Intent data accuracy: how to validate Bombora, G2 and Demandbase signals before you trust them

When a demand generation leader asks whether intent data is accurate, they are really asking two separate questions. First, did the provider attribute the activity to the right company? Second, does that activity mean the company is in a buying cycle for what you sell? A signal can pass the first test and fail the second.

Vendors are open about this. Demandbase describes its intent signals as spikes and trends that indicate potential opportunities "based on probability", and its account identification FAQ says the model "may have some deviations due to dynamic IP usage and remote work". 6sense's own guide to intent data warns that not every spike in research activity means someone is about to buy, and that a competitor doing market research can look like a buyer. None of the major providers we reviewed publishes an accuracy rate or a validation method you could apply to your own market.

That is not a reason to drop intent data. It is a reason to test it like any other channel. If you need the basics first, start with what B2B buyer intent data is, then come back here for the validation method.

Each source sees a different slice of the buying journey, and its failure modes follow from how it collects data.

• Third-party co-op and publisher networks (Bombora): Bombora collects content consumption through a proprietary JavaScript tag on the sites in its data co-op, which it describes as thousands of B2B publishers and websites, and it tracks more than 25,300 intent topics (vendor-reported). Each account gets a Company Surge score from 0 to 100 per topic; Bombora treats 60 or more as surging. RollWorks documents the calculation as account-level intent over a 3-week period relative to a 12-week baseline, updated weekly.
• Review-site intent (G2): G2 Buyer Intent captures activity on G2, Capterra, Software Advice and GetApp, reported at the product level. Signal types include views of your profile, pricing, category, comparison and alternatives pages, plus competitive views in a shared category. G2 scores each company on Buying Stage and Activity Level, updated daily, and your plan tier determines which signals you receive.
• Platform intent (Demandbase, 6sense): Demandbase collects intent through the bidstream of its B2B DSP, which connects to over one million web publishers, and uses natural language processing with a vocabulary of more than 650,000 keywords to match articles to keywords. It flags trending intent when signals rise by two standard deviations in the last 7 days compared with the last 8 weeks. 6sense says it aggregates signals across thousands of sites and that providers resolve anonymous browsing by matching IP addresses and digital fingerprints against company databases.
• First-party website and product signals: your own site visits, pricing and comparison page views, form fills, and product trial or usage events. 6sense's guide calls first-party intent highly accurate because it shows direct interest, with an important limit: it only captures accounts that already engage with you.

Platforms also resell each other. Demandbase supports Bombora, G2 and TrustRadius as third-party intent partners, each under a separate paid license, and notes that Bombora rolls multiple keywords into topics while Demandbase lists keywords individually. When you validate, test each feed separately before you test the blend.

Each failure point below is documented by a vendor or follows from how the data is collected. Knowing which one you face tells you which fix to try.

• IP-to-company matching: anonymous traffic is matched to companies through IP addresses, cookies and device signals. Demandbase says separating parent and subsidiary traffic is challenging when they share IT infrastructure, that mobile traffic is harder to identify, and that its cookie-based integration identifies about 20 percent of B2B traffic, roughly 10 percent of all site traffic, varying by implementation.
• Remote work and VPNs: when employees work from home networks or connect through a VPN, the IP address no longer points cleanly at the employer. In 2020 MarTech reported that 6sense saw matching drop during the shift to remote work, recover after a re-analysis, and moved to confidence levels of high, moderate and low because several people in one household may share an IP address. Demandbase lists public VPN and ZTNA providers among the data points its identity graph uses.
• Small samples: a surge compares an account with its own baseline. For a mid-market company with a handful of researchers, one person reading three articles can move the score. Demandbase scores activity relative to similar-sized accounts for this reason, but the underlying count is still small.
• Topic and keyword taxonomy: a broad topic or keyword catches the wrong readers. Demandbase gives the example of an olive oil article being matched to a company that sells oil field equipment. Bombora recommends a topic threshold of at least 25 percent of the topics in a cluster, so an account must spike on several related topics, not one.
• Review-site context: G2 notes that an alternatives signal does not necessarily mean your product is being compared directly, and that cross-site research can raise signal volume for some products.
• Wrong reason, right company: students, analysts, job seekers, partners and competitors all read about your category. No provider can see intent behind a page view.

