AI Tooling Detector

🤖 What can AI Tooling Detector do?

Give it a company domain and it tells you whether that company just talks about AI, actually runs AI tooling on its site, or charges money for AI. One flat row per domain.

Most AI detectors return a boolean, which is close to useless for targeting. "Published an llms.txt" and "bills $0.99 per AI resolution" are the same answer to a yes-or-no question and completely different sales situations. This actor returns a four-level ai_maturity tier with the evidence behind it.

Bought by teams selling AI infrastructure, AI-adjacent tooling, or anything where "already pays for AI" and "still only talking about AI" are two different pitches.

🚫 This is not a full tech stack detector. It looks only at AI signals. If you want the general marketing and sales stack, use Tech Stack Detector.

💡 Why use AI Tooling Detector?

🪜 The ai_maturity ladder, and why it is strict

Each rung needs its own class of evidence. An llms.txt can never on its own reach deployed, and a chat widget can never on its own reach commercialized. That is what stops the tier from collapsing back into a boolean.

The vendor list covers support agents (Sierra, Decagon, Ada, Intercom Fin, Forethought), AI search (Glean, Algolia, Inkeep, Kapa), personalization (Mutiny, Dynamic Yield), direct inference endpoints (OpenAI, Anthropic, Azure OpenAI, Bedrock, Gemini, Groq, Together, Fireworks), vector stores and app frameworks.

📋 What data can AI Tooling Detector extract?

33 fields per domain. The ones buyers use:

Every row carries every key. A value that is unknown is null, never a missing key.

⚠️ Read confidence and signals_missing before you trust a none. A verdict built on one reachable page should not be read like a verdict built on five. A domain behind a bot challenge comes back blocked: true at low confidence, never as a confident no, because without that flag a challenged domain looks identical to a company with no AI at all.

🛠️ How to detect a company's AI adoption

1. Open the Input tab and put a domain in domain, or a list in domains for batch mode.
2. Leave check_pricing on true. Turning it off makes runs faster but caps every result at deployed, since commercialized can only be proven on a pricing page.
3. Click Start.
4. Read ai_maturity for the answer and evidence for the quote.
5. Suppress rows where blocked is true before you act on a negative.

🧪 Using it in Clay

Add an Apify enrichment column, point it at this actor, and map your domain column to domain. Every output field lands as its own Clay column.

Useful follow-on filters:

• Companies that already buy AI: ai_maturity = "commercialized"
• Companies that talk about AI but have not shipped it: ai_maturity = "declared"
• Suppress unreliable rows: blocked = false AND confidence >= 0.6

📦 Batch mode

Put an array in domains and it takes precedence over domain. batchSize controls how many run concurrently, 1 to 10. The actor is HTTP only, no browser and no proxy, at 256MB, so batches finish in seconds.

Rows are written batch by batch as each one finishes, so a long list never loses the rows already done. The run keeps 90 seconds of its own timeout in reserve and caps every request by the time left. A domain it could not start or finish in time still gets a row, with

blocked_reason: "timed_out"

💵 How much does it cost to detect AI tooling?

You are charged once per domain analyzed (the domain-analyzed event). A domain the actor could not read is not charged: a blocked homepage (blocked: true), a homepage that would not load (fetch_error set), an analysis error, and a domain the run ran out of time for (

blocked_reason: "timed_out"

There is also a start event, actor-start, at $0.00005, charged once per run.

💳 Free Apify plans get 1,000 results per calendar month; paid plans are unlimited. When the free cap is reached the run exits cleanly with a notice row rather than an error. Repeat lookups inside 7 days are served from cache; a cached row is still charged domain-analyzed.

Maximum charge

The actor honors the maximum cost per run you set in Apify: it stops analyzing once another row would not fit, and when the platform reports the limit reached it stops cleanly and keeps every row already charged. You can also set max_charge_usd on the Input tab as a second ceiling.

⌨️ Input

Everything is on the Input tab. The options worth explaining:

🎯 Reporting on one vendor

vendors narrows the answer to the AI tools you care about, which turns this actor into competitive intelligence for any one of the 52 fingerprinted vendors.

{ "domains": ["stripe.com", "notion.so"], "vendors": ["sierra"] }

Selectable vendors, by category:

Three things worth knowing:

• Detection is unchanged. Every fingerprint is still evaluated on every run. The filter shapes the answer, not the work, so a filtered run costs exactly what an unfiltered one costs and reuses the same cache.
• The AI maturity read is never narrowed. ai_maturity, confidence, uses_ai, the llms.txt and robots.txt checks and the pricing score all measure the company's own AI posture, not a vendor list. A row filtered to Sierra still reports honestly that the company is commercialized; it simply does not list the other tools it also found.
• Evidence is filtered in step. Vendor evidence lines for tools you did not ask about are dropped. Every llms.txt, robots.txt, pricing and copy line stays, because those are what the maturity read is built from.

vendors_requested on every row records what you asked for, so a row listing one tool is never confused with a row where only one tool was looked for.

