AI Visibility Audit

Classic SEO asks whether a site ranks. This asks something different: when a buyer puts a question to ChatGPT, Claude, Perplexity or Google's AI, does the engine reach the site at all, can it read what is there, and does it cite the site in the answer?

Those are three separate failures with three separate fixes, and most sites fail the first one without knowing it.

When to use this

Trigger on any of these, however they are phrased:

• "Why doesn't ChatGPT know about us?" / "AI never recommends my business"
• A request for a GEO, AEO, AI-SEO, LLM-SEO or AI-visibility audit
• "Can AI crawlers see my site?" / "Are we blocking GPTBot?"
• "Where does AI get its answers about [industry]?"
• Organic traffic dropping and AI answers suspected
• Someone about to invest in content and wanting to know if AI can read it

Do not use it for conventional rank tracking, backlink audits or keyword research. It measures AI retrieval and citation, not Google's blue links.

How to run it

Option A — MCP tools, if the connector is available

The hosted server exposes three tools:

role accepts Real Estate Agent, Attorney, Founder or Other and tilts a local-service business toward local entity signals, reviews and directories.

Only pass email when the person has asked for the full report. It emails them and registers them as a lead. The snapshot alone needs no email.

Connect the server with:

claude mcp add --transport http specularis-audit https://specularis-mcp-audit-production.up.railway.app/mcp

Option B — direct HTTP, no setup

The same scan engine is reachable without the connector and without a key:

curl -s "https://specularis-mcp-audit-production.up.railway.app/api/scan?d=example.com"

Pass a bare hostname, no scheme and no path. It returns the full structured scan described below. Use this when the MCP connector is not configured, or when scanning several domains in a loop.

Reading the result

Work the fields in this order. Later numbers are meaningless if an earlier one has failed.

1. Did the scan actually complete?

If inconclusive is true, report no score at all. Say the scan could not complete and give inconclusiveReason. A number produced from a failed fetch is worse than no number, because the client will act on it.

2. Can the crawlers get in?

bots.GPTBot        { status, words, served, robotsBlocked, challenged }
bots.ClaudeBot     { ... }
bots.PerplexityBot { ... }

Each bot is fetched with its real published user agent, so this is what the engine itself would receive.

• robotsBlocked: true — robots.txt disallows that agent. Deliberate, usually by a developer who treated AI crawlers as scrapers.
• served: false or a 4xx/5xx status — the origin refused the bot.
• challenged: true — a bot-protection layer returned a challenge page.

Verify any challenged: true before telling a client they are blocked. Challenge pages produce false positives: a WAF may challenge one request and pass the next, and a scan that reports blocking on a site that is actually open destroys trust in everything else in the report. Re-run the scan, and check the site's own WAF rules, before putting "you are blocking AI" in writing.

If a bot genuinely cannot reach the page, stop. Nothing downstream matters. This is the single highest-value finding the audit produces.

3. Can they read what they get?

• words — how much text the crawler received.
• browserWords — how much a real browser receives.
• renderGap: true — the browser sees substantially more than the bot does.

A render gap means the content exists but is assembled by client-side JavaScript, so the crawler receives a near-empty shell. The page looks perfect to the owner and is blank to the engine. Common on JS site builders and SPA frameworks. The fix is server-side rendering or prerendering, not more content.

A words count in the low hundreds with served: true is the same failure in a milder form.

4. Is the entity legible?

• uniq — the JSON-LD @type values found (note: the parser reads inside @graph, so nested schema is counted).
• validSchema, hasOrg, hasPerson, hasAddress
• llmsPresent, llmsUseful, hasSitemap

Schema is how an engine resolves "this page" into "this organisation, in this city, run by this person." Without it the engine has prose it must guess from, and it frequently guesses wrong — resolving a brand name to an unrelated dictionary or historical meaning.

5. The score

pillars   { access, entity, content, offsite, freshness, technical }
pillarMax { access: 20, entity: 15, content: 15, offsite: 30, freshness: 10, technical: 10 }
total, grade, pct

Read the pillars, not the total. The total hides which lever to pull.

Note that offsite is the largest pillar at 30 points — larger than access and entity combined. That reflects how the engines behave: they cite directories, review platforms, trade bodies and forum threads far more often than they cite a company's own marketing pages. A site can be technically flawless and still be invisible because nothing else on the web corroborates it.

Finding what AI cites instead

find_ai_citations(query, domain) takes a question a buyer would actually ask ("best personal injury lawyer in Miami") and returns the sources the engines cite when answering it, each tagged with which engine cited it, plus an appears_in X of N count for the domain.

This is usually more persuasive than the score, because it converts an abstract problem into a named list: these fourteen pages are where your buyers' answers come from, and you are on two of them. That list is the off-site work plan.

Engines are queried according to which API keys the server has configured — Perplexity, Claude, Gemini and ChatGPT are each supported and each gated on its own key, so the engines array in the response tells you which ones actually ran. Report only those. There is a shared daily cap across all users; when it is reached the tool says so rather than returning partial results.

Turning findings into a fix order

1. Unblock the crawlers. robots.txt and WAF rules. Cheapest fix, biggest effect, often same-day.
2. Close the render gap. Server-render the content the crawler is missing.
3. Add entity schema. Organization, Person, LocalBusiness, with sameAs pointing at profiles that corroborate the same facts.
4. Get corroborated off-site. The citation list is the target list. This is the slowest and highest-leverage work.
5. Keep it fresh. Stale pages lose to maintained ones.

Resist reordering by effort. A beautifully structured page behind a blocked robots.txt scores zero.

Honest limits

State these when they apply rather than letting someone discover them later:

• A snapshot reads robots.txt, the homepage and llms.txt. It is not a crawl of the whole site.
• Engine answers are not deterministic. A citation check is a sample at a point in time, not a fixed ranking.
• Challenge pages cause false "blocked" readings. See step 2.
• The citation finder depends on which engine keys the server has configured.
• The audit runs from a URL. It does not integrate with a CMS and will not deploy any of the fixes it recommends.

Credits

Built by Specularis, an AI-visibility agency. Free web version: https://specularisinc.com/free-audit · MIT licensed.