Grounded and training
Every Omnipresence figure reads in three modes. What grounded and training mean, why the same prompt scores differently in each, and which to work on.
Updated October 20, 2026
Ask a model a question with the web attached and you learn what it can find about you. Ask the same question with the web switched off and you learn what it already believes. Those are different measurements, and Omnipresence keeps them apart rather than averaging them into one comfortable number.
The three modes
Grounded. The prompt ran with web search available, so the answer reflects what the model can find today. This is the mode the Visibility Score uses.
Training. The prompt ran with no search tool, so the answer reflects what the model already learned about you during training.
All. Both together. Useful for a volume count, misleading as a rate, because it mixes two populations.
The mode selector changes which runs every figure on the screen is computed from, including share of voice and the per-platform split.
Why they disagree, and what each case means
| Grounded | Training | What it tells you |
|---|---|---|
| Good | Good | You are both findable and known. This is the position to defend. |
| Good | Poor | Retrieval is working and the model has not internalised you yet. Normal for a brand under about two years of consistent presence, and it closes slowly. |
| Poor | Good | Your reputation is ahead of your pages. Something is wrong with what a retriever can find or use right now, which is a retrieval problem. |
| Poor | Poor | Start at the beginning: are you in the sources models read for this market at all? |
Which one to work on
Grounded, almost always. It responds to work on a timescale you can act on: fix what gets retrieved and cited, and the grounded score moves within weeks.
Training moves on the model makers' schedule. You influence it by being cited consistently over a long period, and then only when a model is next trained or refreshed. It is a lagging indicator of everything else you did, which makes it a good thing to report quarterly and a bad thing to chase monthly.
A knowledge cutoff is why a model can be confidently out of date about you in training mode. That is not a measurement error; it is the state of what the model knows.
Search available, search not fired
A grounded probe records whether the model actually ran a search. Sometimes it has the tool and answers from memory anyway.
That distinction matters because it tells you how settled the question is in the model's mind. A prompt where search almost never fires is one where the model thinks it already knows the answer, and influencing it means influencing training-mode belief rather than today's retrieval. A prompt where search fires every time is live and winnable now.
How to use the toggle
- Keep it on grounded for ordinary reporting and for anything you are measuring work against.
- Switch to training when someone asks what models think of the brand by default, or when you want the slow-moving picture for a quarterly review.
- Use all only for counting runs, never for comparing rates.
Questions about grounded and training
- Which mode does the Visibility Score use?
- Grounded. It is the one your work moves within a quarter, so it is the headline.
- Can I choose the mode a prompt runs in?
- The mode is a property of the run rather than the prompt: it records whether search was available and whether the model actually fired one.
- Search was available but the model did not search. Which mode is that?
- Still grounded, and worth knowing about on its own. The product tracks whether a search actually fired, because a model answering from memory when it could have looked is telling you something.