What Is a B2B Data Provider?

A B2B data provider is a company that sells information about businesses and the people who work at them, most often verified work emails, phone numbers, job titles, and company firmographics like industry, employee count and headquarters location.

The reason the term is slippery is that it describes a job rather than a business model. A vendor with a proprietary research operation and a vendor that resells three APIs behind a UI both answer to it, and both will tell you they are a data provider. So will a verification service that holds no contact data at all and only tells you whether the address you already have still works.

The buying question is therefore never just which provider. It is which kind of provider, for which field, against which list. A team that needs firmographics on 40,000 accounts and a team that needs phone numbers on 300 named people are shopping in the same category for two products that share almost nothing except the word data.

The rest of this page is the structure underneath that word: the four types, where the data physically comes from, how coverage and accuracy are actually measured rather than claimed, what the market charges, and the specific test that settles a shortlist.

Four, and they are different businesses rather than different price points: originators that collect data first-hand, database licences that sell access to a stored index, orchestration layers that query many providers per record and store nothing, and verification services that only check data you already have.

Originators. These vendors run their own collection: research teams, contributor networks where users trade their contacts for access, crawlers, or licensed feeds from registries and partners. Their coverage is genuinely theirs, which is why they can be strong in a niche nobody else covers, and why their weak spots are structural rather than temporary. ZoomInfo, Cognism and Apollo are the names most buyers meet here.

Database licences. You buy seats or export quota against a stored index, and you browse it as a research surface. The value is that the whole market is sitting there to be filtered, and the constraint is that everything in it was true on the day it was written down. This is the model whose marketing leans hardest on record counts.

Orchestration layers. No database of their own. A record comes in, the engine asks providers in sequence until one returns a deliverable result, and it stops there. Coverage is the union of the pool rather than any one member's, so a person missing from one source is often found by the next. Clay and Cleanlist AI are both built this way. The tradeoff is that you cannot browse what you do not store: an orchestration layer answers questions about specific records rather than letting you wander a market.

Verification services. They hold no contact data and sell one narrow act: telling you whether an address or a number you already have is still live. ZeroBounce and Emailable are two Cleanlist AI buys from. They are the cheapest useful spend in the category and the most consistently skipped.

Two adjacent categories get filed under the same search term and are not the same purchase. Intent providers sell inferred signals about which companies are researching a topic, which is a behavioural product rather than a factual one and cannot be verified against a second source the way an email can. Web data infrastructure vendors sell the pipes rather than the records, and you build the dataset yourself.

From six sources, and every provider is some blend of them: contributor networks, web crawling, public registries and filings, licensed partner feeds, first-party customer contributions, and inference from patterns. Almost no B2B contact data is collected from the person it describes.

This is the question buyers ask least and the one that predicts the most. Sourcing decides coverage shape, decay speed, and what a provider can do when a regulator or a subject asks where a record came from.

Contributor networks, sometimes called give-to-get. Users grant access to their own address books or inboxes in exchange for platform credits, and the aggregate becomes the database. This is why some databases are enormous and why their strength clusters exactly where their user base works: heavy on North American software sales, thin on European mid-market manufacturing.

Web crawling. Public profiles, company sites, press releases and job boards, parsed at scale. Excellent for job titles and company facts, which are published deliberately. Poor for phone numbers, which almost never are.

Public registries and filings. Company registries, incorporation records, regulatory filings and patent databases. Slow to change, highly reliable, and the backbone of firmographic rather than contact data.

Licensed partner feeds. One provider buying another's data and reselling it inside its own product. This is the reason two vendors can return the identical wrong phone number, and the reason a multi-vendor test sometimes produces less independence than the vendor count suggests.

First-party contribution. Customers uploading their own CRMs into a shared pool. Effective, and the most compliance-sensitive of the six.

Inference. Generating a likely address from a name and a known company email pattern. It is a guess until something verifies it, which is exactly why the verification step below is not a formality.

The practical takeaway is that email and phone behave differently because they come from different places. Emails can be inferred from a pattern and then confirmed against a mail server. Phone numbers cannot be inferred at all, which is why phone coverage is where providers separate most sharply and why phone credits cost more everywhere.

