End-to-end outreach workflow
Table of Contents
- Start with real customer language — pull prompts from support logs or sales interviews to understand what actual buyers are asking.
- Don’t chase the big lists — top cited domains shift dramatically by prompt type, platform, and funnel stage.
- Think beyond citations — track mentions and recommendation rank separately.
- Unify your message across surfaces — get the same answer out on your site, LinkedIn, YouTube, and relevant third-party pages.
- Focus on anchor context, not just anchor text — the editorial framing around a link matters more than the link itself.
A few months ago, I set out to answer a question I kept seeing dodged in the AI SEO conversation: which sites are actually getting cited, and why?
So I partnered with the team at Citation Labs and Xofu: Garrett French, James Wirth, and Valerie Cecil, to dig into a dataset of four million AI citations across platforms, prompt types, and funnel stages.
What came back surprised me.
There’s no universal list.
The top cited domains shift dramatically depending on the prompt, the platform, and the intent behind the query. The stuff being shared on LinkedIn is typically way too broad to act on.
So I brought the team together to talk through what the data actually means, and more importantly, what to do about it.
This is our conversation.
Here’s a slightly-edited transcript:
What were your initial thoughts on the study?
Garrett: The variability of cited domains — yes, that’s a fundamental reality now when it comes to figuring out what work to do. At Citation Labs, where do we start? What thread do we pull?
We’ve seen very consistent high variability by platform, by vertical, by prompt type. B2C vs. B2B, what does this cost, what’s the purchase decision cycle — all of these influence what gets cited. So if you’re taking anything away from high-level aggregate studies other than “I need to figure this out for my specific vertical,” you might be taking the wrong things away.
That said, LinkedIn is one thing we do see consistently, particularly at the bottom of the funnel across multiple verticals and prompt types. We see LinkedIn getting drawn in pretty reliably in bottom-of-funnel queries.
Vince: That’s a good segue. I think your initial comment when I shared the study was that the type of prompt you’re using dramatically changes where the information comes from. Let’s dig into that — you spend a lot of time building prompts and getting very granular.
Can you talk about how Citation Labs thinks about prompt design?
Garrett: For sure. And I want to speak to digital PR practitioners who may have a narrow view of what prompts they should be tracking. The first shift is: think in a multitude of prompt types. And second, look at how cited sources vary as you move up and down the funnel.
If you’ve earned coverage in a major publication — great job — but ask yourself: will that get cited in bottom-of-funnel prompts that actual buyers are using? We don’t have a perfect sense of the exact prompts buyers use, which means there may need to be a more comprehensive approach to the surfaces and publishers who could carry your content.
If you’re in digital PR and you’ve gotten a big piece covered, how should you frame or discuss it so that it’s contextually relevant at the bottom of the funnel? Could a version of it live on LinkedIn as brand content? There’s a lot of support work this kind of study points toward.
The bigger question for any practitioner is: do we have a sense of all the possible prompt types our market might be using? Are we over-indexing on big entity-knowledge prompts?
Or are we modeling the user in the middle of a real working context — trying to solve a problem — and understanding where LLMs draw their information from in that scenario?
James: And are we looking at this for our particular brand? One of the things that jumped out at me as Vince started sharing data was the high level of variability. I was seeing domains I hadn’t seen frequently enough in my own analysis to recognize the pattern. So I went down some interesting rabbit holes comparing his data to client-specific data.
The recognition for me was: this is highly customized — not just by intent type and AI system, but by the brand, the category, the niche, whether it’s B2C or B2B.
The national and international studies are helpful for direction, but are we taking it one step further and looking at this for a given domain, across different AI systems and intent levels? That’s where it gets really actionable — and it’s reasonably accessible.
If the AI results are tailored to personal history, does that make these lists meaningless? Am I being too pessimistic?
Garrett: No, I love the challenge. But I want to reframe “personalization.” What I’d encourage us to think about as an industry is something closer to role-based relevance. What role does this person occupy? What types of information do they need? That’s what personalization should mean for us as marketers.
We can’t know someone’s favorite music. But we can know that there is a logic and structure behind the types of information humans in a given role need in order to make the best decision.
That’s what we ground our prompt design in — a baseline assumption about who’s involved in a purchase decision, what context they’re operating in, and whether we as a business are providing enough information for them to make a great decision.
Valerie: Garrett, let me push back slightly. Would you even look at a top-cited domains list at all? Like when does that come into play?
Garrett: Rarely as a starting point. I’d look at content patterns across surfaces — why is this specific page getting cited? I may not be able to get cited on Forbes, but I can understand what that content does and potentially replicate the approach elsewhere.
The personalization piece — meaning the role-based framing — is the most important starting point.
Valerie: And what I’d add is that rarely are clients already thinking that way.
When you start getting granular about specific roles, specific use cases, specific working contexts, a lot of times marketing isn’t connected to those details. That’s usually the salesperson in the field who knows them.
