Can a vendor case study be used as a benchmark?

No. A case study proves a result happened once, for one customer, under conditions that customer chose and that we chose to publish. The 24 stories in this chapter were published by us, picked because the result was good, never audited, and never recalculated under the benchmark's rules. They are never averaged, never put on a benchmark chart and never used to set an expectation. What tells you what to expect is the range across 153 advertisers, published in the explorer and the methodology.

When may a customer result be compared with a benchmark number?

Only when the two measure the same thing. A cost per lead may sit beside what other advertisers paid for a lead; a cost per click beside the published click rates. A multiple built on every deal the ads touched matches nothing here, so it is shown alone and renamed as what it is: pipeline the ads touched, divided by spend, as the customer reported it. Where a study reports only a percentage improvement on its own past performance, it cannot be placed against our data at all, and the row says so.

2026 B2B Benchmark · The tail chapter

Twenty-four great results. None of them tells you what to expect.

Where these numbers come from. Every figure on this page is a result a customer reported in their own case study. None of them is drawn from the 153-advertiser benchmark, none was computed by us, and none clears the publishing thresholds the rest of this report is held to: no minimum number of advertisers, no spend floor, no concentration cap. They are here as stories about what has happened, not as figures to plan against. How the benchmark numbers are made.

One customer each, published because the result was good.

Every other number in this report describes a group of advertisers we counted before anyone knew how they did. The numbers on this page are the opposite: single customers, written up because the result was good enough to publish. They prove a great outcome is possible; they do not tell you what to expect. So they live here, away from the benchmark. Each one printed with its unit, what it measures, the period, what it was compared against, and a link to the full study, and set beside a benchmark number only where the two are genuinely measuring the same thing. This is what a best-case result looks like when it happens.

Published by us · picked because the result was good · not audited · one customer per story · never averaged · Methodology · What each number means · Numbers we killed

How to read the numbers on this page

Every benchmark figure in this report is locked to a checked list of approved numbers: it is tagged in the page, checked against that list before publication, and blocked if the value drifts or the claim was thrown out. The customer figures on this page sit outside that system on purpose. The list covers numbers we worked out from the data; it cannot vouch for a number a customer worked out about themselves, using their own definition, over a period they chose. Pretending otherwise would dress a marketing figure up as a measured one.

Two kinds of number, treated differently

Customer figures look like this, 122X ROI, amber, underlined, and never on the approved list. They were reported by the customer, picked because they were good, never audited, and Metadata did not recalculate any of them under this report’s rules.

Benchmark figures look like the rest of the report: plain type, on the approved list, with the number of advertisers printed beside them, ads returned 0.56x in the first year across 127 advertisers, at $58,887 per customer and a 21.2% close rate, counting only deals that actually closed or actually died.

If you remember one rule from this page: a number you cannot recalculate is not a benchmark, whoever published it.

The rule this chapter follows

Two kinds of number appear in this report, and they answer different questions. The ranges tell you what to expect, what happened across 153 advertisers who were counted before anyone knew how they did. The stories below are best-case results, picked because they worked, one customer at a time. Putting them on the same chart would compare two different things, so we do not.

“122X is one customer we chose to write up, counting every deal their ads touched. 0.56x is the revenue we can trace back to an ad, per ad dollar, across 127 advertisers whose CRM we could read. Different measure, different companies, different job. They only sit together if you stop treating one great result as the average.”

That sentence is the whole rule in one line, and it is why this chapter exists instead of a “customers see…” strip on the cover.

Standing footnote, wherever an “influenced” number appears

Influenced means the ad touched the account at some point before the deal closed. It does not mean the ad created the deal. It is not comparable to the 0.56x benchmark, which counts only revenue we can trace back to an ad.

When a customer number may sit beside a benchmark number

A customer result may sit beside a benchmark number only when the two measure the same thing. A cost per lead goes beside what other advertisers paid for a lead. A cost per click goes beside the published click prices. A multiple built on every deal the ads touched goes beside nothing here: we relabel it as what it is, pipeline the ads touched, divided by spend, as the customer reported it, and show it alone. Where a study publishes only a percentage improvement on its own past performance and never the actual price, it cannot be placed against our data at all, and the row says so instead of guessing.