A data provider sells you its own data. An orchestration layer sells you access to many providers with the routing, deduplication and verification handled for you, and holds no database of its own. The distinction decides what happens to a record nobody in the pool holds, and it changes what a coverage claim can even mean.

With a single provider, a miss is final. The record is not in that index, and the only fix is to buy a second subscription and run the misses again by hand.

With an orchestration layer, the miss is the trigger for the next call rather than the end of the run. A waterfall walks the pool in cost order, stops at the first deliverable result, and never calls or bills the providers after the hit. Cleanlist AI runs 25+ providers this way and bills per returned field: a verified work email is 1 credit, a phone number 10, both together 11, and a lookup that walks the whole pool and returns nothing is 0.

The honest cost of the model is that you lose the browsable index. A licensed database lets an analyst sit and explore a market, filtering 300,000 companies by six attributes and reading the result. An orchestration layer answers about the records you bring it. If the job is market research rather than list completion, the licence is the right purchase and no amount of waterfall coverage substitutes for it.

The second honest cost is dependency. An orchestration layer's coverage is its suppliers' coverage, so a supplier that degrades, gets acquired or reprices moves the product. The mitigation is pool size and substitutability rather than a promise, and it is a fair thing to ask any orchestrator to describe.

By match rate against your own list, not by the vendor's record count. Match rate is the percentage of the rows you submitted that came back with the field you asked for, and it is the only coverage number that describes your market rather than the vendor's.

Record counts are the loudest number in the category and the least useful. ZoomInfo publicly claims 321M professional profiles and 104M company profiles, Apollo 275M contacts, Lusha 150M, Cognism 70M. Those are cumulative totals of everything ever collected rather than counts of what is current, and none of them tells you whether the 400 people on your list are in there.

A single-source database typically matches 50% to 75% of a real ICP list, while a multi-provider waterfall clears 85%, on Cleanlist AI's own provider comparison published in the State of B2B Data 2026 report. The gap is arithmetic rather than marketing: the moment you query one index you inherit its blind spots, and the recovery from asking a second and a third compounds across a list.

Coverage also varies by segment far more than it varies by vendor. Seniority, region, company size and industry each move it more than the logo on the invoice. Directors at 5,000-person North American software companies are findable almost anywhere. Operations managers at 40-person Central European manufacturers are hard everywhere. A provider that returns 90% on one and 35% on the other is simply describing the internet.

So skip asking a vendor how many records they have. The coverage question worth asking is what match rate do you expect on this list, and what happens to the rows you miss.

Accuracy only means something with a denominator attached, and verified is not a regulated word. The two numbers that matter are the percentage of your submitted rows that returned a value, and the percentage of returned values that survived contact with reality.

A vendor quoting 95% accuracy has told you nothing until you know 95% of what. Of the records it chose to return, or of the records you asked about? Those two can differ by forty points on the same run, because returning fewer rows is the easiest way to raise the first one. Any provider can be 99% accurate by answering only the easy questions.

Verified is similarly elastic. For an email it can mean checked against a mail server in real time, or matched to a known company pattern, or seen recently in a contributor's address book, or simply present in the index. Only the first is a live check. It is worth asking which one a vendor means, in writing, because the word costs nothing and the difference lands in your bounce rate.

Catch-all domains are where the honest vendors separate. A catch-all mail server accepts every address at the domain, so no SMTP check can prove a specific mailbox exists. A provider that marks those as valid is inflating its own accuracy number with rows it cannot stand behind, and a provider that marks them as catch-all is telling you a true thing you may not want to hear.

The measured claim Cleanlist AI makes: 98% of work emails verified and 85% phone number coverage across 500 stratified B2B leads in the Cleanlist AI 500-Lead Enrichment Benchmark, 2026, against 70% to 80% email and 30% to 60% phone from single sources on the identical input. The input list and the denominator are published with it, which is the only reason the number is worth quoting.

