Google Quoted 376 Prices for 42 B2B Data Vendors. We Checked Every One.

Forty-two B2B contact-data vendors had their pricing pages read on September 7, 2026. Five publish a real number for every paid tier. On the same day, 292 Google AI answers stated 471 prices, 376 of which could be tied to exactly one vendor. Of those 376, 48.7% confirm against a vendor page read the same day, 21.8% price a vendor that publishes no dollar amount anywhere, and 14.1% are contradicted by the vendor's own page. Cleanlist is in the sample.

Last updated: September 7, 2026. Every AI answer and every vendor pricing page in this study was read on September 7, 2026. The cross-check index was read on September 1, 2026.

How many B2B data vendors publish a price at all?

Five. Of the 40 vendors that produced a price-publication classification, 5 publish a real number for every paid tier, which is 12.5%. They are Coresignal, Apollo.io, CUFinder, Smartlead and ListKit. Another 27 (67.5%) publish some tiers and put at least one behind contact sales. Eight (20.0%) publish no price at all: Seamless.AI, 6sense, Demandbase, ZoomInfo, SalesIntel, Bombora, Outreach and Salesloft.

Add the partial publishers to the non-publishers and you get 35 of 40, or 87.5%, which is the share of this category that will not show a buyer a complete price. The received wisdom is that four fifths of B2B software hides its pricing. On this sample the real figure is closer to seven eighths.

The demo gate uses 42 as its denominator rather than 40, because a demo wall is observable on a page even when a price is not. The publication classes use 40, because a page we could not read cannot be classified as publishing or not publishing.

What happens to a buyer who cannot see a price before a sales call?

They ask an answer engine, and the answer engine answers. That is the mechanism this study exists to measure, and both halves of it were measured on the same day so the timing cannot be blamed.

Of the 292 AI answers read, 106 state at least one dollar figure, which is 36.3%. One hundred of them contain at least one figure this method could attribute to a vendor. So roughly a third of the answers a buyer meets in this category will put a number in front of them, and the other two thirds describe vendors without pricing them.

Free trials are the other thing a buyer looks for before a call, and the sample is split. 25 of 42 vendors publish a trial. Twelve of those 25 (48.0%) state that no card is needed, 2 require one (UpLead and Swordfish AI) and 11 say nothing either way. Trial lengths run from 7 days, which 9 vendors use, to 90 days at Explorium.

How many prices do Google's AI answers state, and for whom?

471 dollar figures were extracted from the 292 answers. 376 (79.8%) were tied to exactly one vendor and 95 (20.2%) were dropped. The drops are published with their reasons: 69 had no brand in the statement's own segment or the two preceding ones, 15 followed a segment naming more than one brand, and 11 sat inside a brand enumeration such as "tools like NeverBounce, ZeroBounce, or MillionVerifier". Attribution was never guessed to raise the yield.

The 376 attributed statements cover 36 distinct vendors. They are not evenly spread. Apollo alone accounts for 92 of them, very nearly double the next vendor. ZoomInfo takes 47, RocketReach 30, Clay 18, Lusha 16, 6sense 15, NeverBounce 13 and Demandbase 12. The rest of the field splits the remaining 133.

How often is the price an AI answer quotes actually right?

183 of 376 statements confirm against a vendor-published number read the same day, which is 48.7%. That is the honest headline, and it needs its three companion columns to mean anything.

53 statements (14.1%) are contradicted by the vendor's own current page. 82 (21.8%) name a figure for a vendor that publishes no dollar amount anywhere, so no vendor-side evidence can confirm or deny them. 58 (15.4%) could not be checked at all: 13 NeverBounce rows behind a bot wall, 22 rows for vendors with no same-day read, and 23 rows about tiers the vendor prices only on request.

Collapse those into a checkable base and the arithmetic is cleaner. 183 confirmed plus 53 contradicted gives 236 checkable statements, which is 62.8% of 376. Among those 236, the wrong rate is 53 divided by 236, or 22.5%. So roughly one checkable AI-quoted price in five is denied by the vendor it describes, and a further 140 statements sit in a zone where nobody can be shown right or wrong.

Are the wrong prices concentrated on the vendors who publish nothing?

Partly, and the honest answer splits into a robust half and a fragile half.

The robust half is uncheckability. Of the 111 statements about the 8 vendors that publish no price, 82 have nothing vendor-side to check against, which is 73.9%. Of the 243 statements about vendors that do publish a price, 36 are uncheckable, which is 14.8%. For the 5 full publishers alone, the uncheckable share is 0 of 99. Hiding your price does not stop an answer engine from naming one. It removes the only document that could settle the question, and it does so five times over.

