You search your own brand name in DeepSeek. The answer lists five companies in your category. Yours is not one of them. Meanwhile your site still ranks on page one in Google for the same keyword. Both things are true at once, and they are not in conflict: classic search ranks pages, generative search picks sources worth writing into an answer.
That gap is what an AI visibility score tries to capture — turning "how often and how accurately your brand shows up inside AI answers" into a number you can track, compare and take apart. This article covers what such a score actually measures, how it is computed, why it moves, and what to fix once you have one.
An AI Visibility Score Measures Inclusion, Not Ranking
An AI visibility score measures how often your brand is mentioned or cited across a fixed set of real questions, and how accurately it is described when it appears.
That is a different logic from search ranking. A search engine returns an ordered list and lets the user choose. A generative engine retrieves a handful of sources, compresses them into a paragraph, and the user usually reads only that paragraph. Position eight in a list still has a chance of being clicked. A source that never enters the generation context does not exist at all.
Unpacked, visibility is gated at three separate layers, and a break in any one of them makes the others moot:
- Crawl layer. Whether AI vendors' crawlers can fetch your pages at all. How your robots.txt treats user agents such as GPTBot, ClaudeBot, CCBot and PerplexityBot decides whether you are even eligible. Treat each vendor's own documentation as the authority on agent names — for example OpenAI's crawler documentation.
- Retrieval layer. Whether your content gets pulled into the candidate set when someone asks. This depends on semantic match with the question, how clearly your entity is defined, and whether the model's index already knows you exist.
- Generation layer. Once retrieved, whether your paragraph is useful enough to quote. Passages that lead with a conclusion, stand on their own and name their subject explicitly travel better than passages that need surrounding context to make sense. Our piece on how AI search picks its sources goes deeper on this layer.
A score that only counts mentions sees the outcome of the third layer and stays blind to failures in the first. Any usable scoring model therefore has to be multi-dimensional.
Five Dimensions: A Scoring Model You Can Reproduce
A workable AI visibility score splits into five dimensions: mention rate, description accuracy, source citation, crawl accessibility and content structure.
The weights below are this article's reference model, offered to show the reasoning rather than as an industry standard. Adjust them to your own priorities if you like — the only rule that matters is that you keep them fixed once you start comparing periods.
| Dimension | Reference weight | What it measures | How to verify |
|---|---|---|---|
| Mention rate | 30% | Share of questions in the fixed prompt set where the answer names your brand | Ask each question, record presence or absence |
| Description accuracy | 25% | Whether your scope, positioning and capabilities are described correctly when named | Compare the answer against your own site manually |
| Source citation | 20% | Whether platforms that show citations list your domain | Count domains in the citation list |
| Crawl accessibility | 15% | Whether AI crawlers are allowed and whether body copy renders without scripts | Read robots.txt; load key pages with JavaScript off |
| Content structure | 10% | Presence of quotable conclusion sentences, tables, FAQs and structured data | Inspect page structure and JSON-LD |
Normalise each dimension to 0–100, apply the weights, and you have a composite score. The benefit of splitting it this way is diagnostic: when the number drops, you know which layer to repair instead of vaguely resolving to "publish more". For the structure dimension, Google's documentation on Article structured data is the reference worth following.
How the Number Is Computed: Prompt Set, Sampling, Normalisation
The credibility of a visibility score rests almost entirely on three choices: a stable prompt set, repeated sampling and a fixed normalisation rule.
Choosing the prompt set
The prompt set is the foundation of everything downstream. It should cover three kinds of questions: brand questions ("what does AiVisible do"), category questions ("what tools monitor AI search visibility"), and situational questions ("how can an exporter get mentioned in AI answers"). Test only brand questions and the score floats unrealistically high; test only category questions and it stays near the floor for months.
Once the set is fixed, resist the urge to edit it. The whole value of a score is comparability over time, and swapping questions is swapping the ruler.
Sampling and repetition
Generative models are stochastic. Ask the same question twice and the answers may differ. So each question needs to be asked several times within the same window, and what you record is the frequency of appearance rather than a single result. More repetitions mean a steadier number and a higher API bill; fixing the count and documenting it matters more than finding some optimal value.
Normalising to 0–100
Mention rate is already a proportion, so multiplying by 100 is enough. Judgement-based dimensions such as description accuracy need an explicit rubric first — for instance "fully accurate / broadly right with drift / clearly wrong / not mentioned" — then a mapping to points. Write the rubric down so a different reviewer produces the same reading.
Which Score Movements Are Worth Acting On
Visibility scores jitter by nature; only a consistent move in one direction across several rounds is a signal worth acting on.
