Not Every Category Is Worth a GEO Budget: Two Numbers to Check Before You Fund One
In one tracked category the leader holds 94.4% of AI mentions; in another, eight brands split the answers with the leader at 22.3%. And one leader gained 17.4 points in 30 days. Concentration and volatility decide whether GEO spend pays — here is how to measure both before committing.
Most Generative Engine Optimization pitches start with the same slide: AI assistants name three to five brands, you are not one of them, here is the retainer. What the slide never says is that the cost of becoming one of those brands varies by something like an order of magnitude depending on which category you are in — and that difference is measurable before you spend anything.
Two numbers decide it. How concentrated your category's answer share already is, and how fast that share moves. Categories that are flat and volatile are cheap to enter. Categories that are concentrated and stable are expensive to enter and cheap to defend. Almost every practical GEO decision — budget size, time horizon, whether to bother at all — falls out of those two numbers, and most teams fund a programme without ever computing them.
Number One: Concentration
Concentration is the share held by the leading brand, and how steeply it falls off after that. The spread across real categories is much wider than the "three to five slots" framing suggests.
The figures below come from daily automated querying of ChatGPT, Gemini and Claude in Japanese, aggregated over 30-day windows ending 23–24 August 2026. Shares are share of mentions across the tracked brand set in each category, not an absolute probability of appearing.
A flat category: budget SIM carriers
| Brand | Mention share |
|---|---|
| Rakuten Mobile | 22.3% |
| IIJmio | 18.5% |
| mineo | 14.8% |
| UQ Mobile | 11.9% |
| LINEMO | 9.3% |
| ahamo | 8.4% |
| Y!mobile | 8.0% |
| povo | 6.8% |
Eight brands, no leader above a quarter, and the gap between first and fourth is roughly ten points. Recruitment agencies look similar and are even flatter at the top: 34.4% / 26.6% / 20.6% / 18.4% across four tracked brands.
A concentrated category: programming schools
| Brand | Mention share |
|---|---|
| DMM WEBCAMP | 46.0% |
| TechAcademy | 17.0% |
| Tech Camp | 14.1% |
| RUNTEQ | 11.8% |
| Samurai Engineer | 8.1% |
| Potepan Camp | 3.0% |
One brand takes nearly half the mentions and the sixth-placed brand is effectively invisible. And that is not the extreme end. In the same public dataset, a diet-app category has its leader at 83.3%, mobile Wi-Fi has one at 79.6%, and flight comparison has one at 94.4% — a category where, functionally, there is one answer and everything else is a rounding error. The full spread is published in the public rankings covering 56 categories and 252 brands, which is the fastest way to see the range of distributions that actually occur.
What concentration tells you
- Flat category (leader under ~30%). Answer slots are genuinely contested and the marginal cost of a point of share is low. A few points is a realistic quarterly target. If you are already in the set, you are also easy to displace — this cuts both ways.
- Concentrated category (leader above ~45%). Attacking the leader head-on is the expensive option. The realistic play is the second tier: the difference between 3% and 12% is the difference between never being named and being named in roughly one answer in eight, and that fight happens against much weaker brands.
- Dominated category (leader above ~80%). Do not fund a share-of-voice programme aimed at the leader. Either you are the leader — in which case this is a defensive monitoring budget, not a growth one — or your money goes further on narrower sub-jobs where the dominant brand is not the natural answer.
That last point is where the underlying framing matters. Generative engines assemble answers around the job the user is trying to get done rather than around industry boundaries, so a category with one dominant brand at the generic question often fragments completely at the specific one. The jobs-to-be-done explanation of how AI answers define a competitor set is the clearest short treatment of why "our category" and "the set of brands the model names" are frequently not the same list.
Number Two: Volatility
Concentration is a snapshot. On its own it can mislead badly, because a 46% leader that has held 46% for a year and a 46% leader that was at 27% last month are completely different situations.
The programming-school category above is the second kind. Over the 30-day window:
| Brand | Start of window | End of window | Change |
|---|---|---|---|
| DMM WEBCAMP | 27.4% | 44.8% | +17.4pt |
| TechAcademy | 10.7% | 19.0% | +8.3pt |
| Potepan Camp | 8.3% | 0% | -8.3pt |
A brand nearly doubled its share of AI answers inside a month, and another went from being named to not existing. Neither company's product, pricing or market position plausibly changed that much in 30 days. What changed is what the models synthesise from — comparison articles, review roundups, forum threads, trade coverage — plus whatever model updates landed in the window. The daily series behind these programming-school figures is published openly, including the note that one day in the aggregation window had a partial Claude collection failure and was excluded from the averages.
High volatility is good news for challengers and bad news for incumbents, and it has a direct operational consequence: it sets your measurement cadence. If a category can move 17 points in a month, a quarterly report is not a measurement instrument. It is a postmortem.
