The Observer's Paradox In B2B: Why Chasing Clicks Is Training Your Campaigns to Find the Wrong Buyers
We explore how to fix it across multi-channel surfaces where B2B research happens at scale
Welcome back to The B2B Stack. A Substack note I posted a couple of weeks ago on the click-optimisation trap struck more of a nerve than I expected, so I want to properly finish the thought here: where the problem actually comes from, what the evidence says, why it gets worse the moment you step outside display, and what a saner measurement stack looks like in practice.
What you’ll learn in this article:
Why optimising toward clicks trains your system to find clickers, not buyers, and why that’s structurally different from a targeting problem
What the evidence actually shows about who clicks on ads, and how much of it has nothing to do with intent
Why B2C-style click optimisation routes budget toward suspect environments, and why in-app games are the clearest example
How brand-safety filtering and click-chasing bidding combine to filter most real B2B impressions out before they ever reach a DSP
Why clicks don’t exist at all in audio and barely register in CTV, and what that means for attribution design
A practical framework for what to measure instead, channel by channel in order to maximise your investments into programmatic paid media
The pattern, restated
Here’s the setup, because it’s worth being precise about it. A GTM team invests in intent data, which is using an array of observable signals across the internet, job boards, pixel tracking to infer tech stacks and other such sources, to indicate which accounts may be in motion - so far, so good. They layer this into an ABM platform or LinkedIn to run some ads. Builds a high intent TAL. Launches the campaign. Then optimises the whole thing toward whoever clicked.
Not who engaged meaningfully. Not which accounts showed multi-signal buying behaviour. Not which members of the buying committee consumed content across three separate touchpoints over three separate visits. Who clicked.
Its a proxy B2B teams have fallen back to since paid media began, largely because unlike our B2C cousins, we do not have the observable (and short) path to trackable purchase as the ultimate source of truth. In b2C, where we may be selling sneakers, ‘all’ we need to do is pick a handful of data segments and bake them off measured on ROAS. This luxury is not in the B2B building.
So when we deploy the worlds best advertising tech to chase clicks, their elite AI optimisation engines go looking for segments of users who look like prior clickers. The algorithm duly learns to find clickers, based on looking for matching behaviours in the new cohorts of users today which matches those behvaiours shown from previous users who went on to click. You may instantly think this doesn’t really matter if they are within your TALs, but it does. It matters because clickers, it turns out, are very often not buyers. They are not the super-high intent cohort within the TAl that they are often assumed (or hoped?) to be.
This isn’t a targeting flaw you can patch with a better audience or a tighter frequency cap. It’s closer to what statisticians call the observer’s paradox: the act of measuring something changes the thing you’re measuring. Optimise toward what’s countable and the system quietly stops looking for what actually matters, because what actually matters was never the thing being counted.
B2B buying is slow, deliberate, and largely invisible to a dashboard. Gartner’s research (see sources at the end of this newsletter) shows buyers spend only around 17% of their total buying time in direct contact with potential vendors, meaning roughly 80% of the journey is self-directed, and Gartner’s 2026 research puts 70-80% of the B2B buying journey inside this dark funnel before any vendor contact form is filled.
So on B2B’s much [over] cited 95/5 rule, only 5% of your accounts are maybe in market for what you sell today. Of those 5%, 80% are likely still in the dark funnel, and maybe 1% of the overall total TAL is both in-market AND showing observable signal.
Forrester’s 2025 survey found the average B2B purchase now involves around 13 stakeholders inside the buyer’s organisation and nine outside it, spanning three or more departments. None of that shows up as a click. Almost all of it happens before anyone lands on your site.
So when a campaign optimises purely on click-through, it isn’t finding the buying committee. It’s finding whoever happened to be holding a phone at the wrong moment.
What clicking actually measures
If you want to see how weak the link between clicking and intent really is, you don’t need to look at B2B specifically. The clearest evidence comes from the environments where clicks are cheapest to generate: mobile display and in-app gaming, which happen to be a huge share of the inventory that flows through open-exchange, B2C-style DSPs.
