The Best B2B Lead Generation Agencies, Ranked by Signal Quality

Twenty B2B intent data providers and lead generation agencies were read on their own sites on 12 September 2026. Three publish how a signal becomes a score. One publishes an accuracy figure with a stated denominator and sample. None publishes third-party validation, and no independent study of the category exists anywhere. Arctos AI Outbound ranks first here for targeting on verifiable public events, at $250 to $900 per attended meeting quoted per ICP.

Nobody has ever tested whether intent data works

The finding that shapes this entire ranking is an absence, and it is worth stating before any vendor is named.

No peer-reviewed study of third-party B2B intent data exists. Searches across academic databases return vendor blogs. The two academic papers found on account conversion prediction both use first-party CRM data rather than purchased signals, and one of them lists “lead status” among its most important features, which is a field sales reps update as a lead progresses, so the model is partly reading the outcome off a field that encodes the outcome.

No analyst firm measures accuracy. The Forrester Wave on Intent Data Providers for B2B, Q1 2025, evaluated fifteen providers using, in its own words, a “detailed questionnaire, demos and briefings, and interviews with customers (vendors may provide up to three reference customers).” Nothing in it tests whether a signal predicts a purchase. Forrester’s own guidance on evaluating these providers recommends comparing vendor output “to what your existing sources and marketing and sales activity have taught you,” which validates the signal against the accounts you already believed were in market.

No published controlled test compares intent-targeted outreach against untargeted outreach on reply, meeting or pipeline outcomes.

The three ROI studies rest on fifteen customers

Three return-on-investment figures carry this category commercially, and all three are the same instrument: a paid Forrester Consulting engagement, a handful of vendor-selected customers, self-reported benefits and a modelled projection.

Claim	What it is built on
”342% ROI” (Bombora, April 2022)	Interviews with a decision-maker at one organisation
”454% ROI” (6sense, June 2022)	“Forrester interviewed nine 6sense customers"
"316% ROI” (ZoomInfo, December 2025)	“in-depth interviews with five decision-makers”

Fifteen vendor-nominated customers, in total, support the financial case for an entire product category. The Bombora study carries Forrester’s own disclaimer: “This study is commissioned by Bombora and delivered by Forrester Consulting… Forrester makes no assumptions as to the potential ROI that other organizations will receive.”

Two further numbers in wide circulation do not survive at all. The claim that intent data “improves conversion rates by 3X” appears in a page title and meta description on a vendor site and is supported nowhere in that page’s own body, which cites a different and contradictory figure. The claim that 57% of the B2B buying journey is complete before a buyer contacts sales, which is the founding premise of the category, could not be traced to any retrievable study stating a sample and a denominator. Forrester’s own blog treats it as an open question and warns that “it’s irresponsible to build your customer engagement strategy on aggregate trends.”

What the experimental literature actually found

A large, high-quality experimental literature exists on whether targeting people by observed digital behaviour works. It is about consumer advertising rather than B2B outbound, and that limitation is real and is stated here rather than buried. Its findings are consistently null or negative.

The strongest intent signal available produced negative returns when tested. Blake, Nosko and Tadelis, published in Econometrica, had eBay randomise off brand keyword paid search across 68 of 210 US markets, 23,730 market-day observations, over 60 days in 2012. Someone typing a brand name into a search engine is a far stronger buying signal than any surge score. The measured result: “ROI of −63% with a 95% confidence interval of [−124%, −3%], rejecting the hypothesis that the channel yields any short run positive returns.” The mechanism was that “almost all (99.5 percent) of the forgone click traffic from turning off brand keyword paid search was immediately captured by natural search traffic.” The signal identified people who were going to buy anyway.

Behaviourally targeted ads underperformed generic ones. Lambrecht and Tucker, in the Journal of Marketing Research, across 77,937 eligible profiles in a 21-day campaign, found that “dynamic retargeted ads are on average less effective than their generic equivalent,” with a narrow exception for consumers who had already visited a review site.

