Insights · GEO

Generative engine optimization for B2B: build the prompt set, not more content

Most B2B generative engine optimization programmes are content programmes wearing a new label. The work that actually moves AI visibility starts somewhere less comfortable: writing down the thirty questions your buyers put to a model, and finding out how often it names you.

Ask a B2B marketing team what they are doing about generative engine optimization and you will usually hear a content plan. More posts, more FAQs, more schema, a page rewritten to sound quotable. It is not wrong, and it is not enough. The uncomfortable finding when we baseline a client is almost always the same: they cannot say which questions they are trying to win, so they cannot say whether any of it worked. GEO in B2B fails at the definition stage, not the execution stage.

Why B2B GEO programmes stall

B2B has a specific problem that consumer brands do not. The buying question is narrow, high-value and asked by very few people - a handful of procurement leads, a CFO, one committee. There is no head term with 40,000 searches a month to organise the work around. So teams default to what they can measure, which is output: pages published, words written, schema added. Six months later the traffic chart looks flat, the CFO asks what the programme bought, and nobody can answer, because nobody defined the target in the first place. Publishing was never the constraint. Attribution was.

The prompt set is the unit of work

Replace the keyword list with a prompt set: 20 to 40 questions, written in the language your buyer would actually type into ChatGPT, Perplexity or Gemini at each stage of a decision. Not "pricing consultancy" but "who can fix our price-pack architecture before the next round of cost increases". Not "strategy firm" but "alternatives to a big-four strategy team for a mid-cap food business in Poland". These are specific, awkward and few - which is exactly what makes them tractable. A list of thirty prompts is something a leadership team can read, argue about, and commit to owning. A keyword universe of nine thousand terms is not.

The strategic point is that these prompts are category entry points made literal. Ehrenberg-Bass has argued for years that brands are retrieved from memory against buying situations, not against a generic ranking of preference - the cue is "we need to raise prices without losing the listing", and the brands that come to mind in that moment win. What is new is that buyers now type the cue instead of holding it in their heads. The retrieval situation became observable, and therefore auditable. That is a gift to anyone willing to write the list down.

Your buyers stopped keeping their category entry points in their heads and started typing them. For the first time you can read the cue - and check whether you are the answer.

Measure naming rate, not traffic

Once the prompt set exists, the metric follows: naming rate. For each prompt, across each engine, are you named, and who is named alongside you? Run it monthly, keep the answers, and you have a trend line that means something - unlike AI referral traffic, which is a rounding error in most B2B accounts and always will be, because the buyer reads the answer and later arrives via a branded search or a direct approach. Optimising for the click misreads the channel. The value is being in the consideration set that forms before anyone clicks anything.

Track the second dimension too: not only whether you are named, but how you are described. A model that names you as "a Warsaw-based digital agency" when you are a senior commercial advisory has done more damage than one that omits you, because it is confidently telling your buyer something false. Absence you can fix with presence. Misdescription you fix by making the correct description unavoidable - stated the same way, in enough independent places, that the model has nothing else to synthesise from.

What actually moves the number

Three things, in descending order of effect. First, a claim worth repeating: a defined position, phrased the same way everywhere, that a model can attribute to you without hedging. Second, earned presence - third-party articles, comparisons, directories, podcasts, other people's language about you. Models discount your own marketing copy as self-interested, and rightly so; what others say carries the weight. Third, technical hygiene: let GPTBot, ClaudeBot, PerplexityBot and Google-Extended crawl you, and structure your pages so a fact can be lifted cleanly. That third item is the one most programmes start with and the one with the smallest ceiling. It is necessary and it is not the work.

The order matters commercially, because it tells you where the budget belongs. Earned presence compounds and is hard for a rival to copy quickly; another blog post decays and is trivially matched. If you are choosing between a content retainer and getting your point of view into the places your buyers already read, the second is the better asset - it is the one that is still working in eighteen months.

Who owns it

This is the question that decides whether anything happens. GEO sits awkwardly between SEO, PR, product marketing and strategy, so in most organisations it belongs to everyone and therefore to no one. Give the prompt set a single named owner at the level where positioning is actually decided - usually the CMO or commercial director, not the SEO manager - and put naming rate on the same monthly review as pipeline. The mechanism is dull and it is the entire difference between a GEO deck and a GEO programme. A strategy without an owner and a review cadence is a document, which is a failure mode we have written about at length elsewhere.

How we build it

We start with a Brand X-Ray on the prompt set: we draft the 20 to 40 buying questions with your commercial team, run them across the major engines, and hand back the honest baseline - your naming rate, how you are described, and which rivals the models currently prefer to you and why. From there Plan-to-Win turns it into an owned plan: the claim to standardise, the earned presence to build, the crawler and structure fixes to clear, one owner, and a monthly number. If you want the underlying mechanics of how models choose who to cite, we set them out in GEO: how brands get cited by ChatGPT and Perplexity.