Playbook

After 18+ months running AI search campaigns across now 40+ B2B SaaS companies, the biggest mistake we see isn't a missing piece. It's sequencing.

Some companies tweak schema and publish llms.txt files before they have a single page an AI would want to cite. Others chase third-party mentions before the core solution-aware pages even exist.

And plenty pour budget into top-of-funnel content that AI assistants now answer completely, so the buyer never sees the brand.

So in this article, I'll share the full framework we use at Rock the Rankings: the GEO Stack.

Three layers, built in a specific order, because each layer feeds the one after it.

And this time I'm going to get into the actual mechanics. How we build the prompt list. What the listicle looks like on the page. What the outreach email says. What the reporting looks like when a program works, with real numbers from a real engagement.

The GEO Stack is the architecture. But architecture built on sand still falls over, and AI search will surface weak foundations painfully fast. Before this framework receives any investment, three conditions must be met:

Every GEO engagement we run uses the same three-layer architecture. The emphasis shifts by client and by month. The layers themselves don't change.

The layers are built top-down, in order, not in parallel.

Citations pointing at pages that don't exist do nothing. Problem-aware content with no solution-aware foundation bridges buyers to a shortlist you're not on.

The first layer is the set of pages AI assistants pull from when a buyer asks a buying question.

Before anything ships, though, you need to know which questions you're building for. So let's start there, because this is where most GEO programs go wrong on day one.

Here's the uncomfortable truth about prompt research: there is no reliable data source for it.

There's no Ahrefs for prompts. Nobody has search volume for "best conversational AI platform for enterprise deployments with on-prem requirements," because prompts are long, conversational, and nearly unique per buyer.

Any vendor selling you "prompt volume data" is selling you keyword data with a new label.

You have to put yourself in the shoes of a prospective buyer and derive the prompts they'd actually type. We build the list from three inputs, in this order of value.

From those three inputs we build a tracked set of 40 to 50 solution-aware prompts per client.

The tracked set needs to be small enough that movement on it means something and each prompt maps to a page decision, a citation target, or both.

The build in a foundation quarter is at minimum 15 Layer 1 pages, live on the client's domain. The mix skews hard toward three page types:

Alternatives pages deserve special mention. Someone prompting a competitor's name plus "alternatives" is a buyer with a shortlist and a problem with the leader on it. For a challenger brand, these are the highest-intent pages on the internet.

"Write a best-of listicle" is where most teams produce something an AI will skip. So let me walk through the structure we build, using a live page we run for a client in the queue management space as the reference.

Here's the skeleton, top to bottom.

Where does the client rank on their own listicle?

Number one. Always.

You're not going to publish a page on your own domain ranking yourself fourth.

The credibility mechanism isn't fake neutrality. It's the honest machinery around the #1 position: published criteria with weights, real tradeoffs on your own product, and named situations where a competitor is the better answer.

And human buyers, the ones who arrive from the AI's answer to validate it, trust the page for exactly the same reason.

How fast Layer 1 shows up

When the pages are structured this way, first appearances in tracked prompts typically land within days of indexing, not months.

Assistants doing live retrieval reward fresh, extractable, directly-on-topic pages far faster than Google ever rewarded new content. Which is also why this layer goes first: it's the fastest feedback loop in the stack, and everything in Layer 2 points at it.

Presence puts you in the conversation. Citation building is what makes the model confident recommending you over the incumbent.

This is not link building in the traditional sense. The signal you're building is mention frequency: Ahrefs' research shows brand web mentions correlate with AI visibility more strongly than backlinks do.

Building the target list

The target list starts from the citations themselves.

We track the core prompts we want to win and extract every domain the assistants cite when answering them. That extraction alone typically yields a starter list of 500+ domains.

Then we layer on the Google side: every listicle and roundup ranking on page one for the solution-aware keywords in the prompt set. Between the two sources, the working list usually runs to a couple thousand domains before prioritization.

Prioritization is where it gets manageable. The domains that both rank in Google and get cited by the models are the highest-value targets, because one placement works both surfaces. Then we sort by where competitors are present and the client is absent.

At this point the list splits into three types of target, and each gets a different motion.

Editorial targets: the outreach motion

For blogs, publications, and independent roundups, the motion is direct outreach at scale, and I mean actual infrastructure, not a VA with a Gmail account.