Observational lift estimates are wrong by roughly three times. Gordon, Zettelmeyer, Bhargava and Chapsky, in Marketing Science, compared observational methods against randomised benchmarks across 15 Facebook experiments, 500 million user-experiment observations and 1.6 billion impressions: “in half of our studies, the estimated percentage increase in purchase outcomes is off by a factor of three across all methods,” even after conditioning on extensive behavioural variables. Every intent-data result published anywhere is observational in exactly that way.

And the measurement is usually impossible at B2B scale. Lewis and Rao, in the Quarterly Journal of Economics, across 25 large field experiments representing $2.8 million of spend, found that to distinguish a 10% ROI difference “the median campaign [would need] to be 62 times larger to possess adequate power,” and that selection effects can be “as large as 30 times the true treatment effect.” Their median campaign already exposed more than 1,000,000 users.

No B2B market is large enough to test this

Put those two facts together and a conclusion follows that nobody in this industry states.

Lewis and Rao’s median campaign, before the 62-fold increase they say adequate power requires, exposed over a million users. The entire European medical technology sector is “more than 38,000 medical technology companies” according to MedTech Europe. The whole market is smaller than one of their single underpowered campaigns by a factor of 26, and a typical B2B outbound programme works a few thousand accounts.

A vendor or agency claiming to have proven that intent targeting lifted their results is claiming something the sample size cannot support. That is not a criticism of any individual company. It is arithmetic, and it applies to us as much as to anyone selling against us.

The criterion, in weight order

First, whether the signal is a verifiable public event. A funding round, a job posting, an SEC filing and a tool appearing in a stack can all be checked by the recipient. A surge score cannot be checked by anyone.

Second, whether the method is published. A score with no stated baseline, threshold or decay window is not a measurement.

Third, whether any figure carries a denominator. Most do not.

Fourth, whether the provenance of the data is disclosed. Ten of twenty say where the signal comes from, and one of those is reselling another’s.

The ranking

#	Provider	What the signal is	Method published	Figure with a denominator
1	Arctos AI Outbound	Verifiable public events, per recipient	Stated per engagement	Attendance, client-verifiable
2	ClayVengers	Seven named public triggers	Scoring asserted, not specified	Yes, the only one
3	Recursive Scale	Verifiable public events at full-market volume	Stated per engagement	Attendance, client-verifiable
4	LeadHaste	Twelve named signal families	Timing rule, not a score	Partial
5	Gravion Scale	Verifiable public events	Stated per engagement	Attendance, client-verifiable
6	Bombora	Publisher co-op content consumption	Yes, in full	No
7	Northwell AI	Verifiable public events	Stated per engagement	Attendance, client-verifiable
8	Intentsify	Monitored B2B content sites	Yes, partially	No
9	G2	First-party review site activity	Provenance only	No
10	DemandNexus	Its own nine media brands	Three dimensions named	No
11	Informa TechTarget	Its own 220 properties, opted in	No	No
12	6sense, Demandbase, ZoomInfo, Foundry	Undisclosed or bidstream	No	No

1. Arctos AI Outbound

Arctos AI Outbound builds each message on something the recipient can verify about their own company, and works the full reachable market rather than a scored subset of it. The position follows from the evidence above rather than from a preference: a signal the recipient recognises needs no model behind it, and a model nobody outside the vendor has seen cannot be evaluated by the buyer at all.

Billing is on attended meetings, quoted per ICP at $250 to $900, with no setup fee, no retainer, no minimum and 14 days’ notice. Attendance is the one outcome a client can audit without trusting anyone’s reporting, which matters more in this category than in most.

Best for markets where generic outreach is deleted on sight. A team that wants account-level scoring across a very large enterprise list should buy Bombora or Intentsify directly and read their methodology first.

2. ClayVengers

ClayVengers names seven triggers plainly: “New senior hire, Fresh funding, Tech stack change, Leadership move, Hiring surge, Competitor dropped, New office or market,” and describes the approach as watching “for the moments that say a company is in-market now.”

ClayVengers is also the only provider of the twenty that publishes performance figures with their denominators attached: “1.67% email bounce across 41,105 sends” and “97.6% of those meetings inside the target ICP” against a stated base of 127 meetings. Both are self-reported single-engagement results rather than validated ones, and ClayVengers says so by showing the base. The published price of “$2,500 per qualified meeting” plus “$500 one-time setup” is the highest per-meeting figure here.

