How your LLM & SEO data is collected
Every number in GeoBubbles starts with a real question being asked to a real AI engine. This page explains exactly what we fetch, from which engines, how often, and what happens to the answers before they become your scores.
The short version
- Every active prompt is measured on three AI engines: ChatGPT, Gemini and Google AI Mode. This is the standard for all sites and all plans — there is nothing to configure.
- A Google search snapshot (SEO) is taken for every prompt in the same cycle, so AI visibility and classic rankings are always compared on the same day.
- Measurement runs weekly by default. Each site can be set to a 7-, 14- or 28-day rhythm.
- New prompts don't wait for the next cycle — on paid plans they are measured right after you save them.
Which engines are queried
| Engine | What we ask | What comes back |
|---|---|---|
| ChatGPT | Your prompt, exactly as written, in a fresh session with web browsing enabled | The full answer text plus the list of sources ChatGPT cited |
| Gemini | Your prompt, in a fresh session, targeted to the prompt's country or city | The full answer text plus Gemini's cited sources |
| Google AI Mode | Your prompt as a Google search in AI Mode, targeted to the prompt's location | The AI Mode answer plus the references it links to |
| Google Search (SEO) | Your prompt as a classic Google search (desktop, country-level) | The top organic results and your position, the AI Overview with its references, People Also Ask, featured snippet and local pack when present |
Each engine is asked once per prompt per cycle, always in a new, isolated session so earlier questions cannot influence the answer. Sites configured for repeated runs (2 or 3 answers per engine per cycle) have those runs averaged into one score — see LLM measurement runs.
Claude and Perplexity are no longer part of new measurements. Sites that collected data from them in the past keep that history, and it stays visible in their reports.
How often data is fetched
Weekly is the default. A measurement cycle for your site starts automatically in the night before the cycle date (23:30 UTC), runs through every active prompt, and is normally complete within the hour for sites with a few hundred prompts. Larger sites with more than a thousand prompts are processed in chunks of 100 and finish within a few hours.
The rhythm is a per-site setting — choose 7, 14 or 28 days in the site's settings, or ask us via the support chat to change it for you:
| Setting | Measured | Best for |
|---|---|---|
| Every 7 days (default) | Weekly | Active campaigns, competitive markets, sites you optimise continuously |
| Every 14 days | Twice a month | Stable sites where you review results fortnightly |
| Every 28 days | Monthly | Long-term monitoring with a minimal budget |
The SEO snapshot follows the same schedule as the AI engines, so every cycle gives you both views of the same day. (It can be switched off per site if you only want AI data.)
What happens when you add prompts
On a paid plan, saving prompts on the Add prompts page triggers an immediate measurement for the new prompts only — on all three engines, plus the SEO snapshot. You will typically see the first answers in the prompt list within two to five minutes. From then on, the new prompts join your site's regular cycle.
On the free trial, new prompts are queued and measured in the next run we start for your site.
The same applies to prompts generated with PromptDNA once you save them.
Locations: country and city prompts
Every prompt belongs to a location, and the location is sent to the engines together with the question:
- Country-level prompts are asked as a user in your site's home country would ask them.
- City-level prompts are additionally targeted to that city. Gemini and Google AI Mode support native city targeting; ChatGPT is targeted to the country and receives the localised wording of the prompt (for example "… in Los Angeles"). See Prompt Localisation for the details.
A prompt tracked in three cities counts as three prompts — each one is a separate question to each engine.
What we store from every answer
For each prompt, engine and cycle we keep:
- The full answer text, so you can read exactly what the AI said — open any prompt in the cockpit to see it.
- All cited sources (links the engine attached to its answer).
- Brands detected in the answer — your brand and your configured competitors, matched in the text and in the citations (see Brand entity matching and Citations vs mentions).
- Where your brand appears — first mention position and whether it is in the prose, in the sources, or both.
For the Google snapshot we keep the top organic results and your ranking, the AI Overview text and its references, People Also Ask questions, the featured snippet and the local pack.
From these raw facts we calculate the per-prompt metrics — Visibility Score, Share of Voice, Top Mention Rate and AI Impact Score — per engine and per cycle. They roll up into the cockpit, the Insights dashboard, the weekly email and your reports. See the Metrics section for how each one is computed.
How we make sure nothing is missing
Fetching thousands of answers every night is only useful if it is complete, so every cycle is checked automatically:
- Coverage check. When a cycle finishes, the number of stored answers is compared with the number of active prompts × engines that should have been measured. Anything below 99 % is flagged as a partial scan.
- Automatic retry. Prompts that received no answer (an engine timed out, a request failed) are queued again and re-run the same night.
- Alerts. Incomplete cycles raise an alert to the GeoBubbles team, and a cycle that was cut short can never be marked as "completed".
If an engine is unreachable for a whole night, the affected prompts simply keep their previous values and are measured in the next cycle — gaps are never filled with made-up data.
What counts against your plan
Your plan limit counts tracked prompts (a prompt in one location = one tracked prompt). The three AI engines and the SEO snapshot are included for every tracked prompt — adding engines does not consume additional quota, and the measurement rhythm does not change the count either.