AI Search Optimisation for B2B: How to Get Recommended in AI Search

If you’re approaching AI search optimisation right now, you’re probably working from an assumption like “AI search is just newer, so I need to make sure my content is optimised.” On the surface, it sounds reasonable.

We’ve researched this first-hand, and it’s the wrong place to start. You’re taking a playbook built for one search technology and duct-taping it onto something with completely different mechanics.

If the premise is wrong, every tactic built on it inherits the fault. The stakes are highest in B2B, and that’s what this piece is about.

We’re going to walk you through how AI actually works, why B2B is a distinct problem from other markets, and what a genuine AI search optimisation strategy looks like when you start from the right place.

How AI Search Actually Works

AI search looks a lot like the search you already know. You type a query, you get a response. The interface is familiar, the behaviour feels the same, and it’s easy to assume the underlying system is too.

We found it isn’t. When we looked at how AI actually produces a response, there are a few distinct layers to it, each one built on top of the other. And each one is another reason why carrying an SEO playbook into AI search creates problems you can’t easily see coming.

AI Rarely Goes Looking for Content

It’s easy to assume that when someone types a query into an AI, it goes out and looks for information the way Google would.

However, our research shows that’s not what’s actually happening. LLMs only retrieve content 16% of the time. The other 84% of the time, AI is working from its own training data, its own memory.

And even in that 16% where retrieval does happen, it’s not looking for your keyword or your content directly. It uses query fan-out to spawn different sub-queries beyond the original prompt, each shaped by the conversational context of that session.

For example, say someone asks an AI “What project management tool should my team use?” That prompt might fan out into sub-queries around remote collaboration, integrations, team size, and use case. None of those are the original query, and none are what you’d optimise a single piece of content for.

There Is No Ranking to Climb

The second assumption that doesn’t hold is the ranking model.

Traditional search was a directory that would send you website traffic. Getting to the top three results was the holy grail, and your whole strategy was built around climbing and holding that position.

But there is no ranking you’re trying to climb in AI search. It doesn’t rank results like 1, 2, 3 unless you explicitly ask it to.

It’s designed to help people find answers as quickly as possible within the conversation, without routing them elsewhere unless genuinely needed. It runs its sub-queries, synthesises what it finds, and produces a response.

An LLM might respond to a query by saying “Don’t use HubSpot, use Salesforce.” HubSpot may be the number one ranked result in traditional search, but it just got recommended against by an AI making a call.

Every Prompt Carries Its Own Context

Finally, prompts are not keywords. People are not engaging with these systems the way they engage with a search bar.

Research from Semrush found that the average AI Mode query runs 7.22 words, compared to just four words for a traditional search query. And that’s just on Google’s AI product. Conversations on standalone tools like ChatGPT or Claude run even deeper.

Keywords were isolated, predictable events, with a few permutations and that was it. But prompts carry context.

Whatever context someone gives at the start of a session informs the entire conversation in that session. The AI takes that in and shapes every response that follows around it.

Each new prompt carries the weight of everything said before it, and we’ve measured just how far that carry-over reaches. Only about a third of the concepts raised early in a conversation survive into the final recommendation, but they decide which way it goes.

Optimising for an isolated query assumes a system that treats each interaction independently. AI search does not.

This is why AI search behaves more like an advisor than a search engine, one that makes a judgement call based on the full context it has assembled across the session.

Why B2B Is a Fundamentally Different Problem

In most markets, the bulk of the buyer journey is already done by the time someone opens a search bar.

Take toothpaste, for example. Someone knows they need it, they know what it is, and the challenge is mostly picking one and buying it. That’s also the stage, as we saw earlier, where AI actually goes out and searches for current information.

When we tracked this across hundreds of prompt clusters, retrieval ran at 48% at the conversion stage, with the same top brands surfacing in the same positions, run after run.

For those markets, AEO and SEO aren’t wildly different problems. You’re fighting for the moment where AI does its searching, which is also where most of your buyers are.

B2B is the reverse. Think about the last time your organisation evaluated a new platform or brought in a new agency. How long did it take before anyone was even ready to name vendors?

Weeks, probably. Months in a lot of cases. Most of that time was spent understanding the problem itself, before anyone knew what kind of solution they were even looking for.

This is what awareness and consideration look like in B2B.

When we tracked AI behaviour across both stages, running prompts mirroring the questions buyers actually ask during this period, the data looked nothing like the conversion picture. The AI drew entirely from its own training data. It wasn’t querying external sources, so nothing got cited.

Brand mentions were so inconsistent that the same question, run ten times, would surface almost entirely different companies each time.

Your buyer is deep in those conversations, and your analytics see none of it.

We call this Dark AI, the extended awareness and consideration activity that leaves no trace, because the AI at those stages generates none of the signals current AI visibility tools are built to detect.

Every time your buyer uses AI to understand the problem or figure out what matters to them, the AI is building context. It’s accumulating a picture of the situation, the constraints, what a credible solution looks like. This picture shapes which brands it surfaces once the buyer asks for a recommendation.