But I think we would consistently find similar probable surfaces if we start at the bottom of the funnel and work up. We’d arrive at Forbes eventually, but along the way there are a lot of other types of publications and surfaces to interact with. And here’s the thing — you don’t have to be on Forbes to get cited.
If you’re providing genuinely useful information for people trying to make real decisions, you can be drawn into AI answers. That kind of content isn’t commonly created by marketers because we’re typically not on the ground. It’s more field-based.
Vince: So the takeaway is: do your audience research.
Showing up in AI is less about finding exploits or pattern-matching, and more about finding where your customers are and answering the right questions. If you do that well, you’ll influence AI as a byproduct — not as a primary goal. What’s your take, Valerie?
Is showing up in AI more about answering the right questions for your customers?
Valerie: Totally agree. When we first started doing link building — 15 years ago — Garrett’s ethos was always about valuing the reader. Valuing their time. Valuing the person who’s going to read whatever we produce. That’s still at the heart of this.
The closer you get to those unasked questions — the things someone didn’t know they needed answered but are critical to their decision — the better.
If you’re generating content from that place, I believe it gets picked up by LLMs because it’s answering questions that haven’t been well-answered before. It’s genuinely valuable, and the rest follows.
I’m not saying it’s purely magical and you don’t have to think about anything else. But positioning and value are foundational. James would push us to look at query fan-outs as well —
James, correct me if I’m off here — but my understanding is the more fan-outs that run, the less confident the LLM potentially is in its answer, which creates more opportunity for us to get in at the grounding layer?
James: Yes, keep going — that’s right.
Valerie: So combining Garrett’s role-based framing with the QFO data is really valuable. It also gives you guardrails. Without some constraints we’d end up producing 400,000-word guides.
Garrett: It’s a nice constraint. I’ll allow it.
James: And one more reframe on personalization: what we’re really talking about isn’t personalization based on someone’s search history — it’s personalization based on deeply understanding your target customer. Their needs, what they’re up against, their main considerations.
To Vince’s point about whether to throw out big studies — I wouldn’t throw them out, but take them with a grain of salt. They give direction.
The real work is doing your own research, specific to your domain, your ideal customer, and their journey. Because all of those factors determine who and what gets cited in the conversation that’s actually relevant to your brand.
What’s more important, AI mentions or citations?
Valerie: It’s all over the place, honestly. We get client demand for both AI mentions and citations — and even for positioning of mentions within a response.
Like, am I the first recommendation or the fifth? So it really depends on what visibility they currently have within a given use case or product category.
Sometimes the real problem is actually the opposite: they have visibility in all the wrong ways. In those cases you’re almost doing reputation management first, then building from there. And how do you decrease mentions and citations when they’re negative?
That’s its own challenge.
So we really do consider both wins.
Right now, mentions are probably our bigger focus — partly because influencing a mention can sometimes be accomplished by getting other websites cited. Controlling content on a site that impacts mention visibility is more scalable than directly chasing citations.
Citations are harder to crack.
One of the biggest barriers is that we don’t get to control client on-site content. I can’t just tell BuzzStream to rewrite their service page. That’s presented itself as a hurdle — not insurmountable, but real.
Vince: That’s a great transition into owned versus earned. In this study I broke citations into owned — coming from the brand’s own site — and earned — from third-party sources. For the evaluative prompts it was roughly 80% earned, 20% owned. Earned obviously dominates.
When you’re working with a brand, do you focus on getting the owned content right before going after earned?
James: It’s a challenging position. Organizations are largely still running the old playbook, and the person in an SEO or content role — even if they have expanded responsibilities now — often doesn’t control the core messaging. They own the blog. They don’t own the strategy.
And the challenge with that 80% earned side is that most of those surfaces are closed systems. They have their own editorial controls. They’re not handing over a platform where you can say whatever you want.
The other wrinkle is that LLMs aren’t primarily looking at your own website — they’re looking at other websites talking about you. Websites you don’t control. And you may not even be talking about your brand the right way on your own site in the first place, which is part of why owned is only 20% of citations. If the content isn’t useful or germane to the actual prompt being asked, the LLM has nothing to pull from.
What are the actual tactics for getting citations?
Garrett: I’m going to frame it and then hand it to James and Valerie, who are actually running operations.
One of the most exciting things we’ve done recently is grounding our prompt selection in real customer data. We had visibility into HubSpot for a client — specifically leads coming in from ChatGPT. When someone came from an LLM and landed on the site, they’d type what they were looking for.
That language — the actual words real buyers used — became our starting point for the prompts we track.
You work backwards from there. What are people who came in from LLMs actually asking for? Then: how do we support them getting that information?
That’s the most crystallized starting point I’ve had recently. If you can access that data, whether from HubSpot, customer support logs, or sales interviews — start there.