The last accuracy variable is time, and it is not the provider's fault. B2B contact data decays at about 2.1% a month and 22.5% a year on Cognism's figures. Email addresses go fastest at 22.5% to 30% a year on Cognism and SparkDBI, phone numbers at about 18% on SparkDBI, and LinkedIn's Economic Graph puts annual job changes at 10.9% of professionals. Company-level facts move more slowly, around 15% a year on 6sense's figure. A record that was accurate when a provider returned it is a slightly worse record every month afterwards, whoever supplied it.

Between $13.00 and $435.29 per 1,000 verified business emails on annual billing, with a median of $39.13. That is the recomputed range across the 18 discovery vendors that publish an annual price and enough detail to convert it, out of 37 pricing pages Cleanlist AI fetched on September 1, 2026.

A 33.5x spread for a nominally identical unit is the single most important fact about pricing in this category, and quality does not explain it: almost nobody prices the unit directly. Vendors price a plan, a seat, an export cap or a credit, and the credit is the least standardized unit in B2B software: six different nouns are in use across those 37 vendors for the thing you buy, credit, export, lookup, record, action and contact.

Twenty-nine of the 37 define one email as one credit, so that part has converged. What has not converged is everything else the same credit buys. Apollo's published rates are 1 credit for an email and 8 for a phone number, with data enrichment costing up to 9 credits per record. Clay runs two independent meters, Actions for platform usage and Data Credits for the vendor data itself, which is why its entry plan reads $167 a month on annual billing and $185 on monthly: both figures are the sum of two default selectors rather than one price. Cleanlist AI charges 1 credit for a verified work email, 10 for a phone number, 11 for both on one contact, 0.5 to validate an address, 5 for an AI qualification and 0.2 to push a lead to a CRM, with People Search and Company Search costing nothing.

Four structural models sit under all of that. Per-seat subscriptions bill the same whether a rep ran four hundred lookups or none. Annual enterprise contracts trade a floor commitment for a lower unit rate. ZoomInfo publishes no purchasable dollar price at any tier: zoominfo.com/pricing returned HTTP 403 to an automated client on September 1, 2026, including one sending a Googlebot user agent, and served a press-and-hold human-verification challenge in headless Chromium. The roughly $14,995 a year widely quoted for its Professional tier is a buyer-reported figure compiled from G2 and Reddit quotes, not a rate card, and it should carry that label wherever it appears. Six of the 37 vendors in the index publish no purchasable price at all: Cognism, NeverBounce, Outreach, Salesloft, Seamless.AI and ZoomInfo. Credit or usage models bill what you consume. Pay-per-result models bill only when a field comes back.

On that recomputed index Cleanlist AI Starter is $49.33 per 1,000 verified emails on annual billing, which is 11th of 18 and above the $39.13 median, and ten vendors are cheaper. Several of the ten are companies Cleanlist AI buys from, including Hunter, Prospeo, Findymail, Datagma and LeadMagic, which is a conflict worth naming rather than hiding: a study run honestly by a vendor in the category should sometimes put that vendor mid-table, and this one does.

The number to actually decide on is none of these. It is cost per valid record: what you paid divided by the rows that came back correct on your own list. A $0.10 lookup that returns nothing 40% of the time is more expensive than a $0.25 lookup that returns something 90% of the time, and the invoice will never show you that.

More than most buyers assume. Across 37 B2B data and GTM pricing pages read on September 1, 2026, not one stated that a credit card is required to try the product: 14 say explicitly that no card is needed, 18 say nothing at all, and 5 have no trial for the question to apply to.

Fifteen of the 37 run a recurring free plan and 9 offer a time-boxed trial, and those sets overlap rather than partition, so 21 distinct vendors offer one free entry or the other. Among the 9 that publish both a monthly allowance and a per-email credit price, the median free plan is worth 40 verified emails a month.

The column that matters more than the allowance is bulk CSV upload, because a handful of one-off lookups from a browser extension proves nothing about how a provider handles a list. Only 7 of the 37 publish that bulk upload works on the free tier, and 2 of those are hard-capped. This is the gap between what a free plan lets you feel and what it lets you measure.