By the time someone types “[vendor] alternatives” or “best [category] software” into an AI, the shortlist has already been settled. The brands appearing there didn’t earn that position at the moment of citation. The work happened weeks earlier, in conversations nothing was tracking.

The more complex the purchase, the further AEO gets from SEO.

The Principles Behind AI Search Optimisation for B2B

If you’ve read this far, you’re probably expecting the playbook. There isn’t one. No step-by-step playbook for AEO exists yet, not ours, not anyone’s. We’ll let you know when there is one.

The technology is too new and shifts too quickly for anyone to have honestly codified it, which is exactly why we put original research at the centre of what we do. We’re working this out in public, alongside the rest of the market, and we’d encourage you to be sceptical of anyone who claims they’ve already finished.

What we can give you is a set of principles that hold up against everything our research has shown so far.

Each one follows from the mechanics we walked through earlier, which means they should keep holding even as the tactics underneath them change. A checklist would age in months. These won’t.

Make Your Content Worth Sourcing

AI answers from its own memory 84% of the time and retrieves the other 16%. Your content has to earn its way into both, and the filter is the same for each.

During training, your content had to teach the model something it didn’t already know. At retrieval, it has to be worth pulling into a synthesised answer over everything else the fan-out surfaced.

That filter is brutal for the average B2B content library. An LLM has already read close to everything public in your category. Producing a competent summary of known information is the one thing it does best without your help.

So when your blog restates the same definitions, the same best practices, and the same recycled third-party statistics as every competitor, you’ve handed the model nothing new. You don’t get penalised for it. You simply dissolve into the average it was already going to produce.

When we analysed 50,000 pieces of B2B content, most of it read as competent rather than distinctive. Competent is what the model already has.

What earns sourcing is information gain, whether a piece adds new understanding to the problem space or just repackages what already exists.

Original data works hardest, because numbers you generated are numbers the model cannot get anywhere else. Run the study yourself, or publish what you’ve observed across your own client base.

First-hand perspective does a similar job when it keeps its specifics, and so does a strong narrative built on a position you actually hold. Our shorthand for this internally is that content needs to be differentiated, high in information gain, and carried by a real argument.

This matters most in the stages you can’t see. Our Dark AI research showed that during awareness and consideration, the AI draws almost entirely on its training data. There is no retrieval moment to win at those stages.

The only way into those conversations is to be part of what the model learned, and it only learns from content that told it something new.

Build Your Presence Across the Web

Your website is a minority input into how AI describes you.

The mechanics explain why. In the 16% of cases where AI retrieves, query fan-out spawns sub-queries about integrations, use cases, and alternatives, and most of those land on comparison threads, review platforms, and community discussions rather than your domain.

In the other 84%, the model answers from training data built on the entire web’s account of your brand. Your own site contributes one voice to that account, and it’s the voice the model trusts least, because every vendor says the same flattering things about themselves.

You already apply this logic in your own evaluations. When a vendor’s homepage says one thing and the practitioner threads, peer reviews, and analyst notes say another, you believe the outside sources. AI weighs corroboration the same way.

So the work is to show up in the places the model reads. Peer review platforms in your category. The communities where your buyers ask their unbranded, problem-stage questions. Analyst coverage, industry media, podcasts, partner content.

None of this is new activity for a B2B marketer, but the reason for doing it has changed. You’re no longer chasing referral traffic from these placements. You’re feeding the corpus that shapes how AI talks about you.

One warning before you brief the PR agency. Consistency across those sources matters as much as volume. If third parties describe you one way and your own positioning says something else, you’re feeding the model contradictory signals, and a model that can’t reconcile who you are will hedge or leave you out of the answer entirely.

Get Into the Conversation Before BOFU

Most AEO effort today concentrates on the BOFU moment. Brands monitor the “best [category] software” prompts, watch who gets cited, and fight for that citation. We understand the pull, because that’s the only moment current tools can measure.

But you saw what happens in our data. By the time a buyer asks for recommendations, the AI has spent weeks helping them understand their problem, define their criteria, and form a picture of the category. The shortlist has largely converged.

Turning up at the citation moment is turning up after the vote.

Getting in earlier means building for the questions buyers ask before they can name a solution. How to tell whether they’ve outgrown their current setup. What’s driving the problem they’ve just started noticing.

These conversations are unbranded and invisible to your reporting, and they’re exactly where the AI forms its sense of which criteria matter and which brands belong in the conversation. If your thinking shapes how the model frames the problem, you’re inside the context that every later prompt builds on.

The payoff justifies the patience. AI will send you far less traffic than search did, because it was never designed to send traffic at all. But a recommendation carries far more weight than a blue link ever did.

The studies we’ve seen put the conversion rate improvement of AEO-sourced leads over SEO-sourced leads anywhere between 4x and 23x. That range is wide because measurement is still maturing, but the direction is consistent. Fewer visitors, far higher intent.

Don’t Create Content 1:1 for Prompts

The two biggest mistakes we see share the same root. Both come from treating AI search as a search channel and reaching for the SEO playbook.

