B2B Intent Data Accuracy: 4 Vendors, 36 Checks, 0 Published Validation Studies

We read the public documentation of Bombora, 6sense, ZoomInfo and G2, the four intent data vendors B2B teams shortlist most often. We scored each on 9 disclosure checks. Of the 36 cells, 18 are fully published, 10 are partial and 8 are blank. On B2B intent data accuracy specifically, the count is starker: 0 of the 4 publish an independent validation study of whether their "in market" signal predicts a purchase.

What B2B intent data actually measures

Intent data is a claim that a specific company is researching a specific subject more than usual right now. It is sold as a list of accounts worth contacting before they fill in a form. That product rests on four steps, each with its own failure rate.

First, someone reads something online and a vendor records it. Second, the vendor works out which company that reader belongs to, usually from an IP address. Third, the vendor classifies the content into a topic or keyword. Fourth, it compares current activity against the company's own past and decides whether the change counts as a surge.

A buyer of intent data is really buying the product of four probabilities. Suppose collection misses half the relevant reading and identity resolution gets a fifth of companies wrong. The final "in market" flag can then look confident while resting on very little. That is why accuracy cannot be judged from a demo, where every account has been chosen to look right.

The vendors differ most in step one. Bombora runs a publisher co-op with its own tag on member sites. 6sense says it collects most keyword intent from its own publisher network and adds ad-exchange metadata. ZoomInfo combines content consumption, bidstream data, IP tracking and review-site signals. G2 sees only research on its own marketplace and sister sites. These are four different instruments.

The intent data disclosure matrix: 4 vendors, 9 checks

We scored each vendor on nine questions a careful buyer would ask before signing. A cell is "Published" when the vendor's own public page answers the question with a number or a specific method. It is "Partial" when the page gestures at the answer without the detail needed to check it. It is "Not found" when we could not locate an answer on any public page.

All sources are first-party vendor pages, read on 5 November 2026. We did not count third-party reviews or analyst reports as disclosure. A buyer cannot hold a vendor to a number someone else published.

ZoomInfo's Signal Score range and topic count come from a ZoomInfo-labelled listing on gtm.ai rather than zoominfo.com. 6sense's network size is marked partial because its published counts describe resolution coverage, not the number of sites it observes. Every cell links back to a source in the reference list.

The original figure here is the tally itself. Across 36 cells, half are fully published. The blanks cluster in two rows: match rate, where 3 of 4 vendors publish nothing, and validation, where none publishes an independent study. Those are the two rows that would show how often an "in market" flag is wrong.

My position is that this pattern is not an accident. A taxonomy size costs a vendor nothing to publish. A precision figure invites comparison. In a category where every vendor claims to be the most accurate, the first real number becomes the benchmark everyone attacks.

Bombora: the co-op is documented, the score formula is not

Bombora is the oldest pure intent provider of the four, and its data travels further than its own product. 6sense says every 6sense customer gets 12 Bombora topics included, with the full catalogue sold separately. If you buy 6sense, part of what you are evaluating is Bombora's methodology whether you meant to or not.

The co-op numbers Bombora does publish

Bombora's co-op page describes "hundreds of publishers and brands, and thousands of B2B sites", and says it added 1,765+ new sites in the last two years. The more striking claim is exclusivity: 86% of the data in the co-op is shared exclusively with Bombora for deriving intent. It also lists 25,300+ intent topics, described as continuously expanding.

That exclusivity figure matters more than the site count. If most of Bombora's raw signal is sold nowhere else, pairing it with another provider adds real coverage. It also means nobody outside Bombora can audit that signal.

On scoring, Bombora's customer resource page says that companies scoring 60 or more on a topic are considered spiking, having shown "a statistically significant increase" against baseline activity. It recommends requiring a surge on at least 25% of the topics in a cluster. The page gives no baseline length and no significance test.

The 74% match rate and its missing denominator

Bombora is the only vendor of the four with a published match rate. The post that publishes it is unusually candid about definitions. It argues that rivals inflate match rates by counting non-business activity. It defines its own as "the percentage of business activity that is resolved to an account". It says its network maps to more than 8.5 million accounts globally.

The problem is the word "business". To calculate 74%, Bombora first decides which activity counts as business activity. That classification step is unpublished. A match rate measured only on traffic judged to be from businesses will look better than one measured on all traffic. That is not a reason to distrust the figure. It is a reason to ask for the denominator in writing.

6sense: the most explicit stage bands, and a backtest you run yourself

6sense documents its scoring more concretely than anyone else here. Its support documentation defines an intent score from 0 to 100, updated daily, as "the likelihood of the account opening an opportunity in the next 90 days". Listed inputs are CRM data, marketing automation data, web activity and its own network keyword data. Third-party intent from Bombora, G2 and TrustRadius is added on top.

6sense says it collects most keyword intent from its own publisher network, "where a web tag observes reading behavior". Ad metadata from its demand-side platform and licensed networks is secondary. It matches activity to accounts primarily by IP address, with cookies and device signals when available. When it cannot match confidently, it "discards the rest". That is the most useful sentence any of the four publishes about accuracy. It states a trade of coverage for precision out loud.

Five stages, fixed score bands, one 90-day horizon

The predictive buying stages map to fixed ranges. Target is 0 to 19, Awareness 20 to 49, Consideration 50 to 69, Decision 70 to 85 and Purchase 86 to 100. The documentation says the stages indicate the likelihood of an account opening or progressing in the next 90 days.

