AI visibility score: what the number means at Semrush, Ahrefs, Profound and here

An AI visibility score depends on what a vendor asks, which engines it checks, whether it counts a mention or a citation, and how it turns those observations into a number.

The public material does not give us enough evidence to explain Profound’s score fairly. We will say where the limit applies rather than fill the gap with confident arithmetic. The comparison here covers Semrush and Ahrefs, plus CiteGraph’s answer and citation data.

What does your AI visibility score actually count?

Start with the event that makes the score rise. Ahrefs gives the clearest public answer. Its Brand Radar explanation of visibility metrics says visibility comes from how often a brand is mentioned or cited in AI responses. It separates mentions, where the brand name appears in answer text, from citations, where AI links to a page or domain as a source. It also lists AI Share of Voice and Estimated Impressions, with the latter weighted by the search volume behind each prompt.

A mention is not a citation.

Semrush’s public AI pricing page describes the product as being for “businesses that want to track how they’re mentioned in AI search”. It lists AI visibility reports for any domain, brand performance analysis, competitor analysis, prompt research, and mentions from ChatGPT, Google AI, Gemini and Perplexity. The page also shows that the feature set scales with the plan: one tier lists 25 custom prompts and daily AI rankings, while higher tiers advertise custom large-scale prompt tracking, multi-brand and multi-product visibility, and daily or weekly tracking. The page does not give us the complete weighting formula.

Those products may therefore count different events. A company can be named without its site being cited, while a competitor receives the citation. Estimated impressions can also give a common prompt more weight than a niche prompt.

Profound needs the same caution. The available evidence contains no public formula, score definition or quoted product page, so we cannot say what its number counts without guessing.

CiteGraph starts with the answer and the cited page. In the weekly research on who owns AI answers, our scan data records 45,309 citations across 7,260 AI answers. Those answers came from 60 scans of 41 sites in 44 categories, covering 442 distinct buyer questions. The engines cited 7,624 distinct domains and 16,506 distinct pages. We measure a sample; we print its limits; we do not know what we have not measured.

The score is a starting point for inspection.

How does your score reach a number?

Ask the vendor to show its ingredients before you compare the result with another tool. You need the prompt set, the engines, the observation window, the event being counted, and the comparison group. Ahrefs says Brand Radar uses prompt-based indexes across ChatGPT, Perplexity, Gemini, Copilot, AI Overviews and AI Mode. It also describes search-query-based indexes for AI Overviews and AI Mode, says question sets are retested monthly on a 90-day reporting window, and says each prompt is checked daily on every selected platform. That gives you a defined measurement shape, even though the final score still needs interpretation.

Semrush publishes a different shape. Its page lists custom prompts, AI rankings, brand analysis, competitor research and several tracking frequencies. It also advertises “AI visibility reports for any domain” and “custom large-scale AI prompt tracking” at different levels of the product. Those features describe what you can monitor, not a formula we can reproduce from the page.

Use this short check before you buy an AI visibility tool:

1. Write down whether the score counts mentions, citations, rankings, share of voice, impressions, or a mixture.
2. Check which engines and answer types are included.
3. Ask whether the prompts are yours, supplied by the vendor, based on search demand, or sampled from a pre-collected index.
4. Look for the reporting window and retest frequency.
5. Open the underlying answers and cited URLs for a small sample.
6. Record the page change you would make if the score falls.

If a vendor cannot expose the underlying answers, the score is difficult to challenge and harder to improve. It can still support monitoring, provided you know what its sample represents.

The evidence matters more than the neatness of the score.

The 442 buyer questions in CiteGraph’s sample are not every question a buyer could ask. The 41 sites are not every site in those categories. A movement in a score should therefore be read as movement in that vendor’s sample unless the vendor explains why a wider claim is justified. The measurement method guide gives you a practical way to inspect those choices.

Why does your Semrush or Ahrefs score disagree?

Disagreement is normal when products count different things. Ahrefs says its AI Share of Voice is the percentage of AI responses within a topic set that mention or cite your brand against competitors. Its Estimated Impressions metric weights potential visibility by the real search volume behind each prompt. Semrush’s public page foregrounds brand performance, mentions, prompt research, competitor analysis and tracking frequency, without giving us the weighting formula. A result from one dashboard therefore cannot be treated as a failed copy of the result from the other.

The same company can also produce different results inside one product. Change the prompt set and you change the questions. Change the engines and you change the retrieval path. Change the reporting window and you change which answer versions enter the sample. The CiteGraph leaderboards show why comparative results need their underlying question set and evidence beside them.

A concrete diagnosis from the sample looked at an anonymised pricing-comparison page. Before the edit, the page said only that readers could “compare plans and choose the right option”; it did not say who each plan suited or where either option fell short. After the edit, it added a sentence naming the intended customer for each plan and a sentence stating the main limitation of each one. The rerun returned the same page as the citation, but the cited passage now included the suitability and limitation facts. That is a useful result even without a dramatic score change.

The page change is the evidence.

Do not rewrite every page because a dashboard turned amber. Find the answer, the cited URL and the missing fact first. A brand mention may call for clearer product language. A missing citation may call for a page that answers the buyer’s question more directly.

How do you turn your AI visibility score into a page change?

Use the score to choose which answers to inspect. If your brand appears but your page is absent, study the pages that were cited and identify the question they answer. If your page is cited but the brand is not mentioned, check whether the page states the company and product relationship plainly. If competitors appear for a buyer question you care about, compare the page type and the evidence rather than copying the wording.

The change should be specific enough that another person can check it. “Improve AI visibility” is a reporting label. “Add the supported integrations and the plan limit to the comparison section” is a task. Keep the old answer beside the new one so you can tell whether the change affected the evidence or only the dashboard.

The free AI readiness check runs fifteen checks without an account and shows the fix for each. That is useful preparation, but readiness does not tell you whether an answer cited your page. A page can pass access checks and still fail to answer the buyer’s question.

CiteGraph’s paid plans make the workload explicit. Starter is $29 a month for 2 projects, 10 questions each, 3 runs per question on 4 engines, 6 tracked rivals and 30 written pages a month. Growth is $69 a month for 3 projects, 10 questions each on 4 engines, 10 tracked rivals and 60 written pages a month. ChatGPT Only is $19 a month for 1 project and 20 questions, asked on ChatGPT only. Weekly scans ask each question 3 times on every plan; a scan you start on Growth asks it 5 times. Each plan is paid from the day it starts and can be cancelled at any time, according to the CiteGraph pricing page.

Choose the evidence before choosing the score.

Semrush documents prompt and reporting features, but not a complete scoring formula in the material we have. Ahrefs documents mentions, citations, share of voice and estimated impressions. CiteGraph puts the answer and citation sample first, so you can trace a result back to a buyer question and a page.

This work is partly SEO, because access, useful pages and clear facts still matter. It is also measurement design. Once you look beneath the number, you can tell whether the next move is a page rewrite, a missing comparison, a technical fix, or a better-defined sample. The number has become a lead rather than a verdict.