Why Does AI Find a Page but Not Cite It?

Being retrieved is only an intermediate step. A page can be discovered during query fan-out and still fail to appear in the final answer if another source better matches the intent, offers more distinctive information, presents a cleaner answer passage, has stronger corroboration or provides fresher and more defensible evidence. The episode therefore treats GEO as a sequence: match the intent, make the answer easy to retrieve, establish trust, and give the system a reason to select and cite the source.

Editorial note on study figures

Percentages and study findings quoted on this page are presented as statements made by the speakers during the recorded conversation. This episode page preserves the discussion rather than independently re-running or re-validating the underlying studies.

The episode is hosted through Buzzsprout and distributed through the NeuralAdX GEO & AI Search Strategy Podcast. This NeuralAdX page provides the crawlable episode summary, chapter structure, related resources and cleaned transcript.

Hello, and we’d like to say hello today to the wonderful Rhiannon Cecil, who is Head of Content at Digital Authority Partners, a Chicago digital marketing agency founded in 2016. Rhiannon has spent more than 10 years in content, starting in public relations and building brand messaging and communication strategy. She moved into agency SEO at one of the largest digital marketing agencies in the United States, working on Fortune 500 manufacturing and imaging accounts. She began experimenting with AI in 2022 and has carried this experience into her current role. Digital Authority Partners is a MarCom Gold winner for Best SEO Agency and an Inc. 5000 company. She runs DAP’s original research programme on AI search, designed to eliminate the guesswork surrounding what works in search today. The programme runs controlled studies at scale to find out which sources these systems retrieve, which brands they name, and what businesses can do to change either one. Two of those studies are already published: the AI Content Fingerprint and the AI Visibility Gap, with a new study on brand mentions launching soon. Jesus Christ, wow. Now that’s a résumé. Thank you so much for joining us today, Rhiannon. The question we thought would be useful for our viewers, particularly business owners, is: why does AI find a page but not cite it? If a client came to you with that problem, what would be your interpretation and what process would you use to help resolve it?

I think the most important thing to look at when a page is findable but not citable is the content on that page. The first thing I always ask is: are you summarising what everybody else is already saying? If the answer is yes, then AI has no reason to cite the page. The other thing I look at is how old the information is. We all know now that LLMs reward recency. One of our studies found that within a three-week period, only 33% of those citations remained. So you need to keep giving them new information in new formats in order to get cited from that page. What I think may be happening with older content is that the LLMs are using it in their training. They’re browsing these pages and internalising that knowledge, so they don’t necessarily need to cite the page anymore. Essentially, if you are being found and not cited, you need to give LLMs a reason to cite you that they cannot find anywhere else. That might be a downloadable asset. LLMs are designed to be helpful. If somebody asks a question about SEO, for example, you’ve got a better chance of being cited if you have a checklist. The system can give the answer and then offer something useful alongside it. So I would look at recency, whether there is anything unique on the page, and the structure of the page. Formats are very important. In another study, we looked at what types of pages were getting cited. Reference and encyclopedia pages were cited about 75% of the time in that study, followed by editorial content at 58%. If a client is investing heavily in service pages and high buyer-intent pages, I would also advise them to look further into editorial and informational content.

Straight off the bat, one of the things we’ve got to think about is search intent: the particular prompt you’re trying to surface for on behalf of the client. Is there a process you use to establish that throughout the article you’re creating? How do you make sure you’re hitting that search intent precisely?

Yes. If you look at traditional SEO, you had a primary keyword and contextual keywords that helped get your point across. There is a version of that which translates into optimising for GEO. When people ask LLMs questions, it is a conversation rather than simply a query. What we do is try to answer as many of those subqueries as we can on a specific page, so that regardless of where the conversation goes, there is an answer available. A simple example is advertising a forklift. A common prompt might be: ‘Is there one that fits in a 10-foot aisle?’ You can have two sentences that answer that directly. Essentially, we try to answer as many questions, proposed in as many different ways as possible, within a piece of content so that it retains its value.

I think in GEO terms that fits with generative answer coverage. From the initial H1, you move through the H2s and cover relevant questions and semantically related areas around the main intent. That gives you more opportunities when AI engines perform query fan-out for your content to be retrieved and considered during synthesis. What you’re describing is very similar to what we do. It’s partly informed by traditional SEO, but the coverage is broader because AI systems can explore many related subqueries. You want the article to cover as many genuinely relevant areas as possible without drifting away from the main intent. Once search intent is established, the next question is retrieval. The AI has performed its query fan-out and has a large set of potential sources. Is there a particular way you structure content to make it more retrievable or more attractive to the system?

