The answers that never looked you up

Across 1,800 observations, 57.8% of ChatGPT runs answered a gambling question without searching the live web at all. The model reached into training data, named some brands, and stopped. No retrieval, no citations, no chance for anything published after the cutoff to matter.

The split nobody reports

How often ChatGPT does not search at all

A visibility score of 20% can mean two completely different things. It can mean the model searched, found you, and mentioned you in one run of five. Or it can mean the model searched twice out of five runs, found you both times, and answered from memory the other three.

The second is a content and authority problem you can work on. The first is a training-data problem you largely cannot, at least not this quarter. Folding them into one percentage hides which one you have.

Why a model skips the search

• /The question reads as general knowledge. “What is RTP” does not feel like it needs a lookup, so it does not get one.
• /Retrieval costs money and latency. A model that can answer plausibly without it often will.
• /Gambling queries carry safety handling. Some runs return a hedge rather than a search, and a hedge with three brand names in it still counts as a mention in most tools.
• /The phrasing matters more than the topic. Adding “in 2026” or “currently licensed” to the same question moves retrieval rates sharply, which is why a fixed prompt set matters.

What it costs you

If most answers about your category come from training data, then everything published this year is invisible to those answers. The content is not underperforming. It is not in the room.

What actually changes it

Nothing on your own site changes whether a model searches. What changes the unsearched answers is being present in the sources a model already absorbed, which is a third-party presence problem rather than an on-page one. Roughly 68% of AI citations point at domains you do not control, and the unsearched half is worse: it points at nothing you control.

The practical response is to measure the two separately, work the searched half with content and structure, and treat the unsearched half as a two-to-four quarter project of getting named where the training data comes from. Anyone selling a thirty-day fix for the second one has not separated them.

How the split is measured, including the prompt set and the counting rules, is in the published methodology.

Sources for the figures here: the 57.8% non-retrieval rate and the 40–45% range are our own measurement across 1,800 observations per market, taken with the prompt set published in the methodology. The 68% third-party citation share is Seer Interactive, 2026. Retrieval behaviour across the other five surfaces is drawn from the same sampling, run weekly.

The phrasing decides it, not the topic

The same question asked two ways produces different retrieval behaviour, and the difference is larger than anything on your site can influence. In our sampling, adding a time marker or a compliance word to a gambling prompt roughly doubles the chance that the model searches.

None of that is actionable in the way an agency would like it to be. You do not control how a visitor phrases a question. What it does explain is why two suppliers can report wildly different visibility for the same brand in the same month: if one prompt set leans on time markers and the other does not, they are measuring different amounts of retrieval before they measure anything about you.

What to ask a supplier

• /What share of runs retrieved? If the answer is a shrug, the score averages two different problems.
• /Is the prompt set fixed between reports, and can I see it? An unfixed set can produce improvement without your brand moving.
• /Are refusals counted as zeros? They should be reported separately. Folding them in flatters everyone in the set.
• /What happens to the trend when a model updates? The honest answer is that the series breaks and the report says so.

Those four questions take a minute and will tell you more about a supplier's method than any case study. We publish the answers to all four in the methodology, which is the only reason we can be argued with.

One more source note: the 68% third-party citation share is Seer Interactive, 2026, and the surface-by-surface retrieval behaviour is drawn from the same 1,800-observation sampling published in the methodology. Search volume comparisons against Google use SparkToro figures for the same period.

Related on this site

• /The denominator problem — the other way a visibility percentage misleads, and the question to ask a supplier.
• /AI Visibility Index methodology — the prompt bands, the three-run sampling and the counting rules, published in full.
• /LLM visibility monitoring — retrieval, citation and share across six surfaces, reported separately.
• /Generative engine optimisation — the work that moves the runs a model did search.
• /Casino link building — third-party presence, which is most of what moves the unsearched half.
• /Content writing — named authors and checkable claims, which is what gets absorbed in the first place.
• /Free AI visibility audit — your own split, one market, 48 hours.
• /Glossary — retrieval, citation and the rest of the vocabulary, defined.
• /All analysis — five pieces, each built on a figure published somewhere and connected nowhere.

Questions