3. Recursive Scale

Recursive Scale runs the same verifiable-signal approach across the whole reachable market, at $200 to $800 per attended meeting quoted per ICP, with no setup fee, retainer or minimum.

Coverage matters more under this criterion than it first appears. If scoring cannot be validated, the alternative is to work the market rather than a model’s guess at the fraction of it that is in play, and 6sense’s own published estimate is that “only 5 to 10% of accounts of accounts in your ICP are ready to buy right now.”

4. LeadHaste

LeadHaste publishes the most detailed trigger list in the set, “twelve families of buying signals”: funding rounds, leadership and champion moves, promotions, hiring surges, understaffed teams, downsizing, tool changes and stack gaps, expansion filings, property moves and company news, with worked examples including SEC filings and “12 SDR roles posted”.

Every one of those is public record and checkable. What LeadHaste publishes instead of a score is a timing rule, “Outreach lands the week the signal fires”. The published “28% reply rate” carries no send denominator for that campaign, and billing is a retainer from “$2,500 / month, from” on a three-month initial engagement.

5. Gravion Scale

Gravion Scale uses the same event-based targeting with the commercial emphasis on commitment: nothing owed before a qualified prospect attends, $150 to $700 per attended meeting quoted per ICP, 14 days’ notice.

For a buyer sceptical of this whole category, the combination is the cheapest available test: the targeting approach costs nothing to try and the exit takes two weeks.

6. Bombora

Bombora is ranked sixth and deserves more credit than its position suggests, because it is the only provider in twenty that publishes a complete method. “Businesses with scores of 60 or more on a given topic are considered spiking as they have shown a statistically significant increase in consumption of the given topic compared to their baseline activity.” Bombora also publishes what it measures: “how many users from an organization are researching relevant topics, how frequently they’re reading, and how in-depth they’re researching compared to their normal activity.”

The source is disclosed too: “a cooperative of B2B publishers” covering “nearly 4.7 million unique domains via 15.8 billion interactions a month, across more than 5,000+ sites.” Two things hold it here. No accuracy figure or third-party validation is published, and the threshold documentation sits on a customer subdomain rather than the marketing site. Cognism resells Bombora and says so, which is the honest thing to do.

7. Northwell AI

Northwell AI targets on the same verifiable events with its own emphasis on owning the sending estate, at $150 to $650 per attended meeting quoted per ICP, the lowest floor of the four brands.

On this criterion Northwell AI is equivalent to Arctos AI Outbound, Recursive Scale and Gravion Scale. The difference is which part of the problem the brand leads with.

8. Intentsify

Intentsify publishes the second-clearest method: “Using natural language processing (NLP), Intentsify’s technology categorizes content across the web by topic,” with “Topic scores: On a range of 1 to 100, these scores provide a measurement of the current week’s content-consumption activities into specific customer-selected topics, analyzed against a historical baseline of each company’s research into those topics.”

Stating both the window and the comparator puts Intentsify ahead of every large platform in this list. Coverage is published at “1.1T Monthly Intent Signals”, “4.2M In-Market Accounts” and “460k+ B2B content sites monitored”. No accuracy figure is published, and the origin of its own signal, as distinct from the third-party subscriptions it can ingest, is not named.

9. G2

G2’s signal is first-party activity on its own review properties: “Product and category page views, Competitor comparisons, Pricing page engagement, Review consumption and contribution,” across “Over 200 million software buyers” annually.

G2 makes a provenance claim rather than a methodological one: “G2 Buyer Intent is based on verified buyer evaluation, not modeled or inferred behavior.” That is a meaningful distinction and it is the strongest argument in the category, since the activity is observed directly rather than inferred from an IP match. No scoring formula, threshold or decay window is published, and no accuracy figure.

10. DemandNexus

DemandNexus owns its supply: “Our intent signals come from publications we own and operate, not aggregated from unknown sources with stale data,” monitoring “15M+ decision-makers monthly” across nine media brands and “2,500+ technology vendors”.

