—
—Stop guessing
Start optimizing
for the AI/LLM era
Track how ChatGPT, Gemini and Google AI Mode answer for your brand — country by country, city by city, prompt by prompt. Find the markets and cities where a competitor is the answer and you are not, and close the gap.
Your rankings still matter
But the ChatGPT answer
is what people see
ChatGPT, Gemini and Google AI Mode name a handful of brands and move on. GEO Bubbles asks them the questions your customers ask, from the cities your customers are in, on a schedule — and tells you when the name they give is not yours.
LLM tracking that runs itself
Set the prompts once and the platform re-asks every engine on your schedule. Each run is a fresh reading, stored, comparable, and dated — so a drop has a day next to it.
A prompt library, not a keyword list
Curate the questions that actually decide a purchase in your industry, in the language each market uses. Harvest them from Search Console, or generate them and edit down.
Gap detection
Every answer is checked for who was named. When a competitor appears and you do not, that prompt becomes a gap — attached to a city, an engine, and a date.
Recommendations you can hand to a writer
Not "improve your content". A content type, a city, the prompts it would answer, and the expected lift on that city's AI Score.
From prompt to published fix
in four steps
Ask every engine, on a schedule
Your prompts go to ChatGPT, Gemini and Google AI Mode from an IP in the target country. We keep the whole answer — the text, the brands named, the position, the sources cited — not just a score.
City level, because that is where it breaks
The same brand can be the obvious answer in Munich and invisible in Leipzig. A national average hides the one thing you could actually fix this quarter.
See who gets named instead
Competitor mentions are tracked on the same prompts. You get share of voice, mention rate and average position for you and for them, per city and per engine.
Publish, then re-measure
Recommendations say which content type answers which prompts. Publish it, and the next scheduled run tells you whether the engines picked it up.
Four things measured together
because separately they mislead
Multi-model monitoring
ChatGPT, Gemini and Google AI Mode, scored separately before anything is averaged. They disagree with each other far more often than people expect, and the disagreement is usually the insight.
City-level geotargeting
Answers are collected as if asked from the city itself. Munich, Berlin and Zurich get their own readings — and their own bubble on the map.
Competition intelligence
Every brand named in an answer is recorded, not only the ones you listed. Competitors you did not know you had show up in the data before they show up in your pipeline.
One composite score
The AI Visibility Score combines mention rate, position and share of voice into a number you can put in a board deck — with the three components still visible underneath it.
Everything you need
in one view
The map is the front door. Behind it is the view your team works in every week — and the one your CFO asks for once a quarter.
- AI Score & KPI cards. Score, share of voice, top mention rate and visibility, each against your closest competitor, each with the direction of travel.
- LLM visibility positioning. A quadrant of mention rate against share of voice. Leaders top right; if you are bottom left, you are invisible and now you know it.
- Visibility trends. Day by day, so a drop can be lined up against the content change, the site migration or the model update that caused it.
- Mentions and citations. Being named is one thing; being cited as the source is another. Both are tracked, because only one of them sends traffic.
Every plan has every feature
Only the size changes
No feature gates, no AI-provider add-ons, no "contact sales" to find out whether competitor tracking is included. It is. What you pick is how many countries, prompts and sites you need.
Local
For independent businesses and single-location brands — the geographic intelligence without the cost of a tool built for multi-brand operators.
- 1 country of geographic coverage
- 500 prompts across SEO & LLM monitoring
- 5 sites under management
- 15 URLs per LLM audit
Regional
For growing brands running a portfolio of country sites — the tier where geography stops being one map pin and starts being a strategy.
- 10 countries of geographic coverage
- 2,000 prompts across SEO & LLM monitoring
- 15 sites under management
- 30 URLs per LLM audit
Global
For international brands spanning continents — the consolidated view an in-house team or an agency partner needs to make search decisions at scale.
- 25 countries of geographic coverage
- 5,000 prompts across SEO & LLM monitoring
- 35 sites under management
- 50 URLs per LLM audit
Most prompt generators
invent the questions
PromptDNA researches them
Every tool in this category will hand you a list of prompts. Most of that list is synthetic — plausible-sounding questions produced from a category name, disconnected from your business and from the people who actually buy from you. Track those and you learn how visible you are for questions nobody asks. PromptDNA starts at the other end: with research.
First, a business site profile
Before a single prompt is written, a research agent works out what you sell, to whom, in which markets, in which language, and what people already come to you for. It aggregates five sources you already own:
- Google Search Console. The queries that already earn you clicks and impressions — kept past Google's 16-month window, with Core, Spam and Helpful Content updates annotated on the timeline.
- Google Business Profiles. Locations, categories, service areas, opening hours, and the questions customers have already asked you in public.
- Your website. What you actually offer, in your own words, and which pages exist to answer what.
- Keywords you already monitor. The terms you have decided matter — imported, not re-guessed.
- Competition analysis. Who gets named instead of you today, on which questions, in which cities.
Reads the business before it writes anything. Output: a site profile, not a keyword list.
Drafts the questions a real customer in that market would type, in the language they would type them — phrased the way people phrase things when they are close to buying.
Reads the profile and the current spread across the funnel, then recommends where to add weight and which content answers it.
