Why does stitching together five enrichment tools break down?

Best move: stop fanning out to five single-layer tools. Make one call that returns the layers already correlated.

Why it works: fetching each layer is easy. The hard part is joining them, on the same domain, at the same instant, with confidence scores you can line up. A fused source does that join once, so you never write it.

Key takeaways

• Stitching enrichment tools fails on correlation, latency, and schema reconciliation, not on coverage. Adding feeds does not add the join.
• Five tools return five verdicts in five schemas, so comparing “confidence 0.8” against “high” against a bare boolean is a mapping you write and maintain.
• Partial failure is the default: four of five tools returning is common, and a tool that never rendered the page is indistinguishable from one that rendered and found nothing.
• Fan-out costs five auth flows, five rate limits, and five tool definitions in the model’s context, on top of the merge step.
• Bassethound returns tech stack, infrastructure, firmographics, AI-readiness, and security correlated in one keyless call, and reports the gaps its crawl could not reach.

What breaks first when you chain five enrichment APIs?

Each tool owns one layer. Tech stack from one, infrastructure from another, firmographics from a third, AI signals from a fourth, security headers from a fifth. You now run five requests against five services with five error modes. One returns a 429. One times out. One renders JavaScript and one does not, so your five answers describe five different snapshots of the same site, captured seconds apart.

Partial failure becomes the steady state. Your agent has to decide whether four-of-five is a usable dossier or a broken one, and that decision logic is yours to build. Worse, the failures are quiet. A tool that returns “no AI stack detected” because it never executed the page looks byte-for-byte identical to a tool that rendered the page and found nothing there. You cannot tell a real negative from a coverage gap without instrumenting each tool yourself. You added coverage and lost the ability to trust a negative.

Why is correlation the expensive part, not coverage?

Coverage is a purchasing problem. You buy another feed and you have another layer. Correlation is a reasoning problem. Adding a feed hands your agent one more thing to reconcile.

The signal that matters is not “this site loads Pinecone” or “this site runs the Anthropic SDK” in isolation. It is that the vector store, the model SDK, and the Vercel host all belong to the same shipped application, observed in the same crawl. That claim lives in the join. Five separate tools cannot make it for you, because none of them saw the other four layers. They hand you rows keyed to five schemas and five confidence scales, and you have to prove they describe one coherent system. The value you wanted, “this company ships AI, here is the correlated stack,” is the one thing no single tool produces.

How does tool sprawl slow an agent down?

Latency compounds two ways. Sequential fan-out is the sum of five round trips. Parallel fan-out is the max of five, plus five auth handshakes, plus the merge you run after. Parallel hides wall-clock time when every tool is fast and up, but it does nothing for the reconciliation step, and that step is where your agent spends real tokens.

The carrying cost is a separate line item. Five tools can mean five API keys, five billing relationships, five rate-limit budgets to track. In the model’s context you now hold five tool definitions and five response schemas the model has to reason over before it can answer one question. Honest caveat: if all five are fast and healthy, parallel calls feel quick. The cost you cannot parallelize away is the schema-merge logic, and it runs on every request regardless of how fast the fetches returned.

Where do five separate schemas fight each other?

At every boundary you did not think about until it broke. Domain normalization first: www versus apex, trailing redirects, and country TLDs mean two tools can enrich two different canonical hosts and hand you conflicting answers for what you called one domain.

Then infrastructure ambiguity. One tool reports the CDN edge, another reports the origin behind it, and neither labels which is “the host.” Email auth splinters the same way: SPF is a TXT record, DMARC is a policy, DKIM is a selector, and every vendor models the SPF/DKIM/DMARC booleans differently. Timestamp skew adds a fourth axis, because a certificate observed at 12:00 and a header grade observed at 12:04 can disagree after a deploy. None of these are exotic. They are the ordinary tax of merging independent sources, and you pay it on every dossier you assemble.

When is stitching the right call?

Sometimes it is, and pretending otherwise would be dishonest. If you need one layer at extreme depth, a specialist beats a fused generalist. SimilarWeb-grade traffic modeling, Censys-grade or Shodan-grade certificate and port history across the whole address space, WhoisXML-grade historical WHOIS: those are corpora, and a live crawl cannot reproduce them.

Fusion wins on a different axis. It wins when you want the correlated picture of one domain right now, in one call, with the layers already joined. It does not win when you want the deepest possible single layer. Bassethound stays honest about its own edge: a static crawl plus optional JavaScript render can miss fully server-side or proxied components, and the dossier flags those gaps instead of returning a confident-looking blank. Reach for a specialist when depth in one dimension is the whole job. Reach for fusion when the relationship between dimensions is the job.

Bassethound perspective

The industry keeps selling fan-out as a virtue. More tools, more feeds, a marketplace of narrow MCPs you compose at will. We think that is backwards for domain intelligence, and it is a claim a tool-marketplace vendor would dispute out loud.

Composition sounds free and is not. The moment you stitch five single-layer sources, you have signed up to own domain normalization, confidence-scale mapping, timestamp reconciliation, and partial-failure handling, on every call. You call it integration glue. It is the product, and you built it by accident. An agent cannot cheaply buy correlation from five separate MCPs, because correlation exists only after a crawl sees all five layers together.

Fusion is the moat, not fan-out. We own the crawl, so we see the tech stack, the infrastructure, the firmographics, the AI-readiness fingerprint, and the security grade in one pass, and we return them joined, keyless, and stateless. You get the one signal you were stitching toward. You skip the five schemas you would have spent the week reconciling.

Sources

• Model Context Protocol specification: https://modelcontextprotocol.io
• MDN HTTP headers reference: https://developer.mozilla.org/en-US/docs/Web/HTTP/Headers
• IETF RFC 7208, Sender Policy Framework: https://datatracker.ietf.org/doc/html/rfc7208
• llms.txt specification: https://llmstxt.org