Understanding Your AI Visibility Score

An AI visibility score is a single number, usually from 0 to 100, that summarizes how AI engines answered a set of prompts: whether they named a brand and, depending on the tool, where they placed it, how they described it and whether they linked to it. There is no standard formula. Each tool defines its own, so a score means something only next to the formula that produced it.

Your dashboard says 29. There are two tempting readings, and both are wrong. One is that 29% of AI knows you. The other is that you are failing a test with a pass mark. The number is neither. It summarizes the answers to your prompts, on the engines in your audit, through a formula you can reproduce on paper.

Once you can reproduce it, three things follow. You can tell which fix will move it and by roughly how much. You can tell a real change from noise. And you can explain it to your boss or your client without claiming more than it measures.

This lesson takes the score apart: the four metrics, the coverage multiplier that sends an invisible brand to zero, a full worked calculation for Ledgerly, the per-platform view, how other tools calculate theirs, and the limits of any single number for AI visibility.

The formula in one sentence

A weighted average of Mention Rate (40%), Accuracy (30%), Sentiment (15%) and Citation Rate (15%), multiplied by 0.2 + 0.8 × coverage, where coverage is the market-weighted share of the platforms in your audit that mentioned you at least once. Mentioned on no platform, the score is 0.

Each metric runs from 0 to 100 and is computed over every answer in the audit:

• Mention Rate (40%) — the share of answers that name your brand. Each prompt asked on each platform is one answer: 20 prompts on 6 platforms is 120 answers, and if 18 of them name you, your mention rate is 15%.
• Accuracy (30%) — starts at 100 and drops with each statement AI gets wrong against the facts in your brand profile, weighted by how serious the worst error is and how often errors occur. With no facts on file there is nothing to check, and it stays at 100.
• Sentiment (15%) — the tone of the answers, from 0 (negative) through 50 (neutral) to 100 (positive), averaged over every answer.
• Citation Rate (15%) — for each answer, the share of its links that point to your site, averaged over every answer; an answer with no links counts as 0. Every platform can cite sources when it searches, ChatGPT and Claude included; an answer given from memory cites nothing.

Three details change how you should read them.

Mention Rate counts answers, not prompts. One prompt asked on six engines is six answers. A brand that one engine names on every prompt can still post a low mention rate if the other five never do.

An untested accuracy of 100 is not a good 100. With no brand facts on file there is nothing to check, so Accuracy sits at 100 and props up the average at 30% of it. Enter your facts before you trust this metric. The next lesson, Find What AI Gets Wrong About You, shows what to enter.

An answer from memory cannot cite you. ChatGPT, Claude and Gemini decide per prompt whether to search, so some of their answers link nothing, and each of those counts as 0 for Citation Rate. Before you chase a low Citation Rate, open the answers that did not cite you in Query Insights. One marked "Answered from memory" could not have linked anyone. One marked "Searched the web" that linked other sites shows you where the engine looked instead of you.

Coverage: why mentioned nowhere scores zero

That average is then multiplied by your platform coverage: the market-weighted share of the platforms in your audit that mentioned your brand at least once. The weights are ChatGPT 36%, Google AI Overviews 18%, Perplexity 18%, Google AI Mode 10%, Gemini 9% and Claude 9%. The multiplier is 0.2 + 0.8 × coverage:

• Mentioned on no platform: your score is 0, whatever your accuracy and sentiment. There is nothing for them to describe.
• Mentioned on every platform: the multiplier is 1 and you keep the full average.
• In between: mentioned only on Perplexity in a 6-platform audit, your coverage is 18% and the multiplier 0.34, so a weighted average of 50 becomes 17.

If Google shows no AI Overview for a query, that query is skipped, not counted as a miss. The per-platform scores use the same four metrics without the coverage multiplier, so a platform that never mentioned you can still show a score from accuracy and sentiment. Read its mention rate first.

The zero rule exists because the other metrics can look healthy for a brand nobody mentions. An answer that never names you contains no wrong facts about you, so Accuracy stays high, and a neutral answer lands near the middle of the sentiment scale. Without the multiplier, a brand absent from every answer could post a respectable average. With it, absence scores zero, which is what absence is worth.

The floor matters in the other direction. Any coverage at all keeps at least 20% of the average, so a brand named on one small engine is not treated as invisible. Above the floor, every engine that names you adds its weight.

Two things about those weights. First, they are our model's judgment of where buyers ask, not a measured market share. Nobody publishes a measured share that covers all six engines. Second, coverage is relative to the engines actually queried. A two-engine audit and a six-engine audit divide by different totals, so their scores are not directly comparable even when the answers are identical.

Skipped answers change the arithmetic too. A query where Google shows no AI Overview leaves the answer count entirely, so it lowers neither your Mention Rate nor your coverage. If Google shows an overview for none of your prompts, AI Overviews drops out of coverage altogether and the remaining engines share the total between them. The same applies to an engine whose answers all failed to come back. That is the right rule, since an answer that never existed cannot have left you out. It also means the denominator can shift between runs without anything changing on your side.

