Click any text to change it, then click away to save. Numbers are locked. Edits wait here until Claude applies them to the source and republishes: send a comment or a chat message when finished.
Share this figureTraditional attribution only sees what a system logged: 51% of the credit. The other 49% was extracted or inferred from unstructured data like emails, calls and notes. From Upside's study of 500+ new business deals.Traditional attribution only sees what a system logged: 51% of all deal credit in Upside's AI analysis. The other 49% was never logged as a touchpoint, with 25% extracted from unstructured data like emails, calls and notes and 24% inferred when no record holds it. Inferred credit is the softest of the three grades: its size depends on how much of the deal the records cover. Figure 1 from The Untracked Half, Upside's study of 500+ new business deals at 13 B2B companies.
A word from the author
Half the money I spend on advertising is wasted; the trouble is, I don't know which half.
B2B leaders have always known most attribution models miss a big portion of what's really working. The famous line usually credited to John Wanamaker has stayed true, and it's something that still comes up in almost every B2B CMO discussion or community I have been part of.
My intuition has always said that most of that missed tracking, what's really working, sits in marketing. The problem is that there has been no way to actually prove it, so I decided to create a new dataset to measure just how wrong traditional attribution models based on messy data really are.
At Upside we connect to our customers' data sources and clean, enhance and unify their internal data. Then Pipedash, Upside's AI-powered attribution model, has multiple independent agents analyze and reason through every deal to understand what actually mattered. I used an anonymized set of more than 500 new business deals across 13 B2B companies, about 4,000 hours of agent work and 4.4 billion tokens for this analysis.
Here's what traditional models undervalue, overvalue and miss entirely.
What tracked, extracted and inferred evidence looks like on a deal
Seen by traditional attributionHidden from traditional attribution models
TrackedLogged as a touchpoint
CRM activity4 records
EmailRe: proposal2 Marlogged
MeetingIntro call9 Marlogged
CampaignWebinar16 Marlogged
Page view/pricing18 Marlogged
A logged email, meeting, campaign or page view. The only kind of evidence a traditional attribution model can read.
ExtractedEvidence exists in unstructured data
Email thread1 record
Thanks for the walkthrough. Our board member suggested we take a look, so we would like to move quickly on this.
never logged
Found inside another record, such as an email thread, call transcript, Salesforce note or website form fill, but never logged as a touchpoint of its own.
InferredImplied from existing data
Evidenceno touchpoint record
Booked a demo already knowing what the product does
No referrer, UTM, page view or campaign behind the visit
The company runs heavy out-of-home advertising where the prospect works
Inferred: brand awarenessevidence cited
Implied by evidence, e.g. brand awareness, a past relationship or a compliance deadline.
Source: Upside's AI analysis of the study dataset, October 2026.
Traditional attribution reads only part of each deal.
Only the touchpoints that made a difference, as detected by Pipedash, Upside's AI-powered attribution model. Several independent AI agents read every email, call transcript, note and record behind a deal and argue over the credit, a judge weighs their reasoning, and every point of credit cites its evidence (how it works).
The CRM, email, calendar, call transcripts and web activity joined by Upside: the same person resolved across systems, duplicates removed, buying groups detected and touchpoints matched to accounts and deals. More on how Upside cleans and enhances data.
The CRM and the marketing automation tied to it, as they are: the closest to what a company runs attribution on today. It still has some enhancement, because it runs on Upside's data: the CRM's tasks are turned into touchpoints (the emails, calls and meetings reps log, campaign memberships and marketing-automation email activity), credited to the contacts a rep added to the opportunity.
Source: Upside's AI analysis of the study dataset, October 2026.
The study covers every new business deal with a completed AI analysis at 13 B2B companies that use Upside, each read and analyzed at least 3 times by AI agents. Each company's data was anonymized and aggregated on its own before anything was pooled.
521new business deals only; no renewals or expansions, apart from one upsell
13B2B companies: 12 on Salesforce, 1 on HubSpot CRM
1,763independent AI analyses, at least 3 per deal
4.4Btokens read and written, about 4,000 hours of AI agent time
Most attribution models today merge the CRM and marketing automation data, but our research shows that the CRM holds only a fraction of the people, records that are largely a rep's sends and calls nobody answered, and on some deals nothing at all.
On won deals, CRMs only show the buying group as an average of 2.7 people. Upside enhanced data finds 7.9 were actually involved.
Won deals had more people than lost ones at 12 of 13 companies, yet the CRM listed only about a third of those people. Most of the rest were in the CRM but never attached to the deal (45% of everyone involved), and 1 in 5 (21%) had no CRM record at all: Upside found them only in emails, calendars and calls.
Upside enhanced data finds them all by assessing who actually engaged meaningfully during the evaluation, such as in email threads or on calls (more in What enhanced data shows). Without complete buying group visibility, it's hard to know what actually worked to get a company interested in buying a product.
On won deals, the CRM lists about a third of the buying group
Buying group size comparison
The won deals' Upside bar is split by where those people sit in the CRM, as pooled shares (35%, 45% and 21%).
Median deal per company, averaged across the 13 companies by deal count; people Upside matched to the deal from emails, meetings, calls and visits. Source: Upside's AI analysis of the study dataset, October 2026.
Share this figureOn won deals, CRM data only shows an average of 2.7 people in the buying group. Upside enhanced data finds 7.9 were actually involved. From Upside's study of 500+ new business deals.On won deals, CRM data only shows an average of 2.7 people in the buying group. Upside enhanced data finds 7.9 were actually involved. Of everyone involved, 45% were in the CRM but never attached to the deal, and 1 in 5 (21%) had no CRM record at all: Upside found them only in emails, calendars and calls. Figure 4 from The Untracked Half, Upside's study of 500+ new business deals at 13 B2B companies.
How Upside finds the rest of the buying group, on one deal
The CRM lists one contact role on the won deal. Upside picks up the signals in emails, meetings and calls and adds the rest of the buying group.
priya.nadkarni@hooli.comSalesforce · Contact
Nadkarni, PriyaOutreach · Q3 sequence
Priya N.Gong · Call speaker
pnadkarni@gmail.comSalesforce · Lead, duplicate
pnadkarni@hooli.comEmail · 14 threads
Identity resolution joins five disparate records into one person as part of the buying group.
Illustration with a fictional company and people, not study data.
The most meaningful touchpoints are found in unstructured data.
CRM data is full of activity, and little of it changed anything: over an account's life, 68% of what the CRM and marketing automation log has no buyer action at all (sends, unanswered calls and email opens), only logged because they are the easiest thing to track.
On 17% of deals, about 1 in 6, the CRM had nothing logged at all before the deal opened, so first and last touch had nothing to credit. The AI analysis found a meaningful touchpoint before the deal opened on 99% of deals, including nearly all of those with nothing logged.
Records logged per account, touches per deal, and meaningful touchpoints per deal.
Source: Upside's AI analysis of the study dataset, October 2026.
68%of what the CRM and marketing automation log over an account's life has no buyer action at all
17%of deals had nothing logged in the CRM before they opened
In the very early days of Upside, I once spent a full day reading every touchpoint behind one customer's enterprise deal by hand. What activated it was a partner convincing an executive to take a look.
The executive forwarded a note to a manager, who forwarded it on to the team, on top of hundreds of marketing and sales interactions. Some of those were meaningful and built awareness; others went unread into the ether. No rule-based model would have found that chain, and the customer's own team did not even know it had happened.
What fills the records is not what creates the deal.
Marketing email fills the records because it's the easiest touchpoint to track, which is why reports that count touches suggest that sending more email gets more pipeline. Our findings show marketing email does contribute to pipeline, but a lot less than expected.
