80% of what you read about AI SEO is about citations.
“Build citations, get mentioned on the pages ChatGPT cites. Our agency generated a 300% citation boost for our clients.”
Are citations the most important thing to look at? Yes and no. As the adage goes, it depends. But understanding the kind of pages ChatGPT mentions is undeniably worthwhile (as long as you know why you’re doing it, and what you’re looking for).
Citations do play a major role for me when I’m doing AI SEO. It’s simple:
- You want your brand to be mentioned in responses to relevant prompts
- Citations tell you which pages the AI is looking at to get (or support) its data
These citations give us an idea of how ChatGPT/Google etc gather info and make their decisions. Not the whole picture - not even close. But it’s about the best we have to work with.
It’s the 80/20 approach to AI SEO: look at what gets cited, use that information to guide how you build mentions for your brand.
Far from perfect, but there is no perfect AI SEO strategy. Sorry.
The data
Preamble over. Here’s what I brought you here for.
I was putting together a report for our AI SEO, looking into the kinds of sources getting used the most for the prompts we’re tracking.
I use Peec AI for this. There are a hundred AI SEO tracking tools out there, pretty much all provide the same information. But I find Peec does the job very well for me, and has basically all the info I need, easy to find (exception: tracking for Claude prompts. But this seems to be the kind of thing that costs hundreds per month extra, wherever you go).
With their MCP (another reason to use Peec), I was able to pull the citations used across the 50 or so prompts we’ve been tracking on ChatGPT, Google AI Overviews and Google AI Mode. We’ve been working on building mentions online, but it’s still a bit scattergun. I was hoping to get a bit more clarity on the kind of sites, pages and sources AI likes to use in our industry.
The data came to 52,303 citations, pointing at 5,468 pages on 1,419 websites.
It was also very interesting - the kind of thing I’d love to see come across my feed - so I figured, why not share it here as well.
The wrong take on this data
Before I get into the data, here’s what it isn’t.
This data is not a comprehensive breakdown of what AI cites the most. “Here’s what AI actually cites”.
That’s (sorry mum) bullshit.
Too many people parrot this kind of thing online. If I’m giving them the benefit of the doubt, then they just don’t properly grasp the concept. If I’m not, I’d say they’re spewing rubbish for engagement.
LLMs are going to use different types of sources for different types of prompts, in different industries.
The citations for someone asking the best CRM for AI visibility agencies will likely not have any relevance for someone asking how to bake a blueberry pie, or which of Emily St John Mandel’s books are the best (it’s Sea of Tranquility).
Even in the same industry, change the wording of the prompt and you’ll probably see vastly different citations. “What is the best X” vs “How many people use X”, for example.
So, this data covers the specific prompts I’ve been tracking (largely commercial, “what’s the best X” prompts), in my specific industry (ecommerce SaaS, or SaaS-adjacent).
That might be useful to you. It might not be (but hopefully still interesting). Just don’t take it as the ultimate answer to which pages “win” AI search.
The short version of the data
- Listicles were the most-cited page type: 26% of all citations, from 14% of the pages.
- Category vendors took 45% of all citations. These are the websites of companies selling in the category. 40% of their citations went to their own listicles and comparison pages.
- 80% of listicle citations went to lists published by businesses: a vendor in the category (48%) or another company, such as an agency or a tool in a neighboring category (32%). Independent review sites and media got 16%.
- Pricing pages had the most citations per page: 51 on average, against 9.6 across all pages.
- ChatGPT and Google's AI cited different kinds of pages. 64% of ChatGPT's citations went to category vendors, and 0.2% to forums, social and video. In Google AI Overviews and AI Mode, forums, social and video made up 17%.
- Regular blog posts were rarely cited: 6% of cited pages, 2% of citations.
- A year in the title went with more citations for listicles only. Listicles with the current year in the title averaged 23 citations each, against 11 without.
- Citations were concentrated. The top 85 pages, 1.6% of the total, took 42% of all citations.