The first is mapping content one-to-one to prompts. It feels logical if you’ve spent years pairing keywords with landing pages, and prompt-tracking tools encourage it by handing you lists of prompts to target. Build a page for each one, the thinking goes, and you’ve covered the demand.

It falls apart because the AI rarely goes looking for content at all, and even in the 16% of cases where it does, fan-out means it’s running twenty-odd sub-queries rather than the prompt the buyer typed. The exact string you optimised for is the one thing it never searches.

We’ll be blunt. Optimising a piece of content one-to-one for a prompt is optimising for an interaction that never happens.

Prompts also refuse to behave like keywords even when they look similar. A keyword was an isolated event with a handful of permutations. A prompt arrives carrying the whole conversation that preceded it, context you can’t see and can’t replicate.

Two buyers typing identical prompts can get different answers because their sessions primed the model differently. There’s no stable target to build a page against.

What holds up instead is covering the underlying subject with real depth. Content that maps the whole problem space stays relevant across however many sub-queries the fan-out generates, however they happen to be phrased.

Avoid the Urge to Scale AI Content

If you can’t target prompts individually, the reflex is to carpet-bomb the category instead. Stand up an AI content pipeline, publish hundreds of articles, and try to be everywhere at once.

Content influences AI when it’s differentiated, high in information gain, and built on a strong narrative. AI content produced badly is the antithesis of all of those things.

It synthesises what already exists, which is precisely what the model does better than you, for free, on demand. You’d be feeding a synthesis machine more synthesis and expecting it to be impressed.

There’s now hard evidence of how this ends, in a pattern the SEO community has taken to calling Mount AI. Sites publishing AI content at scale watch traffic climb steeply as pages get indexed, then collapse just as steeply once the systems catch on to the fact that the pages add nothing new.

One analysis tracked more than 220 sites using AI content platforms and found over half lost at least 30% of their peak organic traffic, with 39% losing more than half. That’s Google data, and the same logic bites harder in AI search, where the entire system exists to filter for the information gain this content lacks.

AI can still sit inside your production process. Volume just can’t be the strategy. Ten pieces that say something only you can say will do more for how AI perceives you than a thousand that repeat what it already knows.

How to Know If Your AI Search Optimisation Is Working

AI search has a frustrating property for anyone trying to measure their progress in it. The work that matters most happens in the stages that produce the fewest signals.

Start with what to stop watching. Traffic is the first casualty. AI was never designed to send you visitors, so judging your AEO programme on sessions is judging the channel on a job it doesn’t do.

Citation counts are the second. Citations happen reliably at the conversion stage, where retrieval kicks in and the shortlist has already converged. A citation dashboard is a report on a competition that already finished.

Citations are worth tracking. They’re just a lagging indicator, and treating them as your primary KPI means steering by the rear-view mirror.

What you can measure sits closer to the model itself. Take the questions your buyers ask during awareness and consideration, put them to the AI, and repeat the exercise over time. The exercise stays manageable when you group your prompts by the underlying question a buyer is trying to answer, rather than chasing every phrasing of it separately.

You already saw how unstable those answers are, with each run surfacing a different set of companies. That instability is your baseline, and movement against it is your progress. If your brand starts appearing across repeated runs of the unbranded, problem-stage questions in your category, the model’s picture of that category now includes you.

Listen to how the AI talks, as well as who it names. Ask it to describe your category, the problems it solves, and the criteria that separate good solutions from bad ones. If the framing it produces starts to echo the arguments your content has been making, your influence is working at the level that decides shortlists, even in runs where your name never comes up.

You can even put a number on this by tracking how often your brand appears alongside the attributes each stakeholder cares about.

The rest shows up in your pipeline rather than your dashboards. You already saw what a recommendation does to intent. In practice that looks like leads who arrive with the evaluation largely done, close faster, and tell you outright that an AI pointed them your way.

A “how did you hear about us” field on your forms will catch more of your AEO performance than most visibility tools currently can.

We want to be honest about the limits here. None of this adds up to clean attribution, and nobody measuring AI search can claim otherwise right now. The Dark AI phase is dark for vendors’ measurement tools too.

What you’re building with these signals is directional confidence, and for a channel this early, that beats precise-looking numbers that measure the wrong moment.

Start Getting Recommended in AI Search

These principles will hold as the tactics shift, but applying them to your category, your buying committee, and your competitive picture is its own piece of work. That’s the work we do.

We come at it from three directions. We run original, reproducible research into how AI behaves across B2B buying journeys and publish it transparently, which is where every finding you’ve read here came from.

We build our technology directly off the back of that research, so what it measures reflects how AI operates rather than how the industry assumed it would. And we do strategy work that bridges the two, turning what the research and the platform reveal about your brand into a roadmap you can execute against, prompt by prompt, stage by stage.

Somewhere right now, a buyer in your category is asking an AI to help them understand their problem. Come and see what it’s telling them about you, and we’ll show you what it would take to change the answer.