One thing we tested that is working quite well is what we call the TL;DR — ‘too long; didn’t read.’ That little box at the top works really well. Everything changes very quickly in AI, but for now we’ve found that it works particularly well for metrics and data. If you give the system an opportunity to parse the most important citable information first, it appears more likely to cite you. That’s something that has worked really well for us and something we’ll continue doing for now. Then you have your page structure. In our study, we found that how you build a page was a predictor of citation 71% of the time. Proper heading structure matters. Structured data and schema markup are very important, and we’re seeing that increasingly in AI search. Outbound links to authoritative sources have also become more important. Again, there are parallels with traditional SEO. Another thing that works well is being careful about how you position answers to the questions your H2s are posing. The answer needs to be immediate so that, when the AI parses the paragraph, it can clearly see the question-and-answer relationship without defaulting to the vastly overused FAQ-section structure.

What you’re saying about the TL;DR is interesting. I came across research suggesting that implementing a TL;DR can improve the likelihood of AI citing the content. I don’t recall the organisation behind that study, but it supported the same approach. I’ve implemented it myself as well, so we’ve both independently arrived at a similar tactic.

Absolutely. It’s good to know we’re on the same track.

We’re all on an expedition trying to figure this out. Sometimes we’ll test the same thing independently, and once something proves useful, it becomes part of the toolkit for future campaigns. What you were describing — a question followed by a direct answer — also fits passage-level retrieval. You’re structuring the content so it is not only machine-readable but easy for an AI system to extract as a self-contained answer. It might be a short paragraph directly answering the question, potentially followed by evidence or a citation that strengthens citation fidelity. So we’ve dealt with search intent and retrievability. The next stage is trust. Different practitioners approach trust differently, so I’d be interested to hear how you think about it.

I think you’re right. A lot of people are approaching the trust component of AI slightly differently. One thing that really stands out, which one of our studies confirmed, is that the majority of citations that recommend a brand by name are third-party citations. Our study came in at 89% third-party editorial. I think LLMs care more about what other people say about you than what you say about yourself. Because of that, digital PR is going to become increasingly important for building trust. A lot of people also talk about the importance of author bios. What we’ve found, or what I’ve found, is that author bios are important, but not necessarily at page level. They don’t need to be visible on every page. In our content study, we looked at how detailed author bios were, how many credentials they showed and whether they linked through to another page. Pages cited in that study ranged from having no author bio at all to simply saying ‘staff writer.’ Where author bios become important is in overall entity authority. They should be detailed, but at an individual page level I don’t think they are as important. Instead, I would look at PR, conversations with people, roundup articles and comparison content. Comparison content such as ‘this agency versus this agency’ or ‘this tool versus this tool’ directly mirrors questions people ask LLMs. If I were advising a client right now, I would put a large share of the GEO budget into digital PR and the remainder into on-site optimisation.

Those percentages are based on your own accumulated knowledge, but on third-party evidence I completely agree because that’s what I’ve found as well. It makes sense: a company will naturally talk positively about itself on its own website, whereas it is harder to earn independent third-party recognition. So systems can consider what the domain owner says, but third-party corroboration can carry additional weight. At the same time, I’ve found through experimentation that first-party evidence can also be powerful — but it has to be real evidence. For example, NeuralAdX runs an AI Citation Benchmark comparing us with five leading GEO agencies and we update it month after month. That gives AI systems recurring, unique data over a longitudinal period. We also conduct live AI retrieval testing on our proof page. By 19 September, that testing had reached a full year. That gives the site a body of original evidence rather than one-off claims. We’ve seen our own benchmark and proof pages cited in AI answers, which suggests strong first-party evidence can still be selected when it is distinctive, measurable and supported by third-party tracking software. The problem is the workload. Maintaining longitudinal evidence can take 20 or 25 hours a month. But that ongoing work is part of how authority is built. That brings us to topical authority, which is another major part of trust.

Absolutely. It’s been interesting watching the evolution from internal linking being a predominant metric of topical authority to something closer to asking how many unique ideas you can provide in a given amount of content. I think that’s very important. What you’re saying about first-party data also goes hand in hand with digital PR. Nobody is going to pick you up unless you’re studying these things yourself, and that’s one reason we’ve invested so much time in our studies. We’re tracking prompts, we’ve got a team looking at what is happening, and they’re recording what they see. The formula of having first-party data, publishing it and then getting it picked up by third parties is going to matter in GEO. For us, because GEO is what we eat, sleep and breathe, the time investment is something we simply have to accept. But there are also ways to advise clients. A lot of clients are sitting on data they could potentially use and don’t even realise it. It’s then about interpreting that data, drawing useful inferences and turning it into something citable. I think that’s going to become very important.