Never averaged. There is no “our customers average” number on this page, and there will not be one: averaging results that were picked for being good produces a number with no group of companies behind it.
Never a bare multiple. Every figure carries its unit, what it was worked out on, and a link to the study that published it.
Never on the cover. Nothing in this chapter appears in the executive summary, on a benchmark chart, or next to a typical result.
Standing label: picked because they worked, not typical.

These are exceptional results and may not repeat. What tends to drive them: audience quality, how fast the team tests creative, and how well marketing and sales are aligned. One customer per story. Nothing here is adjusted for industry, deal size, or the sales that would have happened anyway.

Group 1, cheaper leads and clicks

Cost per lead, cost per click, cost per MQL, ad spend. These are the only customer figures in the chapter that measure the same thing as something we publish, which makes them the only ones that could sit beside a benchmark number, and even here it mostly does not work, for a reason worth stating plainly.

Cost per lead · cost per click · cost per MQL

What the studies report

Instruqt: 94% lower cost per lead, year over year, and 139% more leads year over year, after replacing an agency. Baseline: their own prior arrangement, the study states $30K a month for fewer than ten leads. The study also reports cost per lead down 88% in the first month.
Zoom: 77% lower cost per click and 24% less total ad spend, against their own prior manually-managed bidding. Period not stated.
Writer: 86% lower cost per MQL. MQL is Writer’s qualification definition, not this report’s lead definition, so it is a different unit from anything we publish.
GoodTime: 13.4% less spend on the same Google Search programs.
Cacheflow: lower cost per click than native Facebook and LinkedIn targeting, and higher deal sizes. Both published without a figure, so both are qualitative only.
Fivetran: 3,300 leads at $115 per lead in the first six months on LinkedIn, the one absolute cost per lead in the whole chapter.

The one comparison this group supports

Fivetran’s $115 per lead is a cost per lead, so it can sit beside what other LinkedIn advertisers actually paid. Across 88 of them, the cheapest quarter paid $196 or less, the middle advertiser paid $376, the priciest quarter started at $658, and the priciest 10% paid $1,340 or more. A lead at that price is cheaper than three out of four advertisers managed. That is a location, not a ranking and not a promise: it is one advertiser, over six months of their own choosing, on campaigns we did not recalculate under this report’s rules.

The others cannot be compared at all, and the reason is simple. A percentage improvement is measured against that company’s own past; our numbers are measured against every other advertiser. Knowing a cost per lead fell by 94% tells you nothing about where it landed. A 94% improvement on a bad number can still be worse than what most advertisers pay. Without the actual price, there is no position to report, so we report none.

For the same reason, the LinkedIn cost per lead of $202 that everyone quotes is cheaper than three out of four advertisers actually paid, only 26% of them paid less. Even our own headline number is a good result, not a typical one. Plan against where you actually sit.

For cost per click we publish no advertiser-by-advertiser range this year, only the going rates: $11.90 on LinkedIn, $3.80 on Facebook, $2.82 on Instagram. Zoom’s reduction measures the same thing as those rates, but there is no range to place it in, so we do not place it.

What this group can proveA lead far cheaper than three out of four advertisers pay is reachable, and big year-over-year cuts in ad cost do happen when an advertiser changes how campaigns are built and bid.

What it cannot proveThat any of these cuts is typical, caused by the platform, or repeatable at your account; that a cheaper lead was a better lead, none of these studies reports cost per customer; or where four of the six advertisers sit against everyone else, because they publish improvements rather than prices.

Group 2, cost per opportunity

Cost per opportunity · opportunity counts

What the studies report

Monotype: 83% lower cost per opportunity, against their own prior manual experiment process.
Webex Events: 44% lower cost per opportunity and 53 opportunities created, on 73% less budget, during a quarter in which the team launched 30 campaigns.
GoodTime: 42% lower cost per opportunity on Google Search.
Nitrogen Wealth: 3X more sales-accepted opportunities on the same budget.