📤 Output

One flat row per domain, exportable as JSON, CSV, Excel, HTML or XML.

{
  "domain": "intercom.com",
  "resolved_url": "https://www.intercom.com/",
  "http_status": 200,
  "ai_maturity": "commercialized",
  "uses_ai": true,
  "confidence": 1,
  "has_llms_txt": true,
  "llms_txt_url": "https://intercom.com/llms.txt",
  "detected_tools": [
    {"vendor": "Intercom Fin", "category": "support", "ai_native": true, "confidence": "high", "evidence": "markup match: fin.ai"}
  ],
  "detected_tool_names": "Intercom Fin",
  "detected_categories": ["support"],
  "ai_copy_weight": 8,
  "ai_pricing_detected": true,
  "ai_pricing_score": 10,
  "pricing_url": "https://www.intercom.com/pricing",
  "evidence": ["AI priced per resolved outcome: \"$0.99 per Fin outcome\""],
  "signals_fired": ["llms_txt", "vendor_fingerprint", "pricing_commercialization"],
  "signals_missing": ["ai_txt", "robots_ai_directives"],
  "blocked": false,
  "is_js_rendered": false,
  "fetch_error": null,
  "checked_at": "2026-08-05T17:20:00.000Z",
  "row_status": "ok",
  "error_reason": null,
  "billable_events_json": "[{\"event\":\"domain-analyzed\",\"count\":1}]"
}

💡 Tips

• ai_maturity = "declared" is the most interesting segment for most sellers. Those companies have publicly committed to AI and have not yet bought the thing that delivers it.
• Always pair a negative with blocked = false. It is the single cheapest quality gate here.
• Keep check_pricing on. It is the only route to the commercialized tier.
• Use vendors when you are building a displacement list for one tool. The row stays narrow and so does the export.
• Paste evidence straight into the email. It is stored precisely so a rep can quote it.

⚠️ Known limits

Vendor recall is capped by what the server sends. Modern chat widgets are injected by client-side JavaScript, so a company running an AI widget that never appears in the served markup will read as declared rather than deployed. Adding a browser would raise recall, and would also raise the memory floor to 1024MB and the price with it. The tradeoff was made deliberately in favor of price.

Anti-bot challenges block a minority of domains, concentrated in large consumer brands and retail. Those rows come back blocked: true rather than silently negative. No proxy is configured by default.

commercialized requires a reachable pricing page. Companies that gate pricing behind "contact sales" cannot be proven to commercialize AI from the outside.

declared is a claim, not a deployment. That is exactly why it is a separate rung, but do not read it as adoption.

llms.txt is a high-precision positive whose absence proves nothing. Most companies have not published one, including companies deep into AI.

Fingerprints need maintenance. Vendors change script URLs, and a fingerprint that has not been refreshed goes quietly stale rather than failing loudly. Read signals_fired to see what actually matched.

❓ FAQ

Why is a company I know uses AI coming back as declared?

Almost always because its AI widget is injected by client-side JavaScript and never appears in the served markup. Check is_js_rendered.

What happens when a site blocks the actor?

The row is still emitted, flagged blocked: true with a blocked_reason, and confidence is pinned low. You never get a silent false negative, and a blocked row is not charged.

Why did /llms.txt not count?

A lot of single-page apps answer /llms.txt with their HTML shell. Those return has_llms_txt: false with an llms_txt_reject_reason, so the strongest signal in the pack stays clean.

Does the run ever return nothing?

No. The run always produces at least one dataset item. If no domain was supplied, or the free-tier cap was reached, you get one notice row with is_summary_row: true. Every domain you sent gets its own row, including one the run ran out of time for, which reads blocked_reason: "timed_out" and is not charged.

Is there an MCP server?

Yes. This actor is a tool on the Mamba Labs GTM Suite server, which exposes the whole fleet through one install.

🧩 Want other GTM data?

Mamba Labs builds a fleet of enrichment actors that share one flat output convention:

Every actor in the suite takes a domain or a company and returns one flat row, so they stack in the same Clay table without reshaping anything.

🆘 Support

Something wrong, or a vendor the actor misses? Open an issue on the Issues tab with the domain and the row, and we will look at it.

ℹ️ Sourcing and legal. Every signal comes from pages the company publishes itself, read directly over HTTP with no proxy: the homepage, the pricing page, robots.txt, llms.txt and ai.txt. The row describes what a company published, not what it runs internally. You are responsible for how you use the output.

Built by Mamba Labs.