side of that same shift is its own topic entirely; this post is the tactical follow-up, what I’d actually change on a page to get cited inside AI Overviews and ChatGPT responses, based on real study data rather than the vendor checklists currently flooding this space.
What Actually Works: The Real Data, Not the Vendor Pitch
A lot of “AI SEO” advice right now is unfalsifiable, add this tag, follow this framework, trust us. I’d rather work from studies that published their methodology. Two are doing most of the load-bearing work in this post: Search Engine Land’s AI Overviews optimization guide, built from an analysis of real AI Overview citations, and Ahrefs’ study of 1.4 million ChatGPT prompts tracking exactly which retrieved pages actually got cited versus ignored.
Both point to the same underlying truth: ranking fundamentals still matter enormously, but a specific, narrower set of content and formatting choices separates a page that ranks from a page that also gets cited.
I’ll refer back to both studies by name throughout rather than re-linking them every time, since it’s worth being clear which claim comes from which dataset, especially in the sections where the two datasets don’t fully agree.
Get the Fundamentals Right First
Before any AI-specific tactic, the data says to fix ordinary SEO first. Search Engine Land’s guide reports that somewhere between 40% and 76% of AI Overview citations also appear in the traditional top 10 organic results for the same query, which means a page that can’t rank reasonably well in the first place is competing from a real disadvantage before a single AI-specific tactic gets applied. The guide also flags Core Web Vitals and HTTPS as baseline technical requirements Google explicitly checks alongside everything else.
Worth noting on keyword selection specifically: the same guide found that roughly 72% of keywords that currently trigger an AI Overview carry a CPC of $0.10 or less, meaning the queries most likely to produce an AI Overview skew toward lower commercial intent, informational, exploratory, often three to five words long.
If a page is built entirely around a high-CPC transactional keyword, don’t expect an AI Overview to be the main opportunity there, that’s still a ranking and conversion play, not a citation play.
Manage Expectations Against the Top Cited Domains
Search Engine Land’s guide also lists the top five domains actually cited in Google AI Overviews across its dataset: YouTube at 23.29%, Wikipedia at 18.41%, Google’s own properties at 16.38%, Reddit at 9.37%, and LinkedIn at 8.80%. Those five alone account for a large share of total citation volume, which means an independent site is realistically competing for the remaining space, not an even playing field against every domain on the web.
That’s not a reason to skip this work, on the international, less-saturated markets I handle for iGaming and Fintech clients, the kind of work I write about across this site, the competition for that remaining space is often thin, but it’s a reason to set expectations honestly on a highly contested English-language head term before promising a client citations there.
Structure Content So an AI System Can Actually Lift It
This is the part that’s genuinely actionable at the page level. Google’s own query fan-out mechanism, which I described in the GEO post, means a page needs to answer more than the one headline question, it needs to cover the cluster of sub-questions a generative system is likely to generate around that topic. In practice that means writing the kind of thorough H2/H3 breakdown that naturally answers “what is it,” “how does it work,” “how much does it cost,” and “what are the alternatives” inside one page, rather than splitting each into a separate thin post.
In practice, I find the fastest way to identify that sub-question cluster isn’t a fancy tool, it’s Google’s own People Also Ask box for the target query, plus whatever your site’s internal search or Search Console’s query report already shows people asking around that same topic, both are a direct window into the exact fan-out territory you’re trying to cover.
Front-Load the Answer, Every Time
Open each section with the direct answer, then explain it, never the reverse. A generative system extracting a passage doesn’t read the whole paragraph looking for your conclusion, it favors the sentence that states the answer plainly. Search Engine Land’s guide specifically recommends using semantic triples, a clear subject-predicate-object structure, while still keeping the actual sentence readable rather than robotic. “RankMath’s redirect manager logs the destination URL automatically” is a semantic triple that’s also just a normal, well-written sentence, and it’s exactly the kind of statement that gets lifted cleanly.
A Worked Example
Take a page targeting “how to choose an SEO plugin for WordPress.” A keyword-first version answers only that headline question, then stops. A fan-out-aware version anticipates the sub-questions a generative system is likely to spin off the same topic: which plugin is free versus paid, does the plugin conflict with page builders, what happens to redirects if you switch plugins later, and how does the schema output differ between them.
Covering four or five of those adjacent sub-questions on one well-organized page gives a generative system several separate, citable passages instead of just one, multiplying the number of ways that page can get pulled into an answer. I used exactly this structure when I sat down to compare Yoast against RankMath feature by feature, and it’s the model I’d replicate for any new page built with AI citations in mind.
Citations, Quotes, and Stats Move the Needle More Than Schema
Here’s the tactic with the best evidence behind it and the least hype around it: Search Engine Land’s analysis found that including citations, quotations from relevant sources, and statistics can boost a page’s source visibility in AI Overviews by more than 40%. That’s a bigger, more consistent lever than most of the technical tweaks getting sold as “AI SEO” right now.
Compare that against structured data, which I covered in more detail in the GEO post, adding JSON-LD schema alone showed no measurable citation uplift in Ahrefs’ 2026 study, and even a slight decline on Google AI Overviews specifically.
The practical implication is to spend the effort on sourcing real numbers and quoting real authorities in the text itself, not on markup that formats an answer you haven’t actually backed with anything checkable yet. Schema is still worth implementing for the reasons I laid out separately, it just isn’t where this particular lever lives.
