If your product category is new, buyers won't type its name into ChatGPT. They describe the problem it solves, and the engines have read almost nothing about either your label or your company. Semrush found that, for most of a study running from October 2024 to February 2026, between 65% and 85% of US ChatGPT prompts could not be matched to any keyword in a database of more than 27 billion (MediaPost, April 2026). So AI visibility for a new product category has to start with problem-language prompts: the how-to and recommendation questions buyers actually ask, measured separately for citations and mentions.
Below: why that happens, which problem prompts actually lead to brand mentions (not all of them do), and how to measure from zero.
Your buyer's prompt doesn't contain your category name
MediaPost's write-up of the Semrush data has a clean example. A traditional Google search might be "best project management software." The ChatGPT version might be "I manage a 12-person remote engineering team and we're constantly missing deadlines," followed by a question about changing their weekly standups. One is a label. The other is a situation.
In an established category, that gap is an inconvenience. The engine knows "project management software" exists and can get from the situation to the label. If you named your category last quarter, there's nothing to get to. Your buyer doesn't know the label, so it isn't in the prompt. The label has no history, so the engine can't connect the situation to it.
Keyword tools show you the same gap, just less helpfully: your category term shows zero volume. Matt Pru at Stackmatix puts it well. Zero volume "means the tool has no reliable estimate, not that nobody searches" (Stackmatix). His framework sorts a startup's terms into four groups. The category name and specific use cases get "zero to tiny" volume today. The problem or symptom, and the incumbent alternative, get real demand. That split also shows you where AI prompts come from. Buyers ask about symptoms and about the tools they use now. They only ask about your category name once someone has taught it to them.
This isn't new. Before law-firm practice-management software was a recognized category, people searched for how to manage a law firm (Helendi). What AI engines change is that the gap compounds. Buyers don't have your label, and the engine doesn't have you.
Why the engines can't know you yet
Two widely cited papers explain why. Kandpal et al. found that a language model's accuracy on a factual question tracks how many pre-training documents relate to it. They estimated models would need to be "many orders of magnitude" larger to handle facts with little pre-training support, and found that retrieval augmentation reduces that dependence (arXiv, ICML 2023). Mallen et al.'s PopQA study (14,000 questions, 10 models) found the same pattern for less popular entities: scaling didn't meaningfully fix it, and retrieval-augmented models beat much larger models without retrieval (arXiv, ACL 2023).
A new category, and a startup inside it, sits at the far end of that long tail: almost no documents at all. Knowledge cutoffs make this concrete. Anthropic's documentation lists a reliable knowledge cutoff of June 2026 for its current Claude models, defined as the date through which the model's knowledge "is most extensive and reliable" (Anthropic). A category named after that date isn't in those models' memory at all. One named shortly before it shows up in a handful of pages at best.
So retrieval, meaning live web search, is your only way in, and it isn't always used. ChatGPT turned on web search for 34.5% of queries as of February 2026, down from 46% in late 2024 (MediaPost). For the other two-thirds it answers from memory, where you don't exist. (We've covered how often ChatGPT searches the web and what that split means.) You can only compete on the searched share, so every page has to be written to be retrieved.
What the engine searches for instead of your prompt
Retrieval has a catch of its own: engines don't search the buyer's raw prompt. Google says AI Mode uses a "query fan-out" technique, "issuing multiple related searches concurrently across subtopics" (Google), and its Search Central docs say AI Overviews may use it too (Google Search Central). ChatGPT works the same way. Peec analyzed over 20 million ChatGPT fan-outs. Their average length roughly doubled, from about 6 words in October 2025 to about 12 in January 2026, while the number per prompt held at around 2.3–2.8. In Peec's words, "ChatGPT makes each query more precise rather than running more queries" (Peec).
To see which searches ChatGPT actually runs when a buyer asks a question, the guide to ChatGPT query fan-out and buyer prompts walks through the pattern with examples.
Apply that to the remote-team example. The buyer describes missed deadlines on a distributed team. The engine writes two or three precise searches of about twelve words each, covering deadlines, remote engineering teams and standups. Your coined label isn't in any of them, because the buyer never typed it. The pages that get retrieved are the ones written in the problem's language. If your site only talks in your category's vocabulary, it's answering a search nobody runs.
The obvious fix is to write for problem prompts. There's a less obvious catch.
