The same question, two very different answers

Ask an AI engine "what is [your brand]?" with no internet access and you are testing one thing: whether the model learned about you during training. Ask the same engine with web search switched on and you are testing something else entirely: whether it finds and cites you when it goes looking. A brand can score zero on the first and win the second. Most AI-visibility tools only measure one of these, then report it as the whole story. That gap is what our Citation Probe now closes.

Two pathways, plainly

Think of an AI engine like a new colleague. Ask about your company off the top of their head (this is parametric recall) and they either remember you or they do not. Hand them a laptop and let them search first (this is grounded retrieval) and the answer depends on what is published, crawlable, and clearly written. These are separate skills. A young brand is usually invisible to recall, because it was not famous when the model was trained. The same brand can absolutely win retrieval today, because retrieval rewards clear, current, well-structured pages rather than fame.

Grounded retrieval is the mechanism behind Perplexity and Google AI Overviews, and behind the web-search modes of ChatGPT, Gemini, and Claude. It is a form of retrieval-augmented generation (what that means). Parametric recall is what you get when the model answers from memory alone.

Why conflating them is dangerous

Here is the trap. If a tool quietly sends a bare question (no web search) to five engines and you are a six-month-old brand, you will see a wall of zeros. The honest reading is "the models have not memorised us yet," which is expected and not very actionable. The misleading reading is "our GEO is failing." Acting on the second reading wastes money. Worse, every lever you can actually pull (clearer pages, verifiable facts, structured data) moves retrieval, not memory. So a tool that measures only recall is grading you on a game you cannot directly play.

What L8EntSpace does about it

The Citation Probe now labels every result by pathway. Each engine shows whether the answer came from parametric recall or grounded retrieval, with a clear marker. There is a Parametric mode, a Grounded mode, and a Both mode that runs the two side by side. A result like "the model does not know us from training, but in live retrieval we are cited and described accurately" is the honest, useful picture: it tells you where you stand and which lever to pull next.

In our own recent probe, that was exactly the pattern: cited and described correctly in grounded mode on brand-name questions, and missing entirely on the broad category questions. That contrast is the roadmap.

What to do next

Run a Both probe on your brand and read the brand-name questions and the category questions separately. If you win grounded retrieval on your own name but miss the category questions ("best tools for X"), you do not have a visibility problem, you have a content-coverage problem, and that is fixable. For why a single run can still mislead you, read One AI Probe Will Lie to You.

Key Takeaways

• Parametric recall (does the model remember you?) and grounded retrieval (does it cite you when it searches?) are different questions.
• New brands often fail recall but can win retrieval today, because retrieval rewards clarity and structure, not fame.
• The levers you control move retrieval, so measuring only recall grades you on the wrong game.
• The Citation Probe labels every result by pathway and offers Parametric, Grounded, and Both modes.
• Read brand-name and category questions separately: they point to different fixes.