Why is my free score different from the paid one?

The free scan draws each prompt once and skips the web-grounded pass, so its share of voice rests on far less evidence and its interval is far wider. The eight sections and their weights are identical. The difference is sample size, and the report states it rather than hiding it.

Why did the score change when nothing on my site changed?

Answer engines are not deterministic. That is why a full audit draws every prompt seven times and reports an interval: a movement inside the interval is the engine talking to itself, and a movement outside it is a change worth acting on.

Do you sell the thing you score?

Yes, and that is a reason to publish the arithmetic rather than a reason to hide it. Every weight, threshold and formula on this page is the one the code uses. If a section looks wrong to you, the numbers are here to argue with.

Why does llms.txt not earn points?

SE Ranking looked at 300,000 domains and found no relationship between having the file and being cited. It may still be a useful directory. It is not evidence of visibility, so scoring it would be manufacturing a signal.

Can I see the raw answers?

The report shows a representative answer per prompt, which engine produced it, and whether the brand and which competitors were named. The finding is the answer, not a number derived from one you never see.

How the AI Visibility Score is measured

Our analyzer returns a number out of 100 and eight bars beneath it. This page is the arithmetic: what each bar weighs, what it is measured on, how wide the error bars are, and what the score does when a measurement fails. Everything here is read from the code that runs the audit, so a weight that changes in the scorer changes on this page in the same commit.

INITE / METHODOLOGY

What the number is measured on

Half the score comes from asking answer engines real buying questions and reading what they say. A visibility figure means nothing without the size of the sample behind it, so every report prints its own.

Prompts: Up to 20 per audit, written from the site itself: what it sells, who it serves, which problems it solves. Prompts that name the brand are dropped before the run, because a question containing the name guarantees the answer contains it.
Draws per prompt: Seven on a full audit, one on the free scan. Answer engines do not repeat themselves; a single draw cannot tell a real change from resampling.
Engines: Every engine with a configured key. The audit supports Claude, ChatGPT, Gemini and Perplexity, and the report states how many actually answered.
Two measurements: Model memory — does the engine recall the brand with no browsing. AI search — the same prompts with the engine's web-search tool on. They are reported separately because they are different facts.
The combined figure: When both ran, share of voice is 60% AI search and 40% model memory. Grounded answers are what a user sees today; recall still matters for the turns that never browse.
Failed cells: A prompt the provider refused or timed out on leaves the denominator and is counted separately. It is not scored as an absence of the brand.

The free scan runs the same eight sections but draws once and skips the web-grounded pass. Its number is a direction; a full audit's number is a measurement.

The eight sections and what they weigh

Weights are relative. A run that measured everything sums to 100; a run that lost a section renormalizes over what remained, and the section is shown with its reason rather than a silent zero.

Section	Weight	What it measures
AI Share of Voice	20	The share of tested prompts whose answer named the brand, averaged over prompts rather than over draws.The citation audit runs on every tier, so its absence is always a failure and never a tier-gated omission.
Retrieval Readiness	16	Whether the pages an engine needs in order to cite anything exist at all, and whether the homepage states plainly what the company is, who it is for and what makes it different.When the page classifier falls back to URL patterns, entity clarity was not measured; the section shrinks to what coverage alone can carry instead of charging the site for the failed call.
Search Presence	14	Where the domain ranks on Google for the same buying questions, and whether Bing has it indexed at all.Google and Bing are scored separately. Whichever one returned nothing leaves the section's scale rather than counting as an absence.
AI Identity & Crawlability	14	Whether retrieval crawlers are allowed in, and whether the page carries the structured data and metadata an engine reads to decide what it is looking at.Fetched directly from the site, so a failure here means the site did not answer.
Technical & Performance	12	Lighthouse over PageSpeed Insights: performance, SEO, accessibility and best practices.A PageSpeed outage is not a slow site. The section leaves the total instead of docking the score.
AI Sentiment	8	How the answers frame the brand when they do name it, judged by a model that is not the one that wrote the answer.No mentions means nothing to assess. Share of voice already prices that absence at full weight.
Mention Position	8	Where the brand falls among the tracked names in an answer, scored as 1/rank so first place is worth twice second.A brand that was never named has no position. Same fact as a zero share of voice, and it is already paid for there.
Competitor Gap	8	The share of the named slots that went to somebody else, over the prompts where anybody was named at all.When no competitor was named either, there is no displacement to measure. An answer that named nobody is not evidence of no displacement.
Total, when every section was measured	100

Only what was measured counts

The earlier version of this scorer treated a missing input as a zero. A PageSpeed outage silently cost a site twelve points, and a free scan — which skips whole sections by design — could not exceed 80 out of 100 no matter how good the site was.

