Method and limits
What the SEO & AI visibility checker measures, and what it cannot.
This page is the contract behind every report. Nothing in a report comes from anywhere but the five steps below.
- Ruleset2026-09-27
- Question protocol2026-09-17.1
- Crawler registry2026-09-13
How it works
One website in, one report out. This is what happens in between.
Auto-plays. Pick a station to stop, or read the full steps below.
Step 01
A bounded read of the public website
The scanner fetches the homepage, then up to eleven more pages chosen by role: location, classes, pricing or intro offer, timetable or booking, contact, about. It identifies itself as DaniyalVisibilityBot and obeys robots.txt rules written for it.
- Pages
- Homepage + up to 11
- Requests
- 30 at most, 2 at a time
- Wait
- 10 s per page, 45 s in total
- Size
- 2 MB per page
- Redirects
- 3 at most
- JavaScript
- Not run
It does not log in, submit forms, open booking flows, or read anything inside embedded frames.
What it records, with the exact excerpt and location as evidence
Google-facing basics
HTTPS and redirects, response time of the check’s fetch, viewport tag, titles and descriptions, canonical tags, noindex directives (meta and header), sitemap, robots.txt.
What a prospective member needs
Studio name and city, street address, class types named in the text, beginner information, an introductory offer and its price, a timetable link, a booking entry point, contact options.
Structured data
JSON-LD nodes and the properties present on the business entity. Microdata and RDFa are not evaluated.
Crawler access
Whether robots.txt blocks the search crawlers behind Google, Bing, ChatGPT search, Claude and Perplexity, shown separately from training crawlers such as GPTBot and ClaudeBot. Blocking a training crawler is never counted as a visibility problem.
Load speed
Google's PageSpeed Insights test on the homepage, on a phone and on a computer: the Lighthouse speed score and its measurements, Google's suggested fixes, and real Chrome visitor data from the last 28 days when Google has enough of it. One test run; scores vary between runs.
Coverage notes
Sparse HTML (rendering not tested), embedded booking widgets, platform hints, pages that could not be fetched.
Every finding carries one of five statuses
- ObservedFound in the sampled HTML.
- IssueFound, and it works against you.
- Needs reviewAmbiguous; a person should look.
- Not measuredOutside what the scanner can see.
- Not applicableDoes not apply to this site.
Step 02
Fixed questions to AI assistants
From the pages the scanner derives a profile: trading name, city, category, and the site's language. If it cannot establish those with confidence the report says so and the AI questions are not run until they are confirmed.
The four templates
Filled in, in the site's language, and sent unchanged. The two unbranded questions never contain the studio's name or domain.
- 1UnbrandedDiscovery
Which [category] studios are there in [city]?
- 2UnbrandedBeginner question
Where can I do [activity] as a beginner in [city]?
- 3By nameLocation and classes
Where is [studio] in [city] and what classes do they offer?
- 4By namePrice and booking
What does a first class at [studio] in [city] cost and how do I book it?
Who is asked, and how it is counted
Providers
OpenAI's Responses API with its web search tool, and Anthropic's Messages API with its web search tool. Both receive an approximate location set to the studio's city and country.
Repeats
Each unbranded question is asked three times per provider in fresh conversations; the two by name once. Three runs show variation; they do not estimate a stable probability, and answers can share cached retrieval.
What counts
A run counts only if the provider actually executed a web search. Runs that failed, timed out, or answered from memory are listed but excluded from the counts.
- Named in x of n
- The studio's name or a close alias appears in the final answer.
- Website cited in x of n
- The provider's citation metadata points to the studio's own host.
- Names appearing as options
- Other business names that appear as list items. Shown as such, not as a competitor ranking.
Step 03
Actions and scores
Recommendations come from a fixed catalogue. Each action has a rule that must match specific findings before it can appear, so no action is shown without evidence.
- Findings match a ruleAn action appears only when the specific findings its rule requires are present in your report.
- Wording is fixedEach action is shown in its catalogue wording, in English or Dutch. No AI model rewrites it.
- Scores follow the findingsThe Google and AI scores and what each fix adds are calculated from the same findings, by fixed rules. Nothing is estimated by a model.
Step 04
What the check cannot know
A report is honest about its edges. None of the following is measured, guessed, or implied.
- Google rankings, impressions, clicks or the list of indexed pages.Those need Search Console access from the owner.
