How do you build a list of companies using AI?

Best move: treat the list as a filter over domains you already have, and filter on the backend stack, not the marketing copy.

Why it works: a loaded SDK, a vector-store call, or an MCP endpoint costs engineering to put in place and is pointless to fake. It survives minification. “AI-powered” in the hero does not.

Key takeaways

• A useful list of AI companies is a filter over domains you already have, not a directory you buy.
• The signal that separates shipping from talking is backend: model providers, AI SDKs, vector stores, orchestration, and MCP endpoints.
• “AI-powered” in the hero copy is a weak proxy for a working stack. The homepage lies more than the network trail.
• Bassethound returns an ai_readiness score, a shipping/experimenting/none verdict, and a hot/warm/cold ICP signal per domain in one call.
• Because the service is stateless and keyless, you refresh the list by re-running it. Nothing rots in a cache you forgot about.

Where do you get the domains to check?

You start with a domain set, not a blank search. Common seeds: your CRM export, a conference attendee list, the portfolio page of a fund, the customer logos on a competitor’s site, a Crunchbase or funding-round pull, or a vertical directory. The list is the input. The AI verdict is what you add on top.

Bassethound does not sell you the seed set and does not pretend to. It takes a domain and returns a dossier. That split matters. Sourcing companies and qualifying them on AI capability are two jobs, and tools that bundle both do neither well. Keep your sourcing wherever it already lives (Apollo, ZoomInfo, a scrape, a hand-built list) and treat sniff_domain as the qualification pass. Feed each domain through your agent, collect the ai_readiness block, and keep the rows that clear your bar. If your seed list is company names rather than domains, resolve the domain first. The detector works on the live site, not the entity name.

Why fingerprint the stack instead of scraping “AI-powered” claims?

Marketing language is the cheapest thing on the internet to produce and the least correlated with a working stack. A landing page that says “AI-powered” tells you a copywriter typed three words. It does not tell you whether an application loads the Anthropic SDK, calls a Pinecone index, or exposes an MCP endpoint.

Those are the signals that survive minification and cost real engineering to stand up. Bassethound reads the live front end and the network trail: model providers (Anthropic, OpenAI, Gemini, Cohere, Mistral), AI SDKs, orchestration (LangChain, LlamaIndex, CrewAI), vector stores, LLM observability (Helicone, LangSmith, Langfuse), AI widgets, llms.txt, and .well-known AI files. It scores the count of distinct signal groups hit, out of six, and returns a verdict: shipping, experimenting, or none. That verdict is defensible in a way a keyword scan never is. Filter a list on “AI-powered” copy and you get everyone with a marketing team. Filter on stack signals and you get the companies that wrote code.

How do you turn per-domain results into a ranked list?

Each sniff_domain call returns a top-level summary you can sort on: primary_stack, ai_readiness {score, verdict}, security_grade, and an icp_signal of hot, warm, or cold. Rank by score to float the deepest adopters. Filter by verdict to split shipping from experimenting from none. Use icp_signal as the fused shortcut when you do not want to weight the layers yourself.

A practical pass: keep verdict == shipping OR icp_signal == hot, sort by score descending, then break ties on the specific signals that matter to you (an MCP endpoint if you sell to MCP builders, a vector store if you sell retrieval infrastructure). Because the dossier correlates all five layers, you can qualify on more than AI in the same row. Firmographics give you the vertical and company context. Infrastructure tells you the CDN and email provider. Security gives you a header grade. One call per domain produces a row wide enough to segment on without joining five separate tools.

How do you keep the list fresh without storing state?

You re-run it. Bassethound is stateless and read-only. Every call is self-contained. For list-building, that is a feature.

A company that had no detectable AI stack in March and shipped a support agent in June shows up when you re-run its domain, with no vendor recrawl schedule to invalidate and no “last enriched” timestamp to distrust. Run your seed list on whatever cadence your pipeline needs (weekly for a hot segment, quarterly for a broad sweep) and diff the verdicts to catch new adopters. The free tier is rate-limited, so a large list on a tight schedule is where the OAuth paid tier earns its place: more depth and volume without a bring-your-own-key setup, because the crawl runs on our side. You own the seed list and the schedule. The detector owns the current read. Refresh is a re-run, not a sync job you maintain.

What will this approach miss?

Be honest about the failure modes. A static crawl plus optional JS render sees what a browser can load. It can miss a stack that runs fully server-side, behind a proxy, or inside a first-party API that never names its vendor on the wire. A company using Claude through its own backend gateway, with no client SDK and no llms.txt, can read as none when it is shipping. The dossier reports these gaps rather than papering over them: a none verdict means no detectable client-side signal, not proof of absence. Treat it that way.

The inverse error is rarer but real. A widget or SDK loaded on a single demo page can register as a signal, even when the company runs it nowhere else. This is why the score counts distinct signal groups, and why two correlated signals beat one. For list-building, that asymmetry is fine. You are ranking, not certifying. Push the confident shipping rows to the top and treat the none tail as unqualified rather than disqualified.

Bassethound perspective

Every list of “companies using AI” on the market is built from words. Press releases, funding announcements, job posts, “AI” in the hero copy. That measures who talks about AI, not who runs it, and the two lists barely overlap. A company can rank high on every intent signal and load no AI stack at all. Another can ship a Claude agent in production and never issue a press release. Build outbound on the marketing signal and you spend the quarter on the first list.

Read the stack instead. A loaded Anthropic SDK, a Pinecone call, an MCP endpoint. Nobody ships those to decorate a landing page. They run in production or they leave no trace. That is the honest tell.

The other half of the conviction is fusion. An agent can pull tech-stack from one tool, infra from a second, firmographics from a third, and then it owns the join. Bassethound returns all five layers correlated in one call, so your list row is already segmentable. The moat is not the fan-out. It is the correlation you did not have to assemble.

Sources

• Model Context Protocol: https://modelcontextprotocol.io
• llms.txt specification: https://llmstxt.org
• Anthropic documentation: https://docs.anthropic.com
• Pinecone documentation: https://docs.pinecone.io