1. Citations by page type
| Page type | Pages | Citations | Overall Share | ChatGPT | AI Overviews | AI Mode |
|---|---|---|---|---|---|---|
| Listicles ("best X" roundups) | 766 | 13,380 | 25.6% | 17.4% | 31.8% | 25.9% |
| Product & feature pages | 887 | 7,278 | 13.9% | 23.1% | 9.0% | 10.8% |
| Marketplace listings | 447 | 4,313 | 8.2% | 10.0% | 7.1% | 8.0% |
| Docs & help center | 644 | 3,490 | 6.7% | 14.2% | 2.7% | 4.0% |
| How-to guides | 573 | 3,249 | 6.2% | 2.2% | 9.5% | 6.0% |
| Pricing pages | 56 | 2,847 | 5.4% | 12.5% | 1.3% | 3.5% |
| Head-to-head comparisons | 291 | 2,791 | 5.3% | 3.3% | 5.3% | 7.7% |
| Forum threads | 240 | 2,713 | 5.2% | 0.1% | 6.5% | 8.9% |
| Alternatives lists | 133 | 2,254 | 4.3% | 3.2% | 5.3% | 4.1% |
| Videos | 256 | 2,182 | 4.2% | 0.0% | 7.0% | 4.8% |
| Homepages | 239 | 2,085 | 4.0% | 6.2% | 2.7% | 3.3% |
| Regular blog posts | 348 | 1,205 | 2.3% | 1.7% | 2.9% | 2.1% |
| Marketplace category pages | 135 | 1,145 | 2.2% | 2.1% | 1.7% | 2.9% |
| LinkedIn articles | 40 | 854 | 1.6% | 0.1% | 2.3% | 2.4% |
| Everything else | 413 | 2,517 | 4.8% | 3.9% | 4.9% | 5.6% |
"Best X" roundups were 25.6% of all citations. With "X alternatives" lists (4.3%), head-to-head "X vs Y" pages (5.3%) and single-product reviews (0.8%), comparison content made up 36% of all citations.
This makes sense - as most of the tracked prompts were some variations of “what’s the best…” “give me some alternatives to…”
Next came product and feature pages (13.9%), app and integration marketplace listings (8.2%), and docs and help center articles (6.7%).
Notable - a big split at the top between ChatGPT and AI Overviews. AIOs (and AI mode, to a lesser extent) prefer listicles a lot more, while GPT prefers product/feature pages.
2. Citations by source type
| Source type | Sites | Pages | Citations | Share | ChatGPT | AI Overviews | AI Mode |
|---|---|---|---|---|---|---|---|
| Category vendors | 231 | 1,619 | 23,798 | 45.5% | 64.4% | 36.0% | 38.2% |
| Other companies’ sites | 846 | 1,530 | 9,321 | 17.8% | 9.9% | 23.3% | 18.9% |
| Marketplaces & app stores | 13 | 582 | 5,458 | 10.4% | 12.0% | 8.8% | 10.9% |
| Social & video | 24 | 437 | 3,655 | 7.0% | 0.1% | 11.2% | 8.6% |
| Forums & communities | 18 | 247 | 2,953 | 5.6% | 0.2% | 7.1% | 9.6% |
| Platform & documentation sites | 62 | 577 | 2,927 | 5.6% | 10.5% | 3.1% | 3.7% |
| Review sites & directories | 116 | 300 | 2,539 | 4.9% | 1.1% | 6.5% | 6.6% |
| News & media | 51 | 71 | 1,139 | 2.2% | 0.8% | 3.1% | 2.4% |
| Other | 74 | 105 | 513 | 1.0% | 1.0% | 0.9% | 1.0% |
Here’s what these source types mean:
- Category vendors: the websites of companies that sell in the category, including my company's.
- Other companies' sites: agencies, consultancies and software companies in neighboring categories.
- Marketplaces & app stores: app stores and integration marketplaces, including their category pages.
- Social & video: YouTube, LinkedIn, Facebook, Instagram and blogging platforms like Medium.
- Forums & communities: Reddit, Quora and community forums.
- Platform & documentation sites: the platforms the category's products run on, and big-tech developer docs.
