Website Intelligence
What happens after you press Analyse
Eight stages, roughly a minute, and a deliberate separation between what code determines, what a model interprets and what a person decides.
Most automated website tools are opaque about what they do. You enter an address, a progress bar advances, and a report appears. Whether anything was actually fetched, what was measured against what was assumed, and where a language model filled a gap are all left to inference.
That opacity is a problem when the output is going to inform spending decisions. This page describes the process precisely enough that you could argue with it.
Sequence
The eight stages
The loading indicator on the homepage names the stage currently running. It reflects genuine progress rather than an animated approximation of it.
Preparing website
Your input is normalised into a canonical address. Tracking parameters are stripped, the protocol is resolved and the format is validated. The target is then resolved and checked against private, loopback, link-local and reserved address ranges; anything that resolves inside a private network is rejected before a request is made.
Reviewing public pages
The home page is requested with a timeout and a size limit. Redirects are followed manually, with the same address checks repeated on every hop. Only textual content types are accepted. A small number of further pages are then requested, discovered from internal links and the sitemap.
Analysing search signals
The returned HTML is parsed without executing any JavaScript. Titles, descriptions, canonical declarations, heading structure, indexing directives, link relationships, image attributes and structured data are extracted as facts. robots.txt and the sitemap are read and checked for consistency.
Assessing GEO readiness
Entity signals are examined: whether the organisation is formally described in structured data, whether services are individually identifiable, whether expertise is attributed to named people, and whether content is arranged so a specific passage can answer a specific question.
Checking performance
Performance is measured through Google PageSpeed Insights for both mobile and desktop where the service is available. Results are cached to avoid repeated calls. If it times out, the analysis continues without it and the report records that the measurement is missing.
Reviewing domain health
Where the relevant modules are selected, public DNS records are resolved, email authentication records are read, response headers are examined and public blocklist zones are queried. These are read-only lookups against published information.
Interpreting findings
Deterministic rules run first and produce findings and scores from the evidence alone. Where an interpretation model is configured, it receives the structured evidence and contributes the executive summary, prioritisation and judgements about clarity and positioning. Its output is schema-validated and filtered.
Preparing your snapshot
Findings are merged, ordered by business impact, and filtered according to what can responsibly be shown at the current verification level. Scores are assembled with their explanations, and the snapshot is rendered.
If the analysis cannot complete a check, it says so. A report that quietly omits what it could not do is worse than one that admits a gap.
Division of labour
What code decides and what a model decides
A language model asked whether a canonical declaration exists will usually be right, and occasionally confidently wrong. Code asked the same question is right every time. So the model is never asked.
Facts are established deterministically before the model is involved at all. What it receives is a structured summary of established evidence, and what it is asked for is judgement: whether the positioning is coherent, whether a first-time reader would understand the offering, which of eleven problems to address first.
Its response must then conform to a strict schema. Anything claiming guaranteed rankings, guaranteed citations, confirmed penalties, exploitable vulnerabilities or active compromise is removed. Scores are clamped to the range the evidence supports.
Determined by code
Presence of titles, descriptions and canonicals. Response status. Indexing directives. Heading structure. Structured data types. Link and image counts. DNS records. Authentication records. Response headers. Performance measurements.
Interpreted by the model
Clarity of the offering. Coherence of positioning. Whether content would be useful to its intended reader. Entity clarity. Trust perception. Priority order. Plain-English explanation of what a finding means commercially.
Decided by a person
In paid engagements: whether the interpretation is right, what the sequencing should be given your constraints, and which findings are worth acting on at all.
Constraints
What the analysis will not do
These are enforced in the fetching layer rather than left to policy, which means they hold regardless of what is entered.
Reach private networks
Addresses are resolved and classified before any request. Private ranges, loopback, link-local, reserved space and cloud metadata endpoints are rejected, and every redirect is re-checked.
Attempt authentication
No credentials are submitted, no login form is completed and no authenticated area is entered. If a page requires a session, it is recorded as inaccessible.
Probe for vulnerabilities
No port scanning, no service enumeration, no payloads and no attempt to trigger error conditions. Security findings come from reading responses your server already sends.
Crawl exhaustively
A handful of pages, requested sequentially. This is deliberately not a site crawler; broad crawls belong in engagements where they have been agreed.
Download large responses
Every request has a byte cap and a timeout, and only textual content types are accepted. Binary responses are discarded rather than processed.
Publish sensitive detail
Exact administrative paths, origin addresses and specific software versions are never shown in a public snapshot, regardless of what was observed.
Frequently asked questions
How many pages do you fetch?
Will this affect my server or my analytics?
How long does the snapshot take?
What does the language model actually see?
Why is some of the report hidden?
Can I run this on a site I do not own?
Watch it run
The loading sequence names each stage as it happens. It is a reasonable way to confirm that this page describes what actually occurs.