AI visibility score: what it measures and what a good score is

What does an AI visibility score actually measure?

An AI visibility score is one number, usually out of 100, that a tool gives you about how your business stands with AI assistants such as ChatGPT, Gemini, Claude and Perplexity. There is no standard behind it. Each tool decides what the number counts, so a 70 from one tool and a 70 from another can describe different things. The same goes for a GEO score or an AEO score: the label does not tell you what was measured.

Most scores belong to one of two families.

• Readiness scores look at your pages: is a site or a page easy for AI to read, recognize and quote?
• Mention scores look at answers: they send questions to assistants and count how often a brand comes back.

The two can disagree, and both can be right. A new local business can have a clear website that no assistant names yet. A famous brand can be named often while its website is hard for bots to read.

Here is how a few tools define their number, read on each vendor’s own pages on November 4, 2026. These products change often, so check the current page before you rely on a line.

A FAQ on Semrush’s toolkit page also describes its score as mentions compared with the median of your top industry competitors. Some vendors run both families: Frase has its GEO Score for pages and a separate Visibility figure for answers.

What four questions should you ask about any score?

Because the definitions differ, a number from one tool cannot be compared with a number from another. Ahrefs says so about its own metric, AI share of voice, in a post dated October 22, 2026: “two tools can report different AI SOV for the same brand without either being wrong.” Four questions tell you what you are holding.

Score of what? One page, a whole site or a brand. Frase’s GEO Score rates a page. MWY’s rates a site. Semrush’s and HubSpot’s rate a brand, and a brand score can move with no change on your website.

Measured how? By rules or by answers, and if by answers, which ones: a model speaking from its training data, or responses captured from real requests? What counts as a hit differs too. Peec counts responses that mention the brand. Frase counts answers that “named or cited you”. Ahrefs weights each prompt by the monthly Google search volume of the keyword behind it, times a platform ratio, so a popular prompt counts for more.

Against which questions? For a mention score, the questions decide the result. Semrush says it sources prompts from AI search clickstream data and Google’s keyword dataset. Amplitude’s FAQ says its prompts are generated by AI models, not based on data from real users, because “OpenAI, Google, and other companies don’t share what users are searching.” HubSpot AEO and Frase work from the prompts tracked in your account. Ask how many there are, who wrote them, and whether they contain your name: a question with your name in it shows how you are described, not whether you are found. MWY’s score asks no questions at all.

On which date? A score is a reading, not a fact about your business. Semrush and HubSpot AEO say they refresh daily. A rules score can move too. We scanned the same sites with MWY’s checker on November 3 and 4, 2026, after changing how one item is judged. Eight sites gained 2 or 3 points from that rule change alone. One site lost 12 points because only its home page could be read that day. Another lost 2 because its server answered more slowly.

Keep one line next to every score you report or pass on:

[number] from [tool and plan], about [page, site or brand],
measured by [fixed rules or asked answers],
[how many questions, written by whom], on [date].

An illustrative example: 62 from a mention tracker’s starter plan, about our brand, measured by asked answers, 25 questions we wrote, on March 3. Compare two scores only when the whole line matches, which in practice means one tool and one setup over time.

What is a good AI visibility score?

There is no honest single number. We found no published benchmark that holds across tools, and the definitions show why: they do not count the same thing. HubSpot wrote in a post updated October 23, 2026 that “there’s no universal standard yet” for what good looks like. Ahrefs, on its own metric: “There isn’t a universally good AI share-of-voice score.”

Some tools publish bands for their own scale, and MWY’s report has cut points too. They do not travel. Here is a score of 70 on three scales.

Each cut point is the vendor’s own choice, ours included. A better question depends on the family.

• A mention score is good compared with the competitors you care about, on the same tool and the same questions. HubSpot says what counts as good “depends on how competitors perform on those same questions”.
• A readiness score is good when nothing blocks the bots and whatever is left is something you chose to leave. The list matters more than the number.
• Any score is most useful against itself. HubSpot calls its score “most useful as a baseline to track improvement over time rather than a fixed benchmark”, and Semrush says “Use your score as a baseline.”

What did 32 real websites score on MWY’s scale?

On November 4, 2026 we ran MWY’s fixed rules over 36 real business websites chosen by hand: local businesses, online shops, SaaS and blogs in the US, the UK, Canada, Australia, France, Germany and the Netherlands. 32 could be read. 4 refused the checker.

The set is hand-picked and small, so it shows a range and not a benchmark. Well-known brands and platform vendors are overrepresented, which pushes the numbers up. We have no figure for the typical small-business website.

The lowest score was 32, the lower quartile 73, the median 82, the upper quartile 90 and the highest 95. The middle three are the 8th, 16th and 24th of the 32, from the lowest.

• Nobody passed everything. Every site had at least 1 item short of a full pass, and the middle site had 5. A score in the 90s still comes with a list.
• The lowest score is not a clean reading. That site sent the checker to a checkout page instead of its home page, so the rules scored a page not written to be read. A low number is a reason to look at what was read.
• These scores say nothing about being named. They show how ready the sites are to be read. This round did not measure how often assistants name them, so a high score here is not evidence that it leads to being named.

A score under 70 on MWY’s scale put a site among 7 of these 32. Treat that as a sign that something basic needs fixing, not as a verdict.

Which items did most sites miss?