Share of touches (everything logged on the account, all systems, in the 365 days before the deal opened) vs share of credit under Layer 3 (AI-detected multi-touch, by Upside). Typical company: the median across the 13 companies
Medians across the 13 companies, so the shares differ from the pooled figures elsewhere. Paid ads are not shown: only 7 companies have meaningful ad data (deep dive into paid ads). Source: Upside account activity and Upside's AI analysis of the study dataset, October 2026.
Share this figureAt the typical company, marketing email is 45% of touches in all systems but 3.7% of the credit under AI-detected multi-touch. What fills the records is not what creates the deal. From Upside's study of 500+ new business deals.What fills the records is not what creates the deal. Marketing email fills the records because it is the easiest touchpoint to track: at the typical company it is 45% of everything logged on the account, all systems, in the 365 days before the deal opened, but 3.7% of the credit under AI-detected multi-touch. It does contribute to pipeline, but a lot less than expected. Figure 7 from The Untracked Half, Upside's study of 500+ new business deals at 13 B2B companies.
Not only does Upside uncover touchpoints no system logged, but those touchpoints were on average twice as valuable to the deal as logged ones.
Only 1.3 of the 6.2 meaningful touchpoints on the average deal were never logged anywhere, but each of them earned about 2 times the credit of a logged one. They are the referral mentioned on a call, the champion who forwarded a note, the peer recommendation inside an email thread: most brand and advocacy touchpoints are found this way.
Rule-based models read only what a system logged, so they miss the touchpoints that count most.
Average credit received per meaningful touchpoint by type.
Share of a deal's credit on each kind of touchpoint (51% logged, 25% extracted, averaged over deals) divided by the average number of each per deal (4.9 and 1.3), all 13 companies. Source: Upside's AI analysis of the study dataset, October 2026.
2Layer 2 · Upside enhanced data
What enhanced data shows
Most attribution models today run on CRM and marketing-automation records as they are. Joining every system behind the deals first discovers what the CRM never logged: website visits, meetings and calls, and people never tied to the deal. In our experience most of what this discovers is marketing.
24%of buyer touches come from systems outside the CRM
29%come from people the CRM never tied to the deal
Upside also classifies each touchpoint as a bid (something the company did) or an engagement (something the buyer did), as well as unifying the CRM, email, calendar, call transcripts and web activity in one timeline per account, matching people to accounts and deals (identity resolution) and finding the full buying group.
Under every model, better data ends up giving marketing more credit. On Upside enhanced data, its share of linear multi-touch nearly doubles, from 20% to 37%.
Linear multi-touch, which splits the credit evenly across touches, moves the most, with Sales going from 58% to 39%. Marketing gains under first touch (38% to 45%) and last touch (15% to 25%) too.
Share of all deal credit by team: Layer 1 (CRM data) vs Layer 2 (Upside enhanced data), same model
Everything else: customer success, partners, people with no known team, and deals with no touch to credit. Sales means reps, SDRs and founders. Source: Upside's AI analysis of the study dataset, October 2026.
Share this figureCRM data gives marketing 20% of linear multi-touch credit. On the same deals, Upside enhanced data nearly doubles that, to 37%. From Upside's study of 500+ new business deals.Under linear multi-touch, CRM data gives marketing 20% of the credit. On the same deals, Upside enhanced data nearly doubles that, to 37%, while sales goes from 58% to 39%. Marketing gains under first touch and last touch too, but linear multi-touch moves the most. Figure 9 from The Untracked Half, Upside's study of 500+ new business deals at 13 B2B companies.
When a measure becomes a target, it ceases to be a good measure.Goodhart's law
In our experience, each type of split should be used as a diagnostic, and no model, including Upside's, should be used for goaling. At Branch, which I co-founded before Upside, the single best source of new deals was VIP dinners that marketing organized and SDRs filled with one-to-one invites; goal each team on its share of the credit and the SDR stops inviting prospects to marketing's dinner, because helping might shrink their own number.
CRM data alone inflates the credit attributed to seller email.
Seller email loses the most because 53% of the seller emails the CRM logged were sends nobody answered, which CRM data alone does not classify and account for.
Better data moves seller email from 59% of last-touch credit to about 40%
Last touch credit using CRM data, Upside enhanced data and meaningful touchpoints only
Paid ads are not shown: fewer than 3 companies have ad data that reaches last touch, and no figure in this report rests on fewer than 3 companies (deep dive into paid ads). Source: Upside's AI analysis of the study dataset, October 2026.
3Layer 3 · Meaningful touchpoints only
What really matters
Most attribution models today only credit what a system logged. The third layer of data keeps only the touchpoints that made a difference, and adds touchpoints found in unstructured data: extracted from emails, call transcripts, notes and form fills, and beyond.
Upside's AI analysis read the full history behind each deal, kept only the meaningful touchpoints and inferred the context. What mattered in getting the deal is more spread out than any rule-based model allows: on average deals had 6 meaningful touchpoints that led to activation, where first and last touch credit one. At the typical company, those touchpoints span 4 channel groups and 3 people on the buyer's side, and 83% of deals had 3 or more channel groups among those touchpoints.
What the AI analysis sees on one deal: every share of credit, with the evidence behind it
Logged by the marketing email system: a product announcement email sent to prospects, including this account.
TeamMarketing
ChannelMarketing email
EvidenceTracked
Why this 8% was decided
Logged by the marketing email system: a buyer at the account clicked the November 2025 newsletter.
TeamMarketing
ChannelMarketing email
EvidenceTracked
Why this 12% was decided
Her specific recall of the product demo 8 months later, including details about pilot launch timing, shows this was a distinct high-impact moment that planted the seed for interest in the product.
Linda stated:I remember looking into your AI tool when it was demoed in Orlando, right? But it hasn't been launched. And I understand that you guys launched it with some select pilot client groups in fall and that tool might be... you know, might provide us with a better approach.
TeamMarketing
ChannelEvents & webinars
EvidenceExtracted
Why this 18% was decided
Found in an email thread: the CSM introduced Linda to an SDR for a demo; no system logged the introduction.
TeamCustomer success
ChannelMeetings & calls
EvidenceExtracted
Why this 0% was decided
The buyer's own demo request. It gets 0% because by then the decision to evaluate had already been made.
TeamCustomer success
ChannelWeb, content & forms
EvidenceTrackedResponseHand-raise
Other customer success touches9%
Why this 22% was decided
Found in the deal's unstructured data, not in any system: the rep met Linda at the conference, the meeting the AI analysis marks as the deal's source.
TeamSales
ChannelEvents & webinars
EvidenceExtractedSource
Why this 3% was decided
Logged in the CRM: a follow-up email from the rep two days after the conference.
TeamSales
ChannelSeller email
EvidenceTracked
Why this 21% was decided
Timing, relationships and readiness the AI analysis inferred from the deal's history rather than from any single record.
Only 38% of the credit lands on the same channel group under both
Layer 1 (CRM data, last touch) vs Layer 3 (AI-detected multi-touch, by Upside)
Source: Upside's AI analysis of the study dataset, October 2026.
Share this figureLast touch on CRM data agrees with AI-detected multi-touch on only 38% of credit. An AI agent learning from that data learns the wrong lessons, and acts on them at scale. From Upside's study of 500+ new business deals.Last touch on CRM data agrees with AI-detected multi-touch on only 38% of the credit. Wrong data used to mean a wrong slide. An AI agent learning from it learns the wrong lessons, and acts on them at scale. Figure 12 from The Untracked Half, Upside's study of 500+ new business deals at 13 B2B companies.
With bad data, last touch over-credits seller email and partners, and under-credits external context (timing and relationships), brand and advocacy, meetings and events.
Because seller email fills the CRM, last touch usually lands on a rep's follow-up, whatever the data underneath. AI-detected multi-touch spreads the credit across every meaningful touchpoint, plus external context: the timing, relationships and readiness no record holds, largely beyond a company's control.