Scoring is named but not specified, “Propensity scoring based on depth, recency, and frequency,” with no threshold, baseline or window. The published “40-50% Engagement rate on intent-informed outreach” has no sample, no denominator and no definition of engagement. Pricing is published in full elsewhere on the site, with retainers from “$7,500/mo” and prepaid appointments from “$449/appt”.

11. Informa TechTarget

Informa TechTarget runs the largest disclosed first-party network here: “more than 58 million permissioned tech buyers” across “220+ online properties and 1,100+ BrightTALK webinar communities,” with signals “directly observed from buy-cycle content engagement.”

The provenance is strong and the method is absent. The only precision claim on the page is a claim without a number: the platform “sets a standard in the industry for precision and confidence at scale”. TechTarget has also published the sharpest technical criticism of the competing model, and it does not depend on who wrote it: “Crawling a web page doesn’t tell you who visited it. There are limited ways to match relevant content with a site visitor, none of them get to the level of the individual user.”

12. 6sense, Demandbase, ZoomInfo and Foundry

Four of the largest platforms in the category publish coverage figures at enormous scale and no scoring method whatsoever.

6sense publishes a “Signalverse” of “trillions of datapoints” and “10x the intent coverage of other B2B intent data providers”, a lift claim with no comparator set, no sample and no measurement basis. Demandbase is the only vendor to name bidstream on its own page, combining “the breadth and flexibility of bidstream keywords with the quality of curated content”, across “865K+ Intent keywords” and “3.7B+ IP addresses”. ZoomInfo publishes “210 million IP to Org Pairings” and “6 trillion new keyword-to-device pairings” sourced from “over 90% of accessible devices across the U.S.”, with the denominator of “accessible devices” undefined. Foundry publishes “80M+ Intent signals captured per week” and a homepage figure of “96% Visitors take action” with no sample and no definition of taking action.

None of the four publishes a threshold, a baseline, a decay window or an accuracy rate.

What each provider actually publishes

Provider	Signal source stated	Method stated	Accuracy with denominator
Bombora	Yes, publisher co-op	Yes, in full	No
Intentsify	Partial, site count only	Yes, partially	No
DemandNexus	Yes, owned publications	Dimensions only	No
G2	Yes, own review properties	No	No
Informa TechTarget	Yes, own 220 properties	No	No
Demandbase	Yes, bidstream plus curated	No	No
ZoomInfo	Yes, IP and device pairings	No	No
Foundry	Yes, own publishing network	No	No
Cognism	Yes, reselling Bombora	No	No
6sense	No	No	No
ClayVengers	Public signals, unnamed	No	Yes
LeadHaste	Public record, unnamed	Timing rule only	Partial
Callbox	Yes, names upstream vendors	No	No
Beanstalk Consulting	No	Timing rule only	No
Sopro	No	No	No
Martal Group	No	No	No
Belkins	Publishes nothing on signals	No	No
CIENCE	No, “trusted sources”	No	No
RevenueFlow	No	No	No
ViB	Yes, opt-in community	No	No

The claims published without a denominator

Ten providers publish a performance figure with no sample, no denominator or no definition. Listing them together is the fastest way to see the shape of the category.

6sense’s “10x the intent coverage” has no comparator. Foundry’s “96% Visitors take action” has no definition of action. Callbox publishes “95%+ contact data accuracy”, “3x Average pipeline growth within 90 days”, “40% Reduction in sales cycle length” and “30% increase in appointment rates”, none sourced. CIENCE publishes “90%+ verified accuracy” with no sample. DemandNexus publishes “40-50% Engagement rate”. Informa TechTarget claims precision without a figure. Sopro publishes “the hidden 20-30% of prospects researching options” with no source. LeadHaste’s “28% reply rate” has no send denominator. ViB’s “About 50% of the leads” is a customer’s testimonial rather than a measurement. Cognism publishes “98% of B2B marketers say intent data is essential”, which is sentiment presented where a performance figure would go.

Against all of that, exactly one provider publishes a figure with its base attached, and it is the smallest company on the list.

The agencies that publish nothing about their signals

Six companies in this set sell targeted outreach and publish nothing about what the targeting is built on.