The fragile half is the wrong rate, and it rests on one company. Among vendors that publish a price, 26 of 207 checkable statements conflict with the vendor page, which is 12.6% (full publishers 9 of 99, or 9.1%; partial publishers 17 of 108, or 15.7%). Among vendors that publish nothing, 27 of 29 checkable statements conflict, which is 93.1%. That looks decisive until you read what the 29 are made of: 27 of them are the ZoomInfo $15,000 family, denied verbatim on ZoomInfo's own page. Remove ZoomInfo and the non-publisher class has 2 checkable statements, both correct, and the comparison collapses.

So the thesis is half confirmed. A category that hides its prices does force the answer engine to reconstruct them, and reconstruction produces figures nobody can audit. The claim that reconstruction is also more often wrong is supported here by a single vendor and should not be quoted as a general law.

What is the single most-repeated price claim, and is it true?

ZoomInfo at $15,000 a year, and no. Across 25 distinct buyer queries the corpus states it 26 times, which is 6.9% of all 376 statements: 15 statements at "$15,000 per year", 8 at "$15k" with no period attached, and 3 at "$14,995 per year". A fourth variant at "$14,900 per year" brings the denied family to 27 rows.

ZoomInfo's own pricing page carries exactly one dollar amount, and it is a denial. Read on September 7, 2026, the page says: "No, despite claims from Cognism, Demandbase, Lusha, Apollo, and others, ZoomInfo's pricing does not start at $15,000." The most repeated number in the corpus is the one number that vendor's page exists to refute. That is what happens when a vendor's only published pricing content is a rebuttal: the rebuttal loses to the thing it is rebutting, 26 times over.

Two answers state the absence of a published price and then supply one anyway. For "salesloft pricing": "Salesloft does not publish official pricing, but third-party procurement data estimates list prices range from $125 to $165 per user per month." For "6sense pricing": "6sense does not publish public list prices, but market data shows that annual costs typically range from $50,000 to $300,000+ per year." Both are accurate about the absence and unfalsifiable about the number.

Which vendor is quoted most often, and how accurate is it?

Apollo, at 92 statements, and it is quoted accurately 83 times, which is 90.2%. Apollo is one of the 5 vendors publishing every paid tier, and at 92 statements against ZoomInfo's 47 it is quoted very nearly twice as often as anyone else in the corpus. Publishing a complete price does not just make you checkable. It makes you the number the engine repeats.

All 9 of Apollo's wrong rows are the same phantom figure, $59 per seat per month, which appears in ranges like "around $49 to $59 per user per month". Apollo publishes $65 monthly and $49 annual for Basic. Neither is $59.

The more interesting Apollo finding is not an error at all. The most-repeated correct figure in the whole corpus, $49 per user per month across 51 distinct queries, is Apollo's annual-billed rate presented by the engine as a monthly one, and the reason is legible in the page itself. The schema.org OfferCatalog delivered in apollo.io/pricing's HTML today publishes Basic at price "49" with a billingIncrement of "Monthly", while the rendered plan table, read the same day, shows Basic at $65 month to month and $49 billed annually. The number the answer engines repeat is the number Apollo's own structured data hands them. Clay's row is the same shape without the JSON-LD: $167 is Launch billed annually, the month-to-month figure is $185, and the answers say "per month".

A buyer budgeting month to month on either answer is short by 32.7% at Apollo ($65 against $49) and by 10.8% at Clay ($185 against $167). A real vendor number on the wrong basis is a harder failure to catch than an invented one, because it survives every check except opening the page and clicking the toggle.

What does a credit actually buy, and who charges more than one credit for one email?

A credit is the unit the price is quoted in and almost never the unit the buyer cares about. 33 of 42 vendors publish a per-email credit rate. Twenty-eight of the 33 charge exactly one credit for one verified email, which makes one credit per email the category norm. Four charge more, which is 12.1% of the 33: Coresignal 20, Enrich.so 10, Explorium 2 and ListKit 2. Clay charges 0.5.

Phones are where the multiplier lives. Twelve of the 33 vendors publishing a credit model charge more than one credit for a phone number or direct dial, which is 36.4%.

Coresignal's row is the one to sit with. It is one of the 5 vendors publishing a real number for every paid tier, and it charges 20 credits for a single verified email. Full transparency at the plan level tells a buyer nothing about the price of the thing they actually buy. Apollo contradicts itself on the same field: two Apollo pages state that emails cost 1 credit and that Apollo never charges more than 1 credit per contact, while Apollo's own pricing calculator prices the email-only option at 2 credits per record. The census records 1, the twice-published explicit rate, and flags the 2.

Where do tonight's reading and the September 1 index disagree?

In six places that matter, and they are published rather than reconciled, because a disagreement between two dated readings of the same page is itself evidence about how legible these pages are.