- Review sites & directories: independent review and comparison sites, and software directories.
- News & media: news sites, trade publications and press releases.
45.5% of all citations came from category vendors. These are the websites of the companies selling in the category: my company, its competitors and every other vendor in the space.
Other companies' sites came second with 17.8%: agencies, and software companies in neighboring categories writing about the category for their own audiences. Then marketplaces and app stores (10.4%), social and video (7.0%), forums and communities (5.6%) and platform and documentation sites (5.6%). Independent review sites and directories got 4.9%, and news and media 2.2%.
Here's how the citations to category vendors split by page type:
| Page type on vendors’ sites | Pages | Citations | Share of category vendor citations |
|---|---|---|---|
| Product, pricing & homepages | 641 | 10,778 | 45.3% |
| Their own listicles & comparisons | 399 | 9,526 | 40.0% |
| Blog posts & news | 366 | 2,332 | 9.8% |
| Docs & help centers | 192 | 1,086 | 4.6% |
| Other | 21 | 76 | 0.3% |
Now, read this before you go and post anything about this data, because it comes with a big caveat.
When this says “category vendors”, it means companies selling in the category the prompt is about - for example, when the prompt is “what’s the best ecommerce CRM”, and the citation comes from an ecommerce CRM tool.
It can also mean homepages, product pages, pricing pages of companies in the category.
With the listicles, I’ve noticed that a lot of these citations don’t appear to be supporting the vendor themselves showing up in the response.
Just from what I see: a lot of self-promotional listicles get cited, often from small or new vendors, but the company itself is not mentioned in the answer.
So it doesn’t seem to be a case where listicles from category vendors are a clear winner in AI. I’d say it’s more a logical correlation - vendors in this category are obviously going to publish content about the category, and thus more likely to show up as citations.
I don’t think there’s a negative causation for promotional listicles either; I highly doubt GPT/Google’s systems are actively excluding the listicle publisher from the response (though you will see a lot of SEO “experts” claiming this online, with nothing solid to back it up).
The takeaway? There’s no takeaway. Other than that listicles aren’t dead.
3. Where the listicles come from
| Published by | Pages | Citations | Share | ChatGPT | AI Overviews | AI Mode |
|---|---|---|---|---|---|---|
| Category vendors | 289 | 7,555 | 48.3% | 74.5% | 42.9% | 38.1% |
| Other companies’ sites | 414 | 5,028 | 32.2% | 17.3% | 35.9% | 36.7% |
| Review sites & directories | 129 | 1,506 | 9.6% | 2.3% | 10.7% | 13.3% |
| News & media | 16 | 946 | 6.1% | 3.6% | 6.9% | 6.3% |
| Other | 11 | 239 | 1.5% | 1.4% | 1.0% | 2.6% |
| Forums & communities | 4 | 237 | 1.5% | 0.2% | 1.6% | 2.4% |
| Social & video | 19 | 77 | 0.5% | 0.1% | 0.8% | 0.4% |
| Platform & documentation sites | 17 | 46 | 0.3% | 0.6% | 0.2% | 0.2% |
Digging deeper into the listicle part. The table above shows where most of the listicles come from.
The majority, across all models, come from category vendors. This challenges the idea that you need “third-party” listicle mentions, or listicles in “trusted publications”.
To be fair, I don’t think ChatGPT or Google really checks, or cares, whether a listicle comes from someone in that category or an “independent” publisher.
There is a notable gap between GPT and Google here, though, on what kind of sites show up most often. Nearly 75% of GPT’s listicles come from category vendors, whereas Google has a much more even split between category sites and independent sites - as well as a lot more from general news & media sites, review sites and directories.