No rule-based model gives external context any credit, yet it is a huge part of why and when people buy. In our experience, companies that work at knowing when an account is in market do better.
Most of a rep's last-touch share belongs elsewhere
Difference in share of credit, in percentage points
Layer 1 (CRM data, last touch) vs Layer 3 (AI-detected multi-touch, by Upside). Positive means last touch over-credits the channel group. Lines show 95% intervals across companies.
Paid ads are not shown: fewer than 3 companies have ad data that reaches last touch, and no figure in this report rests on fewer than 3 companies (deep dive into paid ads). Source: Upside's AI analysis of the study dataset, October 2026.
Share this figureOn CRM data, last touch over-credits seller email by 40 percentage points against AI-detected multi-touch and under-credits external context, brand and advocacy, meetings and events. From Upside's study of 500+ new business deals.Seller email fills the CRM, so last touch usually lands on a rep's follow-up. Compared with AI-detected multi-touch, last touch on CRM data over-credits seller email by 40 percentage points and under-credits external context (timing and relationships), brand and advocacy, meetings and events. No rule-based model gives external context any credit, yet it is a huge part of why and when people buy. Figure 13 from The Untracked Half, Upside's study of 500+ new business deals at 13 B2B companies.
Form fills.Pipedash does not credit the form fill itself. A form is the buyer's hand-raise, the sign that something earlier worked, so the credit goes to the activities that led the buyer to fill it in (see why the hand-raise gets zero credit).
Brand and advocacy start 28% of deals. The CRM's opportunity source field only credits them with 5.6%.
The AI-detected source names the one touch that got each deal started: for brand and advocacy, a customer referral, a review site, a community thread or a champion inside the buyer. On another 1.5% of deals no touch started it, and inferred brand awareness leads. These rarely leave a record: 89% of brand and advocacy's credit was extracted or inferred from unstructured data.
Of the CRM's answers, the opportunity source field comes closest, because people fill it in after looking into the deal, yet it names no channel at all on 2 in 3 deals, and still undercounts brand and advocacy where it does. In our experience, leaders cut social, community, review and referral programs first because their ROI is harder to prove, yet these are sometimes the best-performing channels.
Brand and advocacy programs are severely underreported
The lighter segment is the deals with no starting touch whose inferred credit is led by brand awareness: 1.5% of deals, at 5 of 13 companies. Source: Upside's AI analysis of the study dataset, October 2026.
Share this figureThe CRM's opportunity source field credits brand and advocacy with 5.6% of deals. The AI-detected source shows they start 28%. From Upside's study of 500+ new business deals.In CRM data, the opportunity source field credits brand and advocacy with 5.6% of deals. The AI-detected source shows they start 28%. Referrals, review sites, community threads and champions rarely leave a record: 89% of brand and advocacy's credit was extracted or inferred from unstructured data. Figure 14 from The Untracked Half, Upside's study of 500+ new business deals at 13 B2B companies.
Only 28% of companies can directly tie brand investment to pipeline.Carilu Dietrich, 2026 benchmark report, 168 B2B marketing leaders
Carilu Dietrich, who led marketing at Atlassian through its IPO and now advises CEOs and CMOs of high-growth companies, surveyed 168 B2B marketing leaders for a 2026 benchmark report: 73% believe brand makes demand generation more efficient, yet "only 28% of companies can directly tie brand investment to pipeline", so when budgets are cut 55% would fight hardest for demand spend and 11% for brand.
Nearly 1 in 12 deals came from someone who had used the product at a previous company. Traditional models usually credit sales for those deals.
At 10 of the 13 companies, a person on the buyer's side had used the product in a previous job and brought it to the new one. On another 12% of deals, a former user was in the buying group without starting the deal, so about 1 in 5 deals had one.
About half of them were found only in unstructured data, in emails and calls; the rest by matching people across accounts. No opportunity source field value marks a former user, and on CRM data last touch hands these deals to a rep's email even more often than it does the rest. The groups are small, so read these numbers as directional.
The 1 in 12 is a share of all deals, and 39% is how often those deals were won (against 29% with no former user found), so among won deals the share is higher: 26% had a former user on the buyer's side, and 10% were started by one.
The signal is not only for sales and marketing. In our experience, it shows that customer success teams are also doing a great job, and they should get credit for it and have it as a goal.
Track former users as a source, and share them with sales, marketing and customer success
Deals started by a former user of the product
1 in 12deals came from former users switching jobs7.9% of deals
Won deals with a former user on the buyer's side
1 in 4won deals had a former user on the buyer's side26% of won deals
Started the deal: 10%Involved: 16%
Former users were found two ways: Upside's AI analysis of emails and calls, and matching people across accounts. Source: Upside's AI analysis of the study dataset, October 2026.
Share this figureNearly 1 in 12 deals came from someone who had used the product at a previous company. Traditional models usually credit sales for those deals. Small groups, so read as directional. From Upside's study of 500+ new business deals.Nearly 1 in 12 deals came from someone who had used the product at a previous company, and 1 in 4 won deals had a former user on the buyer's side. No opportunity source field value marks a former user, and traditional models usually credit sales. About half of these former users were found only in unstructured data, in emails and calls. The groups are small, so read these numbers as directional. Figure 15 from The Untracked Half, Upside's study of 500+ new business deals at 13 B2B companies.
Unstructured data matters well beyond marketing: even meetings and calls earned a quarter of their credit from conversations no system logged.
Rule-based models today only read what a system logged. Across every channel, Upside's AI analysis extracted25% of all credit from emails, call transcripts and notes, and inferred another 24% from context no touch records, citing evidence for every point of credit.
How each channel group's credit was found: logged, extracted or inferred
What the inferred causes mean
Readiness / buying window Something changed for the buyer that made now the time to buy: most often a regulatory or compliance deadline, then a problem with the current vendor or a new leader.
Prior relationship Someone on the buying side already knew the seller or the product, such as a former customer who changed jobs.
Brand awareness The buyer arrived already knowing the product, and no touch explains how. Shown as the inferred part of the brand and advocacy bar.
Champion The evidence implies someone inside the buyer's company pushed for the deal, but no message or meeting records it. Champions who show up in an email or call count under brand and advocacy.
Not shown: the 4.5% of credit the AI analysis could not trace to a cause. Source: Upside's AI analysis of the study dataset, October 2026.
Share this figureUnstructured data matters well beyond marketing: even meetings and calls earned a quarter of their credit from conversations no system logged as a touch. From Upside's study of 500+ new business deals.Unstructured data matters well beyond marketing. In Upside's AI analysis, even meetings and calls earned a quarter of their credit from conversations no system logged as a touch. Traditional attribution cannot see that credit. Figure 16 from The Untracked Half, Upside's study of 500+ new business deals at 13 B2B companies.
4Lessons
More than just attribution
Creating a trustworthy GTM data layer
Closed won
What a team can do is no longer limited by the people on it, as AI agents take on more and more of the work.
AI learns from what works, and does more of it.
Every campaign, sequence and meeting is a test of what works. In our experience, AI agents learn the way people do: they look at what led to pipeline, repeat it and drop the rest. They do it far more often, on whatever data they are given.
The problem: the data most teams run on isn't even directionally right.
Chapters 1 to 3 showed how far off it is: last touch on CRM data agrees with the AI analysis on only 38% of the credit (see the gap). Wrong data used to mean a wrong slide. An agent learning from it learns the wrong lessons, and acts on them at scale.
What we found critical while building the system behind this study.
Fixing the data turned out to be most of the work. These are the lessons from building Upside's data layer and the AI analysis behind this report, with what the study's data shows, costs and limits included. Most hold for any team building on GTM data, with or without Upside.
Bringing the records and the people together
1
The CRM alone missed much of what buyers did.