These are the most useful figures in the chapter for running a team, and the least comparable. This report publishes no numbers for cost per opportunity. The customer numbers that survived our publication rules are cost per customer, payback we can trace and close rate, all counted on deals that actually closed in the CRM rather than on opportunities created. An opportunity is also the least standard thing in B2B: what counts as one differs from company to company, and in three of these four studies the comparison is against that same company’s earlier definition. That is the only fair comparison available, and it still is not a market one.

What this group can proveThat moving budget between experiments, keywords and audiences cut one advertiser’s own cost per opportunity by a large factor. The mechanism the rest of this report argues for, seen inside a single account.

What it cannot proveWhere any of them sits against the rest of the market: we publish no cost-per-opportunity benchmark to compare with, and the advertisers in our data do not agree on what an opportunity is. Do not turn these percentages into an expected saving.

Group 3, pipeline and return multiples

This is the group all the rules on this page exist for. It holds the biggest numbers in the chapter, the loosest definitions, and the one figure most likely to be quoted out of context.

Standing footnote, wherever an “influenced” number appears

Return multiples, renamed to say what they actually measure

Firstup: 122X ROI, published alongside $6M revenue, a 24X pipeline increase and 70 new influenced opportunities over six months. Those opportunities are ones the ads touched, and the study never publishes the arithmetic behind the multiple, so we carry it here as what it is, pipeline the ads touched divided by spend, as Firstup reported it, and set it beside no benchmark number.
Nitrogen Wealth: 14X ROI triggered and 23X ROI influenced. The one study in the set that separates the two itself, which is exactly the fine print the other twenty-three are missing. The story text states the triggered figure as 14.45X.
Monte Carlo: 14.5X ROI on $6.1M pipeline. Pipeline divided by spend is a pipeline-to-spend ratio, not a return: pipeline is not revenue.
N-able: 524% pipeline ROI, relabelled here as a pipeline-to-spend ratio, for the same reason. It is the same shape of number as the one above, expressed as a percentage.
Zoom: 9X ROI, up from a stated baseline of 3.26X, with 177% influenced revenue increase and 252% more opportunities.
Pendo: 2X ROAS improvement, return on ad spend as Pendo computes it, on $5M+ pipeline and 32K+ net new leads.
ThoughtSpot: 9.8X ROAS in the first six months on Facebook, alongside 64 opportunities and $2.8M pipeline in that window.

Every multiple above is counted on pipeline, or on every deal the ads touched, or on a basis the study never states. The benchmark’s 0.56x is counted on revenue from deals that actually closed and that an ad set off, inside a fixed 12 months, across 127 advertisers whose CRM we could read, with deals still open on the cut-off date counted as nothing. Those are not the same thing divided by the same thing, and no arithmetic reconciles them. Only the sentence above does.

Pipeline and revenue amounts · opportunity and deal counts

Eightfold.ai: 26 opportunities, 13 closed-won deals and $8M pipeline over one year.
Fivetran: 1,400 opportunities, 94 closed-won deals and $6M pipeline.
Qualified: $6.9M influenced pipeline from 2,000+ surging accounts reached.
ThoughtSpot: 193 opportunities and $5M pipeline overall.
Pendo: $5M+ pipeline generated.
Gainsight: $4.8M pipeline generated within twelve months.
Automation Anywhere: $3.6M pipeline sourced.
Instruqt: $522K pipeline generated.
Webex Events: 60% pipeline increase on a cut budget.
GoodTime: 50% more pipeline and 320% more dollars in pipeline from the same search programs.
Titan: 3X qualified pipeline, against their own prior manual approach.
Zingtree: a closed-won deal within 30 days of launch, the fastest outcome in the set, and a single deal.
Cacheflow: their largest deal to date, closed from a cold-targeting Facebook campaign. No amount published.
Fingerprint: millions in hidden pipeline uncovered in three months. A visibility claim, not a generation claim: the pipeline existed and was not being seen.