Cognism's row changes a class assignment: moved out of the non-publisher group, its 9 quoted statements stop being automatic "vendor publishes no price" rows. NeverBounce's row changes a definition. All 13 of its statements are published as unverifiable rather than as non-publisher rows, because a page we cannot read is not a page with no price, and calling them non-publisher rows would have inflated this study's most interesting category by 13.

A seventh disagreement is a schema difference and not a factual one. Fourteen vendors (Amplemarket, BetterContact, Clay, Dropcontact, FullEnrich, Kaspr, LeadIQ, Ocean.io, RocketReach, Skrapp, Snov.io, Surfe, UpLead and Clearout) read "yes" in the index and "yes_partial" tonight. The index's "yes" means an entry price is published; tonight's "yes_partial" means at least one tier is gated. Do not report those as price changes.

What did we get wrong about our own price?

Nothing in the audit, and that is worth exactly two rows. Cleanlist is named with a price twice, both times at $79 a month, on "lead enrichment tools" and "data enrichment platform". Both confirmed against cleanlist.ai/pricing fetched the same day, where Starter is $79 a month and $59 a month billed annually. Cleanlist is a full publisher: four tiers, every one priced, no contact-sales tier, 1 credit per verified email and 10 per direct dial.

The number to take from our own row is 2 of 376, which is 0.5% of the corpus. On the accuracy measure we score 100%. On the measure that decides whether a buyer ever sees us, we are a rounding error next to Apollo's 92 statements. Publishing every price makes you checkable. It does not make you quoted.

And full publication is not immunity from being misquoted. The 5 full publishers collected 9 wrong statements across 99, which is 9.1%. Being the most legible vendor on the page lowers the error rate by roughly a factor of two against the non-publishers' checkable rows, and it does not take it to zero.

What did this study get wrong?

Four things, all of which moved numbers before publication.

1. The first extraction's vendor attribution was broken and was thrown away. The crude first pass captured words like "Pricing" and "Starting" as vendor names. Attribution was rebuilt around URL blanking with preserved offsets, a published 80-name vocabulary and segment scope. Two guards are load-bearing. Without whitelisting vendor names that are themselves domains (apollo.io, snov.io, seamless.ai, enrich.so) from the bare-domain blanker, Apollo loses 75 of its 92 statements. Without a vendor-position rule on generic-word brands, the ordinary English word "outreach" credited 29 prices to the vendor Outreach.

2. An earlier verdict pass produced 75 false WRONG verdicts on Apollo. Denial detection was reading the census reader's own notes alongside the vendor page quotes, so an analyst's aside about Apollo's markup scored as Apollo denying its own prices. It now reads only the census's evidence_quotes field, defined by schema as text copied verbatim from the vendor page, and the notes field is excluded by name in the source.

3. Two denominators moved while the audit ran. The sprint brief states 230 of 240 buyer SERPs carried a readable AI Overview; the file as read at 05:10 holds 232 and the sprint's own analysis agrees, because it was being backfilled. Every figure here is computed on the 232-answer version, md5 c838060df797688ce9573c05aff9cc32, which is why the corpus is 292 answers and not 290. The vendor census grew the same way: the brief describes 42 attempted, 41 returned, 39 loaded, while the file as read at 05:19, md5 f667f0984f213de6153104ee7b036be3, holds 42 records including Persana AI, with 39 reachable and 40 classified.

4. One AI Overview answered a different question entirely. Google's AI Overview for "clay pricing" is about pottery, quoting "$2 to $8 for small consumer craft packs" and "$10 to $50 for a 25-pound bag". Four statements were vetoed by a per-answer topic guard and appear in no figure on this page. The behaviour is worth reporting on its own: a query naming a funded GTM vendor returned a ceramics supply answer.

What are the limitations of this study?

Nine, stated bluntly.

One snapshot, one day, one country, one device. AI answers are non-deterministic and this corpus contains no repeat pull of the price statements, so do not assume these answers reproduce.

The study covers a third of the corpus. 106 of 292 answers state any price at all and 100 contain an attributable one. The other 186 describe vendors without pricing them, and this page is silent about them.

95 of 471 statements (20.2%) were dropped as unattributable. They are counted and their reasons published. Raising the yield by guessing was available and was not taken.

126 of the 376 attributions (33.5%) come from the query-subject rule, where a query such as "zoominfo pricing" makes the whole answer about one vendor and later bullets stop repeating the name. Strip those rows and the totals fall to 250 statements, with every class moving by under three points.

Zero statements could be shown to be STALE. That verdict requires proving a vendor previously published a number and no longer does. The cross-check index is six days old and every figure matching it also matched a page read today, so the category is empty on evidence rather than by omission.

The wrong-rate half of the thesis rests on ZoomInfo, as described above. The uncheckability half is the robust result.