Not every problem prompt produces a mention
It's tempting to read "problem language first" as "publish lots of 'what is [problem]' explainers." The data says that gets you cited but rarely named.
Semrush's ghost-citations study broke results down by prompt type. It covered 115 prompts in 14 countries, with 3,981 domain appearances across ChatGPT, Google AI Overviews, Gemini and AI Mode (Semrush, June 2026):
| Prompt type | Typical shape | Mention rate | Citation rate |
|---|---|---|---|
| Informational | "what is," "explain," "how does" | 18% | 89.3% |
| How-to | "how do I…" | 42.8% | not reported |
| Comparative | "best," "vs," "recommend" | 43.3% | not reported |
| Commercial | buying intent | 35.6% | 84.4% |
Comparative prompts got about 2.4 times the mention rate of informational ones. Across the whole study, 61.7% of appearances were "ghost citations": the page was cited as a source, but the brand wasn't named in the answer.
For a new-category founder, that changes the job. A page that defines the problem will probably be cited and rarely credited. The problem prompts that lead to mentions look like "how do I stop [symptom] on a [team type]" or "what's the best way to handle [problem] without [current workaround]." Those are prompts where the natural answer is an approach plus a few named options. Aim at problem-language how-to and recommendation prompts, and treat definition prompts as support.
The table also tells you what early progress looks like. A ghost citation isn't a failure. It's the first stage: the engine has started using your page as a source on the problem. To turn citations into mentions, write content that goes from problem to approach to named options, with your product as one of the options.
One limit: we haven't seen a study that directly compares problem-phrased and label-phrased prompts in new categories. The argument rests on mechanics plus nearby data, so test it on your own prompts.
Build the prompt set from buyer language, not your homepage
A workable prompt set for a new category comes together in four moves:
- Collect symptoms in the buyer's exact words. Look through sales call notes, support tickets, demo-request forms and community threads. You want the sentence a buyer used before they'd heard of you, like "we keep missing deadlines," not "we need [your category]."
- Cover the whole cluster, not one prompt. Semrush's topic-authority study used five prompts per category: what the topic is, how options compare, alternatives, use cases, and buying decisions (Semrush, July 2026). Its warning is worth keeping in view: "Winning one prompt in ChatGPT is not the same as owning a topic." Your "alternatives" prompt should name the incumbent workaround, whether that's the spreadsheet or the adjacent tool, because that term has real demand.
- Add context. BusySeed ran six non-branded buyer-intent prompts for each of 47 companies on the AI Utah 100, across ChatGPT, Claude, Gemini and Google AI Overview. Generic prompts surfaced a company in 29.6% of answers. Utah-qualified prompts surfaced one in 44.7% (BusySeed). It was a single pass, so treat it as a hint, but the direction makes sense. Narrow problem-plus-context prompts ("for a 10-person agency," "for a five-location dental group") leave more room for smaller players.
- Keep one clean definition page for your label. It's cheap, and it's there for the day a buyer does use the term. Stackmatix's advice fits: rank for it early and cheaply, but don't forecast pipeline from it.
Step one has a trap that applies to tools like ours, so we'll name it. Geoptimizer drafts suggested prompts from your own site: "Accept the ones that fit, add your own, pin the priorities." If your site speaks in your coined label, the suggestions will too. BusySeed's study had the same bias, since its prompts were written from each company's own site description. As its authors note, if a site's language differs from buyers', "models may never connect you to the query." In a new category, any generated set is a draft to rewrite in buyer language, not a list to accept as-is. That includes our free Prompt Ideas Generator, which drafts 20 buyer-intent prompts by funnel stage.
Give the engines someone else's word for it
Content on your own site gets you retrieved. On its own, it doesn't get you recommended. BusySeed puts it plainly: "Models do not trust vendor claims about vendors." They look for third-party validation. In that study, 12 of the 47 companies (25.5%) never appeared in any answer on any engine. The engines named a small set of national or global alternatives instead.
Ahrefs studied 75,000 brands and found that branded web mentions correlated 0.664 with AI Overview mentions. That was ahead of branded anchors (0.527) and branded search volume (0.392), and 26% of the brands had no AI Overview mentions at all (Ahrefs). It's a correlation, and it covers only Google's AI Overviews, but it points the same way.