Cleanlist's Free plan gives 50 verified emails a month for up to 3 seats, above the 40 median, and the trial is where the bulk motion is easiest to test: 14 days of Pro with up to 3 seats, 250 credits and every Pro feature, which is enough to run a control list of a few hundred rows both ways.

The general lesson from the census is about silence rather than generosity. Eighteen of the 37 vendors leave the credit-card question unanswered on the page where buyers look for it first, which is why so much search volume in this category is people asking whether a trial needs a card rather than simply starting one.

Run the same control list from your own CRM through every finalist on the same day, and decide on cost per valid record. Everything else on a comparison page is downstream of that one test, and it takes an afternoon.

The method: pull 100 rows you already know the answers for, ideally people you have actually reached this quarter, and strip the fields you are buying. Submit the identical file to each vendor within the same 24 hours, because coverage moves. Then count four things. How many rows came back at all, which is match rate. How many returned values match what you already knew, which is accuracy. How many bounced when you sent to them, which is the only email number that is not self-reported. And what each vendor billed for the run, including the misses.

Use your own list rather than a vendor's sample. A vendor's sample is chosen, and every provider in this market can assemble 100 records it is excellent at.

The questions worth getting in writing before a trial rather than after one, because each has been the reason a tool that demoed well was the wrong purchase: what is the literal list of fields the response returns, how is a lookup that finds nothing billed, does verified mean a live server check or a pattern match, how are catch-all domains labelled, is the CRM integration two-way or export-only, is there a per-seat charge on top of usage, what is the minimum commitment and the auto-renewal notice period, and can you export your enriched records if you leave.

Two numbers to discount entirely. Database size, because a hundred million verified records beats three hundred million unverified ones on every metric a rep experiences. And any accuracy percentage quoted without the input list it was measured on.

This page is published by a vendor in the category, which is a reason to run the test rather than a reason to skip it. Every new Cleanlist AI workspace starts with 14 days of Pro so Cleanlist AI can be one of the columns in it.

Four, and provenance is the one that predicts the others: where did this specific field come from, can you suppress a person across your own sources on request, what does your data processing agreement actually commit you to, and how quickly do deletion requests propagate. This section describes how the obligations usually fall and is not legal advice.

B2B enrichment in the EU and the UK is normally carried out under legitimate interest rather than consent. That is a lawful basis you have to document and defend, in a written balancing test covering the purpose, why the processing is necessary for it, and why the person's rights do not override it. Keep it with your records of processing, because the moment it matters is the moment somebody asks.

Provenance is a purchasing decision rather than a policy one. A provider that cannot tell you where a field came from cannot help you answer a regulator, and cannot honour a suppression request at the source either. It is a fair thing to ask for at the field level rather than the vendor level.

The duty that gets missed operationally is propagation. A deletion or objection request has to reach every copy, which in most stacks means the CRM, the sequencer and the enrichment tool rather than only the CRM. Under CCPA and CPRA the disclosure duty is similar in shape: say what categories you collected and from where, and honour deletion and opt-out requests. Under CAN-SPAM every commercial email needs a working unsubscribe honoured within ten business days.

Sourcing model changes the risk profile. Contributor networks and first-party contribution pools carry the questions that publicly crawled and registry data mostly do not, because a person who never interacted with either company is in the database because a third party's address book was uploaded. That is a question to ask before signing rather than after an inbound request arrives.

Yes, visibly, and it matters mostly as portability risk. Four verified changes inside 2026 alone: Zoom acquired Common Room, HubSpot acquired Warmly, Apollo acquired Pocus, and Persana sunset its platform on May 2, 2026 after joining Rox, deleting all remaining customer data at the end of a 30-day export window. The persana.ai marketing site is still live and still published an $85 Starter price on September 1, 2026, which is why 2026 roundups still list it as available. ZoomInfo also took a goodwill impairment in its Q2 2026 results.

The pattern is that standalone signal and intelligence products are being absorbed into the platforms their buyers already run, which is rational for the acquirers and inconvenient for anyone mid-contract with a tool that just changed owner.