We pull email, calendars, call transcripts and web visits into one timeline per account with the CRM. On Upside enhanced data, 24% of buyer touches on the deals came from systems the CRM never logged (see layer 2). Each connection is one more system to keep in sync, and the same meeting can show up in several of them, so we count each touch once per person per event. Recording calls and keeping every thread and note on the account is where any team can start.
2
One buyer was spread across many records, and many buyers had none.
Before counting anything, every email, meeting, call and visit has to be matched to the right person, account and deal, with duplicates merged (identity resolution; how we do it). Of everyone involved in won deals, 21% had no CRM record at all: we found them only in emails, calendars and calls (see where they were). The matching has to run every day to catch new data fragments; our data pipelines run daily.
3
The deal's contact list was the wrong place to find the buying group.
We detect the buying group from everyone on the buyer's side who was actually involved in the buying process, whether or not they're associated with the opportunity. On won deals, the CRM lists 2.7 of the 7.9 people involved (see the gap). The tradeoff: someone who never showed up in an email, meeting, call or visit, and was never added to the deal, is still missing. Their influence shows up only when someone mentions them, or as inferred credit, such as a champion with no message on record.
4
Buyers bring deals with them when they change jobs.
Nearly 1 in 12 deals came from someone who had used the product at a previous company, and no opportunity source field value marks one (see the deals). Our identity resolution follows an email address, and a new job means a new one. Matching people across accounts and the AI analysis of emails and calls each found about half of these deals, and few were found by both. The groups are small, so we read this as directional. Customer success flagging champions who leave customer accounts is a start any team can make.
Sorting the touches and grading the evidence
5
Not all touchpoints matter.
Over an account's life, 68% of what the CRM and marketing automation log has no buyer action at all: sends, unanswered calls and email opens (see the data). We count a touch only when a buyer did something, so seller email stops taking credit for sends nobody answered. Our AI analysis still takes those sends into account: they can quietly build awareness, so our agents reason through the whole journey to see whether that's the case.
6
Not giving credit to hand-raise events.
We sort each touch into what the company did, how the buyer responded and what a third party did. Responses, like a demo request, get 0%: they show that something earlier worked, and the credit goes to the activities a team can do more of. A marketing or sales team can't just "get more demo requests": when they try, they often end up running paid campaigns that bring in spam demos that never convert.
On one of our own deals, the hand-raise was a text to me saying "we need Upside". The AI analysis gave it 0%, and most of the credit to internal championing (47%) and a relationship built years earlier that moved with the champion into a new role (28%).
7
Half the credit was never logged as a touchpoint.
We grade every point of credit as tracked by a system, extracted from unstructured data such as an email, transcript or note, or inferred when no record holds it, so whoever acts on a number can see how it is known. In this study, 49% of the credit was extracted or inferred (Figure 1). Extracted credit can only come from what a company keeps, and inferred credit is the softest grade: its size depends on how much of the deal the records cover.
Checking the AI and agreeing on definitions
8
One AI analyst, asked once, sounded just as sure when it was wrong.
Every deal gets at least three independent AI analysts, each defending a credit split with citations, and a judge who calls in five, then seven, when they disagree too much (how we built it). Another agent rejects any citation to a record that doesn't exist. It took about 4,000 hours of agent time for this study (see the dataset). Anyone using AI to read deals must not trust the reasoning to one agent alone: ask the same question several times, compare the answers and check that every cited record exists.
9
We let the AI analysis say when it could not tell.
The AI-detected source can be "none" when the evidence points to timing or a relationship, and the analysis reports the credit it could not trace to any cause: 4.5% of all credit. It is still a judgment, not tested against an outside truth, and it explains deals after they close, too late for day-to-day campaign decisions (see the limits). No model, limited by the size of B2B datasets, can prove cause, ours included.
10
The same question can get different answers unless the definitions are written down.
For this study, an agent mapped every touch onto each company's own teams and channels, so every deal adds up the same way. The model is a definition too: on the same Upside enhanced data, marketing email gets 1.9% of the credit under last touch and 5.7% under linear multi-touch (compare the models). Upside's librarian applies the same idea: before a person or an agent queries the data, it hands them the company's own definitions and the right way to answer, and it learns from how the team asks. The low-cost version is a written glossary that every dashboard and agent reads.
CRM data, one piece at a time: what a deal's records hold, the opportunity source field teams rely on, the buying group the CRM lists, deals with nothing logged, and why paid ads go missing.
Every deal was read on the three layers of data: CRM data, Upside enhanced data and meaningful touchpoints only.
74%of a deal's CRM records before it opened are a seller's one-to-one email
68%of what the CRM and marketing automation log per account has no buyer action: sends, unanswered calls and email opens
17%of deals had nothing logged on their own contacts before they were created (why it matters)
Only about a third of deals have a usable channel in the opportunity source field; the rest are blank or name only a team
The opportunity source field is usually set after someone investigates what started the deal, or by a sync or web form that records only how the buyer arrived ("Direct Traffic"). So it names a channel on only about a third of deals: 28% are blank or just say "Other" and 38% name only a team. The AI analysis names a source on 79% of deals, because it reads the unstructured data where the answer usually sits: emails, call transcripts and notes. The AI analysis is very good at detecting what started a deal; on most deals the field simply doesn't hold it.
Layer 1 (CRM data, opportunity source field): what the field holds on each deal, share of deals
"Names a channel" counts any specific channel, including partners and market and relationship; a field that only says "Other" counts as blank. Source: Upside's AI analysis of the study dataset, October 2026.
Camille Ricketts remembers all-hands sessions at Notion called HTDWW, How the Deal Was Won, where a salesperson walked through the steps that had landed the business; the walkthrough "relied entirely on their own analysis and intuition, and was seldom systematized", and it hugely undervalued brand marketing because brand was so hard to measure. The source field is that story, written down.
Where people do fill it in, they are right more often for some channels than others. Among deals where the field names a channel and the AI analysis names a single source, the two agree on 37% overall: 65% when the field says events and webinars and 57% when it says brand and advocacy, but 15% for partners and other and 11% for web, content and forms, its most common answer. Counting every deal it labels web, including those where the AI analysis names no single source, the AI-detected source is brand and advocacy on 39%.
When the field and the AI analysis both name a team, they agree on 42% of deals. At 5 of 13 companies the field never once said referral, and the primary campaign, the other field teams lean on, was empty on 56% of deals at the 9 companies that have it.
When people fill in the opportunity source, they're mostly right about events and brand and advocacy, and mostly wrong about web and forms
Layer 1 (CRM data, opportunity source field) vs Layer 3 (AI-detected source, by Upside): share of deals where both name the same channel group
Deals where the field names a channel and the AI analysis names a single source, by the channel the field names. Channels the field names at fewer than 3 companies are not shown. Source: Upside's AI analysis of the study dataset, October 2026.
When it does name a channel, the opportunity source field still undercounts brand and advocacy: 16% of the deals it labels, against 34% for the AI-detected source
Web includes demo and contact forms, "Direct Traffic" and "Organic Search". Groups fewer than 3 companies name are folded into Partners & other. A row with one bar: the other source never names that group, or fewer than 3 companies do. Source: Upside's AI analysis of the study dataset, October 2026.
Buying group by deal outcome
On won deals, 35% of the people Upside found were contact roles on the deal, 45% were in the CRM but not attached to the deal, and 21% had no CRM record at all (11% at the typical company). Most of the people in the CRM but not on the deal got their CRM record only after it opened.
Before the deal opened the gap is different: of the people active on won deals by then, 52% were contact roles, 43% were in the CRM but not on the deal, and only 5% had no CRM record at all. Most people with no CRM record joined the conversation after the deal opened.
On won deals, the CRM lists about a third of the buying group
Buyer-side people count if Upside matched them to the deal from their emails, meetings, calls or visits, before or after it opened. Source: Upside's AI analysis of the study dataset, October 2026.