A widely cited Retale survey found that 60% of clicks on mobile banner ads are mistakes, most often happening while people are checking the news, scrolling social media, or playing games. This, in my opinion, does not render the impression value-less, I think there IS still strong value in the wider, persistent muscle memory building of repeated exposure across environments. When a buying journey does begin, being front of mind is critical to being part of the small segment of prospective vendors that the researching buying committee will speak to - but the click in this instance is not valuable. Its indeed likely to be annoying. To the extent that I sometimes wonder if the fact internet ads can be clicked is a disadvantage.
An earlier Online Publishers Association study put the figure at roughly half of all mobile ad clicks being accidental, and the mechanism is exactly what you’d expect: touchscreen games place control buttons close to the ad unit, and the closer the ad sits to something a person is actually tapping to play, the higher the accidental-click rate climbs.
So when you’re training your DSP to chase clickers, they are going out there actively trying to buy these ad placements in a way they would not be doing if you were not chasing clicks to begin with.
It’s got worse, not better, as interfaces have evolved. Speaking at App Promotion Summit London in 2025, a Kayzen partnerships lead presented data showing video ad click rates on mobile climbing from 10-20% three years earlier to over 80% today, particularly on iOS, driven by interface changes that make accidental taps more likely.
In one heatmap study, 75% of the unintentional clicks clustered around skip buttons were responsible for only 13% of the resulting app installs. Publishers know exactly what’s happening here: the shift from clickable buttons to full-screen clickable ad units, combined with smaller and slower skip buttons, has been a deliberate industry-wide trend.
The TLDR? the publishers know the game the advertisers are playing and they are gaming the system to get the advertising dollars.
Then there’s the deliberate version of the same problem. A meaningful share of app inventory monetises through rewarded formats, where users are paid in coins, tokens, or in-app currency to tap an ad. Security researchers describe these as incentivised clicks: technically generated by real humans, but with artificial intent, since users are clicking to earn a reward rather than out of interest in the advertiser’s offer.
Even by the ad tech industry’s own accounting, more than half of the top US non-gaming apps in a major performance index carry some incentivised ad format, and the practice is especially entrenched in gaming, where roughly 15 of the top 25 media sources sell incentivised inventory alongside their other formats
Put these together and you get a click-through rate that looks encouraging on a dashboard and means almost nothing. A high CTR in this kind of environment isn’t a signal of engagement. Security vendors flag it as one of the clearest tells of a low-quality or fraudulent placement: an abnormally high click-through rate paired with poor post-click engagement is itself a warning sign, not a result to celebrate
This is the part that should worry anyone running B2B programmatic through the same infrastructure built for B2C performance marketing. If your bidding algorithm is trained to chase clicks, and a large share of the cheapest, most abundant clicks in the ecosystem come from someone’s thumb missing a “skip” button in a mobile game, your model is not learning what a buyer looks like. It’s learning what a thumb looks like.
This is all a totally avoidable longtail of online display, which sullies its name and drives B2B brands away from the space. The good news though is that when you do it right, you are buying high intent, high quality attention at significantly better rates then in environments like LinkedIn, Meta and Google - so understanding this, working around it and building better systems is disproportionately worth your time.
Filtered out before you ever see it
Here’s the part of the story that gets missed almost entirely, and it’s the reason so much genuinely B2B-relevant inventory never shows up in a DSP like The Trade Desk or StackAdapt in the first place.
Programmatic brand safety is mostly still run on blunt keyword and category blocklists. It’s designed to protect brands from genuinely harmful adjacency, but it’s calibrated so conservatively that it takes out enormous amounts of perfectly legitimate content along with it.