What that means for a purchase is narrow and practical. A point tool acquired by a platform tends to get bundled into that platform's pricing rather than kept as a standalone SKU, so a two-year commitment to a small vendor carries a real chance of ending up inside a suite you did not choose. Ask what happens to your contract and your data on a change of control, and read the export terms specifically.

For orchestration layers the consolidation risk points at the supply side rather than the front door. A supplier that gets acquired can be repriced or withdrawn, which is one of the arguments for a wide pool: with 25+ providers behind a single lookup, no single supplier's fate is the product's fate.

The honest counterweight is that consolidation is not decline. The category is growing and the acquisitions are being paid for. It just means the specific vendor list any article publishes has a shelf life measured in months, which is another argument for learning the structure rather than memorising a ranking.

One provider is enough when your ICP sits inside its strong segment and the volume is small enough that misses can be researched by hand. You need several the moment either of those stops being true, and the crossover comes earlier than most teams expect.

One is the right answer for a narrow, well-covered market, for a team doing tens of lookups a week rather than thousands, and for anyone whose motion depends on browsing a market rather than completing a list. A single subscription is simpler to administer, simpler to reason about on compliance, and there is nothing wrong with it.

Several becomes correct when the misses stop being researchable. At 2,000 rows a month, a 60% match rate means 800 records a human is expected to chase, which is where the arithmetic stops being about data quality and starts being about payroll. It is also correct when phone matters, since phone coverage is the field where single-source ceilings bite hardest, and when your ICP spans regions with genuinely different provider strengths.

The part worth being honest about is that several does not have to mean an orchestration platform. Running two subscriptions and reconciling by hand works and is sometimes cheaper, particularly at low volume. What you are buying with an orchestration layer is the routing, the deduplication, the verification and the single bill, and if you would not use those, the platform is overhead.

And where a single provider is simply the right purchase, it is worth saying which. If the motion depends on knowing which accounts are in market, an intent platform is the correct tool and enrichment will not substitute for it. If the team needs a research surface to explore, that is a database licence. If all you need is to know whether the addresses you already hold still work, that is a verification service at a fraction of the price of either.

As an orchestration layer rather than a data vendor. Cleanlist AI owns no database. It runs a waterfall across 25+ providers, verifies each answer before accepting it, and bills per returned field: 1 credit for a verified work email, 10 for a phone number, 11 for both, and 0 when nothing comes back.

The Free plan is $0 for 50 credits a month with up to 3 seats. Starter is $49 per seat a month with 750 credits per seat. Pro is $89 per seat a month with 1,500 credits per seat, credit levels up to 100,000 a month, the REST API, agents and two-way CRM sync. Enterprise is custom. Every seat adds its credits to one shared balance, and annual billing takes 25% off. An enriched company record returns up to 180 firmographic properties.

The boundaries are the useful part of this section. Cleanlist AI does not sell intent data. It does not sell technographics as a stored field. It does not sell an owned database to browse, because there is not one. It sends email from your own inbox through Sequences on paid plans, places no calls, and still writes into Outreach, Salesloft and Lemlist for teams that already run them. On the published price index it sits 11th of 18 with ten vendors cheaper per 1,000 verified emails.

The suppliers are named rather than hidden, because a page arguing that provenance is the question a buyer should ask cannot then decline to answer it about itself. Cleanlist AI buys from Hunter, Prospeo, Findymail, Datagma, LeadMagic, Dropcontact, Lusha, ZeroBounce, Emailable and others in the pool, and on their own published pricing several of them charge less per verified email than Cleanlist AI does. What Cleanlist AI sells on top of them is the routing, the verification, the reconciliation into one record with its source attached, and a single bill rather than ten.

What is left is a narrow job: take a list, resolve who the people are, find and verify how to reach them, and put the result where the team already works. Every claim on this page is about that job and nothing else.

