Need help with B2B Marketing?
Let the smarketers’ team drive your pipeline with data-led campaigns and AI-powered growth strategies.
Most answer engine advice assumes you have thousands of queries to work with. Contract development and manufacturing does not. Neither does HL7 integration, SAP S/4HANA migration, or industrial motion control. A category with roughly 200 relevant words cannot be measured with share-of-voice tools built for consumer search, because the sample is too small for a percentage to carry meaning.
The approach that works is to enumerate the full prompt set by hand, then measure citation coverage as the share of prompts where you appear at all. That number is stable, auditable, and it moves when you do the work.
We ran this for a client selling into a single enterprise software ecosystem. Their keyword research returned zero monthly volume for every term that actually described what they sold. The buyers existed. The queries existed. The tools just could not see them.
Why every GEO framework assumes volume you do not have
Read the popular measurement frameworks and you find the same architecture underneath all of them. Pick a keyword universe. Sample it. Report the percentage of answers where your brand appears. Compare that percentage to competitors. Track it weekly.
That works when the universe has 50,000 queries and your sample of 2,000 is statistically meaningful. It falls apart at 200, where a single prompt is half a percentage point and normal week-to-week model variance swamps any signal from your actual work.
There is a second problem. Most platforms build their prompt sets from keyword tools, which means a category with no recorded volume produces an empty prompt set. The tool then reports that you have no visibility problem, because it cannot see the category at all.
STAT
Median enterprise B2B brands rank for around 9,700 keywords but appear in just 3% of relevant AI Overviews. The top quartile reaches only 4.5%, and 4.6% of enterprise B2B companies receive zero citations. Source: Walker Sands B2B AI Search Visibility Benchmark, 828 companies across 14 industries and 45 million queries, published June 2026.
Those numbers are the ceiling for well-resourced brands in categories that do have volume. In a 200-prompt category the maths is different, and in one respect much better. You are not fighting for a sliver of a huge answer surface. You are fighting for a defined and countable set of answers, most of which nobody has deliberately optimised for.
How to enumerate your real prompt set
This part is manual. It takes a working week and it cannot be automated well, because the whole point is that no tool has this data.
Start with four prompt families. Build them separately, because they behave differently and they convert differently.
Category prompts. How a buyer describes the thing they need before they know vendor names. “Who can run a sterile fill finish at commercial scale?” “What is a ServiceNow data integration partner?” These are the highest-value prompts and usually the smallest family, often 20 to 40 in total.
Competitor prompts. Every named competitor, crossed with the qualifiers buyers actually use: alternatives, versus, reviews, pricing, implementation partners. With eight competitors and six qualifiers you have 48 prompts before you write anything.
Problem prompts. The symptom the buyer types before they know a category exists. “Our integration keeps failing during upgrades.” “How do we validate a supplier for GMP manufacturing?” These rarely convert directly, but they are where the front-runner preference forms.
Procurement prompts. The compliance and diligence questions that get asked late, often by someone who is not your champion. “Is this vendor SOC 2 Type II certified?” “Who has ISO 13485 for medical device contract manufacturing.” Almost nobody optimises for these, and they sit closest to a signature.
| Prompt family | Typical count | Who asks | Conversion proximity |
|---|---|---|---|
| Category | 20 to 40 | Buyer defining the need | High |
| Competitor | 40 to 80 | Buyer building a shortlist | Very high |
| Problem | 60 to 100 | Practitioner with a symptom | Low, but shapes preference |
| Procurement | 15 to 30 | Security, legal, QA, procurement | Very high |
Source the prompts from three places: your own sales call recordings, the questions your support and solution architects field, and the actual autocomplete and follow-up suggestions the answer engines generate when you start typing a category term. The third source is the one people skip and it is the most useful, because it shows you how the model itself expands the topic.
VISUAL 1 · CAPTURE THIS
Screenshot of Perplexity showing the follow-up question suggestions after a category-level B2B query, with the sources panel visible. Annotate which suggestions became prompts in the enumerated set. Capture at desktop width, no login required.
Why share of voice breaks and what replaces it
Once you have the list, the temptation is to calculate what percentage of mentions across those prompts belong to you. Resist it. With 200 prompts and five competitors, that percentage is noise.
Measure coverage instead. Coverage is the share of prompts in which your brand is cited at all, regardless of position or prominence. It is a binary per prompt, which makes it far more stable than a share metric, and it maps directly to work: every uncovered prompt is a specific piece of content you have not built yet.