Use this table to plan the test for each feed you pay for. The last column is the specific check we run before a source goes into scoring or sales alerts.

The goal is a fair comparison: accounts that show a signal and receive action, against similar accounts that show the same signal and receive none, against fit accounts with no signal at all. The outcome is meetings, opportunities and pipeline, not engagement.

1. Freeze the rules before day 1: write down the ICP-fit account list, the surge or trending definition for each source, the stage definitions for meeting and qualified opportunity, and who owns the readout. Changing any of these mid-test breaks the comparison.
2. Backtest against closed-won: take deals won and lost in the last four quarters and check whether each source flagged those accounts in the 90 days before the opportunity was created. Demandbase keeps up to 13 months of intent history, depending on availability; Demandbase notes that Bombora data inside its platform has no history, so you may need Bombora directly or a forward-only test.
3. Build three groups from ICP-fit accounts: surging accounts that get the full play (sales alert, ads, outreach), surging accounts held out with no extra action, and a baseline of fit accounts with no signal. Assign surging accounts to action or holdout at random, not by rep preference.
4. Run the programs for 90 days without moving accounts between groups. If an account in the holdout books a meeting on its own, record it; that is exactly the evidence you need.
5. Collect sales feedback every week: for every alert, the rep records a reason code such as relevant, wrong company, wrong person, competitor or partner, existing customer, or no response. Rejections by reason tell you whether the problem is matching or meaning.
6. Audit the keyword sets and topics at day 30 and day 60: remove terms that only appear on rejected accounts and add terms seen in won deals, and log every change so the readout knows what was tested.
7. Read out on day 90: compare meeting rate, opportunity rate and pipeline per account across the three groups and for each source. Keep the sources and topics where the action group beats both the holdout and the baseline; fix or cut the rest.

Be careful with small groups. If each group holds a few dozen accounts, a difference of two or three opportunities can be chance. Widen the test or extend it a quarter before you cut a source.

Measure at the account level and use the same stage definitions as the rest of your reporting. Our guide to ABM conversion rates by stage explains how to set a baseline from non-target accounts, the same logic the holdout uses.

Four results come up often. If surging accounts beat the baseline but the holdout performs as well as the action group, the signal predicts buying but your play adds little; fix the play. If surging accounts do no better than the baseline, the source or the topic set is not predictive for your market; narrow the topics or cut the feed. If sales rejects alerts mainly as wrong company, the problem is identity matching, so look at subsidiaries, regions and remote-heavy segments. If the rejections are mainly wrong reason, the problem is the taxonomy.

Report precision as well as lift: the share of flagged accounts that sales accepted as relevant, and the share that reached a meeting. Reps keep acting on alerts they trust, so a precise, smaller feed often beats a noisy one.

Validation is not a one-time project. Topics drift, keyword sets grow and territories change. We keep three routines running after the first readout.

• A weekly account review where sales and marketing look at the alerts, the reason codes and the next action, so turning intent signals to sales action stays a habit rather than a dashboard.
• A keyword and topic audit every 30 days. For Demandbase customers, our method for auditing Demandbase intent keyword sets covers duplicates, off-topic terms and testing a refined set against a control.
• A quarterly backtest that repeats the closed-won check with the new quarter of deals, so the scoring weights follow what actually predicted pipeline.

A disclosure: PlanConversion is an official Demandbase partner, and our team has implemented 6sense for more than five large enterprises. The method above is vendor-neutral on purpose; we run the same holdout whichever source a client pays for, and we recommend cutting feeds that do not beat the baseline.

Our ABM services include intent validation, keyword set design, CRM routing and the reporting that shows whether intent spend creates pipeline. If you want a second opinion on your intent data before a renewal, talk to our team.