AI Visibility Check: can ChatGPT cite your site, and does it?

Two questions about the same website, answered in one run.

1. Can an AI assistant read you? The site is scored out of 100 on whether ChatGPT, Perplexity, Claude, Copilot and Google's AI answers can reach it, understand what it is, quote it and date it.
2. Does an AI assistant name you? Give the questions your buyers actually type, and the actor collects the answers Google's AI Overview, Perplexity and ChatGPT give, then checks whether your brand appears in them, and which rivals appear instead.

Most tools do one or the other. On its own, neither is useful: being told "the AI never mentions you" does not say why, and being told "your llms.txt is missing" does not say whether it costs you anything. Together they say which of the two is the thing to fix first, and the actor writes that verdict in plain words on every row.

Public signals only. No login, nothing written to the sites it reads.

Half one: can an AI read you

Area	Out of	What it reads
Can AI crawlers read it	25	robots.txt, evaluated separately for fourteen AI crawlers and fetchers (OAI-SearchBot, ChatGPT-User, GPTBot, PerplexityBot, ClaudeBot, Claude-SearchBot, Google-Extended, Bingbot, Applebot-Extended, CCBot, DuckAssistBot and more)
Is there a map for agents	15	llms.txt: present, readable, pointing at the site's own pages
Is the entity clear	20	Organization, LocalBusiness, FAQPage and Article schema on the sampled pages
Is there something to quote	25	Answer-first opening paragraphs with a figure, page length, section structure, tables
Can it be dated and trusted	15	Publication and modification dates, phrases that name a source, a sitemap with lastmod

Each site returns a score, a grade from A to E, the five area scores, every finding in plain words, the three fixes that move the score most, and a verdict per AI crawler.

Half two: does an AI name you

Fill in Questions your buyers ask and the actor asks each engine you selected, reads the answer, and records:

• whether your brand name or your domain appears in it
• which of the rivals you named appear in it
• where your site sits in the Google top ten for the same question
• the answer text itself, so you can read what the AI actually told your buyer

Engine	Needs a key from you	What it costs
Google AI Overview	No	Runs the Apify Google Search Scraper on your own account, so its charge lands on your Apify bill, roughly a third of a cent per question
Perplexity	Yes, your own Perplexity key	Whatever your Perplexity plan charges
ChatGPT (OpenAI)	Yes, your own OpenAI key	Whatever your OpenAI plan charges

Keys are pasted into encrypted fields, used for the run and never written to the output.

What you get back besides the numbers

• A report page you can forward. Every run writes a single self-contained HTML page to its storage: all the sites, their scores, the AI answers word for word, the rivals named instead of you, and the fixes in order. Open it in a browser or send it to a client as it is.
• The change since last time. Each run remembers one line per site, so the next one shows which way the score and the mention share moved, and when the previous check was. Turn it off with trackChanges: false.
• Bot walls named for what they are. A site behind a Cloudflare browser challenge is reported as exactly that, separately from a plain refusal, because it is the more expensive problem: an AI fetcher never solves a challenge, so every assistant is turned away from that site no matter how good its pages are.
• Questions asked twice when Google stays silent. Google serves an AI Overview inconsistently, so a question that comes back empty is asked once more before it is reported as unanswered, and unanswered questions are left out of the share instead of counted as a miss.

Input

{
  "urls": ["https://example.com", "competitor.com"],
  "questions": ["best property agency in phuket", "where to buy a condo in phuket"],
  "brand": "Example Group",
  "competitors": ["FazWaz", "CBRE"],
  "engines": ["google-ai-overview"],
  "countryCode": "th",
  "languageCode": "en",
  "samplePages": 3,
  "budgetSeconds": 20
}

• urls: one or more sites. A bare host works; the actor adds https and reads the homepage, robots.txt, llms.txt, the sitemap and the sampled pages.
• questions: leave empty and only the score is produced. Every question is asked once and the answer is reused for every site in the run, so checking yourself against five competitors costs no more than checking yourself.
• brand: how your company is written in ordinary text. Left empty, the domain is used, so fazwaz.com becomes Fazwaz. It applies to the first site in the list; the others are read off their own domains.
• competitors: names or domains. Every answer is checked for these too.
• countryCode and languageCode: the country and language the Google question is asked from. AI answers differ by country, so set the market you sell in.
• samplePages: content pages to read beyond the homepage (1 to 10, default 3).
• budgetSeconds: hard limit per site (5 to 120, default 20). Raise it for slow sites.