We enrich contacts with Clay to get the right author or editor with a verified address, and send through Instantly on a fully warmed-up sending infrastructure with rotating inboxes, because outreach at this volume on your main domain is how you end up in spam folders permanently.

The sequence is four touches. Here's the first email, genericized from a live campaign we're running for a conversational AI client:

A few things worth noticing about why this converts.

The ask is small and concrete: an addition to an existing article, not a new post.

And the value exchange doesn't require the client to have proprietary data to trade. The trade is reciprocal placement: a feature on the client's high-authority blog, a contextual dofollow link, and the AI-visibility benefit of being cited by an established brand in the space.

That last one has become the most persuasive line in the email over the past year, because editors have started caring about their own AI visibility.

The follow-ups get progressively shorter. Touch two is two lines. Touch four is "still interested in a content swap? No worries either way." Every email closes with a one-line PS telling them a single reply of "stop" ends the sequence.

Low pressure, easy out, and it keeps the sender reputation clean.

Review sites: the paid placement calculation

Some of the most-cited domains in any B2B category can't be pitched. G2, Capterra, Software Advice. You don't outreach these, and most agencies treat that as the end of the conversation.

We treat it as arithmetic.

The citation data tells you exactly which review-site pages the models pull from. Cross-reference the Google side: if Capterra's category page ranks #1 on Google for the money keyword and shows up in the extracted AI citations, a paid placement on that page is doing two jobs at once.

It drives deal flow from the Google ranking, which is how you justify the spend the way you always would have, and it puts the brand inside a page the models already trust and retrieve.

When the overlap is there, paying for the placement is a no-brainer. When it isn't, skip it.

Communities: participation, not spam

Reddit and niche communities get cited constantly, and the agency response to that fact has been to spam them, which is why half of B2B Reddit now reads like a drop-shipping forum.

That's not the motion.

The motion is being genuinely active: posting, engaging, answering questions in the subreddits and Slack groups where the ICP actually discusses the problem, over months, from accounts with real history. Not replying to threads with your business domain.

The mentions that models pick up from communities are the organic kind, the ones that happen because a brand is a known participant in the conversation. There is no shortcut here, which is precisely why it works and why spam doesn't.

Why the layer compounds

Each placement raises the probability the model treats your brand as a credible answer on the next adjacent prompt.

Citations stack on citations: the model that saw you in three trusted sources last month weighs the fourth mention differently than it weighed the first.

The final layer targets the moment before the buyer knows the category exists.

A VP of Finance types "how do I reduce invoice errors across 3 entities?" The AI bridges that process pain to a software category. If your content is the source it pulls from, you enter at the exact moment of realization.

In practice this layer is two content types.

Why this layer goes last

Not every problem prompt bridges to a product. Some return pure process advice, and content built for those prompts produces visibility with no pipeline behind it.

Run it after Layers 1 and 2 are producing, so that when the bridge does fire, it lands on a shortlist you're already on. Its payoff is the highest-intent bridge in the stack: the buyer AI just made aware they need a tool.

Same layers, same sequence, different emphasis. Your position in the category decides where the weight goes.

Some rules apply to every engagement, every layer, every category position. Get them wrong and the program can't prove itself.

1. Baseline your prompt tracking before any work ships
   No baseline, no before-and-after, no proof. You want to know exactly your (rough) visibility and more importantly existing demand converting into demos and pipeline from AI search to measure efforts and improvement against.
2. Split branded from non-branded in your visibility reporting
   Branded prompts name you in the question, so presence on them measures brand recall. Non-branded prompts are the acquisition metric: on open category questions where no brand is named, who gets recommended. A blended number can look great while the acquisition number is flat. Report both, lead with non-branded.
3. Don't block the AI crawlers
   Blocking GPTBot to "protect content" forfeits citation eligibility on the fastest-growing discovery surface.
4. Keep entity naming consistent everywhere
   Your brand name, category label, and product names should read identically across your site, review profiles, and placements, so the models tie the entity to the category.
5. Put the open-text HDYHAU field on every high-intent form
   It's the only witness that catches this channel. Add a single "How did you hear about us?" free text field on your conversion forms so you can later walk that data back against core CRM source-field data. Otherwise, most AI search conversions will fall into direct and be invisible.