Ranges count as two statements. "$27 to $33 per month" produces a row for $27 and a row for $33, and the two can take different verdicts, as they do for Lusha's "$29 to $36" and RocketReach's "$27 to $33".

Three vendors defeated the live fetch. Apollo and RocketReach render prices client-side and ZoomInfo returns 403; all three fall back to same-day census reads, and Apollo also to its own JSON-LD. HubSpot Breeze Intelligence returns 404 on both candidate pricing URLs, so its 4 statements are unverifiable. Unreadable and unpublished are kept apart throughout: NeverBounce, Nimbler and Swordfish AI could not be read, and their statements are unverifiable.

Currency is assumed. Zeliq, Adapt.io and Amplemarket print a bare dollar sign and never name USD. LeadIQ's pricing HTML carries an unrendered "$200" placeholder inside the Pro card that a naive scraper reads as a plan price; it is excluded here.

Where can I download the dataset?

The full open dataset is at /data/ai-quoted-price-audit-2026-09.csv, published under CC BY 4.0. Attribute it to Cleanlist and link back to this page.

One file, two sections, split at the blank line before the ## SECTION 2 marker row. Section one is 376 rows, one per attributed price statement: query, surface (ai_overview or ai_mode), vendor, quoted_price_usd, quoted_unit, quoted_context (verbatim), vendor_page_price_usd, vendor_page_url, verdict, source_used, attribution_scope, notes. Section two is 43 rows, the 42-vendor supply-side census plus Cleanlist's own row: vendor, in_42_vendor_sample, pricing_page_url, pricing_page_reachable, publishes_a_price, demo_required_to_see_pricing, has_free_tier, free_trial_terms, free_trial_length_days, free_trial_card_required, credits_per_verified_email, credits_per_direct_dial, charged_on_no_match, annual_term_only. Cleanlist carries in_42_vendor_sample=no, so every census denominator on this page stays recomputable without us.

To reproduce it. Take column one of section one as the query list. Call a live Google SERP API with location_code 2840, language English, device desktop, requesting the AI Overview element with the asynchronous flag set, plus an AI Mode endpoint for the 60 buying questions. Extract every dollar figure with its segment. Before matching brands, blank every markdown link target, URL, www host and bare domain with a run of spaces of identical length so offsets stay aligned, whitelisting the vendor names that are themselves domains. Match against a fixed published vocabulary, scope attribution to the markdown segment, drop anything ambiguous. Then fetch each vendor's pricing page the same day in a headless browser with the billing toggle operated in both positions, and apply the four verdicts in the order printed in the method box. Expect your counts to differ, because AI answers regenerate per request. What should survive is the shape: about half confirmable, a fifth unfalsifiable by construction, and the uncheckable share rising with how much of its price a vendor hides.

What should a buyer do with this?

Five checks, each of which costs less than the mistake it prevents.

1. Never carry an AI-quoted price into a budget. Half of them confirm and you cannot tell which half from inside the answer. Open the vendor's page, and if there is no page, treat the number as an estimate nobody stands behind. 82 of the 376 statements here describe a vendor that has published no dollar amount at all.

2. Click the billing toggle before you write the number down. The single most-repeated correct figure in this corpus, Apollo at $49, is an annual-billed rate presented as monthly. Same at Clay, where $167 is annual and $185 is monthly. A buyer budgeting month to month on the quoted figure is short by 32.7% at Apollo and 10.8% at Clay.

3. Do the cost-per-verified-contact arithmetic yourself, in the unit you actually buy. Take the plan price, divide by the credits it includes, then multiply by the credits your work consumes. At one credit per verified email, a $99 plan with 170 credits is $0.58 per email; at Coresignal's 20 credits per email, the same nominal credit price is 20 times that. If you need phones, run the multiplier. Twelve of the 33 vendors publishing a credit model charge more than one credit for a direct dial, and the observed multipliers run 2, 3, 5, 6, 8, 10 and 500. A 1,000-contact month at 1 email credit plus 10 phone credits costs 11,000 credits, not 1,000.

4. Ask what a miss costs, in writing. Only 24 of 42 vendors (57.1%) state anywhere on their pricing page that a no-match is free. The other 18 say nothing either way, which means the answer is in a contract you have not read yet. It is one line in an email and it changes the effective rate by whatever your match rate is.

5. Treat "contact sales" as a data point about the vendor. Nine of 42 vendors require a demo before showing any pricing, and the 8 non-publishers attracted 111 AI price statements, 82 of which nobody can check. A vendor whose only findable number was written by somebody else has outsourced its price list to an answer engine and to whatever blog that engine read.

Cleanlist publishes four tiers, prices all four, and charges 1 credit for a verified work email, 10 for a direct dial and 11 for both, with search free and unlimited on every plan and no charge for a miss. Those figures are in section two of the dataset above, on the same rules as everybody else's.