The total now divides by the weight of the sections that actually ran. A full audit that lost PageSpeed and had no brand mentions to judge divides by 80, not by 100, and the report says which two are missing and why.

This is also why the report prints a coverage figure. A 72 measured over the whole model and a 72 measured over three quarters of it are different claims, and the reader is told which one they have.

One fact, one charge

A brand that appears in none of the tested answers is invisible in three of the eight sections at once: it has no share of voice, no position to rank, and every named slot went to a competitor. Scoring all three would charge one finding three times and drive an ordinary result to near zero.

So share of voice prices that absence at its full weight and names it as a critical finding. Sentiment, position and competitor gap report it and step out of the arithmetic.

The same rule runs the other way. An answer in which nobody at all was named is not proof that competitors are absent; it is the absence of a measurement, and an earlier version handed out full marks for it.

How sure the number is

Every headline figure here is a proportion estimated from a handful of prompts on one afternoon. Zero out of six prompts and zero out of six hundred print the same 0% and mean entirely different things, so share of voice carries a 95% interval and the report shows it.

The interval is a Wilson interval rather than the textbook normal one, which collapses to zero width exactly at rates of 0 and 1 — the most confident possible statement from the least informative possible sample.

Repeats of one prompt are not independent observations. Seven draws of twenty prompts is twenty looks at the world, not a hundred and forty, and the interval is computed over prompts for that reason. Treating draws as independent would make it about two and a half times too narrow.

Confidence is the weaker of two readings: how much of the model was measured, and how wide the interval on share of voice came out. Broad coverage does not get to hide a sample of six.

What the confidence label means

High: At least 90 of the 100 weight points were measured, and the interval on share of voice spans no more than 15 points.
Medium: At least 70 points measured, interval no wider than 30 points.
Low: Anything below that. The number is a direction, and the report says so rather than dressing it up.

Google — AI features and your website (the guidance this audit does not try to outsmart)
SE Ranking — 300,000 domains, no relationship between llms.txt and AI citation frequency
seoClarity, October 2025 — 362,000 queries: 56% of AI Overview citations come from the top 20 organic results
Schulte et al. 2026, and the 2026 critical survey of generative-engine optimization, on intervals and on clustered generations

From points to a verdict

The word beside the number comes from bands, not from a single finding.

70 and above: Looking good
45 to 69: Needs attention
Below 45: Critical issues found

Two or more critical findings demote the verdict by one band. One finding does not: a single problem should not flip a 90-out-of-100 site to critical, which is what the earlier rule did.

Inside AI Identity

Fetched from the site directly and scored out of 100, which then becomes the section's share of the composite.

25AI crawler accessWhether the retrieval crawlers can read the site, parsed from robots.txt properly: user-agent groups with their own rules, most specific group wins, otherwise the * group.
25Structured dataJSON-LD or microdata, and whether it carries an Organization or WebSite entity rather than only page-level markup.
15Social & AI previewsThe tags an engine and a social card both read to build a preview.
15Meta descriptionPresent, and not the same one repeated across the site.
10Robots & sitemapA robots.txt that exists, a sitemap that resolves, and ideally the sitemap declared in the robots file.
10Technical SEOCanonical URLs, and hreflang where a site publishes in more than one language.

The crawlers that are scored, and what each is worth

Retrieval crawlers decide whether the site can appear in an answer. Their weights sum to the 25 points crawler access carries.

6BingbotPowers Bing search — the index behind ChatGPT browsing and Copilot answers.
5OAI-SearchBotBuilds the ChatGPT Search index.
4PerplexityBotIndexes content for Perplexity answers.
3ChatGPT-UserFetches pages live when ChatGPT users open or cite links.
3Claude-SearchBotBuilds the Claude search index.
2Perplexity-UserFetches pages live on behalf of Perplexity users.
2Claude-UserFetches pages live when Claude users open or cite links.
25Allowed earns the full weight. A path-scoped partial allow earns 70% of it. Blocked earns nothing.