What Actually Counts as a Real Citation in the Text
“Add statistics” gets thrown around as advice without specifying what qualifies, so to be concrete: a real citation is a specific number, attributed to a named, checkable source, that a reader could go verify themselves, your own Search Console or analytics data if you’re citing your own results, a named industry report, a government or trade-body dataset, or a study from a publication like the two behind this post.
It is not a vague “studies show” line, and it isn’t a number you half-remember from a conference talk two years ago. If you can’t link to where a number came from, don’t use it, both for the reader’s trust and because an unlinked, unverifiable claim is exactly the kind of thing that damages E-E-A-T signals rather than helping them.
Platform Differences: Google Isn’t ChatGPT
Treating “AI search” as one target is a mistake the data doesn’t support. Ahrefs’ 1.4 million-prompt study found that pages with natural-language URL slugs got cited 89.78% of the time versus 81.11% for pages without one, a gap specific to how ChatGPT’s retrieval step matches a page’s URL and title against its own generated sub-queries; a slug like /how-ai-overviews-work aligns with that matching far better than a numeric or keyword-stuffed one.
The same study found cited pages’ titles were meaningfully closer, by cosine similarity, to ChatGPT’s internal fan-out queries than non-cited pages, 0.656 versus a noticeably lower score for pages that didn’t get picked, which reinforces the same lesson as the fan-out section above from a different angle: title and content should match the underlying question cluster, not just the headline keyword.
Where the Two Platforms Actually Diverge
| Signal | Google AI Overviews | ChatGPT |
|---|---|---|
| Adding JSON-LD schema | No uplift; slight decline (Ahrefs, 2026) | No measurable effect (Ahrefs, 2026) |
| Natural-language URL slugs | Not isolated in this dataset | 89.78% vs 81.11% citation rate (Ahrefs) |
| Overlap with top-10 organic rankings | 40-76% of citations also rank top 10 (Search Engine Land) | Search-sourced citations dominate at 88.46% (Ahrefs) |
| Content freshness | Roughly 44% of citations came from content published that same year (Search Engine Land) | Cited pages skew older, median around 500 days, likely reflecting accumulated authority (Ahrefs) |
That freshness row is worth sitting with. It looks contradictory until you separate the two mechanisms: Google’s AI Overviews seem to reward recency more directly, while ChatGPT’s citations in this dataset lean toward pages that have had time to accumulate backlinks, mentions, and general authority.
Neither number is wrong, they’re measuring different systems with different retrieval logic, which is exactly why I wouldn’t run one universal “AI SEO” playbook across both. If a page is aimed specifically at ChatGPT visibility, I’d weight the URL-slug and title-alignment work higher; if it’s aimed at Google’s AI Overviews, I’d weight publishing cadence and update frequency higher instead, since that’s the lever the freshness gap actually points to.
How to Tell Whether Any of This Is Working
Google Search Console’s Performance report, filtered to the “Web” search type, is the closest thing to first-party data on how a page performs specifically inside AI Overviews and AI Mode, and it’s worth checking before and after making any of the changes above rather than guessing at impact.
Neither ChatGPT, Perplexity, nor Claude report citation data back to Search Console at all, so tracking those requires either a paid citation-monitoring tool or a manual habit: run your target prompts through each platform yourself on a fixed schedule and log whether your page shows up.
It’s tedious, but for the international, less-saturated markets where most of the work I do lives, a manual monthly check across ten or fifteen core prompts is genuinely enough to spot a trend, you don’t need enterprise tooling to notice whether a change moved anything.
Rank-tracking tools that flag AI Overview presence in the SERP are the other practical signal worth watching, most major platforms added this over the past year specifically because so many site owners were asking for it. Pair that SERP-level view with the Search Console data above and you get a reasonably complete picture of the Google side even without a dedicated GEO dashboard, which, as I mentioned in the GEO post, still doesn’t exist as a single unified tool.
A Practical Checklist
Everything above compresses into six checks, and I’d run them in this order rather than tackling whichever feels easiest first. Skipping the ranking check to jump straight to citations and statistics is the most common mistake I see, because the citation tactics only pay off once the page can already compete for a top-10 position in the first place.
- Confirm it already ranks reasonably well organically, or fix that first, before anything else on this list.
- Rewrite the opening sentence of each major section to state the answer directly, not build up to it.
- Map the realistic sub-questions a reader in this topic would ask next, and make sure the page actually answers at least three or four of them, not just the title’s headline question.
- Add at least one real statistic or quoted source per major section, with a link to where it actually came from.
- Check the URL slug reads as a natural phrase, not a keyword string or an ID.
- Leave schema in place if it’s already there for its real benefits, but don’t treat adding it as the fix if citations are the actual goal.
What I’d Actually Prioritize This Quarter
If I only had time for one change, it’s the citations-and-statistics tactic, the evidence behind it is the strongest of everything covered here, and it’s the one most sites are still skipping in favor of markup tweaks. Ranking in AI Overviews and getting cited by ChatGPT both reward the same underlying habit: say something specific, back it with something checkable, and structure the page so a system built to extract exactly that kind of sentence can find it fast.
What I’d deliberately avoid doing is chasing every new tactic a vendor announces the moment a platform update ships. Both studies behind this post took months to compile and still only cover two platforms out of the several that now matter, Perplexity and Copilot aren’t in either dataset, which means some of this will need revisiting as more independent data comes out rather than treated as settled forever.
That’s not a new discipline bolted onto SEO, it’s the same standard I’ve been applying to the international client sites I work on for years, just with a more literal payoff now that something is actually reading the page and deciding whether to quote it.
Dimitar Kondev is an SEO expert working on the International iGaming market with more than 6 years of experience.