4. Citations per page
| Page type | Pages | Citations per page |
|---|---|---|
| Pricing pages | 56 | 50.8 |
| Listicles ("best X" roundups) | 766 | 17.5 |
| Alternatives lists | 133 | 16.9 |
| Forum threads | 240 | 11.3 |
| Marketplace listings | 447 | 9.6 |
| Head-to-head comparisons | 291 | 9.6 |
| Homepages | 239 | 8.7 |
| Videos | 256 | 8.5 |
| Marketplace category pages | 135 | 8.5 |
| Product & feature pages | 887 | 8.2 |
| Review-site profiles | 84 | 7.7 |
| News & press | 64 | 5.7 |
| How-to guides | 573 | 5.7 |
| Docs & help center | 644 | 5.4 |
| Social posts | 89 | 5.2 |
| Case studies | 75 | 4.1 |
| Regular blog posts | 348 | 3.5 |
| All pages | 5,468 | 9.6 |
This data shows which type of pages got cited repeatedly, across numerous prompts. The clear winner: pricing pages.
56 pricing pages were cited 2,847 times: 51 citations per page, against 9.6 across all pages. Listicles (17.5), alternatives lists (16.9) and forum threads (11.3) were also above the average.
At the bottom were regular blog posts (3.5 per page), case studies (4.1), docs (5.4) and how-to guides (5.7).
This makes sense - again, before you go away and tell your CEO “We need to publish more pricing pages”, think about what the data means.
Most sites just have one pricing page. If someone is asking for product recommendations, it makes sense for the AI to provide pricing information. The best place to get that pricing information is the vendor’s pricing page.
So, obviously, that page would show up in citations across a number of different prompts, especially for category leaders that show up regularly.
Other page types are swapped out more regularly. The AI platforms don’t reuse the same sources for every response.
It is interesting that a lot of listicles get cited regularly across different prompts. This checks out from the eye test too. The takeaway - I’m not sure that super-specific listicles are necessary. Like “Best Shopify Page Builder for $5M Beauty Stores”. From my experience, and from what the data shows, AI will probably just use “Best Shopify Page Builder” listicles as sources for this.
5. ChatGPT vs Google's AI
| Page family | ChatGPT | AI Overviews | AI Mode |
|---|---|---|---|
| Comparison content (listicles, alternatives, X vs Y, reviews) | 23.9% | 43.5% | 38.6% |
| Company site pages (product, pricing, home) | 44.3% | 13.2% | 18.0% |
| Marketplace listings & review-site profiles | 12.9% | 10.0% | 12.6% |
| Docs & help centers | 14.2% | 2.8% | 4.0% |
| Blog posts & news | 4.4% | 13.1% | 9.1% |
| Forums, social & video | 0.2% | 17.1% | 17.3% |
The split between GPT and Google is quite interesting.
To me, the data shows that Google’s citations seem to be a lot more centered around sourcing their decisions. “This is the best X for Y, per this article.”
ChatGPT, however, has a lot more citations leading directly to vendors’ sites. “Company X offers a free trial and prices starting from $39/month, according to their pricing page.”
It’s also notable that there’s a massive gap between the two in how often they cite forums, social and video content. For ChatGPT, it’s next to nothing (once again, remember this is just one dataset. Your results may look a lot different). For Google, over 17% of citations come from these types of sources.
A number of these citations come from YouTube, which definitely makes sense that Google would favor more heavily; YouTube of course being a Google property.
ChatGPT, on the other hand, uses documentation and help center articles a lot more.
Regular blog posts, overall, made up 6.4% of the cited pages and 2.3% of citations. By regular blog posts, I mean opinion, trend and news-style posts, not how-to guides, listicles or comparisons. Google tends to favor these a lot more than GPT.
6. Listicle length and the year in the title
| Products promised in the title | Pages | Citations per page |
|---|---|---|
| 3 to 5 | 78 | 21.7 |
| 6 to 10 | 251 | 29.6 |
| 11 to 20 | 102 | 14.3 |
| 21 or more | 30 | 6.1 |
| No number | 438 | 11.1 |
| Page group | Year in title | Pages | Citations per page |
|---|---|---|---|
| Listicles | Current year | 517 | 22.7 |
| Listicles | None | 335 | 10.9 |
| Everything else | Current year | 594 | 7.8 |
| Everything else | None | 3,893 | 8.1 |
Not sure there’s anything really actionable to take away from this, but it’s worth looking at.
