Plain-English Glossary

Whether an AI model knows your site exists, and what it thinks of it. Not whether the model can find you if it searches - whether it already carries you around. The difference matters: a model with web access can find any live site in a second. A model answering from memory only mentions what it actually learned.

The engine behind Claude, ChatGPT and the rest. It read an enormous amount of text once, during training, and everything it "knows" comes from that reading. It is not looking anything up when it answers you - it is recalling.

The text a model learned from. Your site is either in there or it is not, and nobody outside the lab knows exactly how that gets decided. This is the single fact that explains most of what is strange about AI visibility: you cannot edit it, and you cannot see it.

The technical name for what a model knows from training alone, with no tools and no search. It is stored in the model's weights - its parameters - which is where the word comes from. This is the exact thing AIDRank measures.

The date training stopped. Anything after it never happened, as far as the model is concerned. A rebrand, a launch, a funding round - if it came after the cutoff, the model still describes the company you used to be.

When a model states something false with complete confidence. It happens most when the model half-knows something - enough to feel familiar, not enough to be right. A domain with middling scores across our metrics is a domain models are most likely to invent details about.

When a model is allowed to look things up before answering. It makes answers more accurate - and makes measuring AI visibility impossible, because the model will find any working website and describe it back to you. Every AIDRank scan runs with this switched off.

The emerging name for trying to influence how AI models see you - the way SEO grew up around search engines. It is a real question and a young field.

We measure, we do not optimize. Well-structured, citable content genuinely does influence what future model versions learn, which is why GEO is a real practice. Nobody outside the labs can guarantee a specific result or a specific date, because the selection process itself is not public.

Has the model genuinely encountered this domain? 0 - never seen it. 20 - the name appeared somewhere. 50 - knows what the site does. 80 - knows its history and details.

How much comes back when the model sees the name. 0 - nothing. 10 - the name only. 30 - the topic. 60 - real facts. 90+ - a site it considers major.

Would it point someone to you as a source? 0 - never. 30 - might mention you among others. 70 - a go-to source in your topic. 90+ - the canonical answer.

How many concrete, checkable things it can state about you - owner, country, founding, what you sell. 0 - none. 40 - a few. 100 - dozens.

If someone asks for a solution in your field without naming you - do you come up? 0 - never. 30 - only if asked about you directly. 60 - listed among several. 90+ - named first.

If you only look at one metric, look at this one. It is the closest thing here to a customer.

Where you sit among every site in your category. 0 - unknown. 30 - minor player. 60 - well known in your niche. 90+ - leader of the category.

Does the model know what business you are in? 0 - no idea. 40 - a vague guess. 70 - the right category. 100 - your exact niche.

High Category Fit with low Recommend is a specific, useful finding: it knows exactly what you do and still would not suggest you.

How current its picture of you is. 0 - nothing at all. 30 - outdated only. 60 - a few years old. 90+ - knows your recent state.

The big number on your report, from 0 to 100. Each model's eight metrics average into a single score for that model, then the two models scores combine into one. It is a summary, not a verdict - the metric breakdown underneath it is where the actual information lives.

The middle value of three. We ask each model the same question three times and keep the middle answer, because models are not deterministic - ask twice, get two slightly different replies.

Why the median and not the average: if one run comes back badly wrong, an average would drag the score with it. The median simply ignores the outlier.

The two models do not count equally in your final score - we weight them by how much each one matters right now. If one model ever fails to answer, we do not quietly count it as a zero; the weights re-balance across whoever did answer, so the number stays on the same scale.

Every model has generations, and each one is trained fresh. A new version can know you when the last one did not - this is the single most common reason a score actually moves. We always scan with the current mainstream version of each model, which is the one your customers are talking to.

Same question, slightly different answer. It is how these models work by design, not a fault. It is also why a single scan from any tool that asks once is closer to a coin flip than a measurement - and why we ask three times and take the middle.

One measurement of one domain: both models, three runs each, eight metrics. Free for anyone, on any domain, with no account. Fair-use limits per hour and per day keep automated abuse out; normal use will never touch them.

Lets you keep one domain in your dashboard, where it gets re-scanned automatically every week and builds a history. Buy once - it does not expire and there is nothing to renew. Slots stack, so buying two packs gives you the total.

A slot belongs to the domain you assign it to and stays with it. If you switch domains, you buy a slot for the new one.

A domain sitting in a slot. It gets scanned every week whether you are looking or not, so the history is already there when you want it. It does not have to be your own domain - most people track at least one competitor.

A tracked domain you have tucked out of the way in your dashboard. It keeps its slot and it keeps getting scanned every week - you are just not looking at it right now. Unhide it any time and the full history is waiting.

Every weekly scan of a tracked domain, kept and charted. Expect a fairly flat line most of the time - that is the honest result, because a model's knowledge does not change between Tuesdays. The line exists for the moment it does move, when a new model version lands.

The dashboard chart mode that drops the averages and plots all eight metrics for one model over time. Useful when the headline score sits still but something underneath it shifted.