The branded-versus-non-branded split deserves a beat, because it's the most common way AI visibility numbers mislead. Do the same split on your organic search reporting, because rising branded search against flat non-branded is the downstream fingerprint of upstream AI mentions.

Your sales team is already hearing "we found you on ChatGPT." Your CRM is not recording any of it, and that is not a misconfiguration.

Between 35 and 70% of AI referral sessions arrive with no referrer at all, because the in-app browser strips it. Those sessions land in Direct. A buyer who asks an assistant on Monday and searches your brand name on Tuesday gets filed as Organic or Paid. AI creates the demand and another channel collects the receipt.

The fix is not a tracking product. It is two fields and a reading rule.

We call this the Two-Witness Rule, and the full method, including the completion-rate correction and how conflicts get resolved, is written up separately. What matters here is that it goes in before the first page indexes. Set up in month four, it cannot tell you anything about months one to three.

Frameworks are cheap, so here's what one quarter of this system produced for a client of ours, an enterprise AI platform in a category with an entrenched leader.

The program launched in March: eight new Layer 1 posts (best-of listicles and competitor alternatives pages) on top of an existing commercial page set, with citation outreach running behind them.

Google side

Six of the eight new posts held average top-10 positions within the quarter, and the set went from near zero to roughly 400,000 Google impressions per 28 days.

Organic sessions to the commercial page set rose from 245 in Q1 to 2,214 in Q2, up 804%. The pre-launch baseline was around 35 sessions a month. May and June both cleared 900.

AI side

Visible AI referral sessions rose 732%, from 25 to 208, with ChatGPT growing 9x and Claude going from a single session to second place. And those are floors, since most AI visits arrive with the referrer stripped.

A reminder on the above: we know that this is certainly not all AI referral sessions due to the nature of headers being stripped when someone is sent on-site in many cases. A lot of the traffic will also be falling into the direct bucket, so we want to be tracking direct traffic increases as well.

Pipeline, which is the point

Demo requests attributed to SEO and AI search more than doubled, from 11 to 25 in the fully tracked segment.

A note on the trackers, and why pipeline is the KPI

Anyone who's run prompt tracking will tell you the answers aren't stable. Ask the same prompt on different days and the shortlist shifts.

That's normal. It's also why no AI search tracker is perfect, ours included, and why we hold visibility and citation metrics as directional.

The number we actually manage to is pipeline generated from AI search and attributed back through the recovered-attribution motion above.

I'll be direct about the failure mode, because it's not what most people expect.

These programs don't usually fail because the framework was wrong, or the category was too hard, or the model wouldn't cooperate. They fail on content execution.

Specifically: a client says they have the internal bandwidth to produce the Layer 1 content, and then they don't produce it.

We've lived this. The strategy is approved, the prompt research is done, the page briefs are written, and then the pages sit in a queue behind a product launch and a rebrand and a content manager's fourteen other priorities. Two months in, three of fifteen pages are live.

And without Layer 1, nothing downstream can work and the whole engagement stalls at the foundation.

The lesson we took from it is now part of how we scope. Content production capacity gets pressure-tested before kickoff, not discovered mid-engagement. If the client's team can genuinely ship, great.

If there's any doubt, production comes to us, because the fifteen pages aren't a nice-to-have attached to the strategy. They are the strategy, and every week they don't exist is a week the compounding isn't happening.

If you take one operational thing from this article: be brutally honest about who is writing the pages and whether they'll actually ship. Everything else in the stack is downstream of that answer.

The GEO Stack isn't complicated. Solution-aware presence first, citation building second, problem-aware capture last, weighted by your category position, measured against pipeline instead of vanity visibility.

What it demands is discipline: the discipline to build in sequence when every vendor is selling you a shortcut, the discipline to actually ship the pages, and the discipline to wait out the four-to-eight-week lag between visibility and demos.

The window matters more than the difficulty. Every citation, listicle inclusion, and brand mention trains the next answer. Early presence becomes a moat.

We run AI search and SEO programs exclusively for B2B SaaS, measured against pipeline. If you want a senior operator to audit your category's AI search landscape, map the prompts that matter for your ICP, and hand you a prioritized 90-day plan that drives actual pipeline and demos, let's chat.