Total pipeline dollars cannot be compared with anything in this report, by design. We publish rates only, costs per unit, percentages and multiples, and never a group’s total dollars, because a total in a small group would let a reader work back to an individual advertiser. So there is no range of pipeline dollars to place these against, and building one would break the rule that protects the advertisers in our data.

What this group can proveThat paid programs on this platform have produced multi-million-dollar pipeline and won deals for individual advertisers, and that returns far above the report’s payback floor have been reported at least once each.

What it cannot proveWhat you should expect to get back. Not one of these multiples is worked out the way the benchmark works out payback; most divide pipeline, or revenue the ads merely touched, by spend, over a period the advertiser chose. Averaging them, quoting a range across them, or reading any of them as what a new advertiser should expect is the exact mistake this chapter exists to prevent.

Group 4, time and workload

Experiments run · hours automated · cost avoided

ThoughtSpot: 1,600+ experiments run and 2,795 hours automated.
Fivetran: 6,967 hours automated across 2,215 campaigns tested, the largest operational figure in the set.
Eightfold.ai: 1,414 hours automated.
Qualified: 703 hours automated.
BigID: 300+ experiments launched and 632 hours automated after moving campaign management in-house.
Monte Carlo: 621 hours automated and 3X less time on campaign management, by a one-person paid media team.
Gainsight: 70% of manual execution eliminated and $100K annual savings.
Pendo: 3X faster campaign building with a two-person team.
Zoom: campaign setup from 3.5 weeks to under 10 minutes.
Docebo, LaunchDarkly and Fingerprint: reporting, attribution and account-visibility outcomes published without figures, qualitative only.

Hours saved is not a return. It is a labour number, and it belongs in a different conversation from cost per customer: an hour saved is worth whatever the person freed by it does next, which no case study measures. We publish these because capacity is the mechanism behind several of the pipeline stories above, a two-person team running 30 campaigns in a quarter is a different team from the one running three. Not because hours turn into dollars at any rate we can defend.

What this group can proveThat the number of experiments and the day-to-day campaign workload changed by large factors for these advertisers. A claim about what the platform can do, checkable in their own accounts.

What it cannot proveAny revenue, pipeline or efficiency result whatsoever. Hours saved must never be turned into a return, and a count of experiments says nothing about how good those experiments were.

All 24 stories, with the fine print attached

One row per published customer story. Every figure is reproduced exactly as the study states it, with its unit and what it measures attached, and the last column says either which benchmark number it may be read against or why it may be read against none. Four pieces of fine print are the same for all 24 rows, so they are stated once here rather than repeated 24 times:

How these were picked: we published them, and we published them because the result was good. These are the stories we chose to write up out of more than 200 customers. We do not publish the ones that did not work, and no reader should assume they do not exist.
Who checked them: the customer reported them and nobody audited them. Metadata did not recalculate a single figure below under this report’s rules, publication limits or 12-month period.
How credit was assigned: the studies do not say, and it is not the even credit split used everywhere else in this report.
Who is in it: one customer per row. Nothing in this table is adjusted for industry, deal size, sales cycle or the sales that would have happened anyway.

Seven columns of fine print, scroll the table sideways to read them all.

