Enrichment Desk

Explainer / Technical due diligence

Technographic data: how stack detection works and where it fails

A detected technology shows that a script or record was there on scan day. It does not prove a company uses or pays for it. Here is how to read the data.

Enrichment Desk · · 8 min readShare

Technographic data is a list of the technologies a crawler found traces of on a company's website and in its domain records. Each entry shows that a script loaded, a header was sent or a DNS record pointed somewhere on the day of the scan. It does not show that the company uses the product, still pays for it or ever bought it. Read every row as a dated signal, and ask the provider how it saw it.

This matters most when an agent does the qualifying. An agent that scores accounts on "runs Shopify" inherits every weakness of the detection beneath that label.

What is technographic data?

Technographic data describes what sits behind a company's domain: its content management system, ecommerce platform, analytics tools, CDN and email provider. Most providers collect it the same way. BuiltWith's FAQ says every website gives off signals of the technologies it uses. It says the company indexes the web the way a search engine does and takes no data from third parties.

TechLookup's methodology gives the most detailed account in public. Its crawler loads the home page in a full Chromium browser and lets the scripts run. It then reads the rendered HTML, script URLs, response headers, cookies, background requests, the TLS certificate issuer and the domain's DNS records.

How do tech stack detection tools work?

They match what a page exposes against a catalog of fingerprints. A fingerprint is a set of patterns for one technology, such as the path of a script file, a response header, a named cookie, a meta tag or a JavaScript variable. The open-source Go port of Wappalyzer's fingerprints lists these fields in its code. It also carries an "implies" list of technologies that one detection adds automatically.

Some signals are blunt. Wappalyzer's own guide shows three of them:

  • WordPress writes its name and version into a generator meta tag.
  • Shopify leaves a global JavaScript object on the page.
  • Web servers and CDNs often name themselves in a server header.

TechLookup's catalog is built on the community continuation of those fingerprints and covers more than 7,000 technologies. Each pattern that matches adds weight to a confidence score, which is capped at 100%. Some technologies are inferred. A site running Next.js is recorded as running React even if React left no trace of its own.

What can detection not see?

Detection cannot see anything that never reaches the browser. TechLookup lists its own blind spots plainly:

  • It scans only the home page. Checkout pages, a help centre at a different path and anything behind a login go unseen.
  • Server-side software that leaves nothing in the page, headers, cookies or DNS cannot be detected. That includes a database, a back-office tool and most programming languages behind a proxy.
  • A tool that waits for a click or for consent before loading may never show up.

Scripts that other scripts load are a partial exception. TechLookup's browser runs the page's code and records the background requests it makes, so by our reading a tool injected by a tag manager can still show up if it loads on the home page without a click. Even then, the scanner reads at most 20 external scripts and stylesheets, and only the first 64 KB of each.

The result is a structural bias. The detected stack is the visible layer, meaning whatever the home page loads. A missing technology is not evidence that the company lacks it.

Sites also get in the way. Researchers at Stony Brook University found that caching proxies which minify static files change the file hashes some tools match on. They also found that bot detection such as Cloudflare can send a scanner to a CAPTCHA page, which rules out fingerprinting entirely. TechLookup notes that a blocked scan still detects something real, often the protection service itself, while the rest of the stack goes missing.

Signal What it can show Where it goes wrong
HTML and meta tags A CMS's generator tag, with name and version The version string can be turned off in settings or by a plugin
JavaScript variables A platform's global object, such as Shopify's A tool loaded only after a click or consent banner may not appear
HTTP headers The web server or CDN A bot check can return a challenge page instead of the site
Script and file URLs A version number in a script's file name, like jquery-3.7.1.js A caching proxy that minifies files breaks hash-based matching
DNS records An email provider in the MX records Shows services the organisation uses, not what the site runs on
Implied technologies React added because Next.js was found Inferred, with no trace of its own

Why is a detected technology not proof of use?

A detection is a pattern match on a date. TechLookup warns that a fingerprint can match a site that merely mentions a product. Its "first seen" date records when the scanner first detected a technology, not when the company adopted it.

Nothing in a fingerprint separates a tool under contract from a snippet left in a template after the contract ended. Both load, and both match.