- Google Maps positions and the private side of your Google Business Profile.Views, calls, direction requests, posts. The public listing (review count, rating, website, hours) is read through Google Places and compared with the nearest same-service studios; that comparison set is not a ranking.
- Google AI Overviews and AI Mode answers.There is no public API, and the check does not scrape them.
- What a person sees in the consumer ChatGPT or Claude apps.The APIs share the search tools but not the apps’ instructions, memory or personalisation. Reports say “measured via API”.
- Perplexity and Gemini answers.Not included in this version.
- Traffic, bookings, conversion, revenue.Or whether any change caused any result.
- Content rendered by JavaScript or inside embedded widgets.Nor whether a crawler that is blocked actually obeys the block.
- More than one language at a time.A report runs in one language; Dutch and English answers are separate experiments.
Step 05
Cost, identity and data
Cost
A full check costs roughly half a dollar to a dollar in provider fees, paid by me.
Identity
The crawler's identity and how to block it are on the crawler page.
Data
What is stored, for how long, and who processes it is on the privacy page. Reports are private links that expire.
Appendix: crawler registry 2026-09-1316 agents
| Agent | Vendor | Purpose |
|---|---|---|
| Googlebot | search, Google Search, including AI Overviews and AI Mode, uses Googlebot-crawled content. | |
| Google-Extended | training, A permission token for Gemini training/grounding; does not affect Google Search inclusion. | |
| Bingbot | Microsoft | search, Bing index also feeds several AI assistants. |
| OAI-SearchBot | OpenAI | search, Surfaces sites in ChatGPT search results. |
| ChatGPT-User | OpenAI | user |
| GPTBot | OpenAI | training |
| Claude-SearchBot | Anthropic | search, Indexes content for Claude search quality. |
| Claude-User | Anthropic | user |
| ClaudeBot | Anthropic | training |
| PerplexityBot | Perplexity | search |
| Perplexity-User | Perplexity | user |
| CCBot | Common Crawl | training |
| Applebot | Apple | search, Siri and Spotlight suggestions. |
| Applebot-Extended | Apple | training |
| meta-externalagent | Meta | training |
| Amazonbot | Amazon | search, Alexa answers. |
Sources: OpenAI bots, Google special-case crawlers, Perplexity crawlers, Google: AI features and your website, Schulte et al., "Don't measure once" (2026) on repeated measurement of AI search visibility.
FAQ
Questions about the method
Which AI assistants are asked?
OpenAI's Responses API with its web search tool, the model family behind ChatGPT, and Anthropic's Messages API with its web search tool, which is Claude. Both are given your studio's city and country as an approximate location. Answers can differ from the consumer apps, which have their own instructions, memory and personalisation, so reports say "measured via API".
Which questions are asked, and how often?
Four fixed templates, sent unchanged in your website's language. Two are unbranded and never contain your name: which studios of your category there are in your city, and where a beginner can start. Two are by name: where you are and what classes you offer, and what a first class costs and how to book. The two unbranded questions are asked three times per assistant, the two by name once, each in a fresh conversation: sixteen answers in all.
What does "named in 2 of 3" mean?
In two of the three runs of that question, the answer contained your studio's name or a close alias. "Website cited" means the assistant's citation metadata pointed to your own site. Three runs show variation; they are not a probability, and answers can share cached retrieval.
What can the check not tell me?
Google rankings and Google Maps positions, impressions or clicks, the private side of your Google Business Profile (views, calls, direction requests; the public listing and its review count are read when the local stage is on), Google AI Overviews and AI Mode answers, what a person sees in the consumer ChatGPT or Claude apps, Perplexity or Gemini answers, traffic, bookings or revenue, and anything rendered only by JavaScript or kept behind a login.
Is there a score?
No. Every finding carries one of five statuses with the exact excerpt and its location as evidence. Missing evidence is reported as "not found in the sampled HTML", never as proof that the studio lacks something.
How is this different from an SEO audit?
An SEO audit is about ranking on Google. This check is about what AI assistants say when asked, and whether the facts they need are readable on your site in plain text. It records the Google-facing basics it can see, such as titles, canonical tags and robots.txt, but it does not measure rankings.
What does a check cost, and who pays?
Roughly half a dollar to a dollar in provider fees for a full check, paid by me. For the studio it is free.