These items most often did not pass across the 32 sites, named as the report names them.

At the other end, nearly every site passed the basics: AI search bots are allowed in (31 of 32), AI training bots (31), Language is declared (31), HTTPS and short redirects (30), Not hidden from search results (30).

The pattern: these sites let AI in, but say less about who they are. The group “Does AI know who you are?” holds 37 of the 123 points in the score, yet it accounts for 414 of the 780 points the 32 sites lost between them. No site fully passed all three of business details, address and phone, and profile links.

Two cautions. The contact, headings, service page and questions items are signals, read from the words and structure of a page, so they can be wrong. And some misses are choices: a business that prefers not to publish contact details may leave that item as it is.

Which items carry the most weight in MWY’s score?

Every item has a weight. A pass earns all of it, a partial half and a fail nothing. An item that does not apply, such as a map link for a shop that only sells online, leaves the total. The score is the points earned divided by the weight of the items that apply, times 100. With all 24 applying, the total is 123 points. The methodology page lists every item and its weight.

The weights follow one idea: how directly an item decides whether an AI can read, identify and quote a site. Five items hold 52 of the 123 points.

• AI search bots are allowed in and Content is in the HTML, 12 points each: a bot that may not enter, or finds an empty page, gets nothing else your site offers.
• Business details as structured data and Not hidden from search results, 10 each: one tells machines who you are, the other keeps your pages in the search engines AI tools rely on.
• The firewall lets AI bots through, 8: a site can allow bots in robots.txt while its firewall still refuses them.

An llms.txt file sits at the other end, with 1 point, the least. It is a summary of a site written for AI, and it is worth having, but there is no evidence yet that the big assistants use it. Adding one moves a score by less than 1 point.

These weights are our judgment, not a measurement of what assistants reward, and we found no AI company publishing weights like them. And a sum can hide a blocker: a site that blocks AI search bots and passes everything else would add up to 90. So while robots.txt keeps out Google, Bing or two of the assistants’ search bots, the score stops at 49. None of the 32 sites hit this. Other blockers have no such limit, though none of the 19 sites at 80 or higher failed any of the first four items in the next table. Read the items, ordered by points gained, before you trust the number.

What should you fix first when a score is low?

Go by what blocks first, then by points for the effort. The points below are what a fix adds when an item goes from needs fixing to pass on a site where all 24 apply; a partial gains half. Your own report shows the exact gain. Who and effort are as MWY rates them.

First, open the door. In our round, all 7 sites under 70 had at least one of the first four items short of a pass. Part of that is arithmetic, since those items carry the most weight. It is also the sensible order: a page a bot cannot open gains nothing from better writing.

Then say who you are. These were among the most common gaps in our round, and MWY rates all three low effort.

Then write the answers. These take longer, and the numbers in them have to come from you.

Scan again after each fix. The same pages get the same result, so a change in the score comes from a change in what the checker read. Clear any site cache first. To work through the whole list by hand, see our guide to running an AI visibility audit.

What does an AI visibility report contain?

The phrase covers anything from a one-page grade to a weekly dashboard, so look for the same four answers in any report: what was scored, how, against which questions and when. MWY’s free report needs no sign-up, has its own link, and shows when the site was scored and how many pages were read. Pro opens soon, and adds the rest of the right-hand column.

The weekly answers belong to the other family of measure, so they are counted on their own and do not change the score.

The sample report is a real Pro report that the checker ran on mwyconsulting.com: the score, all 24 items, the fixes and the fix plan in order. Open it to see what an item looks like.

Where the figures come from

• The 32 sites: MWY’s own scan rounds of November 3 and 4, 2026, recounted from the raw result tables. The sites are not named.
• Semrush: knowledge base pages on its AI Visibility Toolkit and on where the toolkit’s data comes from, and its free AI visibility checker page.
• HubSpot: the post “AI visibility score: How to summarize your AI visibility” (updated October 23, 2026), the FAQ of its AI Visibility Tracking Tool page, and its AI Search Grader page.
• Amplitude: documentation pages “AI Visibility” and “AI Visibility FAQ”.
• Peec AI: help article “How is visibility calculated?” (updated October 29, 2026).
• Frase: the GEO Score checker page and the documentation page “AI Visibility”.
• Ahrefs: the post “AI Share of Voice” (October 22, 2026).

All vendor pages were read on November 4, 2026. They describe their inputs, but they do not all give a full formula or every weight; Semrush itself says “no platform can provide exact numbers on visibility”. And we found no AI company that publishes how it picks the businesses it names, so no score, ours included, can be checked against that.

What the MWY checker does and does not do with a score

The MWY AI visibility checker gives a readiness score. It reads the raw HTML of your home page and up to 7 inner pages, plus robots.txt, sitemap.xml and llms.txt, skips anything robots.txt closes to MWYGeoCheck, and scores 24 items with fixed rules and no AI in the scoring. The free report arrives in a few seconds with no sign-up.

It asks no assistant anything to make that score, so the score does not say how often you are named. Pro, which opens soon, adds the weekly answers and reports them separately. The checker cannot read a site that refuses it, as 4 of the 36 did in our round, and it says so.

It does not promise that an assistant will name your business: nobody outside the AI companies controls that. A high score says the site is ready to be read. What the checker gives you is a list of what to fix and how.