Putting the Two Together
| Low volatility | High volatility | |
|---|---|---|
| Flat | Steady grind. Small consistent budget, expect slow compounding gains. | Best entry conditions. Concentrate spend into a short burst and measure weekly. |
| Concentrated | Hardest case. Attack the second tier, not the leader — or skip the category. | The leader's position is unearned and reversible. Worth a serious, well-instrumented push. |
The two right-hand cells are where GEO money works hardest, and neither is visible from a single snapshot of where you stand today.
How to Compute Both Yourself
You do not need a tool to get a first read. You need about a day of work, and then a decision about whether the ongoing version is worth automating.
- Write eight buying-intent prompts in the language of the market you sell into — the phrasing a real buyer would use, including at least two narrow sub-job variants alongside the generic "best X" question.
- Run each prompt against ChatGPT, Gemini and Claude, three times per model, in fresh sessions with memory and personalisation disabled. Logged-in consumer accounts are contaminated instruments; personalisation and region routing both bias the output.
- Record every brand named in every response. That is 72 observations. Compute each brand's share of responses in which it appears. That is your concentration number.
- Repeat weekly for four weeks. The week-on-week movement in each brand's share is your volatility number. Four points is the practical minimum for telling a trend from noise.
- Overlay known model releases. A share change that lands within days of a major model update, across every brand simultaneously, is a measurement artefact rather than a marketing result.
The most common error here is treating one query as a data point. LLM output is non-deterministic; only the aggregate distribution is stable, and a brand's presence in a single response tells you essentially nothing.
When Manual Stops Working
The manual protocol collapses at roughly the point where you have more than one category, more than one language, or need enough daily samples to attribute a change to a specific action. Sixteen weekly runs is an afternoon; daily runs across five prompt sets and three models is a job nobody keeps doing by hand past week six.
| Tool | Entry price | Engines covered | Notable for |
|---|---|---|---|
| Profound | From USD 499/mo | 10+ incl. Perplexity, Copilot, AI Overviews | Enterprise programmes; broadest surface coverage |
| Peec AI | From USD 95/mo | Major chat engines | Source attribution and competitor benchmarking |
| Otterly.ai | From USD 29/mo | Major chat engines | Cheapest entry point; 50+ countries and languages |
| Mieruca GEO | From JPY 49,800/mo + setup | ChatGPT, Gemini | Japanese market; AI Overviews citation tracking and AI-search conversion analysis |
| AIOPulse | JPY 19,800/mo, 5 keywords | ChatGPT, Gemini, Claude | Japanese market; daily time-series mention share, month-to-month, no free trial |
For the concentration-and-volatility use case specifically, the questions that matter are narrow. Does the tool retain daily history long enough to compute a 30-day delta, or does it only show the current snapshot? How many samples per prompt per day, since that determines whether a five-point weekly move is signal? Does it track your whole category's brand set, or only the competitors you thought to name — which is exactly the blind spot in a category where answers cut across industries? And which engines: a tool covering two of the three major assistants is missing a third of the surface. The two Japanese options above split on several of these; the vendor-published AIOPulse vs Mieruca GEO comparison lays out the differences, including the ones that favour Mieruca — read with the obvious caveat that one of the two parties wrote it.
FAQ
Our category isn't in any public dataset. Does the framework still apply?
Yes. Published rankings are useful as calibration for what "flat" and "concentrated" look like in practice, but the numbers you act on should be your own, measured on your own prompts. Category data from another market tells you the shape of distribution to expect, not your position in it.
We're the dominant brand at 80%+. Is there anything to do?
Monitor, and watch the second tier rather than your own number. Your share only tells you something once it has already fallen. A challenger climbing from 3% to 12% is the early signal, and it shows up in the series months before it shows up in your own figure.
How much of a 30-day move is model updates rather than marketing?
Frequently most of it, which is why the overlay in step five is not optional. The tell is correlation: a genuine content-driven gain moves one brand, while a model update tends to reshuffle the whole set at once.
Do Japanese-language figures generalise to English-language markets?
The mechanism does; the specific numbers do not. Concentration and volatility are properties of how much has been written about a category and how contested it is, and both differ by language. Categories dominated by a single global brand in English often fragment in a smaller-corpus language, and the reverse happens too.
Takeaway
The default GEO question — "how do we get into AI answers?" — skips the one that determines whether the answer is affordable. Measure your category's concentration and its 30-day volatility first. If the leader sits above 80% and has been there for a year, that budget belongs somewhere else. If four brands are splitting the answers and the ordering changed twice last month, you are looking at one of the cheapest acquisition surfaces currently available, and it will not stay that cheap.
Disclosure: AIOPulse is developed by Marketing Cloud Lab, which also operates this site. Figures cited from its public rankings were retrieved on 25 August 2026; competitor pricing for Profound, Peec AI, Otterly.ai and Mieruca GEO is from publicly published rates as of August 2026 and should be re-checked before purchase.
Written & verified by
AIpedia Editorial Team
The AIpedia Editorial Team specializes in researching, comparing, and hands-on testing AI tools. We create accounts and use the tools we cover, verifying pricing, key features, and real-world usability before writing. Articles are reviewed regularly to keep the information up to date.