Customer	Figures, as the study publishes them	What the study calls it	What kind of number	Period	Compared against	Can be read against
Automation Anywhere	$3.6M pipeline sourced	“Pipeline Sourced”	A pipeline total they say the ads sourced	Not stated	None stated	Nothing, we publish no pipeline totals
BigID	300+ experiments launched; 632 hours automated	“Experiments Launched”, “Hours Automated”	Workload	Not stated	Own prior agency-managed process	Nothing, we publish no number like it
Cacheflow	Largest deal closed from cold targeting; higher deal sizes; lower cost per click vs native	Stated without any figures	Described, not measured	Not stated	Own prior native targeting	Nothing, qualitative only
Docebo	6 audience segments built; real-time account insights	“Audience Segments Built”	Workload / described, not measured	Not stated	Own prior reporting process	Nothing, we publish no number like it
Eightfold.ai	26 opportunities; 13 closed-won deals; $8M pipeline; 1,414 hours automated	“Opportunities”, “Closed-Won Deals”, “Pipeline”, “Hours Automated”	Counts of opportunities and deals; a pipeline total; workload	One year	None stated	Nothing. These are counts and totals, not rates
Fingerprint	Millions in hidden pipeline uncovered; hours saved each week	Stated without a figure	Visibility / described, not measured	Three months	Own prior fragmented tooling	Nothing, qualitative only
Firstup	122X ROI; 24X pipeline increase; $6M revenue; 70 new influenced opportunities	“ROI”. The sum behind it is never published; the opportunities are ones the ads touched	Deals the ads touched, pipeline divided by spend, as the customer reported it	90 days, six months and overall milestones	Own prior paid-social performance	Nothing. A number built on every deal the ads touched gets a warning label, never a benchmark comparison
Fivetran	1,400 opportunities; 94 closed-won deals; $6M pipeline; 3,300 leads at $115 per lead; 6,967 hours automated	“Opportunities”, “Closed-Won Deals”, “Pipeline”, cost per lead	An ad-cost number (the cost per lead); counts and totals elsewhere	First six months, then overall	Own prior LinkedIn performance	Cost per lead only: cheaper than three out of four of the 88 LinkedIn advertisers paid
Gainsight	$4.8M pipeline generated; $100K annual savings; 70% manual work eliminated	“Pipeline Generated”, “Annual Savings”, “Manual Work Eliminated”	A pipeline total; workload	Twelve months	Own prior manual process	Nothing, we publish no pipeline totals and no hours-saved numbers
GoodTime	50% more pipeline; 13.4% less spend; 42% lower cost per opportunity; 320% more dollars in pipeline	“More Pipeline”, “Less Spend”, “Lower Cost Per Opportunity”	An improvement on their own past	Not stated	Own prior Google Search programs	Nothing, we publish no cost-per-opportunity numbers
Instruqt	94% lower cost per lead; 139% more leads; $522K pipeline; 10X faster launch time	“Lower CPL”, “More Leads”, “Pipeline Generated”, “Faster Launch Time”	An improvement on their own past; a pipeline total; workload	Year over year; first month reported separately	Own prior agency arrangement	Nothing. An improvement with no actual price cannot be placed against what others paid
LaunchDarkly	100% accurate pipeline attribution; zero UTM parameter errors	“Accurate Pipeline Attribution”, “UTM Parameter Errors”	Workload / data quality	Not stated	Own prior tracking setup	Nothing, we publish no number like it
Monotype	83% lower cost per opportunity	“Lower Cost Per Opportunity”	An improvement on their own past	Not stated	Own prior manual experiment process	Nothing, we publish no cost-per-opportunity numbers
Monte Carlo	$6.1M pipeline; 14.5X ROI; 621 hours automated	“ROI”. The sum behind it is never published, and it is reported beside a pipeline total	Pipeline-to-spend ratio	Not stated	None stated	Nothing. Pipeline divided by spend is not how the benchmark counts payback
N-able	524% pipeline ROI	“Pipeline ROI”, renamed here as pipeline divided by spend	Pipeline-to-spend ratio	Not stated	None stated	Nothing, pipeline is not revenue
Nitrogen Wealth	14X ROI triggered; 23X ROI influenced; 3X sales-accepted opportunities	“ROI Triggered” and “ROI Influenced”, the only study that separates them; the story text states 14.45X triggered	Deals the ads set off, and deals they merely touched, stated separately	Through and after a rebrand; dates not stated	Own prior performance, same budget	Nothing. That figure was not worked out on this report’s 12 months, credit split or publication rules
Pendo	$5M+ pipeline; 32K+ net new leads; 2X ROAS improvement; 3X faster campaigns	“Pipeline Generated”, “Net New Leads”, “ROAS Improvement”	A pipeline total; a return the ad platform reported; workload	Not stated	Own prior two-person workflow	Nothing, the platform reported this return, not the CRM
Qualified	$6.9M influenced pipeline; 2,000+ surging accounts reached; 703 hours automated	“Influenced Pipeline”, “Surging Accounts Reached”	Deals the ads touched	Not stated	None stated	Nothing, pipeline the ads merely touched matches no number we publish
ThoughtSpot	1,600+ experiments; 193 opportunities; $5M pipeline; 2,795 hours automated; first six months on Facebook 64 opportunities, $2.8M pipeline, 9.8X ROAS	“Experiments Run”, “Opportunities”, “Pipeline”, “Hours Automated”, ROAS	Counts and totals; a return the ad platform reported; workload	First six months, then overall	Own prior intent-data activation	Nothing, the platform reported this return, not the CRM
Titan	3X qualified pipeline	“Qualified Pipeline”	An improvement on their own past	Not stated	Own prior manual approach	Nothing, a multiple of a starting point the study never states
Webex Events	60% pipeline increase; 73% less budget; 44% lower cost per opportunity; 53 opportunities created	“Pipeline Increase”, “Less Budget”, “Lower Cost Per Opportunity”, “Opportunities Created”	An improvement on their own past; a count of opportunities	One quarter (30 campaigns launched in it)	Own prior lead-gen-heavy budget	Nothing, we publish no cost-per-opportunity numbers
Writer	86% lower cost per MQL	“Lower Cost Per MQL”, MQL as Writer qualifies it	An improvement on their own past, on a unit only they define	Not stated	Their own earlier, evenly split budget	Nothing. An MQL is not what this report calls a lead
Zingtree	30 days to a closed-won deal; ~100% targeting confidence	“Days to Closed-Won Deal”, “Targeting Confidence”	How fast one deal closed; described, not measured	30 days from launch	Own prior paid-search performance	Nothing. One deal is not a rate
Zoom	77% lower cost per click; 9X ROI up from 3.26X; 252% more opportunities; 177% influenced revenue increase	“CPC reduction”, “ROI”, “More opportunities”, “Influenced revenue increase”	An improvement on their own past; revenue from deals the ads touched	Not stated	Own prior performance, including the stated 3.26X	Cost per click only, against the going rates, we publish no advertiser-by-advertiser range for cost per click