Do you also find that one of the hardest things with clients is explaining that this has to be calm, methodical work built on accurate information? Clients naturally want immediate results, but AI visibility does not always work that way. You have to identify useful data and evidence the client may already possess, work out where it belongs, structure it properly and then execute. It isn’t simply a case of publishing several 2,000-word blog posts every day. You have to think about where the evidence will have the greatest effect. Do you find clients sometimes struggle with that because they want the result immediately?

AI becoming so dominant has huge advantages, but it also has very marked disadvantages. We see clients wanting to produce multiple blog posts at scale all the time. What I always say is that you may be focused on GEO, but there’s absolutely no reason to destroy your SEO in pursuit of what may be a fleeting citation. There is this idea that if we publish hundreds of blog posts, we’ll get a citation somewhere. But if you’re simply summarising what everybody else is saying, you can do that a hundred times and the outcome may be the same. Some clients understand that it takes time and would rather do it properly. I think it’s best to caution clients: let’s take this carefully and position it in the right way. Whatever you publish needs to be defensible. You need to be able to show your methodology, particularly when you’re talking about first-party research. I’m very direct with clients. If something isn’t going to work, I’ll tell them. We can go after citations, and citations are useful, but what I think we would prefer to achieve is brand mentions. Being recommended by name is another level altogether. Then you want content that withstands the churn — URLs that survive over time. Those are the important things.

That’s one of the hardest levels of GEO. Earning a domain citation is valuable, but the next stage is brand visibility — being named directly in the answer. That’s arguably one of the most important outcomes.

I have no idea — you just froze there for a second.

Your screen just vanished for a moment and I was wondering what had happened.

That’s what happened on my end as well. I thought you’d gone.

Coming back to brand visibility, when the company’s name is being mentioned, that’s top-tier GEO. But consistency over time is another level again. It’s possible to appear briefly because of recency effects, but maintaining citation stability and brand-mention stability is much harder. That’s what I think of as elite GEO. I’m seeing that in our own live AI retrieval testing. For the first nine or ten months we surfaced very strongly for particular queries, but by months 11 and 12 some results began to dip. That’s the reality: another agency can publish a newer study that temporarily commands more attention. So durable GEO requires ongoing effort, precision and fresh evidence. One analogy I use with clients is a judge in court. You can’t simply make claims and expect them to be accepted. You need evidence — facts, records, witness statements, something that validates what you’re saying. That analogy helps clients understand why we can’t promise immediate results just because money has changed hands. We can work towards the result, but the system still has to be persuaded by the evidence.

Absolutely. If an agency says it can guarantee you a fixed number of citations within the first few months, I think we all know that’s unrealistic. I compare it to traditional PR. PR also took time because you had to build relationships. This is another form of that. We have to convince LLMs that we, or our clients, are entities they can recognise and trust. It’s exactly what you said: the LLMs are coming back because they know your first-party data is good quality and that you update it regularly. Once they recognise you as an entity, the work isn’t finished, but you’ve overcome the first hurdle. Citations are useful because they show visibility and help build recognition. But what we’re really after is the bigger picture — brand recognition. We want your name to appear when people search for something and for the AI to recommend you by name. That’s the level we aim for with clients.

Here’s my view of where this may go. The importance of appearing in AI answers could increase dramatically. Google already blends AI-generated answers into the search experience, and the direction of travel is towards more conversational search. If traditional blue-link results become less prominent, then being one of the few brands actually named in an AI answer becomes much more valuable. For a high-intent commercial query, an AI system might mention only a handful of companies even though thousands want that visibility. That makes brand mentions and AI selection strategically important. Then take it one stage further and look at AI agents. As agents become more capable of interacting with websites and completing tasks, the question becomes not only whether the AI mentions you, but whether it chooses your website or service as part of the action.

Yes.

Exactly. GEO then also becomes about making your website one that AI systems can understand, trust and choose to interact with. Further down the line, that could involve APIs, agent transactions or other machine-to-machine interactions. That is why I think practitioners who are doing the testing now — learning what works, what fails and how systems behave — are building useful baseline knowledge for what comes next.

Absolutely. Now is the time to experiment. It really is the opportunity to make mistakes, and anybody going after GEO needs to look at it that way. Every mistake is a lesson. We need to learn as much as we can before AI Mode becomes the default experience. You can see what the search results look like now: more and more queries trigger an AI Overview, and users can increasingly interact with the answer directly. The more high-quality information we can give these LLMs while these systems continue developing, the better positioned we may be. I also think citations could become less important over time. LLMs are showing that they’re increasingly comfortable answering certain queries without citations. They may name a brand without necessarily citing a webpage. I think brand mentions could therefore become the real currency we’re working with going forward.