Four layers
so you can aim at a phase instead of a keyword
Prompts are classified across the customer journey. That turns one visibility number into four, and makes an obvious question answerable: are we invisible to people who have never heard of us, or to people who are comparing us right now?
The question asked before anyone knows you exist.
Comparing the options, narrowing the shortlist.
Ready to act, deciding where to go.
Support, renewal, and the answers that earn a recommendation.
Not one assistant
A crew
each doing one thing well
Every agent is specialised, trained on its own material and limited to its own job. Nothing an agent writes reaches your site without passing a second agent that runs on a different model — and then a person.
They also work as a crew. Agents hand work to each other, watch each other's output and step in when something looks off: the Brand Monitor Agent checks every claim against your brand guidelines, the Watch Agent reports what actually changed after a publish, and the Site Profile Agent is the only source of facts the others may use. No agent acts unchecked — and every check lands in the Agent Log you can watch on the map above.
How the agents learn
Every agent starts with an onboarding training on your own material: your cities and services, your pages and the facts on them, your brand's tone and voice, the claims you must never make, and the rules of your industry. That is how it understands the context before it writes a single word.
Then comes a calibration run. On a first batch of real uncited prompts and content briefs the agents draft everything — and nothing is published. You approve or decline each piece and, where something is not as expected, say why: too promotional, wrong term, never quote a waiting time, not how we talk about price in Denmark. Every decision feeds the agent's next learning round, and the confidence shown next to a draft is calibrated on exactly this feedback. The more detail you give, the faster the drafts sound like you — the agents are eager to learn ;-)
The crew, and what each one is for
Site Profile Agent
Reads the business before anything else does, and builds the profile every other agent works from.
- Trained on
- your Search Console queries, Google Business Profiles, your own pages, the keywords you already monitor, and who gets named instead of you today
- Never does
- invent a fact about you that no source supports
- Hands off
- a business site profile — what you sell, to whom, in which markets and in which language
SEO · GEO · LLM Agent
Turns the gap between the prompts people ask and the answers you appear in into a ranked list of moves.
- Trained on
- your visibility history per city and engine, competitor share of voice, your organic rankings, and which past changes actually moved a number
- Never does
- promise a placement — no engine sells one, and it will say so
- Hands off
- the prompts to answer, the pages to refresh and the cities to prioritise, each with the lift it expects
Watch Agent
Watches the engines for the changes you did not cause — the ones nobody can catch by hand across every city.
- Trained on
- your answer history per prompt, city and engine, the shape of normal run-to-run variance, and the published algorithm-update calendar
- Never does
- call a single bad run a trend, or raise an alert inside the noise floor
- Hands off
- lost citations, hallucinated facts, newly appearing competitors and annotated updates — straight into the Agent Log
Brand Monitor Agent
Catches the answers that describe your brand wrong — then traces the claim back to the page that caused it.
- Trained on
- your brand guidelines — positioning, approved claims, product and service naming, the things that must never be said about you — and the facts your own pages support
- Never does
- call a mention misleading without the source to prove it, or contact anyone on your behalf
- Hands off
- the off-guideline answer, where it appeared, the sources the engine drew it from, and which site owner to contact first to get it corrected
The questions
people actually ask us
ChatGPT, Gemini and Google AI Mode, each scored on its own before anything is averaged.
Answers vary run to run. Every prompt is asked repeatedly so the number is a reading, not a lucky draw.
No engine publishes one. Everyone in this category is inferring from answers — we would rather say so.
Nobody can buy a citation. What you can do is remove the reasons you are not cited.
Every tracked prompt is re-asked weekly, so a change shows up in days rather than months. Add a prompt and it runs within 15 minutes — you are not waiting for the next cycle.
How is this different from a rank tracker?
A rank tracker tells you where a page sits in a list. An answer engine does not show a list — it names two or three brands in a sentence. We measure whether one of them is you, in which position, and what the model cited when it decided.
How can you measure something that changes every time you ask?
By asking repeatedly and reporting the spread. Each prompt is put to each engine several times per run, from the right country, and the score is the result across those runs. Variance is shown, not smoothed away.
Can you guarantee we get cited?
No, and nobody can. There is no paid placement and no ranking API in any of these engines. What we can do is find the prompts where you are absent, show what the model cited instead, and tell you what to publish.
Do you need access to our website?
Not to measure. Measuring needs your brand, your cities and your prompts. Search Console is a read-only connection you can revoke, and publishing a recommendation is something you do in your own CMS.
Which engines, and how often?
ChatGPT, Gemini and Google AI Mode today, on a schedule you set. More engines are added as their answers become stable enough to measure honestly.
Is our data ours?
Yes. Prompts, collected answers and the whole history stay in your workspace and export at any time. Nothing is sold or pooled.
What we are learning
about GEO, in public
Algorithm updates: Google, ChatGPT, Gemini, Perplexity & AI Overviews
One running log of every update that moves answers, with the dates you need to annotate your own charts.
TechnicalWhy technical SEO matters more in the LLM era than it ever did for search
Crawlability, structure and speed decide whether a model can quote you at all.
PlanningThe GEO Bubbles Strategy Planner: one board for your 2027 AI-visibility strategy
A single board for prompts, markets, content and owners — free to use.
ProductWhat GEO Bubbles does: a guided tour of the features that make AI visibility measurable
The full walkthrough, from the prompt library to the competitor quadrant.