Track it three ways. Overall coverage tells the board whether the programme is working. Coverage by prompt family tells you where to spend next. Coverage on the procurement family alone is worth watching separately, because it is the cheapest to fix and the closest to revenue.
KEY TAKEAWAY
In a small prompt set, coverage answers a question a CFO can act on: of the 200 questions our buyers ask machines, how many do we currently show up for, and which ones are we fixing this quarter.
What good looks like at 200 prompts
Expect to start low. Most B2B companies in niche categories begin somewhere between 5% and 15% coverage, and almost all of it comes from competitor prompts where a third party has listed them.
A reasonable first-year target is 40% overall coverage, with procurement prompts pushed above 70% because they are structurally easy. Category prompts are the hardest and the slowest, since they depend on the model associating your brand with a concept rather than a name.
PROOF POINT
Perspectium, selling into a single enterprise software ecosystem with almost no recorded keyword volume, grew organic visibility by 66.52% and moved 25 keywords into the top ten. The work was depth in a narrow topic set, not breadth across a large one.
The lever that matters most is not on your own site. Roughly 84% of AI citations come from earned and third-party sources, according to research presented at MozCon in 2026. In a small category that concentration is even sharper, because there are only a handful of directories, review platforms, community threads and trade publications the models trust. Getting listed accurately in six places will usually outperform publishing twelve blog posts.
Instrumenting this without buying a platform
You do not need a tool for 200 prompts. You need a spreadsheet, a rota, and discipline.
- Put every prompt in a sheet with its family, the date last checked, and a yes or no for citation.
- Check the full set monthly across the engines your buyers actually use. For most B2B categories that is ChatGPT, Perplexity and Google AI Overviews, with Copilot added where the buyer is a Microsoft enterprise.
- Log the cited sources every time, not just whether you appeared. The source list is your third-party target list.
- Add one question to your discovery script asking where the prospect first encountered you and what they asked. Tag those opportunities in the CRM.
- Review coverage and pipeline tags together, quarterly.
Step four is the one that converts this from a marketing metric into a revenue conversation. Analytics will rarely attribute an answer engine visit, so the conversation on the call becomes your evidence.
Where this does not work
This method assumes a genuinely small category. If your enumeration passes about 1,000 prompts, the manual approach becomes impractical and a sampling platform is the better tool.
It also assumes the models have some data about your space. In categories so new that the training data is thin, coverage will stay near zero regardless of effort, and the work shifts to entity definition and third-party presence before anything else is worth measuring.
Finally, coverage is a visibility metric, not a demand metric. If nobody is asking about your category at all, being cited in every answer will not create buyers. That is a positioning problem, and no amount of answer engine work will solve it.
Frequently Asked Questions
How do I do GEO when my keyword tool shows zero search volume?
Enumerate the prompt set manually rather than deriving it from keyword data. Build four families of prompts from sales call recordings, support questions and the answer engines’ own follow-up suggestions, then measure the share of those prompts where your brand is cited. Zero recorded volume means the tool cannot see the queries, not that buyers are not asking them.
What is a realistic AI citation rate for a niche B2B category?
Most niche B2B companies start between 5% and 15% coverage of their enumerated prompt set. A credible first-year target is around 40% overall, with procurement and compliance prompts pushed above 70% because those are the easiest to fix.
Why does share of voice not work for small prompt sets?
With 200 prompts, one prompt is half a percentage point, so ordinary model variance between checks produces swings larger than any real change. Coverage, measured as a yes or no per prompt, is far more stable and points directly at the specific content gaps to close.
Which answer engines should a B2B company track?
Track the engines your buyers use rather than the ones with the largest consumer share. For most B2B categories that means ChatGPT, Perplexity and Google AI Overviews, with Microsoft Copilot added where the buying organisation is a Microsoft enterprise.
Do I need a paid AI visibility platform?
Not below roughly 1,000 prompts. A spreadsheet checked monthly, with the cited sources logged each time, gives you the same signal and the additional benefit of a ranked third-party outreach list. Platforms become worthwhile once manual checking stops being practical.
What matters more, my own content or third-party mentions?
Third-party sources, by a wide margin. Research presented at MozCon in 2026 found roughly 84% of AI citations come from earned and third-party media. In a narrow category only a small number of directories, review sites and trade publications carry weight, so accurate listings in those places usually outperform additional blog output.
Indrani Gope
Content Head