Influenced means the ad touched the account at some point before the deal closed; it does not mean the ad created the deal. It is not comparable to the 0.56x benchmark, which counts only revenue we can trace back to an ad. Twenty-four stories are published and Metadata has more than 200 customers, so this table is a hand-picked set of successes and cannot be read as a success rate.

How do we know?

Differently from every other page in this report, and that is the point of the chapter:

Where they come from: every figure above is copied word for word from the published study on this site, linked in its row. Nothing here was worked out for this report, and nothing here cleared the report’s publication rules, because those rules apply to groups of advertisers, not to individual customers.
How they were picked: because they worked. A case study exists because the result was good. The 153 advertisers exist because they spent money in 2025. Who was in was fixed before anyone knew how they did. That single difference is why one set can tell you what to expect and the other cannot.
How credit was assigned: unknown, story by story, and certainly not the even credit split the benchmark uses. Where a study says influenced, the standing footnote above applies; where it says nothing, assume nothing.
Over what period: whatever the advertiser chose, or unstated. The benchmark uses the same 12 months for everybody, with every deal’s status frozen on one stated date.
No averaging, ever: there is no arithmetic on this page that combines two customers. An average of results picked for being good is not an estimate of anything.
What would change our mind: a customer result becomes quotable the moment it is recalculated under the rules in the Closed-Won Protocol, same 12 months, same definition of a sale an ad set off, same even credit split, same limit on any one advertiser, and survives them. Until then it is proof that it happened once, not proof that it is likely, and it is published as exactly that.

The rules behind every other number in this report: the Closed-Won Protocol, what each number means, how to check our working, the questions this data cannot answer, and the twelve claims we killed.

Cite as: Metadata 2026 B2B Ad Spend Benchmark, Version 2026.1 (n=153 advertisers, 2025), metadata.io/benchmark-report-2026. Do not cite this page as a benchmark: it is the tail chapter, and none of its figures is one.

The other kind of number

Open the range these stories are not in

What every advertiser in the data actually paid and produced, filterable by company size, industry, channel and format, no email gate. The numbers we held back stay visible, labeled with the rule they broke.