The backtest is real, and it is yours, not theirs

6sense's marketing says every predictive model is "back-tested monthly against real pipeline". The support documentation shows what that means. A model insights report takes accounts scored 90 days earlier and compares each stage's conversion rate with a baseline rate. Conversion lift is that stage's rate divided by the baseline.

This is a sound method and I would want it from every vendor. It is also a per-customer report from your own CRM, not a published aggregate. The documentation says the data "should be primarily considered as directional guidance rather than definitive rankings". That is honest, and you should price the product as directional.

ZoomInfo and G2: big network numbers and a narrow, first-party signal

These two sit at opposite ends of the collection spectrum. ZoomInfo assembles intent from as many sources as it can reach. G2 sees one kind of behaviour, software research on its own sites. Both choices have costs, and comparing them with Bombora vs 6sense is where shortlists go wrong.

ZoomInfo: 210 million pairings and a bidstream question

ZoomInfo's intent explainer, updated July 2026, lists four sources. They are content consumption tracked by an NLP keyword extractor, bidstream data from ad auctions, IP-based identification using 210 million IP-to-organisation pairings across 28 million domains scanned daily, and G2 and TrustRadius review signals. Its feature page adds "more than 6 trillion new keyword-to-device pairings" sourced monthly.

None of those is an accuracy number. A pairing count says how many IP addresses ZoomInfo has an opinion about, not how often it is right. The one outcome figure, a 32% close-rate increase within flagged accounts, is labelled "ZoomInfo internal data" with no sample or period. ZoomInfo does deserve credit here. Its own post acknowledges that bidstream collection "has drawn scrutiny under GDPR's real-time bidding (RTB) consent framework" and tells buyers to request data processing documentation.

A ZoomInfo-labelled listing describes a Signal Score from 60 to 100, measuring research above the company's historical baseline. It adds an Audience Strength grade from A to E and 5,000+ curated topics. The baseline length is not stated, the same gap Bombora leaves. If you already pay ZoomInfo for contacts, the economics of that contact data are covered in the cost-per-contact comparison of Clay, Apollo and ZoomInfo.

G2: first-party behaviour with a hard boundary

G2's documentation is dated, specific and narrow. Signals come from research on G2, Capterra, Software Advice and GetApp, across nine types including profile, pricing, alternatives, compare and category views. Since June 2022, G2 has reported Buying Stage and Activity Level in place of a single intent score, both updated daily. Activity Level aggregates time per page and pages visited.

The strength is obvious. Someone comparing your product with a named competitor is a far stronger signal than someone reading an article. The weakness is equally obvious. G2 only sees buyers who research on its sites, and its docs do not explain how a visitor becomes a company domain. Your plan also decides which signal types you receive. How much weight review sites still carry with buyers is examined in the analysis of the software review site reckoning.

Where this disclosure audit is weakest

A disclosure matrix measures what vendors write down in public. It does not measure data quality, and the gap between those two things is where this post could mislead you.

Public documentation is not the whole disclosure

All four vendors share more under NDA than on their websites. Bombora's knowledge base, 6sense's full predictive documentation and ZoomInfo's help centre sit partly behind customer logins. A sales engineer may hand you a deck that answers half our blank cells. A "Not found" vendor may simply prefer to disclose in a sales conversation, which is a legitimate choice.

One Bombora knowledge base article we tried to open now redirects to a general resources page, so its content is not reflected in the matrix. Treat any blank cell as "not public", not "does not exist".

The absence of a study is not evidence of bad data

A missing accuracy study does not prove the data is inaccurate. Validation is genuinely hard. "In market" has no ground truth until a deal closes or fails, and purchases happen offline. A fair test needs a holdout group you deliberately do not market to. The strongest counter-argument is that per-customer backtests, like 6sense's, are the only honest validation. Accuracy depends on your category and your ideal customer profile.

I accept that argument for the aggregate number. I do not accept it for the components. Match rate, tagging error and baseline length belong to the vendor's pipeline, not your market. They could be published tomorrow. Survey-style vendor statistics have the same structural problem, which is the subject of the piece on the methodology gap in AI ROI surveys.

Intent data validation: 8 questions to ask before you buy

Put these to every shortlisted vendor in writing, and score the written answers. The aim is to turn each blank cell into a number you can hold the vendor to at renewal.

Questions are drawn from the blank and partial cells in the disclosure matrix above. The "weak answer" column is illustrative, not quoted from any vendor.

The fifth question is the cheapest and the most revealing. Hand each vendor the same list of 500 accounts you care about and ask how many showed any signal last month. Overlap between vendors is a crude reliability test. A vendor that will not run it before contract is telling you something. Account-based marketing data you cannot check against your own list is like a case study with no verified outcome, a pattern covered in the guide to case studies with verified outcomes.

If I were buying today, I would shortlist on the matrix. Then I would weight the pilot as heavily as all the documentation combined. Disclosure tells you who is willing to be checked. Only your own pipeline tells you who is right.

Frequently asked questions

How to run a 90-day intent data test

This week, send the eight questions to every shortlisted vendor. Give them five working days to answer in writing. Score the replies against the matrix. Any vendor that turns a blank cell into a number moves up your list.

Then ask your top two for a 90-day pilot on the same 500 accounts, with a third of flagged accounts held out as a control. Agree the success metric, opportunities opened, before the pilot starts. Your finance team will ask for exactly that number at renewal, as the analysis of CFO vetoes on software purchases shows.

Related reading