Model-training crawlers — GPTBot, ClaudeBot, Meta-ExternalAgent, Bytespider and their kin — are reported and carry no weight at all. Blocking them is a legitimate business decision with no cost to visibility, and a score that punished it would be pushing an opinion rather than taking a measurement.

Inside Retrieval Readiness

Two questions. Do the pages an engine would want to cite exist, and does the homepage say plainly enough what this company is for a model to place it in a category.

Archetype coverage is worth 60 of the section's 100 points. Entity clarity, itself scored out of 100, contributes the remaining 40.

The eleven page types

A homepage that states who you are, what category you are in and who you serve
A page per product or service, with what it does and what it is for
Pricing, plans, or an explicit pricing model
Pages for specific problems or industries served
Comparisons and alternative-to pages
Structured question-and-answer content
What the product works with
Team, mission, certifications, testimonials
Articles, guides and original research
Clear definitions of the terms in your field
Case studies, benchmarks, original data

The five entity signals

Each is judged strong, weak or missing. Strong earns 20 points, weak earns 8, missing earns none — so five strong signals make 100 and five weak ones make 40.

What category the company is in, stated rather than implied
Who the product or service is for
What makes it different from the alternatives
Visible proof: numbers, case studies, references
What it actually does, and how it works

20 / 8 / 0 · 60 + 40

Inside Technical & Performance

One PageSpeed Insights run against the site, weighted: performance 40%, SEO 25%, accessibility 20%, best practices 15%.

A category PageSpeed did not return leaves the average rather than scoring zero, and the divisor is the surviving weight. A response carrying no Lighthouse category at all is treated as no measurement, not as a site that failed every check.

Inside Search Presence

The same buying questions, run against Google and checked against Bing's index.

Google is worth 10 points: 70% of them for the share of queries where the domain appears in the top 10, 30% for the share where it appears in the top 3. Bing is worth 4, for being indexed at all.

The denominator is the queries the provider actually answered, not the queries that were asked. A site is never told it is absent from a result nobody read.

What this does not measure

The honest boundary of the number.

It is not a ranking. There is no index of every company against which a position could be given; the score describes one site against the prompts its own content implies.
It is a moment, widened by repetition. Seven draws separate a real change from an engine's own variance, and they do not turn one afternoon into a trend. That is what re-running it does.
It sees no personalization. Answer engines vary their answers by account, history and location, and an audit sees one unauthenticated view.
Prompts come from the site. A site that describes itself vaguely gets vaguer prompts, and its score reflects that honestly rather than compensating for it.
It does not score AI convention files. ai.json, llms.txt and their relatives are reported as emerging standards and carry no weight, because the evidence for their effect on citation does not exist yet.
It does not score training-crawler policy. Opting out of model training has no measured effect on being cited.

FAQ

Why is my free score different from the paid one?: The free scan draws each prompt once and skips the web-grounded pass, so its share of voice rests on far less evidence and its interval is far wider. The eight sections and their weights are identical. The difference is sample size, and the report states it rather than hiding it.
Why did the score change when nothing on my site changed?: Answer engines are not deterministic. That is why a full audit draws every prompt seven times and reports an interval: a movement inside the interval is the engine talking to itself, and a movement outside it is a change worth acting on.
Do you sell the thing you score?: Yes, and that is a reason to publish the arithmetic rather than a reason to hide it. Every weight, threshold and formula on this page is the one the code uses. If a section looks wrong to you, the numbers are here to argue with.
Why does llms.txt not earn points?: SE Ranking looked at 300,000 domains and found no relationship between having the file and being cited. It may still be a useful directory. It is not evidence of visibility, so scoring it would be manufacturing a signal.
Can I see the raw answers?: The report shows a representative answer per prompt, which engine produced it, and whether the brand and which competitors were named. The finding is the answer, not a number derived from one you never see.

Run it against your own site

The free scan is the same eight sections, drawn once. No account is needed to see the score.

Start the free scan