AI Visibility Audit: What We Check and What You Receive

Machine readability is the property of a page whose main content, headings and entity references resolve from the raw HTML response, with no JavaScript execution required. Crawler access is the permission layer above it: the robots rules, server responses and edge settings that decide whether an assistant's agent retrieves the page at all. The audit tests access first and readability second, because a blocked page never reaches the stage where its markup matters.

Five categories hold the 40 checks: 7 on crawler access, 14 on rendering and machine readability, 8 on structured data, 6 on entity clarity, and 5 recording current assistant mentions. Every check is numbered and every client receives the same 40, so a finding can be requested back by its number. A check returning a measured value records the number itself — the status code, the render difference, the count of entity references — rather than a verdict.

Crawler access

Crawler access is tested first because a blocked page never reaches the rest of the audit. Six agents are requested by name, and the status code each one receives is recorded beside the URL that returned it.

• Robots directives covering the retrieval agents OAI-SearchBot, PerplexityBot, Claude-SearchBot and meta-externalagent, and separately the training agents GPTBot and ClaudeBot
• Edge rules and bot-protection settings that refuse an AI agent while admitting Googlebot
• Server responses on the homepage, the service templates and the sitemap
• Redirect chains that consume a crawl before it reaches a live URL

Each fail is recorded with the exact directive or firewall rule that causes it, which reduces the repair to one edit.

Rendering and machine readability

Rendering is the step where a browser executes JavaScript and assembles the visible page. Machine readability is what survives when that step never happens, and AI crawlers largely do not execute JavaScript. Each template is fetched twice, once with scripting enabled and once without, and the two results are compared.

• Main content, checked for presence in the raw HTML response
• Headings, checked for a single H1 and an unbroken H2 order
• Navigation links, checked for presence as anchors rather than script handlers
• Boilerplate ratio, recorded as main content measured against template chrome

A page that assembles only after scripting is recorded as a fail, with the element that depends on the script named beside it.

Structured data and entity clarity

Structured data is the JSON-LD layer that states in machine terms what a page is about. Entity clarity is the degree to which a business resolves to one identity across its own site and the wider web: one name, one address, one telephone number, one set of external references. Assistants resolve an entity before they cite it.

• Schema types present, validated against the page's visible content
• Organization identity, checked for a name, address and telephone string matching the visible page byte for byte
• External references, checked for a sameAs set pointing at resolvable records
• Wikidata identifiers, checked for presence — zero of the nine competitor pages audited in this metro carry one

Markup describing something the page does not show is recorded as a fail, because an unsupported claim damages the identity it asserts.

Current assistant mentions

Current assistant mentions establish the citation baseline: a record of which domains seven assistants name today in response to a fixed prompt set. Seven assistants are queried — Google AI Overviews, ChatGPT, Perplexity, Gemini, Claude, Grok and Meta AI. The baseline enables later comparison, because a citation gained is provable only against a citation count recorded first.

• Prompt set, fixed at the audit and reused unchanged in every later run
• Cited domains, logged per assistant per prompt
• Competitor domains, logged where they hold a citation this business does not
• Factual errors, logged where an assistant states something wrong about the business

The baseline is a single capture rather than a trend, and the monitoring engagement repeats the same prompt set against it.

A prioritised repair list is a numbered sequence of fixes ordered by effect over effort rather than by discovery order. Effect is the reach of a fix across pages and surfaces; effort is the working time and the access it needs. A robots directive unblocking a whole domain sits above a schema property on one template.

The findings document contains four parts: the failed checks with their evidence, the repair for each, the ordered sequence, and the citation baseline table. Every value in that table is restated as a sentence, so the document reads correctly as plain text.

Sample page from a findings document. Each entry line carries five elements: check name, URL tested, evidence, repair, and rank.

You own the document. In-house repair is a supported outcome, and the same order runs inside the delivery stages we perform ourselves.

Ten business days is the promise attached to the deliverable, counted from the day the request arrives, not from a kickoff call. Three passes fill that window.

1. Fetch and validate — crawler requests by agent, scripted and unscripted renders, schema validation
2. Query and log — the seven assistants against the fixed prompt set, every cited domain recorded
3. Prioritise and write — failed checks ordered by effect over effort

The prompt runs set the pace: the seven assistants named above are queried one at a time, each answer recorded by hand.

Three limits sit outside any auditor's control. Competitor behaviour changes the citation set continuously, because an assistant names the sources available at the moment of the query. Index refresh timing governs when a repair becomes visible, and that clock belongs to the search index. Market demand sets how often a prompt is asked at all.

One further boundary deserves naming. A model that names a business with no web search running is drawing on training data rather than on retrieval, and no change to a website produces that. The audit measures retrieval and extraction — the two layers a website controls — and reports the third rather than selling it.