Methodology: how I measure AI visibility
This page documents exactly what I measure and how, when I check what ChatGPT and Gemini say about a business. Almost nobody in this market publishes their method. I do, because a measurement is worth exactly as much as it can be verified. Here are the seven dimensions, the weights, and how a probe runs — including where this very site is still weak.
I measure two different things and keep them strictly apart. One is readiness — the seven dimensions below, largely in the owner's hands. The other is recommendation: whether the models actually say your name to a real buyer question. The latter is partly decided outside your business, and it cannot be promised. Anyone who conflates the two is selling something that isn't theirs to sell.
The seven dimensions
Where the measurement is machine-run, the row carries an [AUTO] mark: those checks run as plain requests, without a browser, and are cheaply and traceably repeatable.
| Dimension | Weight | What it checks | Verifiable? |
|---|---|---|---|
| D1 · Reachability & crawlability | 20% | Can the AI crawler reach the site, and is the content visible without JavaScript. | [AUTO] |
| D2 · Structured data | 15% | Does the site identify itself machine-readably: Organization, sameAs links, FAQ markup. | [AUTO] |
| D3 · Answer-ready content | 15% | Can a standalone, quotable answer be lifted from the first sentences after a heading. | [AUTO] |
| D4 · Entity & NAP consistency | 15% | Do name, address and phone match across the site, the markup and the directories. | [AUTO] + manual check |
| D5 · External presence | 25% | Is the business talked about elsewhere: reviews, directories, independent mentions, press. | Partly machine, partly manual |
| D6 · E-E-A-T signals | 5% | Is credibility visible: HTTPS, named author, legal pages, cited sources. | [AUTO] |
| D7 · Content depth | 5% | Is there substantive, on-topic content and internal linking between key pages. | [AUTO] |
The heaviest weight sits on external presence — deliberately. Most of the checks, though, are on your own site, because that is what an owner can fix fastest. The weights aren't arbitrary: they lean on correlation studies and industry consensus, and I review them quarterly because crawler names and model behaviour change fast.
The most important sentence on this page. A high readiness score means the AI can find and read your site — not that it will recommend you. Recommendation is dominated by external presence (D5), review volume and brand age, and is largely decided off your website. A well-known clinic surfaces even with a weak site, because it has been present in many places for years; an immaculately built new site stays invisible if nobody talks about it elsewhere. Technical readiness creates eligibility, not victory.
The data backs this. A large-sample analysis found external brand mentions are the strongest predictor of whether an AI cites a company — stronger than backlinks (see Ahrefs' correlation study). A report examining seven thousand citations found a large share of named sources are Wikipedia and forums, not companies' own websites (see Digital Bloom's 2025 citation report). That's why I never promise "the AI will recommend you." The honest frame is different: your competitors are visible to AI by accident — I make you visible on purpose.
How a measurement runs
Readiness is scored from the site and its public signals along the seven dimensions. Recommendation I test the way a buyer would: I put the same question — say, "which is the best dental clinic in this town?" — to several AI models, and record who gets named. Every answer is dated and stored, so it can be re-examined later.
Never a single model, because that would distort the picture. Each model answers differently: one recalls a known name from its training data, another assembles results from live search, a third leans on reviews. Only the multi-model picture is honest. A single screenshot misleads easily — a dated, multi-model measurement does not.
My own dated probe from May 2026 was unambiguous: I put identical buyer questions to four publicly available models — 48 queries at national, district and city level. Of fifteen Budapest dental clinics tested at local level, not one was reliably recommended by name by any model. The models mostly listed differing names, often disagreeing with each other.
That is not a bug in the models. It is the starting position for most SMEs. SOCi's 2026 Local Visibility Index found ChatGPT recommends only 1.2% of local businesses (Gemini 11%, Perplexity 7.4%, vs 35.9% in Google's local pack), and that business-profile accuracy on AI platforms runs at 68% on ChatGPT and Perplexity versus 100% on Maps-grounded Gemini (source: SOCi LVI, 2026). The named sources are mostly third-party: Wikipedia alone is 47.9% of ChatGPT's citations, forums nearly half of Perplexity's (Digital Bloom). And language matters: sites that translated their content earned up to 327% more AI citations on searches in the added language (source: Weglot's language study).
This site holds itself to the same rubric: a live, dimension-by-dimension self-scorecard — including the honestly weak external-presence ring of a newborn site — is published on the Hungarian methodology page and updates from the build pipeline, not by hand.
Why publish all this?
An AI-visibility score is worth exactly as much as its owner can verify. If I hid the weights, the checks and the probe procedure, you'd be buying a number on faith. Instead I write down what I look at, so anyone can check it — and argue with it. It binds me, too: if my own score drops, the page prints the drop. A claim your customer can disprove in minutes would poison not just the measurement but everything after it.