Removals lag by design. TechLookup waits until two full scans in a row miss a technology before it marks it removed, which protects against one flaky scan. Sites in the Tranco top 100,000 are rescanned about weekly and other sites about every 30 days, except those found only through links with no Tranco rank, which wait about 90 days. By our arithmetic, a tool dropped just after a scan can stay listed for about two months on the 30-day schedule and about six months on the 90-day one. When a later scan cannot read a version number, the last version seen is kept.

The strongest objection is that confidence scores already handle this. They help, but they measure how sure the pattern match is, not whether the company is a customer. In the Go port, an implied technology is added with the same confidence as the technology that implied it, even though it left no trace of its own. A high score on an inferred row is a statement about the parent detection.

How accurate is technographic data?

Accuracy is rarely checked by anyone without a stake in the answer, and the independent evidence that exists covers detection tools, not vendors' databases.

The Stony Brook study tested six fingerprinting tools against 1,360 releases of five open-source web applications. In lab conditions, at least one tool labelled 94.8% of releases correctly. Against 726 real websites, accuracy fell by 20% to 80% from the lab results:

  • BlindElephant identified only 46% of real WordPress sites.
  • Wappalyzer degraded least but still lost double digits on nearly every application.

The authors would not treat third-party lists such as BuiltWith's as ground truth, because those lists rely on fingerprinting themselves. To learn which sites really ran each application, they mainly used the software makers' showcase pages, supplemented by a small dataset from third parties they contacted who knew first-hand what software ran on their sites.

Vendor figures need their source attached. HG Insights, which sells technographic data, published a report in which an outside evaluator puts the category's typical hit rate at 50 to 60 percent. In the same report, a blind sample of about 100 accounts scored 87 to 88 percent for one provider.

That is a vendor's account of an unnamed evaluator's numbers on a small sample. The report itself says the result moved one evaluation to its next stage and did not close a deal. Its procedural advice holds up anyway: test against accounts where you already know the answer, and ask for the false-positive rate as well as the hit rate.

How should you judge a technographics provider?

Ask for the evidence behind each row, not just the row. A provider worth relying on should answer five questions:

  • Which signal matched? It might be a header, a script URL, a cookie, a DNS record or an inference from another technology. TechLookup's reports list this for each technology, with confidence and first-seen and last-seen dates.
  • What gets scanned? Ask how many pages per site, and which lists of sites the crawl starts from.
  • How fresh is it? Ask how often a site is rescanned, and how many misses it takes before a technology is marked removed.
  • Where else did data come from? Check whether any record came from something other than a scan of the site. BuiltWith, for one, says it takes no data from third parties.
  • How often is it wrong? Ask for the false-positive rate on accounts you already know.

Cost per record belongs in the same check. Wappalyzer's pricing page lists current plans. Its Business plan at $450 a month includes 20,000 API credits, and a lookup typically costs one credit per website. Used in full, that is about 2.3 cents a site by our arithmetic. Credits included in a plan expire after 60 days, so unused credits push the real cost per record higher. Most self-serve purchases are by credit card, and larger purchases are arranged with sales.

The next step is a blind test. Take 50 accounts whose stack you know from closed deals or your own CRM. Send them to the provider without saying which ones you have verified, and count both the misses and the false hits. That count will tell you more than any accuracy claim, and it shows which signals your agent should never qualify on alone.

Frequently asked questions

What is technographic data?

Technographic data is a record of the technologies a provider detected on a company's website and in its domain records, such as its CMS, analytics tools, CDN and email provider. Each entry is evidence that a signal was present when the site was scanned.

How do tech stack detection tools work?

They load a site and match its HTML, scripts, headers, cookies, JavaScript variables and DNS records against a catalog of fingerprints. Each fingerprint holds the patterns that one technology leaves behind.

Can technographic tools detect back-end software?

Not by scanning a website. Software that leaves no trace in the page, headers, cookies or DNS, such as a database or a back-office tool, cannot be detected that way.

How accurate is technographic data?

It is rarely measured independently. In a Stony Brook University study, fingerprinting tools lost 20% to 80% of their lab accuracy on real websites, so test any provider against accounts whose stack you already know.

How often is technographic data updated?

It depends on the provider and the site. TechLookup rescans the Tranco top 100,000 about weekly, other sites about every 30 days and sites found only through links with no Tranco rank about every 90 days, and it marks a technology as removed only after two scans in a row miss it.

Read this article as Markdown Open in ChatGPT Open in Claude