Industry research on this shows that aggressive blocklists and broad category exclusions routinely knock out legitimate news publishers because news is treated as a blanket risk category, health and wellness content because medical terminology trips keyword filters, and large swathes of non-English inventory because classifiers are less confident outside core markets. The same research puts the scale of the problem starkly: keyword-based brand safety tools and overly strict filters can eliminate 40% to over 60% of otherwise high-quality inventory, and up to 70% of what gets blocked later turns out to have been unnecessarily restricted
Trade press, business analysis, financial commentary, cybersecurity reporting, defence and industrial coverage: this is precisely the category of content a B2B buying committee actually reads, and it’s also precisely the category most likely to trip a legacy blocklist. A programmatic sales lead described the mechanism bluntly: requests for a news private marketplace regularly fail to scale once you trace the problem back to a keyword list, and in some cases buyers are simply blocking news as a category outright, using an old legacy filter that was one of the IAB’s original presets.
The absurdity of the approach became visible during Euro 2024 coverage, when roughly half of one major publisher’s tournament coverage was blocked for containing football terms like “shoot” and “attack”. The same crude logic runs across B2B-adjacent verticals every day: earnings terminology, security incident coverage, macro and geopolitical commentary, all of it liable to trip a filter built for a completely different risk profile.
So the inventory most likely to carry a genuine business audience is disproportionately filtered out before the auction even happens. What’s left in the open exchange, disproportionately, is the cheap, high-volume, high-accidental-click inventory that survives the blocklist precisely because it’s bland enough not to trigger it: casual games, low-quality utility apps, content farms.
Then click-optimised bidding does the rest. Even where quality B2B-adjacent inventory does clear the brand-safety filter, it rarely wins a click-chasing auction against a mobile game with an 80% click rate, because the algorithm was never taught to tell the difference between a genuine click and a thumb slip. Two filtering layers, stacked, and the result is the same: a meaningful share of the impressions your buying committee is actually exposed to never gets bought at all, while budget concentrates in exactly the environments the evidence above describes.
This is the mechanism worth sitting with. It isn’t that B2B advertisers are choosing bad inventory. It’s that the combination of blunt brand safety and click-chasing optimisation systematically filters out the inventory that would have reached real buyers, and systematically rewards the inventory that generates the most low-intent taps.
Clicks don’t exist everywhere else
Assume, for a moment, that you fix all of the above within display. You still have a channel problem, because most of a modern B2B media plan doesn’t run through a clickable surface at all.
Audio has no viewability standard and no clickthrough benchmark to speak of, which is one reason WARC found that while listeners spend around 31% of their time with audio, brands allocate less than 9% of total media budget to the channel. This is changing fast and something we see at FunnelFuel is a rapidly increasing appetite from leading vendors to buy the underpriced audio inventory representing industry podcasts and other quality media. Research shows that the c-suite spend 51 minutes a day on audio - often whilst in the gym, commuting or walking the dog. The lack of clicks is no longer holding back the biggest vendors, hey want in, especially now ABM can be run end to end in audio environments.
CTV is worse still on this click predicated dimension. A viewer watching a spot on a connected TV device simply cannot tap a link, and the default attribution windows on most DSPs are still built around click-through behaviour that the channel structurally cannot produce. Video completion rate tends to run high in CTV precisely because most ad formats there are non-skippable, which makes it a useful signal for detecting gross problems but a poor one for anything more granular. B2B buyers are not going to be scanning QR codes for offers either, so we have to think differently now.
The channels that carry the most weight in a considered B2B buying journey, audio and long-form video among them, are exactly the channels where a click-based measurement stack goes silent. If your attribution model can’t register anything in the channel, the channel gets underfunded relative to its actual influence, and budget drifts back toward the clickable surfaces described above by default, not by decision.
The industry’s answer here is more mature than in B2B, and worth borrowing wholesale.
CTV outcome measurement typically leans on some combination of view-through attribution, household-level matching that links exposure to downstream activity, modelled or probabilistic attribution, and lift testing that compares an exposed group against a genuine control. I know CTV can be bought targeting and reporting against an account, because I helped design the design that enables it. The rest can therefore be achieved once the account signal is known. This is doable today.