Worked example: Ledgerly's score by hand

Now ask which fix moves it. Run three versions of next month, each changing one thing:

• Accuracy fix. Ledgerly removes the old price from three pages and accuracy rises from 82 to 94. Nothing else moves. Score: 32 (+3).
• One Claude mention. Claude names Ledgerly in a single answer. Mention Rate rises to 18.2 and coverage gains Claude's weight, to 72%. Score: 33 (+4).
• One AI Overview mention. Google's overview names Ledgerly once instead. Coverage reaches 81%. Score: 36 (+7).

Read that the right way. In this example a single answer on a silent engine moves the score more than a 12-point accuracy gain, because it flips that engine from zero to mentioned and adds its whole weight to coverage. That makes coverage the most sensitive part of the score, and also the noisiest: one lucky answer can add points the next run takes away. Whatever the accuracy fix does to the number, it changes what every buyer reads about your price. Do not chase the multiplier by tuning prompts. Fix what decides whether engines can reach you, covered in Module 3, and what they say about you, covered in Module 4.

Per-platform breakdowns

Read the per-platform columns answer by answer, because one column can hold two kinds of answer.

Perplexity, Google AI Overviews and Google AI Mode retrieve live on every answer. ChatGPT and Claude are asked with web search on and decide per prompt whether to use it; Gemini decides the same way with Google Search grounding. Every answer in the audit is marked "Searched the web" or "Answered from memory."

An answer from memory reports what the model learned in training. A fix to your site reaches it only when a newer model is trained on newer data, and nothing you do this quarter speeds that up. An answer that searched can change once the engine fetches your corrected page.

So two low columns can be two different problems. High on Perplexity but low on ChatGPT and Claude, where their answers came from memory, often means you are newer than their training data. That is a waiting problem plus off-site corroboration, not something a page edit fixes. If those low answers searched, live retrieval is not finding or reaching your pages: check crawler access first.

Do not average the columns in your head. The engines do not count the same units.

Reading a change over time

The score is built for one job: comparing your audit with itself. To compare two scores, hold four things fixed: the same prompts, word for word; the same engines; the same brand facts; and the same scoring method. Change any one of them and the difference partly measures your change, not the engines. When 99Visibility changes how a score is computed, it stores a version with each result for exactly this reason.

Then take the change apart before you celebrate or panic:

1. Did coverage change? An engine that flipped between named and not named moves the multiplier, and can account for the whole difference by itself.
2. Which metric moved? Put the four metrics side by side with the previous run.
3. How many answers is that? On Ledgerly's 55-answer audit, one answer moves Mention Rate by 1.8 points, about 0.5 points of score at its current coverage. A swing that size is one answer, and one answer is noise until it repeats.
4. Did an engine change rather than you? Answers from memory that jump while nothing changed on your side may reflect a new model. Look for the same jump on the other engines before you take credit.

Check the answer count as well. If Google showed overviews for six of your prompts this month and four last month, the two audits hold different numbers of answers, and Mention Rate is a share of that number. A rate can rise while the count of answers that name you stays flat.

Trend beats snapshot. Three runs moving the same way tell you more than one big jump.

How other tools calculate an AI visibility score

Every vendor publishes its own definition, and they divide by different things. This table was read from each vendor's documentation, checked 26 October 2026. Open the linked page before you quote one, because these change.

The first four count the same thing, mentions, over different denominators, and the denominator decides the number. Take 100 answers: 30 name you, 30 name only competitors and 40 name no brand. Divide by answers, as Peec does, and you score 30%. Leave out the answers that named nobody, as Profound does, and you score 30 of 60: 50%. Count prompts, as Otterly and HubSpot's knowledge base do: if those 100 answers came from 20 prompts and 12 of them named you at least once, you score 60%. Same answers, three scores. Keyword.com and Semrush then fold in placement or topic spread, and HubSpot's free grader gives sentiment the largest share of its 100 points.

So a score from one tool cannot be compared with a score from another. To get a number for your own brand, the free AI visibility checker asks up to 3 keywords on ChatGPT and Google AI Overviews without an account and scores the answers with the formula in this lesson.

What the score is not

The 99Visibility score is a consistent yardstick: the same formula, applied to the same prompts, on the same engines, run after run. That makes it good at showing change and useless as a universal grade. Two companies' scores are not comparable unless they ran the same prompts on the same engines, and even then brand size dominates.

Semrush's index of 126 million prompts (January to April 2026) found 36 brands visible on all four surfaces it tracked every month, among them YouTube, Google, Reddit and Amazon. That is a measurement of brand size. Compare yourself with competitors of your own size, on your own prompt set, or you end up reporting your headcount under a GEO label.

The score is also not traffic. Google's generative AI report in Search Console shows impressions only, with no clicks, click-through rate or position, so nobody can tie a visibility number to visits directly. Monitoring and Measuring GEO ROI covers what you can measure. The dashboard's labels, from Critical to Excellent, are bands for reading at a glance, not benchmarks against other brands.

Common mistakes

• Comparing your score with another brand's when the prompts or engines differ.
• Trusting an accuracy of 100 with no brand facts entered. Nothing was checked.
• Reading one flipped answer as a trend. Coverage is sensitive by design. Wait for a second run.
• Changing the engine count and reading the result as progress. More engines means a different denominator.
• Chasing citations on answers from memory. An answer marked "Answered from memory" links nothing on any engine. Work on the answers that searched.
• Rewriting prompts to raise the score. You moved the instrument, not the answers.