It’s interesting because this connects with what I think of as confident content: areas where an AI system has so much established information that it may not need to retrieve externally for every answer. That can be concerning for business owners because the system may no longer need to pull their website into the conversation. One thing that still encourages retrieval is the need to reduce hallucination risk and ground answers in evidence. But if models become much more reliable over time, they may feel less need to retrieve external pages for familiar questions. That would make genuinely new, useful information even more important.

I also think it depends on the type of query. For many ‘what is’ and ‘how-to’ queries, I’m not seeing much RAG happening anymore unless the topic is particularly obscure. ‘What is SEO?’, ‘What is GEO?’ and ‘How do I optimise my website?’ are questions LLMs can often answer from their existing knowledge. So it comes down to optimising for queries they can’t easily answer on their own, or providing genuine additional value. Something we’ve experimented with is a value-based click strategy: give people something that encourages them to click through from the AI Overview itself. That might be a study, white paper, checklist, e-book or another useful asset. If the LLM is confident the asset genuinely contains what you say it contains, you’re giving people an additional reason to click through.

That brings me to knowledge graph enrichment. One of the biggest things I’ve learned from our own testing is the value of introducing genuinely new information that the AI has not seen before. If the model already feels confident about a topic, simply repackaging existing information doesn’t add much. But if you introduce new evidence, especially when it can be backed by third-party software or independent sources, the system has a reason to investigate it and potentially synthesise it into the answer. The difficulty is that this takes real work. You can’t just repurpose what is already out there; you have to go looking for new knowledge.

Absolutely.

That’s another way to get noticed by AI systems, but again, it requires real effort.

It is. I think information gain is going to be a major differentiator. That’s essentially what we’re discussing with studies, knowledge graphs and new research. Producing verifiable information these LLMs don’t already have is going to become increasingly time-consuming, and I think that’s the challenge going forward. The fact that we’re already experimenting and publishing what we’re finding puts us in good stead for the future, when everybody has done the obvious research and tried the obvious unique angles. We’re going to have to be increasingly creative. If you genuinely want to produce something rankable and valuable, you need to think of something an AI hasn’t already thought of, focus on that and report on it. That’s exciting. You’ve really got to stay on your toes.

I think the same thing. The frustration comes back to clients because producing new evidence often requires their participation. Take a construction client: you may need them to record an installation, explain what they’re doing and provide photographs. Then we can build a page around that evidence, support it with a transcript and connect it to other projects and areas of expertise. That creates an ecosystem of fresh, real-world information. But the client is also running a business and may wonder whether they have time to do it. I think this kind of evidence production is going to become increasingly important because AI systems need real signals that distinguish one business from another.

I think that’s right. If proximity happens to work in your favour, that’s great. But outside of that, everybody is going to have to put in the work.

Another thing I like about AI search is the pressure it can put on empty marketing claims. Systems are not perfect yet, and weak claims can still get repeated, but I think the direction of travel is towards greater scrutiny. The more these systems improve, the more valuable precise, accurate and evidenced information should become. That excites me because there is so much marketing content online where it’s difficult to know what is genuine and what is simply promotional. If AI systems increasingly reward verifiable information, that’s a positive development.

It’s the great leveller.

It definitely is. I think we’ll bring things to a pause there because we’ve done around 45 minutes and I’m sure you have plenty to get on with. Rhiannon, sincerely, thank you so much for coming on. I really value your expertise. It’s been fascinating hearing about your experiences in South Africa, your work with clients in America and the testing and research you’ve been doing. It helps me, and hopefully it helps everybody listening and absorbing this information. Thank you so much. I really appreciate it.

Thank you for having me. It’s been really great. I’ve learned a lot as well. Let’s see where all of this goes. Thanks very much.

No problem at all. Thank you so much indeed. God bless. I’ll see you soon. Take care. Bye.

Evidence, methodology and further learning

Research, Methodology and AI Citation Evidence

Follow the themes discussed in this episode into NeuralAdX’s methodology, academic research base, longitudinal AI citation benchmarks, live retrieval evidence and wider AI search resources.

Continue exploring

Explore NeuralAdX GEO Research and More Podcast Episodes

Return to the main podcast hub for every episode, or continue into the NeuralAdX methodology, benchmark evidence and Generative Engine Optimisation service.