Multi-touch attribution across devices tends to be more accurate than last-click for capturing CTV’s upper-funnel contribution, even though it isn’t foolproof, and brands running incrementality tests have consistently found CTV view-through contributing 15-25% lift to site visits and conversions that a last-touch model would miss entirely. For a channel with no click, the answer isn’t to force a click-shaped metric onto it. It’s to build a measurement stack that was never expecting one.
What attribution should actually look like
So: what would you optimise toward if clicks didn’t exist? A workable answer looks less like a single metric and more like a layered confidence model, built around three shifts.
Move the unit of analysis from the user to the account. A click happens to one device, at one moment, possibly by accident. A buying decision happens across a committee, over months. Any signal you weight has to be capable of aggregating across a real buying group, not a single cookie or device ID, or you’re structurally incapable of seeing the thing you’re trying to measure.
Weight engagement depth over engagement occurrence. Whether someone clicked is a binary. Whether they read to the end, returned within the attribution window, consumed content across formats, or showed up across multiple members of the same account is a much richer signal, and it’s available in the same data most teams are already collecting but not weighting. Dwell time, return visits, and cross-format consumption within an account correlate with buying-stage progression in a way a single click never will.
Replace channel-specific proxies with channel-appropriate ones, then unify them at the account level. View-through and household matching for CTV. Completion rate and frequency for audio. Genuine multi-touch weighting, not last-click, for anything with a clickable surface. None of these individually solves attribution. Together, rolled up to the account rather than the device, they build something closer to a confidence score than a click count, which is the actual point.
Practically, this means auditing where your current optimisation signal is actually coming from before you touch anything else. Pull the placement-level breakdown behind your best-performing “engagement” segment and check what proportion of it is running in-app, on mobile, in casual gaming or utility categories. If a disproportionate share of your best-performing inventory by CTR is running there, you already have your answer about what the algorithm has actually learned to find. Then look at what’s been filtered out on the other side: run a domain and category report on your blocklist exclusions and ask, honestly, how much of what’s excluded is genuinely unsafe versus simply unmeasured by an outdated keyword list.
None of this requires new inventory. It requires a different question at the point where the bid is placed and the model is trained: not “did this generate a click,” but “does this look like the kind of exposure a buying committee accumulates on the way to a decision.” That’s a harder thing to build. It’s also the only version of the question that was ever worth asking.
If this resonated, the honest next step is auditing where your own optimisation signal is actually coming from, not adding another tool on top of it. Subscribe below for the next instalment, or reply and tell me what your placement-level breakdown looks like once you actually pull it.
Sources referenced:
Here’s the full source list with URLs, matched to what each one supports in the piece:
Accidental clicks
Online Publishers Association study (via VentureBeat) — ~50% of mobile ad clicks accidental: https://venturebeat.com/mobile/conversions-over-clicks-how-mobile-app-developers-can-drive-results-for-advertisers/
Retale survey (via MediaPost) — 60% of mobile banner clicks are mistakes: https://www.mediapost.com/publications/article/268266/60-of-all-mobile-banner-ad-clicks-are-accidents.html
Pixalate accidental click data (2017 programmatic analysis by device/format): https://www.pixalate.com/blog/accidental-clicks-mobile-ads-data