Output

One dataset row per site:

{
  "url": "https://example.com/",
  "host": "example.com",
  "brand": "Example Group",
  "score": 74,
  "grade": "B",
  "mentionRate": 25,
  "namedInAnswers": 1,
  "answersCollected": 4,
  "verdict": "Named in 25% of answers despite a weak site score. Fix access and structure and the share should rise.",
  "access": 25, "index": 15, "entity": 17, "content": 14, "trust": 3,
  "topFixes": ["0 of 3 sampled pages open with an answer-first paragraph...", "3 sampled pages expose no publication or modified date..."],
  "rivalsAhead": [{ "rival": "FazWaz", "answersNaming": 3 }],
  "questionsWithoutBrand": ["where to buy a condo in phuket"],
  "byEngine": { "google-ai-overview": { "asked": 2, "mentioned": 1 } },
  "answers": [{ "engine": "google-ai-overview", "query": "...", "answer": "...", "brandMentioned": false, "citedHosts": ["fazwaz.com"], "organicPosition": 4 }],
  "findings": [{ "area": "Can AI crawlers read it", "level": "pass", "text": "All 14 AI crawlers and fetchers are allowed." }],
  "crawlers": [{ "label": "ChatGPT search (OAI-SearchBot)", "kind": "search", "verdict": "allowed" }],
  "sampledPages": [{ "url": "...", "words": 875, "answerFirst": false, "dated": false, "schema": ["Organization"] }],
  "sitemap": { "found": true, "count": 41, "lastmod": true },
  "checkedAt": "2026-09-11T02:15:00.000Z"
}

Three ready views in the dataset tab: Verdict for the one-line answer per site, Site scores for the five areas, AI answers for who gets named instead of you. Export to CSV, Excel or JSON, or read it through the Apify API.

Uses

• Your own site, monthly. Watch both numbers move after each fix.
• You against your competitors. Put every rival in urls and every rival name in competitors. One run shows who the AI can read and who it actually recommends.
• A pitch that is hard to argue with. Run a prospect's site on the questions their customers ask, and the output is the whole pitch: here is the answer your buyer got, here is who was named, here is why it was not you.
• Before and after a migration. A blocked crawler or a broken llms.txt shows up immediately.

What it is not

It does not measure traffic, that is in your analytics. It does not log in, and sites behind a login or a bot wall cannot be scored. AI answers are not stable: the same question can be answered differently an hour later, so read a share across several questions, not a single yes or no.

Billing

Pay per event: one charge per site scored, one per AI answer collected. Sites that are skipped or unreachable and engines that fail are never charged, and neither is a question no engine answered. The Google Search Scraper run sits on your own Apify account and is billed there.

Because the answers are collected once per run and then checked against every site in it, adding rivals costs nothing beyond their site score. Checking yourself against nine competitors on twenty questions is the price of twenty answers plus ten site scores, not of two hundred brand checks.

Who made it

OperStack: tools for content sites that have to rank, be cited and convert. The single-site version of the score runs free at oper-stack.com/ai-visibility; this actor is the batch version with the citation half added. The open-source quality gates behind the fixes are at github.com/oper-stack/gates.

Local development

npm install
npm run verify           # unit tests, then a real end-to-end run of the actor
apify run --input '{"urls":["https://example.com"]}'

npm test covers the scoring and the citation logic. npm run smoke runs src/main.mjs itself and checks the row, the verdict and the report page, because a unit test passes happily on an entry point that cannot execute. Run npm run verify before every push.

src/ai-visibility.mjs is vendored from oper-stack-website/src/lib/ai-visibility.mjs. Keep the two identical: both take { budgetMs, samplePages }. src/citations.mjs is actor-only.

Asking each question more than once

In plain words. An AI does not answer the same question the same way twice. Ask "where do I buy X" today and it names three companies; ask again an hour later and one of them is gone. So a report built on one answer tells you almost nothing: you could be missing from that one answer and present in the next four. Set How many times to ask each question to three or five, and the report stops saying "you are not named" and starts saying "you were named in two of five answers", which is a number you can watch move.

What it does. Each question is asked runsPerQuestion times per engine, one round after another. The report then shows, per question, how many runs named you and the share, plus how often the runs agreed with each other. When the runs disagree a lot, the overall share is shaky, and the report says so instead of hiding it behind an average. Asked once, the report says plainly that this is a snapshot rather than a measurement.

Every round costs one more answer, so the free look starts at one and you raise it when the number starts to matter.