Former users: how they were found and what the records say
Upside found former users two ways: its AI analysis of emails and calls, and a match of people across accounts. Each found about half of the deals a former user started, and only 10% were found by both, because identity resolution follows an email address, and a person who changes jobs gets a new one. Deals a former user started were won 39% of the time and deals a former user joined later 45%, against 29% with none. The gap comes from the matches across accounts; on the AI analysis alone, deals a former user started were won about as often as the rest. The AI-detected source names brand and advocacy on 34% of the deals a former user started, mostly internal champions and referrals. On CRM data, last touch credits seller email on 73% of the deals a former user started, against 58% of all other deals.
Deals with nothing logged before they opened, by layer
Before a deal opened, the CRM had nothing logged on 17% of deals and only the company's own outreach, with no buyer action, on 8%; 75% had at least one buyer action. On enhanced data, 17% still had nothing. The AI analysis found a meaningful touchpoint before the deal opened on 99% of deals.
Better plumbing alone doesn't find the missing start of a deal
Layer 1 (CRM data), Layer 2 (Upside enhanced data) and Layer 3 (Meaningful touchpoints only): share of deals with no known touch before they were created
Layer 1 counts any CRM or marketing-automation record, one-way sends included; layer 2 counts only buyer actions, so the deals with nothing logged are not the same on both layers. Layer 3 is the share where the AI analysis found no credited touch. Source: Upside's AI analysis of the study dataset, October 2026.
Paid ads
Most of what the ad platforms record never reaches the CRM.
Among the 7 companies with a meaningful paid-media program, most of what the ad platforms record never reaches the CRM. LinkedIn engagement, for example, arrives as a monthly rollup per account and campaign, stored as one outbound touch with no engagement: a hundred impressions and four clicks look like nothing happened.
83%of ad impressions never matched an account in the CRM (median of the 7; 50 to 91%)
93%of ad clicks never reached the CRM as a person or an account (85 to 98%)
0.9%of deal credit went to paid ads under AI-detected multi-touch, and under 1% under first touch, last touch, linear multi-touch and W-shaped on CRM and enhanced data
Where paid campaigns are synced into CRM campaign membership, paid ads do show up as where deals started.
Part 2Layer 2 · Upside enhanced dataaccount and deal
What enhanced data changes
Layer 1 against layer 2: any two models and layers side by side, marketing's share under every model, what each model answers, and every model on every layer.
130buyer actions per account over its lifetime, across all systems
24%of buyer touches on the deals come from systems the CRM never logged
29%come from people the CRM never tied to the deal
Every system joined in one place and people matched to accounts and deals (how Upside cleans and enhances data). Packaged tools such as Salesforce Campaign Influence and Marketo Measure credit only the contacts a rep attached to the opportunity, so they miss most of these people (the buying group).
Compare any two models and layers of data, by team or by channel
Comparison by layer and model
Share of credit
By
Rule-based models on layer 3 run on the meaningful touchpoints the AI analysis found; the rules every chart follows are in How to read the charts. Source: Upside's AI analysis of the study dataset, October 2026.
Marketing's credit
With better data, the same models give marketing far more of the credit.
On Upside enhanced data, with every system joined and people matched to accounts, marketing's share goes up under every model we ran on both it and CRM data.
AI-detected multi-touch gives marketing less than any model on enhanced data, because it looks beyond teams alone and adds external inferred influences that on average end up being 21% of the credit.
On CRM data alone, marketing gets a lot less credit.
On enhanced data, marketing's share rises under every rule-based model
Marketing's share of all credit: Layer 1 and 2 under last touch, linear multi-touch and first touch, then Layer 3 (AI-detected multi-touch, by Upside) with and without external context
The external context excluded bar is a view of the same AI analysis with the external context left out and the rest of the credit re-shared across teams. Source: Upside's AI analysis of the study dataset, October 2026.
The model still matters
Even with the best data, each model answers a different question.
Better data gives marketing more credit, but it does not make the models agree. On Upside enhanced data, first touch gives marketing 45% of the credit and last touch25%. Each model answers a different question.
The AI analysis itself gives two answers. AI-detected sourcenames the one touch that got the deal started, or declines when the evidence does not point to one. AI-detected multi-touch splits the deal's credit across every meaningful touchpoint, including external context.
Five models, five questions
Model
The question it answers
Credited touch days before, median
First touch
What first brought this account to us? It credits awareness and early interest.
253 days Layer 1 (CRM data)
Last touch
What happened right before the deal opened? It credits conversion.
6 days Layer 1 (CRM data)
Linear multi-touch
What did the buyer touch along the way? Every touch counts the same.
every touch
AI-detected source
If we could pick only one, what got this deal started?
30 days Layer 3 (Meaningful touchpoints only)
AI-detected multi-touch
How much did each thing matter, including external context?
every touch
Source: Upside's AI analysis of the study dataset, October 2026.
First touch often lands a year or more before a deal opens. Last touch happens in the final week. The AI analysis finds most deals actually started in the three months before the deal was opened
Layer 1 or 2 (first and last touch, select below) vs Layer 3 (AI-detected source, by Upside): share of deals by how long before the deal was created the credited touch happened
"No touch to credit" for the AI-detected source: the deal started from external context, with no single touch to name. Source: Upside's AI analysis of the study dataset, October 2026.
First touch answers when someone first entered the database, usually months before anyone was buying. Last touch lands on whatever was logged in the final week: usually a rep's email, sometimes the meeting that opened the deal. Both are real questions, and they leave a third one open: what started this deal?
The AI-detected source answers that one question, and it is still just one model. Its multi-touch credit answers another. On the same layer 3 data, brand and advocacy get 28% under the AI-detected source but 6% under last touch, and events and webinars get 15% under first touch but 3.3% under last touch.
Every model on every layer, one team at a time
Every model on every layer: a team's share of all credit
Each cell is one layer and one model: one team's share of all credit
Team
Part 3Layer 3 · Meaningful touchpoints onlyper deal
What really matters
Any model on any layer against AI-detected multi-touch. For how marketing email's volume compares with its credit, see what fills the records in chapter 1.
6meaningful touchpoints on the typical deal, 97% of them dated before it opened
14%of deals have brand and advocacy as their first touch once first touch runs on meaningful touchpoints only (the AI-detected source says they started 28%), against under 1% on the enhanced data
The rule-based models run on layer 3 too. Touchpoints extracted from unstructured data are fewer but weigh more: they carry 25% of the credit against 51% for the tracked ones, about 2 times as much per touch (1.3 of the 6.2 meaningful touchpoints on the average deal).
What each model over- and under-credits, by channel, against AI-detected multi-touch
Every rule-based model credits what gets logged over what gets said
Difference in share of credit, in percentage points: the selected model and layer vs Layer 3 (AI-detected multi-touch, by Upside). Positive means the model over-credits the channel group. Lines show 95% intervals across companies.
A channel fewer than 3 companies feed on the model's side counts as 0 there, so its credit shows as missed. Source: Upside's AI analysis of the study dataset, October 2026.
One model on all three layers of data, next to the AI analysis
The same model, layer by layer
One model (select below) on Layer 1 (CRM data), Layer 2 (Upside enhanced data) and Layer 3 (Meaningful touchpoints only), next to Layer 3 (AI-detected multi-touch, by Upside): share of all credit by channel group or team
Model
By
Part 4Method
Dataset, method and limits
How we measured, how to read the charts, how Upside's AI analysis reads a deal, one deal as an example, the dataset, the method notes and the limits.
How we measured
We ran the models teams use on three layers of data, and Upside's AI analysis read every deal.
The study covers every new business deal with a completed AI analysis at 13 B2B companies that use Upside, each read and analyzed at least 3 times by AI agents. Each company's data was anonymized and aggregated on its own before anything was pooled.