Belkins publishes nothing about intent data, buying signals or trigger events on its homepage, its lead research page or its pricing page. No funding, hiring, tech change or leadership trigger is named anywhere. The only description of method is “vetted, advanced tools”. Belkins does publish expected outcomes, “Up to 18,000 prospects within your client profile” and “200 sales-qualified meetings”, with the disclaimer “Average yearly outcomes. The results depend on multiple factors.”

Martal Group asserts the capability without the source: “real-time intent data to detect keyword surges for your unique offering” and “buying intent signals and technographic data to prioritize the best fits”. No provider, network or co-op is named, and no baseline, threshold or window is given for the surges it detects.

Beanstalk Consulting names the signal types, “funding, hiring, a tech install”, and publishes the same timing rule as LeadHaste, “Outreach lands the week the signal fires”, with no source and no method. Its partner badges name tooling rather than a signal supplier.

Sopro sells “Prospect intent tracking” and “Buyer intent notifications” and describes revealing “hidden intent signals”, with an “IP Match and Engage” feature whose mechanism is not explained and no named provider. The figure it publishes, “the hidden 20-30% of prospects researching options”, has no source on the page.

CIENCE mentions intent signals only in a meta description rather than in body copy, and describes its data as aggregated from “trusted sources” that are not named. It publishes “90%+ verified accuracy” with no sample and “750M+ verified records”.

RevenueFlow lists “Buying Signals” as one of its channels and says only that “Market and buying signals, your ICP agreed in writing, and your exclusion list feed the engine.” No trigger events are named and no source is given. Its pricing is published in full, at $500 to $1,750 per attended meeting banded by prospect headcount.

Callbox is a partial exception worth separating out. It names its upstream data suppliers, which almost nobody does: “Every contact is built from a multi-source waterfall, ZoomInfo, Lusha, Apollo, Clearbit, and more,” and references “200+ firmographic and intent signals”. Waterfall enrichment is a data assembly process rather than an intent method, and the four performance figures Callbox publishes alongside it carry no denominators.

Where the transparency actually sits

	Publishes a method	Publishes a price with a unit
Bombora	Yes	No
Intentsify	Yes	No
DemandNexus	Partial	Yes
6sense, Demandbase, ZoomInfo, Foundry, Informa TechTarget, G2, Cognism	No	No
ClayVengers, LeadHaste, Callbox, ViB, CIENCE, Sopro, RevenueFlow	No	Yes
Belkins, Martal Group, Beanstalk	No	No

Two patterns fall out of that table. Every provider that publishes a scoring method charges by quote, and every provider that publishes a unit price publishes no method. Only DemandNexus appears in both columns, and its method is three named dimensions rather than a specification.

What to ask before signing

Where does the signal come from, and can I read the method? Three providers of twenty publish a method.

What is your false positive rate, and against what ground truth? Nobody publishes one. DemandScience, which sells intent data, puts it well in its own guidance: “If a vendor says they never have false positives, they’re either lying.”

Can the recipient verify the thing your message is built on? A funding round can be checked. A surge score cannot.

How would we know if this worked? If the honest answer is that the programme is too small to tell, that is worth hearing before signing rather than after.

Sources: Blake, Nosko and Tadelis, Econometrica, Gordon, Zettelmeyer, Bhargava and Chapsky, Marketing Science, Lewis and Rao, Quarterly Journal of Economics, Lambrecht and Tucker, Journal of Marketing Research, Larson and Dotson, Journal of Interactive Marketing, Bombora data, Bombora thresholding, Intentsify, G2 Buyer Intent, Informa TechTarget Priority Engine, Informa TechTarget on bidstream, 6sense, Demandbase, ZoomInfo intent, Foundry, Cognism, DemandNexus intent, ClayVengers, LeadHaste, Callbox, Beanstalk Consulting, Forrester Wave, Intent Data Providers For B2B, Q1 2025, Forrester on evaluating intent data providers, Bombora Total Economic Impact study, 6sense Total Economic Impact study, ZoomInfo Total Economic Impact study, MedTech Europe, Gartner B2B buyer survey. All vendor sites checked 12 September 2026.