James: Exactly. Whatever the data source — HubSpot, customer support logs, sales interviews — deeply understand who your customer is and what they’re looking for.
Then make sure your own content reflects that. Then go do outreach to that 80% earned side. And if you can’t get the right message out through owned or earned surfaces, create new surfaces.
That’s the toolbox. We can help with all of it, but that’s the framework.
Valerie: It does depend on what visibility gap you’re trying to close and which surfaces you use to close it. Garrett built a process that starts with a prompt and builds out content that addresses visibility gaps across specific surfaces — the client’s website, a LinkedIn article, a YouTube video, a Chamber of Commerce site.
The goal is that the answer is unified across multiple surfaces.
We have seen relatively fast uptick from this approach. I’m not talking about citation change necessarily — I’m talking about mention change.
We’ve seen citation of a placed article within a week when we narrowly and verbosely address a single query fan-out tied to a specific bottom-of-funnel prompt. A week.
I’m not saying do that a million times and you’re set. How long does that last? Does it translate to your own site getting cited?
Those are the next questions. But every step gives us more to test and learn from. And the clients who are willing to go along for the ride — those engagements are genuinely exciting.
James: One thing I want to flag for people watching: this may sound ambiguous, and we don’t mean it to be.
It’s just that every engagement is highly customized to that client’s specific visibility challenges, KPIs, and goals.
It’s hard to say “do X, Y, Z and get ABC” because we’re still in early days of figuring this out.
Citations, mentions, and recommendation rank are all distinct. Recommendation rank — actually being named and recommended in an AI response — is often a shorter list than traditional search.
Maybe three to five brands that are genuinely recommended. Getting cited is different from being mentioned. Being mentioned is different from being in that recommendation set.
Each is its own challenge.
Garrett: And we’ve been doing some of this work at the page level — finding articles that are already being cited, reaching out, asking for a mention insertion or a link. Some are pay-to-play, some aren’t. We call it anchor context rather than anchor text, because what matters isn’t just the link but the editorial framing around it. The AI system can find your website on its own — what it needs is the context and descriptors that properly frame your brand in relation to the query.
What kind of backlinks actually move the needle for AI citations?
James: We start by doing a deep dive into what’s currently showing up in the answers for the prompts that matter to us.
Who’s being mentioned? What sites are being cited?
Then we figure out what message we want to get out there and which surfaces to use. It’s rarely about raw link quantity.
There’s almost always an editorial component — some contextual value being added.
That editorial aspect may actually be more important than the link itself, because the AI system can already discover your brand.
What it needs is the right framing around that discovery.
Garrett: Valerie coined the term “anchor context” for this — how are we structuring the information we want ingested? What impact are we trying to have, on which needle, measuring what outcome? That’s been one of the core driving questions at Citation Labs. Links are still part of it — we still build a lot of them — but we’re also thinking about how to affect rankings for query fan-outs and support all the different types of visibility opportunities that exist now.
Quick one: which tools are best for identifying AI citations?
James: We built Xofu because that was a hard question to answer. That’s our answer.
Valerie: Xofu.
Can you give a quick overview of how to do an AI audit?
Valerie: We actually have that functionality built into Xofu.
It starts with a competitor landscape comparison, then lets you drill down by URL — like a specific product page — and build out prompt sets from there.
You could run a thousand prompts for a single product, or ten prompts across a thousand products. We can also track the output over time: how many offsite placements have we built addressing these prompts, what gap are they closing, and against which competitor?
It’s genuinely hard to give a one-size-fits-all answer because it’s such a moving target, and the specifics vary so much by brand and category.
But the Xofu audit is built to handle that variability.
Vince: The takeaway from this whole conversation, for me, is that there’s an inherent specificity to this work that gets lost in the LinkedIn discourse — where someone says “I did this one thing and increased visibility 1,000%.”
It’s never that simple.
And even metrics like “number of citations” can be vanity metrics if you’re not tracking them against the right audience.
If BuzzStream is getting tons of citations around affiliate marketing use cases, but our core audience is digital PR and link builders — there’s a disconnect.
You have to work backward from the questions your actual customers are asking.
Garrett: We haven’t always been invited into that room as an industry, Vince.
We were held accountable for traffic, for top-of-funnel keywords, for visibility. We weren’t invited into the conversations where buyers are actually making decisions, interacting with sales, wrestling with real working problems.
Google never invited us in either. They said rank for this, make non-commodity content — but didn’t tell us how.
Then it was chase the keywords that drive the most traffic. As an industry, we haven’t been working at this long enough to have a fully functional model for thinking at the individual user level the way AI systems do. That’s what’s shifting.
Vince: It’s a mindset shift. Round peg, square hole — we’re trying to solve new problems with old frameworks.
There will definitely be a part two of this. Stay tuned and subscribe to our respective newsletters.
Thank you all — Garrett, James, Valerie — for your time and insights. And thank you to everyone who listened.
Check out the BuzzStream Podcast