Every deal counts once and is worth the same whatever its size, under every layer and model, and every layer and model runs on the same set of deals. Credit is a share of all deal credit.
Bars need not add up to 100%. The rest sits with teams or channels the chart leaves out, or on deals with no touch to credit: on about 17% of deals, on CRM and enhanced data alike, nothing was logged before the deal opened. Team charts show sales, marketing and, on layer 3, third parties and external context; customer success, partnerships and internal people with no known team are left out.
A missing bar means fewer than 3 companies feed it. No figure in this report rests on fewer than 3 companies. "Under 1%" marks a share that is real but below 1%, or fed by too few companies to print.
Timing. First touch, last touch and the multi-touch rules count only touches before the deal was created, with no lookback limit. AI-detected multi-touch can credit a touch dated after; 3% of the touchpoints it credits are.
Brand awareness is inferred and counts as marketing in team charts and under brand and advocacy in channel charts. The rest of the inferred credit stays apart as external context.
How the AI analysis works
How Upside's AI analysis reads a deal
Rule-based models weigh touches by when they happened. Pipedash reads what happened, the way an analyst would, and has to defend every point of credit with evidence.
Unify the data. Upside pulls the CRM, email, calendar, call transcripts and web activity into one timeline per account, merges duplicate people and records and finds the buying group. This is layer 2.
Sort every touch. Each action is something the company did (an email, an event, a post), something the buyer did in response (a form fill, a reply, showing up to a meeting), or something a third party did about the company (a referral, a review, a mention). Responses get 0%: they are the sign that something earlier worked. That is why a demo request form earns nothing.
Grade the evidence.Tracked touches have a record. Extracted touches exist only in unstructured data and are cited from emails, transcripts and notes, like "great meeting you for coffee" in a thread; they appear from layer 3 on. Credit that can't be tied to any touch is inferred and goes to a named cause: a buying window, a prior relationship, brand awareness, a champion. All but brand awareness is external context. Its size depends on how much of the deal the records cover.
Three+ analysts, one judge. Three independent AI analysts read the full deal history and each defends a credit split with citations to records, transcript quotes and emails. A judge decides on the strength of their reasoning and calls in more analysts, five and then seven, when they disagree too much. One more agent rejects any citation to a record that doesn't exist; another maps every touch onto the company's own teams and channels so deals add up.
Two answers per deal. A split of 100% of the credit across everything that mattered (AI-detected multi-touch), and the one touch whose removal would have broken the chain (AI-detected source). The buyer's hand-raise is marked, with no credit.
The method is slow and costly per deal (the dataset card has the counts), needs the unified data underneath, and explains deals after they close, too late for day-to-day campaign decisions.
One deal, two answers: the form that got all the credit earned none of it
A deal described by Camille Ricketts: last touch on the CRM's own records vs AI-detected multi-touch, by Upside, as share of the deal's credit
Last touch
Demo request form100%
The last thing logged before the deal opened gets everything.
AI-detected multi-touch
Webinar program a champion joined months earlier30%
Demo request form0%
Everything else on the deal, not itemized in the post70%
The form is the buyer raising a hand. The credit goes to what made them ready.
Camille's post also lists touches a CRM rarely records: a dinner that changed a buyer's mind, an exec-to-exec intro logged only as "referral", a teardown emailed to champions, a peer's recommendation. Source: Camille Ricketts, "Finally a better way to measure and manage ABM", The Upside Down Funnel, September 15, 2026.
The dataset
521new business deals only; no renewals or expansions, apart from one upsell
13B2B companies: 12 on Salesforce, 1 on HubSpot CRM
64 daysfrom creation to close on a typical closed deal, 107 on average
28%won, 60% lost, 12% still open
6.2meaningful touchpoints on the average deal (median 6): about 4.9 tracked and 1.3 extracted from unstructured data
$81Kmedian deal size across the 475 deals with an amount (middle half $50K to $156K). The typical company's average deal was $126K. Amounts are Upside's normalized deal value in USD; 11 deals were booked in AUD or GBP and converted
3 or moreindependent AI analyses of every deal, 1,763 in all
4,000hours of AI agent time across the dataset
4.4Btokens read and written for the analyses in the study, about 8.4 million per deal (7.1B including earlier and repeat runs on the same deals)
Companies
13 B2B companies that use Upside
Deals
A sample: the new business deals with a completed AI analysis at those companies, no cap, plus one upsell that was analyzed too. Someone chose to analyze them, so they are a convenience sample (see Caveats). 86% were created in the 24 months to October 2026, most of them in 2026; the rest earlier
Scope
The AI analysis can credit touchpoints before or after the deal was created (97% of the touchpoints it credits are dated before). First touch, last touch, linear multi-touch and W-shaped count only touches before creation. The question throughout is what started the deal, not what won it
Account view
Separate from the deal analysis. Layers 1 and 2 count everything ever logged on the accounts behind the deals, up to October 8, 2026 (at the median company the records start in 2019; 11 of 13 reach back to 2021 or earlier).
Unit
Every deal is worth the same (100% of its own credit) under every layer and model. Credit is a percentage of all deal credit, with every deal counted once
Tools
Upside's AI agents read and credited each deal. The pooled analysis ran in Claude and Codex, and Claude handled the final revisions
Privacy
Each company's data was anonymized and aggregated separately, then pooled. No names and no figure tied to an identifiable company. The only dollar figures are the deal-size median and middle half across all deals and the typical company's average
Does deal size or cycle length change the answer?
Deal size made little difference in this sample. Small, medium and large deals split their credit almost identically between tracked, extracted and inferred evidence.
Deal length changes who is involved; the credit split holds. Across all deals, buying groups grow from 1.9 people on the shortest deals to 8.2 on the longest (the shortest group spans 12 companies, the longest 11). The fastest deals were the most likely to start in a conversation: 50% of the shortest, against 33% of the longest.
Across all deals, inferred credit looks higher on long deals, 30% against 22% on the shortest. That is company mix: within the typical company it stays flat.
The cohort: most deals close within a quarter, a long tail runs past a year, and won deals are not systematically bigger
Most deals close within a quarter; a long tail runs past a year
Closed opportunities by cycle length and by size relative to the company's median deal
Days from creation to close
Amount ÷ the company's median deal
Closed deals only; open deals come from 4 companies. The cycle chart leaves out deals closed before their creation date; the size chart leaves out 46 deals without an amount. Source: Upside's AI analysis of the study dataset, October 2026.
Table view
Won and lost deals have the same median size and differ only among the very largest deals. Won deals take longer to close, 82 versus 60 days at the median. Half of all deals sit between roughly three-quarters and one-and-a-half times their company's median deal; 17% are more than twice it and 5% more than five times it.
Deal size does not change where the credit goes. Across small, medium and large deals (thirds of each company's deals by amount), inferred credit is 25% / 26% / 24% of all credit, extracted credit is 26% / 25% / 22%, and third parties hold 14% / 13% / 12%.
Method notes: views, models, lookback, pooling and anonymity
Views of the data
Account view (layers 1 and 2 in the drawing, and the account tiles on the layer 1 and 2 cards). Layers 1 and 2 count every touch ever logged on the accounts behind the deals, up to October 8, 2026. Layer 1 counts CRM and marketing-automation record types (emails, meetings, calls, campaign memberships, marketing email). A buyer action is a click, reply, attended meeting, connected call, visit or form fill, never an open. Figures are the typical company's (median of 13). It never feeds a credit figure.
Layer 1: CRM data, per deal. People: the CRM's Opportunity Contact Roles. Activities: CRM-native tasks, events and campaign members, every logged record regardless of direction. This is an activity-timeline rule over the CRM's own records, what a Salesforce activity report or a home-built SQL model over CRM objects would see; no packaged attribution product counts an unanswered send.