The names that appear on nearly every published list are ZoomInfo, Apollo, Cognism, Clay, UpLead, RocketReach, People Data Labs, HubSpot Breeze Intelligence, Lusha, 6sense and Bombora. Sorted into the four types above, that list stops being a ranking and becomes a menu, because most of these vendors are poor substitutes for one another. Every price below comes from the 37 vendor pricing pages Cleanlist AI fetched on September 1, 2026, and each one states its billing basis, because a price with no basis attached is the most common error made about this category.

Originators. ZoomInfo: best for large North American enterprise teams that need org charts and technographics alongside contacts. It publishes 321M professional profiles and 104M company profiles, and it publishes no purchasable dollar price at any tier. Apollo: best for SMB teams that want the database and the sequencer on one bill. It publishes 275M contacts, charges 1 credit for an email and 8 for a phone number, grants 30,000 credits per seat per year, and lists Basic at $65 a month or $49 a month per seat on annual billing, which converts to $19.60 per 1,000 verified emails and is the seventh cheapest of the 18 discovery vendors in that index. Cognism: best for outbound into the UK and Europe where phone-verified mobiles decide the quarter. It counts 1 credit as one revealed contact, includes 5 seats on both Standard and Pro, and publishes no price.

Database licences. UpLead: best for buyers who want one credit to return an email and a mobile together. Essentials is $99 a month, or $74 a month on annual billing, for 170 credits, and that bundling is most of why it converts to $435.29 per 1,000 verified emails, the highest published rate in the index. RocketReach: best for teams that read far more records than they export. Essentials Email Only is $27.42 a month on annual billing for 1,200 exports a year, or $274.17 per 1,000, the second highest rate in the index, because the export cap rather than the lookup is what prices the plan. People Data Labs: best for engineering teams building a product on raw records, sold as bulk files and an API rather than seats, and outside the price index.

Orchestration. Clay: best for RevOps teams that want to chain providers themselves. Launch is $185 a month and $167 on annual billing, and that $167 is two meters added together, 180,000 actions a year at $54 plus 30,000 data credits a year at $113.

Native CRM enrichment. HubSpot Breeze Intelligence, formerly Clearbit: best for teams already standardised on HubSpot. It has no standalone pricing page, because that URL redirects to a credits page which redirects again into a knowledge base article, and the one readable rate is the overage at $0.010 per HubSpot credit, invoiced in increments of ten.

Adjacent rather than comparable. 6sense and Bombora sell intent, which the types section above deliberately excludes, and for a structural reason: an intent score has no second source to check it against, so it cannot be measured the way an email address can. Lusha appears on most of these lists as the fast browser-extension lookup, and it is also one of the providers Cleanlist AI buys data from, which is the kind of overlap a ranked table hides.

The questions of which B2B platform is best and which B2B data provider is the best have no single answer for the same reason: those lists rank four different businesses against each other. The answer depends on whether the job is completing a list, browsing a market, building a product, or timing an account.

Cleanlist AI measured the live Google US desktop result for all nine keywords in this cluster on September 7, 2026, through the DataForSEO SERP API at depth 50 with the AI Overview body loaded. All nine carry an AI Overview and all nine returned a readable answer. Together those nine terms are worth 3,060 US searches a month on Google Ads volume, at $50.35 cost per click on the three highest-volume ones.

Three brands are named in all nine answers: ZoomInfo, Apollo and Cognism. Lusha, Clearbit and HubSpot Breeze Intelligence appear in seven, Clay, 6sense and Bombora in five, People Data Labs and UpLead in two, and BookYourData in one. Cleanlist AI is named in none of the nine and cited as a source in none of the nine, which is the honest starting position for a page arguing that structure beats a ranking.

The answers repeat far more than the query set suggests. One answer body is served byte for byte identical across five terms (b2b data provider, b2b data providers, b2b database providers, b2b data companies and b2b data vendors), and a second identical body covers best b2b data providers and top b2b data providers. Four distinct answers cover the whole cluster.

Twelve domains fill the 59 citation slots: autobound.ai 16, YouTube 12, demandbase.com and enginy.ai 8 each, reddit.com 5, and seven others once or twice. The division of labour is the finding worth keeping. The vendor roundups supply the record counts and the prices, and the Reddit thread is cited only for buyer sentiment about cost, specifically the observation that some users report high costs relative to lead quality.