Listicles promising 6 to 10 products averaged 30 citations per page. Lists of 3 to 5 averaged 22, lists of 11 to 20 averaged 14, and lists of 21 or more averaged 6.
Pages with the current year in the title or URL got 31% of citations from 20% of the pages. Almost all of that difference comes from listicles. Dated listicles averaged 23 citations per page, undated ones 11. For every other page type, a year in the title made no difference: 7.8 citations per page with it, 8.1 without.
7. Concentration
| Most-cited pages | Share of all cited pages | Share of all citations |
|---|---|---|
| Top 85 | 1.6% | 41.6% |
| Top 274 | 5% | 61.5% |
| Top 547 | 10% | 72.3% |
| Top 1,094 | 20% | 82.6% |
Does AI cite the same pages over and over? There certainly is a bit of a power law at play here.
The 85 most-cited pages made up 1.6% of all cited pages, and 42% of citations. Like everything, this is prompt-dependent. More varied prompts will likely produce more varied sources. But it does indicate that some pages, for whatever reason, get cited significantly more often than others.
There is a lot of noise here, and this data is very much surface-level. But it checks out with what I’ve seen.
What to do with this information
Like I said at the top; don't take any of this as proof of the sources AI uses for your industry or your prompts.
It’s the data for our industry, for the prompts we’re tracking.
If you want to know what gets cited for your business, measure it yourself:
- Set up a tool that tracks AI answers, like Peec AI.
- Load the prompts that matter to your business. Tip: use Claude or ChatGPT with the Peec MCP server to build and add them, unless you want to spend a week adding them by hand.
- Once answers come in, look at the sources: what kinds of pages and sites get cited for your prompts.
There are two more parts where you might trip up, though - these might be the most common mistakes in AI SEO.
Tracking the right prompts
If you’re monitoring things that no one asks, you could be the only answer that ever shows up, and it won’t matter.
About 57% of AI SEO is just understanding what it is your ideal customers are searching for. It’s a black box. And if you’re tracking the wrong things, you’re basically just screaming into a void.
And don’t get AI to come up with the prompts, or have an AI tool create the prompt suggestions from your site. Easy - but big - mistake to make.
Why? It’s inherently biased towards showing your brand. It’s influenced by your messaging. It builds prompts based on what you say on your site, and how you say it, not necessarily what your actual customers are saying.
If your site is fully aligned with how your customers talk and search, it might not be a problem. But if you’re not dialed in, you’re going to end up in a loop where you’re optimizing for the wrong language.
Building mentions
Here’s the part that I think the most people get wrong.
They say:
“Here are the articles that get cited on ChatGPT. You need to get mentioned in these articles.”
Maybe, maybe not.
Is it valuable to get your brand on these pages? Likely yes. Is it the main thing you should be focusing on? I don’t think so.
For one thing, you’re not going to be able to get mentioned on a lot of these. You’ll have locked threads, blog posts that no one updates, sales pages. Videos (I have actually had people ask me before if they can get a mention in one of our existing videos. Like… how do you propose I do that?)
There’s some value to reaching out to roundup posts and asking them to include your brand. But overall, chasing the existing citations is more effort, less return.
What you should be doing: look at the type of pages getting cited, and replicate them.
- If they’re listicles on third-party sites, publish new listicles on third-party sites
- If listicles from category vendors get cited (the horror- that one would claim that their product is the best at what they do), publish your own
- If social content shows up a lot, post your own
- If videos show up a lot, post your own
You get it.
Most of the time, it’s not that ChatGPT specifically likes that site, and that site alone, that it used as a citation. The content just, for some reason, fit the bill for them to use as a source.
Focus on replicating whatever they do, building your own sources, with the narrative you want to be repeated.
IMO that’s 97% of AI SEO. Building the narrative, putting it out into the world. It’s not easy. But it’s simple.
Want to chat about your own AI SEO? I don’t have full-time availability, but I generally have some openings for consultation. If you want to talk, send me an email at hello@andrewbuck.me