Layer 2: Upside enhanced data, per deal. People: contact roles plus the buying group Upside detects from activity (employees are never buying-group members). Activities: every system Upside unifies, filtered to buyer engagement, counted once per person per event (how Upside cleans data). Closest to Campaign Influence, Marketo Measure or HubSpot attribution run on a complete activity graph.
Layer 3: Meaningful touchpoints only, per deal. The touch inventory from the latest AI analysis: logged and extracted touches with per-touch credit, plus the inferred credit with its causes and evidence.
Models
Only touches dated before the opportunity was created count. First touch: earliest eligible touch. Last touch: latest. Linear: even split. W-shaped: 30 / 30 / 30 on first, hand-raise and last, 10 over the rest, renormalized when a position is missing; when two positions fall on the same touch (a hand-raise that is also the first touch), it receives both weights, as in HubSpot and Marketo Measure. Hand-raise (Layers 1 and 2): first buyer-initiated inbound action by rule; Layer 3 uses the AI-detected marker. Opportunity source and primary campaign: the CRM fields. AI-detected source: 100% of the deal's credit to the touch marked as what started the deal; the AI analysis declines to name one on 21% of deals. AI-detected multi-touch: the per-touch credit from the AI analysis including the inferred credit. Same-timestamp ties split credit evenly. A deal with no eligible touch puts its 100% on "no touch to credit", so every layer × model column sums to 100%. U-shaped, time-decay and full-path models, which HubSpot Enterprise and Marketo Measure shops also run, were not computed; they sit between linear and W-shaped in how they treat position.
Lookback. The headline in this paper is an unlimited window; 365-day and 90-day variants exist for first touch, last touch and linear at every layer. Limit the model to the last 90 days and 25% of deals have no CRM touch at all; only 4% lack a meaningful touchpoint. Per-company reports use 365 days as the headline; the pooled paper uses unlimited because several 365-day cells for other models are single-company cells. In common tools: Campaign Influence has an auto-association time frame with no default; Marketo Measure has no attribution window (touchpoints are kept for the life of the contact, and its 90 days is an import limit for ads and forms); HubSpot's attribution reports have no window either (the 90-day lookback people cite belongs to its ads attribution). So the unlimited headline is what HubSpot and Marketo Measure actually compute, and the 90-day variant approximates a Salesforce org with a 90-day time frame. The Layer 2 hand-raise cell, pooled over 12 companies, agrees with the AI-detected hand-raise marker on the channel on 57% of deals.
Inferred credit and external context
Both describe the same credit. By kind of evidence, 24% of all credit is inferred, and that is the figure in the hero and the main read. In every team and channel view the 3.1% of it that is brand awareness is counted with marketing and with brand and advocacy, so the external context kept apart is 21% of all credit, and that figure includes the 4.5% the AI analysis could not trace to a cause.
External context excluded
Upside's product also offers an "all assigned" view of the same analysis, shown in this report as "external context excluded". Brand awareness goes to marketing and champion relays to third parties; the rest of each deal's external context (buying window, prior relationship and causes it could not trace) is spread over the teams that already hold credit on that deal, in proportion to that credit, so every point lands on a team. Pooled, the 21% kept apart as external context falls to under 1% of team credit. The view is off by default in the product, so AI-detected multi-touch, with that external context kept apart, stays the reference everywhere else in this report.
Pooling and anonymity
Every pooled percentage is Σ(pct × deals) / Σ(deals) over the companies whose export contains that cell, so a company is weighted by its deal count and never by revenue. A channel group is shown when any channel in it is fed by 3 or more companies (the fewer-than-3 rule is in How to read the charts). Every model charted here was run for all 13 companies; the one figure pooled over 12, the layer 2 hand-raise agreement, says so where it appears above. Each company's team and channel vocabulary is mapped into one frozen taxonomy and the mapping decisions are written down per company. The exports carry no names, IDs, emails or amounts. Every outside source cited in this report was fetched and checked against the quoted figure before publication.
Caveats and methodology limitations
These deals are not a complete sample of the market. They are the deals with a completed AI analysis at 13 companies that use Upside, which chose to buy an intelligence layer product. Comparisons between models still hold, but no channel mix here is any company's real mix.
The AI-detected source and AI-detected multi-touch are judgments by the AI analysis. Every percent cites evidence readers can check, and every deal had at least three independent analyses, but we did not test it against an outside truth.
No model can prove cause, the AI analysis included. Sales calls the accounts most likely to buy, and nobody holds out a control group. At eBay, a model said non-brand paid search returned over 4,100%; a randomized test put it at minus 63%.
The account views (lifetime, for layers 1 and 2) and the per-deal numbers cover different windows and people. Account counts also include activity after the deal was created, from every contact on the account; deal counts mostly stop at creation (97% of meaningful touchpoints are dated before it) but reach back without limit. Read them side by side for scale, never as a share.
At the typical company, 55% of the credit was extracted or inferred from unstructured data, above the 49% total, and 11 of 13 companies had 40% or more.
Ad data in this sample group might not be representative of the broader market. Among the 7 companies with a meaningful paid-media program, a median 83% of impressions never matched an account and 93% of clicks never reached the CRM. At one, the ad-click feed stopped in late 2025 while spend continued; at another, only a handful of person-level clicks appear against hundreds of thousands on the ad platform. Paid ads' small share of credit partly reflects that gap.
External context and inferred touchpoints are a reading of the evidence by the AI analysis, and 4.5% of all credit could not be traced to any cause.
A touch with a system record: a logged email, meeting, call, campaign membership, page view or form.
Extracted
Found in unstructured data but never logged as a touchpoint of its own, e.g. a referral mentioned on a call.
Unstructured data
Emails, call transcripts, notes and form fills: text that holds evidence no system logs as a touchpoint.
Inferred
Implied by evidence when no record holds it, e.g. a buying window or a prior relationship. The credit is external context or brand awareness.
External context
Factors behind a deal that no touch records, largely beyond a company's control, e.g. a buying window, a compliance deadline, a prior relationship or a champion with no message on record. It is the inferred credit minus brand awareness: 21% of all credit, against 24% inferred (method notes). Only AI-detected multi-touch credits it.
External context excluded
A view of AI-detected multi-touch with the external context left out: nearly every point of it is re-shared across the teams, in proportion to the credit each team already holds on the deal. Used only for team shares (method notes).
The activities, campaigns and contacts logged in Salesforce or HubSpot and the marketing automation tied to them, as they are, scattered across systems.
Upside enhanced data (layer 2)
CRM data with every connected system joined in one place by Upside: duplicates merged, people matched to accounts and deals, keeping only touches where a buyer did something (how Upside cleans data).
Meaningful touchpoints only (layer 3)
All the touchpoints that made a difference: tracked, or extracted and inferred by Upside's AI analysis from unstructured data. The rule-based models run on it too. It is the AI analysis's reading of the evidence (limits).
Identity resolution
Matching the emails, meetings, calls and web visits in every system to the right person, account and deal, so activity the CRM never tied to a deal counts.
Buying group
The people on the buyer's side who acted on the deal, before or after it opened, detected from their emails, meetings, calls and visits. Contact roles are the subset a rep linked in the CRM. Found in Upside enhanced data.
Models and credit
Credit
The share of each deal a model hands to a channel or team. Every deal counts as 100%, whatever its size.
Deal-weighted
Every deal is worth the same, regardless of size, so no large deal dominates.
Typical company
The median across the 13 companies, each company counted once. Pooled figures count every deal once instead, so the two can differ.
First touch
Gives 100% of a deal's credit to the earliest logged touch before the deal was created, usually whatever first put the account in the database.
Last touch
Gives 100% of a deal's credit to the latest logged touch before the deal was created, often a rep's follow-up email.