The prices in those answers are the weak point. Apollo is quoted at $49 a month with no mention that this is per seat on annual billing while Basic is $65 a month. UpLead is quoted at around $99 a month, which is the monthly rate, against $74 on annual. ZoomInfo is quoted at $15,000 a year by one source and at $15k or more a year by another, and neither is a ZoomInfo rate card, because there is no readable one to quote. Stating the billing basis and the fetch date on every figure is the only part of this that is hard to copy.

One verified business email costs between $0.013 and $0.435 at entry-tier annual pricing, with a median of $0.039. That is the same recomputed set as the pricing section above, the 18 discovery vendors in the September 1, 2026 index that publish an annual price and enough detail to convert it, restated per contact rather than per thousand: Datagma at the floor, UpLead at the ceiling, a 33.5x spread on a nominally identical unit.

A direct dial is the expensive record everywhere, and the published credit ratios are where you see it. Apollo charges 8 credits for a phone number against 1 for an email. Cleanlist AI, LeadIQ, Findymail, FullEnrich, Ocean.io, BetterContact and Persana all charge 10 against 1. LeadMagic charges 5 against 1. Lemlist charges 20 against 5, the same 4x shape at a different scale. UpLead is the structural exception, bundling an email and a mobile into one credit, which is most of why its per-email figure is the highest in the index.

Four billing shapes sit under all of it. Per-seat subscription, where Apollo Basic bills $49 a month per seat on annual whether that seat ran 400 lookups or none. Annual enterprise floor, where the number is a commitment rather than a rate and is often not published at all. Credit or usage, where you buy a pool and spend it. And pay-per-result, which is worth asking about by name: nine of the 37 vendors publish that a lookup returning nothing is not billed. Seven state it on the pricing page in so many words, LeadMagic, BetterContact, FullEnrich, Findymail, Datagma, Dropcontact and RocketReach, and on the verification side Bouncer refunds duplicates and unknowns while Emailable states that unknown results are free.

The meter buyers miss is search. Four vendors state outright that finding a record is metered separately from enriching it. Snov.io charges 1 credit per prospect search, Ocean.io 0.2 credits per search result, Lusha 1 credit for an API call or a CSV export of 1 to 25 results, and LeadMagic excludes search from its entry tier entirely. RocketReach prices its plan by the export cap rather than by the lookup.

Cleanlist: best for teams whose reps use data unevenly, because every seat's credits go into one shared balance. It runs waterfall enrichment across 25+ providers and charges 1 credit for a verified work email and 10 for a phone number, with search costing nothing and a miss costing nothing, starting at $49 per seat a month, or $37 per seat on annual billing. Every new workspace starts with 14 days of Pro: 3 seats, 250 credits, every Pro feature.

None of these per-contact figures is the number to decide on. That is cost per valid record, which the pricing section above defines, and which only a control list from your own CRM can produce.

Thirty of 37. Cleanlist AI fetched 37 B2B data and GTM vendor pricing pages on September 1, 2026 and recorded what each one published: 30 show a purchasable dollar price, 6 show none at all, and 1 is partial. The six publishing no price are Cognism, NeverBounce, Outreach, Salesloft, Seamless.AI and ZoomInfo. The partial one is HubSpot Breeze Intelligence, whose standalone pricing URL redirects twice and lands in a knowledge base article. Amplemarket publishes a number on its Startup tier only, with Growth and Elite quoted as Custom.

Two of the six go further than quiet and are unreadable to a machine. zoominfo.com/pricing returned HTTP 403 to an automated client sending a Chrome user agent, HTTP 403 to one sending a Googlebot user agent, and a press-and-hold human-verification challenge in headless Chromium. neverbounce.com/pricing returned 403 to Chrome and rendered only the same challenge.