Kayzen / App Promotion Summit London 2025 findings (10-20%→80%+ CTR trend, skip-button heatmap study) via Business of Apps: https://www.businessofapps.com/news/mobile-video-ad-click-rates-surge-raising-questions-about-user-intent/
Pontiflex/Harris Interactive survey (47% of app users click ads by mistake) via MarketingProfs: https://www.marketingprofs.com/charts/2011/4365/mobile-app-users-click-on-ads-mostly-by-mistake
Tolomei et al., Yahoo Research, accidental click detection paper: https://arxiv.org/abs/1804.06912
Incentivised / fraudulent clicks
HUMAN Security click fraud industry guide: https://www.humansecurity.com/learn/resources/click-fraud-industry-guide-protection-prevention/
TrafficGuard on click fraud types (rewarded/token clicks): https://www.trafficguard.ai/blog/types-of-click-fraud
AppHarbr on forced/deceptive click ads: https://appharbr.com/mobile-forced-clicks-ads/
AppsFlyer on rewarded advertising prevalence in gaming apps: https://www.appsflyer.com/blog/mobile-marketing/rewarded-advertising-good-bad-ugly/
ClickGuard on fraudster click monetisation mechanics: https://www.clickguard.com/blog/how-fraudsters-make-money/
Wikipedia, Click fraud (background/definitional): https://en.wikipedia.org/wiki/Click_fraud
B2B dark funnel / buying committee stats
Gartner/Forrester figures via Similarweb: https://www.similarweb.com/blog/marketing/marketing-strategy/dark-funnel-b2b/
Gartner/Forrester figures via Medium (Sterling Phoenix): https://medium.com/@IamSterlingP/the-dark-funnel-is-not-a-measurement-problem-it-is-a-buying-system-shift-ed03081327d3
ENaiBLD on Gartner 2024 buyer research (17% direct contact time, stakeholder counts): https://enaibld.com/resources/buyer-behavior-and-the-modern-sales-cycle/gartner-forrester-research-confident-misunderstanding-b2b-buying/
Iliana AI on Forrester 2025 Buyers’ Journey Survey (13 internal/9 external stakeholders): https://ilianaai.com/dark-funnel-explained/
Brixon Group on Gartner’s 80% rule: https://brixongroup.com/en/the-modern-b2b-buying-journey-why-buyers-complete-80-of-their-journey-alone-and-how-you-can-still-remain-visible
Brand safety / keyword over-blocking
MGID on brand safety over-blocking (40-70% figure): https://www.mgid.com/blog/brand-safety-in-programmatic-advertising
AdMonsters on Jana Meron / news over-blocking: https://www.admonsters.com/rethinking-brand-safety-lessons-from-jana-meron-on-news-advertising-in-2024/
Digiday, programmatic sales lead confessional on news blocking: https://digiday.com/media/they-are-blatantly-blocking-news-confessions-of-a-programmatic-sales-lead-on-brand-safety-filters-impact-on-publishers-direct-sold-ads/
Peer39 on the cost of keyword blocking: https://www.peer39.com/blog/costs-of-keyword-blocking
ExchangeWire on 2024 Euros “shoot/attack” over-blocking: https://www.exchangewire.com/blog/2024/12/10/rewriting-the-rules-how-2025-will-redefine-brand-safety-in-media/
Audio / CTV clickless attribution
Adweek/Adweek partner content on audio measurement (31% time / 9% budget): https://www.adweek.com/partner-articles/when-it-comes-to-audio-measurement-think-programmatic/
AI Digital on CTV measurement methods: https://www.aidigital.com/blog/ctv-measurement
Adtaxi on proving CTV performance without last-click: https://www.adtaxi.com/blog/proving-ctv-performance-without-last-click-attribution/
Singular on CTV attribution: https://www.singular.net/ctv/
Influencers Time on CTV attribution windows / DV360 config, 15-25% lift figure: https://www.influencers-time.com/creator-content-on-programmatic-ctv-specs-and-rights-guide/
Strategus / Gigawatt Media on VCR/PCR as CTV metrics: https://www.strategus.com/blog/ctv-measurement-performance-metrics-attribution / https://gigawatt.media/the-basics-of-analyzing-programmatic-campaign-performance/
This really gets to the heart of the issue. Optimizing for clicks can improve campaign metrics while moving further away from real business outcomes. As measurement evolves, the focus has to shift from volume to signal quality because better decisions start with more trustworthy signals.