Roughly ordered by how urgently they deserve attention:
- Act immediately. Crawl accessibility drops to zero. Usually a changed robots.txt, a hosting or CDN default that blocks AI agents, or a switch to pure client-side rendering. This one slowly bleeds every other dimension.
- Schedule the work. Mention rate on category questions declines over several rounds while brand questions stay healthy. Your relative position in the candidate pool is slipping; usually the answer is more substantive content plus more third-party mentions.
- Log it and watch. Small round-to-round swings. Model randomness and routine index refreshes both produce this.
- Ignore it. Absolute gaps between different platforms. Retrieval mechanics differ, so comparing raw numbers across engines says little; comparing each engine's own trend says more.
Alerting on the first category and reviewing the rest weekly is a sustainable rhythm. Our guide on running a full AI visibility audit sets out the diagnostic order in more detail.
From Score to Fix List
A score earns its keep only by pointing at actions: each dimension maps to a concrete set of repairs you can finish within a fortnight.
- Low crawl accessibility. Check the user-agent rules in robots.txt line by line and confirm your hosting platform is not blocking agents by default. Open key pages with scripts disabled and see whether the body copy survives.
- Low mention rate. Close the gap on category coverage. Build proper topic clusters around the questions people actually ask rather than scattered keyword pages, and pursue natural mentions in trade media, directories and communities so retrieval has more paths to you.
- Low description accuracy. Standardise how you describe yourself. Keep brand name, positioning line and scope identical across the homepage, the about page and your structured data, so there is less room for a model to reassemble your business incorrectly.
- Low source citation. Publish verifiable, traceable material — your own data, explicit methodology, quotations with named sources — so citing platforms have a reason to list your domain.
- Low content structure. Make the first sentence of each section a standalone conclusion, add tables and FAQs, and ship Article and FAQPage structured data.
This breakdown follows the broader GEO methodology; if the framework is new to you, start with what GEO is and how GEO differs from SEO. On the academic side, the paper "GEO: Generative Engine Optimization" treats "making a source more likely to be used by a generative engine" as an optimisation problem in its own right.
What the Score Cannot Tell You
A visibility score measures exposure structure, not business outcomes — it is neither traffic nor conversion.
Three limits are worth stating up front.
First, it does not measure commercial value. Being named in eight questions out of ten says nothing about whether those eight reach people who buy. The commercial worth of the number is set by the quality of the prompt set behind it.
Second, it does not explain causation. In the same week the score rose you may have published content, edited robots.txt and been mentioned by a trade publication. Attribution requires changing one thing at a time, or observing long enough to see a pattern.
Third, it does not cover every AI surface. What is measurable is what has a public interface or can be reproduced reliably. Entry points locked inside apps or heavily personalised stay unmeasured, and one platform's score should never be presented as the whole picture.
Saying all of that out loud keeps the number from becoming a KPI to be inflated and leaves it as what it is useful for: a map for locating faults. AiVisible exists to fix this measurement process in place, run it weekly and lay out how each of the five dimensions moves; you can equally build your own version from the model above. What matters is not changing the ruler.
If you want a baseline before deciding which dimension to start with, run one through the free visibility check.
FAQ
Is an AI visibility score the same as an SEO ranking?
No. A ranking describes where your page sits in a list that users can scroll through themselves. A visibility score describes whether your brand made it into the generated answer at all. A source that never enters the generation context is effectively absent in AI search. The two metrics can easily move in opposite directions.
How often should AI visibility be measured?
Weekly suits most teams. Generative output is stochastic, so measuring too frequently turns noise into apparent trends, while long gaps let urgent failures such as a blocked crawler go unnoticed. A practical split is to alert on crawl accessibility continuously and review the remaining dimensions on a weekly trend line.
Which dimension should be fixed first when the score is low?
Start with crawl accessibility. If AI crawlers are blocked by robots.txt or by a hosting platform's default rules, no amount of content investment can reach the score. Confirm the agents are allowed and that body copy renders without JavaScript, then move on to mention rate and content structure.
Can scores from different AI platforms be compared directly?
Comparing absolute values across platforms is not advisable. Retrieval scope, corpus freshness and citation behaviour all differ, so cross-engine comparisons invite wrong conclusions. The useful comparison is longitudinal: same platform, same prompt set, same weights, tracked over time.
Can this be measured manually without a tool?
Yes. Fix a prompt set, ask each question a few times on each platform, and record whether the brand appears and whether the description is right. That gives a rough but usable baseline. The constraint is consistency over time, since repeated sampling and week-over-week comparison grow linearly with the number of platforms.