New categories have it harder here, because two of the third-party sources engines rely on are closed to you for a while. G2 won't create a category for a market of one. It usually needs "a minimum of 10 unique, competing products" and checks search volume first (G2). Wikipedia requires "significant coverage in multiple reliable secondary sources that are independent of the subject," and doesn't count press releases or funding coverage (Wikipedia).
That leaves independent mentions you can earn: podcast and newsletter appearances, founder posts in communities where buyers complain about the symptom, integration partners' pages. The strongest lever a small team controls is original data. A small benchmark of the problem you solve gives other writers something to cite and link to. That's why original research earns an outsized share of AI citations.
Press distribution is one route to those independent mentions, and this guide to whether press releases actually help AI visibility looks at what the data shows about them.
Should you name the category at all?
Positioning and GEO lead to the same answer here. April Dunford estimates only "about 10%" of companies are best positioned in a newly created category, says category-first results range "from sorta bad to absolute disaster," and advises: "Even aspiring category creators shouldn't start with a market category." Many winners started in underserved subsegments of existing categories (Dunford).
The same logic applies to retrieval. "Project tracking for remote engineering teams" gives the engine a parent category it already knows, plus a qualifier, so it can be retrieved today. A coined label can't be, until the market adopts it. And markets adopt labels; vendors can't decree them. Our own field shows this. The 2023 paper that coined "generative engine optimization" (Aggarwal et al.) didn't settle the name. Wikipedia lists AEO, LLMO, AIO and "AI SEO" as competing terms with "no consensus definition," and notes that "GEO" gets confused with geography (Wikipedia). Buyers still find tools in this space, mostly by asking some version of "why doesn't ChatGPT recommend my company?" That's a problem prompt.
Measuring AI visibility from zero
When you start at zero or one mention, it's easy to misread the numbers. Four habits help.
Track mention rate and citation rate separately. If citations climb while mentions stay flat, your problem content is working at stage one. A single blended number hides that, and split metrics show it. The engines also behave differently: in the ghost-citations study, ChatGPT cited sources 87% of the time but mentioned brands only 20.7%, while Gemini mentioned brands in 83.7% of appearances but cited only 21.4%. A four-engine view keeps one quiet engine from hiding progress on another.
Judge trends, not single scans. Near zero, one scan going from 0 to 1 is noise. This is where a tracker earns its keep. Geoptimizer reports mention and citation rates per engine on a 7-day rolling window with a confidence band, and lets you mark the day you published a problem guide, so you can see whether the line moved after it.
Expect an unsettled field, and use it. In Semrush's study, 53.7% of categories were unsettled in ChatGPT. Low-demand categories were more likely to have a clear owner (19%) than high-demand ones (11.3%), and owners kept first place in 90.4% of month-over-month comparisons. The study doesn't break results down by brand size or age, so it doesn't show that startups win. It does suggest a small category isn't locked up, and that showing up consistently across the whole cluster is what holds a position.
Compare against your own baseline, not the category average. Our 2026 B2B SaaS AI visibility benchmark breakdown is useful context, but a new-category company should measure progress against its own prompt set over time.
FAQ
Why doesn't ChatGPT mention my new product category?
Usually it's two problems at once. The model has almost no training data about a recently named category, and buyers don't type the label anyway. ChatGPT searched the web on 34.5% of queries as of February 2026, so on most prompts it answers from memory, where a new category doesn't exist. Your realistic route is the searched share, using pages written in the buyer's problem language.
Should I optimize for my category name or the problem it solves?
Both, but weight them differently. Keep one clear definition page for your label so you're covered when buyers start using it. Put most of your content and tracking into problem-language how-to and recommendation prompts, which drew 42.8% and 43.3% mention rates in Semrush's study, compared with 18% for informational prompts.
Does Google require special optimization for AI Overviews?
No. Google's Search Central documentation says there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." That's consistent with this playbook, which is mostly fundamentals. The new-category difference is which queries you write for.
Start with the problem, then measure it
A new category is invisible to AI engines for an understandable reason: nobody has written much about it yet, and buyers don't know what it's called. You don't fix that by repeating the label. You fix it by showing up where the buyer's problem gets searched. Write for how-to and recommendation prompts in your buyers' own words, earn independent mentions, and expect citations before mentions.
To see where you stand today, run the free Instant AI Visibility Check on Claude Haiku 5.5. It runs three buyer-intent prompts, no signup needed. Then rewrite your prompt set in problem language and start your chart.