Linear multi-touch
Splits a deal's credit evenly across every touch before it opened.
W-shaped
Gives 30% each to the first touch, the hand-raise and the last touch before the deal was created, and splits the remaining 10% across every other touch.
Opportunity source field
The CRM's own answer to what started a deal, held in the field some CRMs call lead source and usually entered by hand or by a sync or web form. Most companies keep only a team-level value ("Outbound", "Marketing"); on two deals in three it holds no channel.
AI-detected source
The single-touch answer from the AI analysis: the one touch that got the deal started, or none when the evidence points to external context instead. "Started" in this report always means this source.
AI-detected multi-touch
The AI analysis splits 100% of a deal's credit across every meaningful touchpoint, including external context. It is the reference this report compares every other model with.
No touch to credit
Credit a model can't give to any team or channel because it finds nothing on the deal to credit. For first touch, last touch, linear multi-touch and W-shaped on CRM data or Upside enhanced data: no touch logged before the deal opened. For the opportunity source field: a field that is blank or names no team. For the AI-detected source: a deal that started from external context, with no single touch to name. For the hand-raise: a deal where the buyer never visibly raised a hand.
Hand-raise
The buyer's first visible signal of intent: a form, a reply, an inbound call. The sign that something earlier worked, which is why it gets no credit.
Lookback
How far before a deal was created a model looks for touches. This report uses no limit; 365 and 90 days are in the method notes.
Teams and channel groups
Sales
The company's sales reps, SDRs and founders. Team charts show their credit together.
Third parties
People other than the seller's own teams who helped start a deal: customers who refer, review sites and marketplaces, communities, and champions inside the buyer's company.
Brand & advocacy
Demand built by a company's brand and by the people who vouch for it: customer referrals and recommendations, review sites and marketplaces, community and social mentions, and internal champions at the buyer. Under AI-detected multi-touch it also includes inferred brand awareness.
Seller email
A rep's own one-to-one email to the buyer.
Meetings & calls
Meetings, phone calls, and sales conversations no one logged that the AI analysis found in unstructured data.
Events & webinars
In-person and virtual events and webinars.
Marketing email
One-to-many email from marketing automation. A click counts as a buyer action; an open does not.
Web, content & forms
Website visits, chat, form fills and content such as guides and posts.
Paid ads
Paid media such as LinkedIn, search and display ads.
Partners & other
Partner-sourced touches and the few touches that fit no other group.
References
Cite this paper. Seghete, M. (2026). Traditional attribution models can't see half of what really creates pipeline. Upside Research.
Numbered in order of first citation, then further reading.
Matt Heinz, "A postmortem for the B2B attribution software category", May 2026. mattheinz.substack.com
Jeff Pedowitz, "Pipeline isn't revenue: the accountability gap killing marketing", The Pedowitz Group, August 2025. pedowitzgroup.com
Randall A. Lewis and Justin M. Rao, "The unfavorable economics of measuring the returns to advertising", Quarterly Journal of Economics 130(4), 2015, 1941–1973. econpapers.repec.org
Brett R. Gordon, Florian Zettelmeyer, Neha Bhargava and Dan Chapsky, "A comparison of approaches to advertising measurement: evidence from big field experiments at Facebook", Marketing Science 38(2), 2019. SSRN 3033144; the factor-of-three finding is on pp. 1–2 of the MSI working paper (18-113)
Brett R. Gordon, Robert Moakler and Florian Zettelmeyer, "Close enough? A large-scale exploration of non-experimental approaches to advertising measurement", Marketing Science 42(4), 2023; arXiv 2201.07055. arxiv.org/abs/2201.07055
Thomas Blake, Chris Nosko and Steven Tadelis, "Consumer heterogeneity and paid search effectiveness: a large-scale field experiment", Econometrica 83(1), 2015, 155–174. Econometrica; the comparison of a 4,100% estimated return with an actual 63% loss is on p. 5 of the working paper
Brian Dalessandro, Claudia Perlich, Ori Stitelman and Foster Provost, "Causally motivated attribution for online advertising", ADKDD 2012. ACM DL
Ron Berman, "Beyond the last touch: attribution in online advertising", Marketing Science 37(5), 2018. SSRN 2384211
Refine Labs, "Refine Labs study confirms measurement gap in software-based attribution", April 2023. refinelabs.com
Refine Labs, "The attribution mirage", October 2024 (the same 620-conversion sample as the 2023 study). refinelabs.com
Upside, "How we built Pipedash: an attribution engine that argues with itself", July 2026. upside.tech
Mada Seghete, "Running our Pipedash attribution model on our own GTM data", Upside, August 2026. upside.tech
Upside, "Four ways Upside cleans your data out of the box", August 20, 2025. upside.tech
Gartner, "Win More B2B Sales Deals" (ed. Brent Adamson), 2018, reporting the 2017 Gartner Digital B2B Buyer Survey. gartner.com (PDF)
Katie Fabiszak and Kaitlyn Bretagne, "Three realities about B2B buying networks", Forrester, February 2026, citing Forrester's Buyers' Journey Survey, 2025. forrester.com
Gartner, "Gartner sales survey finds 74% of B2B buyer teams demonstrate 'unhealthy conflict' during the decision process", press release, May 7, 2025; 632 B2B buyers surveyed August to September 2024. gartner.com
Gartner, "Gartner B2B Buying Report: how to adapt your sales and marketing strategies to the current state of B2B buying", 2023, drawing on the 2022 Gartner B2B Buyer Survey. gartner.com (PDF)
Philipp Stelzer, "Only 24-52% of sales activity reaches Salesforce when logging is manual", Weflow, September 2026; observed capture rates across the Salesforce orgs Weflow connects to, no sample size stated. weflow.ai
Dreamdata, "Dreamdata LinkedIn Ads Benchmarks Report 2026", press release, March 10, 2026; 3.5 million B2B customer journeys, 66 million sessions. prnewswire.com
HockeyStack Labs, "HockeyStack's Self-Reported Attribution Report 2024", 8,528 demo-form responses from dozens of B2B companies. hockeystack.com
Forrester, "The State of Business Buying: companies still struggle to meet buyer expectations", 2024. forrester.com
Validity, "State of CRM Data Management in 2025", July 2025 (602 CRM users and administrators). prnewswire.com
6sense, "The B2B Buyer Experience Report for 2025", November 2025. 6sense.com
6sense, "The timeline for influencing B2B buyers is shrinking: insights from 6sense's 2025 Buyer Experience Report", press release, November 12, 2025; more than 4,000 buyers. businesswire.com
6sense, "B2B buyers are even less of a blank slate than we thought", 2025. 6sense.com
Belkins, "What are B2B cold email response rates? (2026 study)", 7.5 million cold emails sent in 2025. belkins.io
Camille Ricketts, "Finally a better way to measure and manage ABM", The Upside Down Funnel, September 15, 2026. aigtm.engineer
Carilu Dietrich, Alex Bauer and Mada Seghete, "How AI is changing marketing attribution", Hypergrowth Leadership, January 28, 2026. carilu.com
Carilu Dietrich, "Brand vs. demand isn't a budget fight. It's a measurement crisis", Hypergrowth Leadership, January 7, 2026; 168 B2B marketing leaders surveyed with BenchMarkit, Jon Miller and Bill Macaitis for the 2026 B2B Brand vs. Demand Marketing Benchmark Report. carilu.com
Mada Seghete, "Attribution didn't just die. It was murdered by your comp plan", The Upside Down Funnel (formerly The AI GTM Engineer), July 2026. aigtm.engineer
Version 5, October 16, 2026: former-user finding, ten rules, new cross-checks of the opportunity source field and of deals with nothing logged. Earlier editions: October 5, 6, 8, 12 and 14, 2026. Corrections will be listed here with their date.
