GEO / AI Visibility Audit

Find technical obstacles to AI search access and extractable website content, for SEO teams, agencies and website owners.

Audit public HTML without a browser or paid language-model calls. Each page receives a transparent technical score, observations and concrete fixes. Check search crawler rules separately from training permissions, initial-HTML readability, JSON-LD, answer structure and factual/source signals.

Price and costs

$4.99 per 1,000 audited HTML pages, plus Apify's standard start event ($0.00005 per run). Paid Apify plans get tier discounts down to $3.49 per 1,000 pages. 100 audited pages cost about $0.50.

Only successfully audited HTML pages are charged, even when their score is low. Failed, blocked, non-HTML and truncated responses and the summary are free.

Tested on Apify on November 4, 2026: one website, 60 page attempts, 50 audited pages, 10 non-HTML sitemap entries skipped free of charge, about 2 minutes at 512 MB, without proxy.

Proxy use is off by default. A blocked URL is reported without retrying under another identity. If you enable Apify Proxy, Apify bills the proxy traffic separately.

Input

{
  "startUrls": [{"url": "https://example.com/guide"}],
  "maxPages": 20,
  "maxDepth": 2,
  "useSitemap": true,
  "crawlLinks": true,
  "proxyConfiguration": {"useApifyProxy": false}
}

maxPages limits attempted pages across the whole run, including errors; maximum 1,000. maxDepth applies to HTML links; sitemap discoveries enter at depth 0. Disable both discovery options to audit only the supplied URLs. An empty input uses the schema's example.com sample. Explicit invalid types and null settings are rejected.

Requests use the identified StevenKrampBot, a minimum 2-second host interval, published Crawl-delay and one attempt per URL/hop. A hard request limit, response-size limits and a timeout reserve bound work. Starts, files named by robots.txt and every redirect target are checked. Unknown robots access stops fetching; missing robots.txt allows access. Denied robots.txt responses also stop this Actor.

Output and privacy

The dataset contains page rows plus one summary row. Each successful page has seven scored checks and fixed English fix hints. This fictional excerpt shows selected fields:

{
  "type": "page",
  "pageId": "SHA256_OF_NORMALIZED_INPUT_URL",
  "inputIndex": 0,
  "statusCode": 200,
  "error": null,
  "billable": true,
  "score": 75,
  "scoreVersion": "technical-1",
  "checks": [{
    "code": "ai-search-access",
    "passed": false,
    "weight": 25,
    "fix": "Review the reported search-crawler rules in robots.txt. Allow only the search access you intend; training access is a separate choice."
  }]
}

pageId, finalPageId and siteId are SHA-256 fingerprints, not raw URLs. inputIndex is the zero-based seed position, or null for discovered pages. To map a URL locally, normalize scheme/host, drop the default port and fragment, retain path/query and hash UTF-8 bytes. The dataset omits URL paths, query strings, titles, author names, contacts, snippets, raw schema values and source URLs. For easier mapping, supply an explicit URL list and disable discovery.

The SUMMARY key-value record includes attempted/audited/failed/skipped counts, physical HTTP requests, average known score, stop reason, sitemap results and billablePagesStored. A page may be audited but not stored when the SDK cost limit rejects it; these counts are distinct. Queued aliases can already have been visited. Downloaded HTML and llms.txt are not persisted.

Score

Check	Points
Both observed AI search crawler rules allow the page	25
No applicable noindex/none/nosnippet restriction	10
At least 800 visible characters in initial HTML	20
Structurally valid schema.org JSON-LD, no invalid block	10
Heading followed by a 40–700-character paragraph/list item	15
Numerical facts with units or date signals	10
At least one external HTTP(S) link	10

Unknown access stays null instead of becoming denied or zero. A complete score is null if a scored observation is unknown; observedPoints and assessedWeight preserve the known portion. Failed fetches have no score. Schema validity is structural, not full schema.org validation; microdata is counted separately. Facts and source checks count patterns without verifying truth, relevance, authority or whether a link supports a claim. Navigation and boilerplate may contribute.

Training and user-requested agents do not affect search-access points. This is a static robots interpretation, not a live test from provider IPs. OpenAI documents distinct bot purposes; user-triggered requests can behave differently. Anthropic also separates search, training and user-requested agents.

Root /llms.txt is optional and carries zero score weight. Successful non-HTML Markdown with a title and an absolute link passes this narrow heuristic. Missing, invalid, denied, truncated and unavailable files are distinguished. Scoped files, relative-link variants, Markdown alternates and llms-full.txt are outside this check. The llms.txt proposal is not a universal provider requirement. Nonstandard Content-Signal values retain group scope and do not change scores.

API

After publication, set APIFY_TOKEN in your own environment and call:

curl --request POST \
  --header "Authorization: Bearer $APIFY_TOKEN" \
  --header "Content-Type: application/json" \
  --data '{"startUrls":[{"url":"https://example.com/"}],"maxPages":1,"useSitemap":false,"crawlLinks":false}' \
  "https://api.apify.com/v2/acts/stevenkramp~geo-ai-visibility-audit/run-sync-get-dataset-items?memory=512&timeout=120"

This endpoint and MCP identifier are reserved for the new Actor; they are not currently published.

Limits

A high score does not guarantee mentions, citations, rankings or recommendations by ChatGPT or any other AI service. These checks establish no such relationship. No model is called, no prompt visibility is measured and no JavaScript or browser rendering is evaluated. Low text density alone is not a JS shell: that warning requires very little visible text plus scripts. Access can differ by IP, provider policy, authentication and bot protection.

Use only public pages you are entitled to audit. There is no login, CAPTCHA solving or bot impersonation. Cookies are cleared between requests. Public-destination checks are not a guarantee against DNS rebinding. This is not a legal-compliance or security audit.

Use with AI agents (MCP)

Once published, connect an MCP client to: https://mcp.apify.com?tools=stevenkramp/geo-ai-visibility-audit

Example: “Audit these public documentation URLs for technical AI search readiness. Explain failed checks and distinguish search crawling from model training.”

Local checks

Install bounded requirements in an isolated environment, then run:

python -m unittest discover -s tests -v
python -m tests.lauf_auswerten dataset.json
python -m tests.kosten --pages 1000 --sites 1
python -m tests.kosten --run-json run.json --summary-json summary.json

Tests use fictional HTTP responses without contacting websites. The cost helper leaves missing cost inputs null.