That is the mechanical reason the AI Overviews measured on September 7, 2026 quote Apollo and UpLead by the dollar and never quote ZoomInfo except through a number somebody else compiled. A price a crawler cannot read is a price an answer engine cannot cite, so vendors that publish get their own figures repeated back and vendors that hide get whatever a third party estimated. For a buyer the practical consequence is narrower: roughly one in six of any shortlist needs a sales call before a number exists, so budget the calendar time. The full 37-row dataset, with a source URL and a fetch date on every row, is published under CC BY 4.0 and linked below.

Ask what the AI produced before asking which vendor is best, because the label covers three different products and only one of them changes what you actually receive.

AI as the research step. An agent reads public sources at query time and writes an answer into a column: whether a company is hiring, what a page says about its stack, how a job description reads. That output is generated rather than looked up, it carries no verification, and its worth is the worth of its date. Cleanlist AI prices this as an AI qualification at 5 credits and returns the date alongside the answer.

AI as the interface. Natural-language search over a structured filter set, which changes who on the team can run a query rather than what a query can reach. This boundary is worth stating plainly because vendors blur it. Clu in the Cleanlist AI app and the MCP server take natural language. The public REST API takes structured filters and has no natural-language endpoint.

AI as the sourcing claim. Inference from patterns, which the sourcing section above already covers. A generated address is a guess until a live check confirms it, and the check is the part that costs money.

Two questions settle the shortlist. Did the AI produce a field or an answer, and does that answer carry a date. Then a third that buyer checklists have started asking: is there an MCP server or an open API an agent can call, or does the AI only work inside the vendor's own screen. Cleanlist AI probed 59 GTM and B2B data vendors on September 1, 2026 and 43 of them ship a product MCP server, so the useful follow-up is what that server can reach and whether it can spend money without asking first.

A B2B contact database is a stored index you license and browse, and it is sold in three shapes: a self-serve credit plan, a seat licence, or a bulk file and API feed. Which shape you need depends on whether the job is exploring a market, working a list, or building a product on the records.

Self-serve credit plans are the shape most buyers meet first, and every figure here comes from the 37 pricing pages fetched on September 1, 2026. UpLead Essentials is $99 a month, or $74 on annual billing, for 170 credits, with one credit returning an email and a mobile together. RocketReach Essentials Email Only is $27.42 a month on annual for 1,200 exports a year. Apollo Basic is $65 a month, or $49 per seat on annual, with 30,000 credits per seat per year. Snov.io Starter is $39 a month, or $29.25 on annual, for 1,000 credits, and it spends 1 credit on a prospect search as well as 1 on the email.

Seat licences and annual floors are where the number stops being published. Six of the 37 vendors publish no purchasable price at any tier, so part of any shortlist cannot be compared until a sales call has happened.

Bulk files and API feeds are the developer purchase, priced per record rather than per seat. People Data Labs is the name that comes up most often in that shape.

Before comparing any of them, read the coverage and accuracy sections above. A database is worth its match rate on your list, and every vendor here will offer you a record count instead.

Six kinds, and the split that matters runs between the first four and the last two. Contact data is the person: name, work email, phone number, title, seniority. Firmographic data is the company: industry, employee count, headquarters, revenue band. Technographic data is the software a company runs. Engagement data is what your own systems recorded about a person. Intent data is an inferred signal that an account is researching a topic. Signal or event data is a dated public occurrence: a funding round, a leadership change, a hiring surge, a tool going live.

The distinction worth owning is that a fact can be verified against a second source and a signal cannot. An email address either resolves at a mail server or it does not, so a bounce rate audits it. A claim that an account is in market has nothing to check it against, which is why intent is priced and sold differently and why it cannot be benchmarked the way the accuracy section above benchmarks a match rate. That makes it a useful product for timing and a poor product to buy on an accuracy promise.

This is also the cleanest place to state what Cleanlist AI does not sell, because a reader sorting six categories will want it. Cleanlist AI covers contact and firmographic data, up to 180 firmographic properties on an enriched company record. It does not sell intent data, does not store technographics as a field, and does not run job-change alerting. Funding stage is a live filter on Company Search, written snake_case as series_b, and People Search rejects the key. Where the job is timing rather than reach, an intent platform is the correct purchase and enrichment will not stand in for it.