Skip to main content
Blog

How to Evaluate a Technographic Data Provider

Technographic data is inferred, so the question is not how many companies it covers but how often it is right about yours. Five criteria (detection method, evidence, recency, ICP coverage, catalog) and a precision-and-recall test you run on companies whose stack you already know.

Christian PalouChristian Palou•November 19, 2026

Technographic data is not observed but inferred, whether from job postings, from web scans or from both, and every inference is sometimes wrong. So the headline numbers a provider leads with, the millions of companies and tens of thousands of technologies, answer a question you are not actually asking. The question is not how much they cover. It is how often they are right about the companies you sell to, and how badly they are wrong when they miss. The only honest way to find out is to test them on companies whose stack you already know, which is the method at the end of this page. These five criteria tell you what to test and how to read the result. For job data specifically, the same discipline is in how to evaluate a jobs dataset.

1. How the signal is detected

Everything downstream depends on where the data comes from, and there are two main methods, each blind in a different place.

Web scanning looks at a company's website and public footprint for the technology's fingerprint, such as a tracking tag, a JavaScript library or a DNS record. It is good at client-side and marketing tools that leave a mark on the page, and blind to anything internal: the data warehouse, the ERP, the CRM a company runs but never ships to its homepage.

Job-posting inference, which is how we source our tech stack data, reads what companies say they hire for, such as a "Senior React Developer" title, a "we use Snowflake" line in a description, or an ATS domain in the job URL. It sees internal and back-office tools that never touch the website, and it is blind to companies that are not hiring, or that use a technology without mentioning it. The rule to keep in mind is that a technology that is not mentioned is not a technology that is absent. An inferred dataset reports presence, never absence, and a blank means "we have no signal", not "they do not use it".

Ask which method a provider uses, and match it to what you sell. If your product integrates with a data warehouse, a web-scan provider will barely see your market. If it is a front-end analytics tool, job-posting inference will under-report it. Neither method is better in the abstract. The best fit is the one that can see the category you care about.

2. Evidence you can trace

An inferred signal is only as trustworthy as the evidence you can inspect behind it. A provider that hands you a flat list of technologies per company, with nothing underneath, is asking you to take every call on faith, and you will not know a true match from a parsing error until it costs you a bad outreach.

Ask whether each signal links back to its source. For us, every technology card traces to the original job postings that mention it, with the number of jobs, the first and last dates the technology appeared, how often it shows up relative to others in its category, and a link straight to those postings. Ask too for a confidence score and what drives it. Ours is low, medium or high, computed from how many times the technology is mentioned, how recently, where in the posting, and how it stacks up against similar tools, so 40 mentions of one CRM against 4 of another is read for what it is.

3. Recency

A technology a company mentioned once in 2021 is not evidence of what runs there today, and a provider that cannot tell you when a signal was last seen is selling you a stack that may be years stale. Adoption and churn are the whole point of technographics, and a dataset frozen in time cannot show either.

Ask how fresh the signals are and whether each one carries a last-seen date. Because our data comes from job postings processed as they are collected, every technology-to-company link carries a first-found and a last-found date, so you can weight a stack that shows up in last month's hiring differently from one last seen three years ago, and watch a technology enter or leave a company over time.

4. Coverage of your ICP, not the catalog size

"33k technologies across 13M companies" is a real number and a poor basis for a decision. What matters is coverage of the companies and tools in your ideal-customer profile. A provider can track a vast catalog and still be thin on the specific technology you sell against, or on the country, size and industry segment you go after.

Ask for coverage where you sell, meaning how many companies in your target segment have the technology you care about, and how many carry any stack at all. An inferred dataset inherits the coverage bias of its source, so be concrete about it. Job-based data leans toward companies that are actively hiring, which is a strength for finding growing accounts and a limitation for mapping a sleepy incumbent. For context, we have mapped 54M technology-to-company signals, with over 5 million companies carrying at least one technology. The number that should decide your choice is how many of those are yours.

5. Catalog definition and breadth

Two providers can both claim a technology and mean different things by it. One treats a whole product family as a single entry. Another splits it into editions and modules. Before you compare counts, check that you are comparing the same thing.

Ask how the catalog is organized and whether the technologies you care about are tracked as distinct entries. Browse our 33k-technology catalog across languages, frameworks, databases, ERPs and CRMs, and confirm the specific tools in your competitive set are there and defined the way you need them.

The checklist

Line the providers up against the same questions:

  • Detection method, whether web scan, job-posting inference or both, and whether it can see the category you sell.
  • Whether every signal traces to source evidence, and what the confidence score is built from.
  • A last-seen date on each signal, so you can tell a current stack from a stale one.
  • Coverage of your ICP segment and your target technology, not the global catalog size.
  • How the catalog defines the specific technologies in your competitive set.

Test precision and recall on companies you know

Headline accuracy claims are unfalsifiable until you bring your own ground truth. You already have some: your current customers and their stacks, companies you have lost deals to and know what they run, and public case studies and engineering blogs that state a technology outright. Build a list of 30 to 50 companies where you know the answer, split into ones you know use a technology and ones you know do not.

Then run two measurements against each provider.

Precision, how often a match is real. Take the companies the provider says use the technology and check them against your truth. The share that are genuinely correct is its precision. The rest are false positives, and false positives are what burn a list, because every wrong match is an outreach that lands wrong.

Recall, how many real users it finds. Take the companies you know use the technology and check how many the provider caught. The share it found is its recall. The ones it missed are the deals you will never see, because they never show up in a search.

The two trade off, and which one you weight depends on the job. For outbound where a bad match costs credibility, lean on precision and, with inferred data, stick to medium and high confidence signals. For market sizing or research where you would rather see everything and filter later, weight recall and keep all confidence levels. Either way, do the counting on companies you can check, not on the provider's own demo.

Finish by tracing a few matches to their evidence. Open the proof behind three or four signals and read it. If the evidence is thin, or there is none to open, the headline accuracy number was never yours to verify.

Running it on TheirStack

You can run the whole test before you spend anything. Searches that only count matches are free and unlimited, since credits are consumed only when you reveal a company, so you can size coverage of your ICP, check precision and recall against your known companies, and open the evidence behind every signal, all with a free account. We also document that our technographics inherit the coverage bias of our job sources, because that is the first thing to weigh before you trust any inferred dataset, ours included.

TheirStack: integrations with AI assistants, automations and CRMs

TheirStack works with the AI assistants, automation tools and CRMs you already use: ask Claude or ChatGPT through the MCP server, send new matches to Clay, n8n, Make or Zapier by webhook, and create leads in Salesforce or HubSpot.

Test it on companies whose stack you know

Search by technology, see the evidence behind every match, and check it against the companies you already have the truth for, before you spend a credit.

Start free: 200+ free credits.