Why B2B marketing attribution is degrading, and what to measure instead

A dashboard can allocate 100% of conversion credit while observing only a fraction of the buying process. The arithmetic is complete because the attribution rule distributed all the credit it received. It says nothing about the colleague who suggested the supplier, the AI answer that put it on the shortlist or what would have happened without the campaign.

Attribution records a rule applied to visible interactions. It is often useful. It is also routinely asked to explain why a deal happened, which it cannot do on its own.

Some of the missing signal can be modelled, and parts of AI discovery are now measurable. Google Analytics models certain unobserved key events when its quality thresholds are met, ChatGPT adds a source parameter to outbound publisher links, and Google has begun testing dedicated reporting for impressions in generative AI search features. There is still no complete chain joining every influence to every opportunity.

A credible B2B report keeps three views separate: recorded activity, estimated incremental effect and buyer-reported influence.

Three claims hiding in one chart

Google Analytics defines attribution as assigning credit to ads, clicks and other factors along a user path. Assigning is the important word. The model applies a rule to available events. It does not establish what would have happened without the marketing activity.

Consider a buyer who hears a supplier discussed by a colleague, sees its name in an AI answer, searches for the company and later completes a form. Web analytics may record the final search and form submission accurately. A last-click rule can reasonably say that organic search receives credit for the recorded conversion. It cannot conclude that organic search created the demand, because the model never observed the colleague’s recommendation, the AI answer or a comparable buyer who did not receive those influences.

Google makes the same distinction in its measurement documentation. Standard attributed conversions follow the campaign’s tracking settings and credit rules. Conversion Lift compares treatment and control groups to estimate conversions that would not otherwise have occurred. Reporting both under the same label hides what each one means.

Where the visible path gets thinner

The size of the gap varies with browser mix, consent implementation, sales cycle, channels and the amount of research that happens outside company-controlled systems.

Browser and consent controls change what can be joined

Safari’s tracking prevention blocks third-party cookies by default and reduces third-party referrers to their origin. Firefox isolates third-party cookies and blocks known trackers through Enhanced Tracking Protection. These protections make cross-site identity and detailed path reconstruction less complete, although they do not make all first-party website measurement disappear.

Consent adds another boundary. With basic Google consent mode, for example, Google tags do not load before consent and send no data when consent is denied. With advanced consent mode, cookieless signals may support advertiser-specific modeling. Google’s consent documentation is explicit that the observed and modeled data depend on which implementation is used.

Separate directly recorded events from platform-modelled events. Combining them may produce a workable planning number; calling the total fully observed gives it false certainty.

Buyers use sources your analytics system does not own

A B2B purchase is rarely contained within one browser history. Buyers move among supplier websites, search, video calls, in-person conversations, colleagues and third-party experts. McKinsey’s 2026 Global B2B Pulse, based on nearly 4,000 decision-makers, found an average of ten channels across the buying journey. Gartner separately reported that buyers used an average of seven information sources in a recent purchase.

A private message may prompt a direct visit. An analyst conversation may prompt a branded search, and a webinar recording can circulate internally without its original link. Grouping all of that under “dark social” gives the gap a catchy name without turning it into one measurable channel. Unobserved influence is a plainer description of what the evidence lacks.

AI discovery is no longer a complete blind spot

AI discovery is not entirely invisible in 2026. A recommendation can still lead to a branded search with no trace of the original answer, but impression and referral data now exist in some products.

Google announced dedicated generative AI performance reports in Search Console in June 2026. The reports show impressions, pages, countries, devices and dates for appearances in AI features on Search and Discover. The initial rollout is limited to a subset of websites, so this is not yet a universal source of data.

OpenAI tells publishers that links from ChatGPT automatically include utm_source=chatgpt.com, which means a clicked source link can appear in web analytics. ChatGPT Search also uses factors such as location and, when enabled, relevant memory to improve results, according to OpenAI’s search documentation. A manual citation audit therefore needs controlled sessions and a recorded test context if the team wants to compare results over time.

An impression in a Google AI feature, a clicked ChatGPT source and a brand mention in a controlled test are different events. None proves that the appearance caused an opportunity.

Citation visibility can show whether the company appears, which claims are repeated and which third-party sources shape the answer. Treat it as a diagnostic measure. For comparisons over time, record the question, location, session conditions, engine, model and date. The separate guide on testing whether AI search recommends you describes the baseline method.

Use three questions, not one attribution model

Ask three questions instead of forcing everything into a single attribution model.

Observe the trace

Observed data shows whether a campaign delivered, which pages people used, whether a form worked and how an opportunity moved. It also exposes mundane defects: missing campaign parameters, duplicate events, inconsistent CRM source fields and broken offline conversion imports.

An incomplete trace still needs to be clean. An incrementality study cannot rescue unreliable CRM outcomes, and a buyer survey will not reveal that the form was broken for a week.

Attribution belongs in this layer as a reporting convention. Name the model, scope and lookback window, then call the result assigned credit. Pipeline-influence reports also show association: an account had a recorded interaction within a defined period. That may help account planning without establishing incremental effect.

Estimate the incremental effect

Incrementality asks what would have happened without the activity. The design needs a credible comparison group that resembles the exposed group apart from the marketing treatment. Depending on the channel, that might be a platform user holdout or a randomised geographic test.

A two-week study of form submissions may say little about qualified pipeline in a long enterprise sales cycle. Spillover between regions, low conversion volume, simultaneous campaigns and changes in sales coverage can all weaken the comparison.

Google’s documentation says geography-based Conversion Lift typically requires a higher budget. Its research on randomized paired geo experiments explains the difficulty created by small numbers of heterogeneous regions. If a study lacks enough power to separate an effect from normal variation, “inconclusive” is the honest result.

Learn from buyers

An open-text “How did you hear about us?” question can recover influences that event logs miss. Keep the original wording, then categorise the responses later.

Call this buyer-reported influence. Recall is incomplete, respondents may name the most memorable interaction, and the person filling in the form may not know what influenced the rest of the buying group. Interviews and win-loss research add context a short field cannot hold.

Evidence becomes more persuasive when interviews repeatedly mention a source, referral data shows relevant visits and a well-designed test detects lift when the activity changes. Keep the records separate so their different limitations remain visible.

Choose the method that fits the decision and the data

There is no universal replacement for multi-touch attribution. Choose a method from the question, available variation, outcome volume and whether a credible counterfactual is possible.

A marketing mix model estimates channel effects without reconstructing individual paths. Google’s Meridian documentation says its goal is causal estimation, while noting that causal quality is difficult to validate without well-designed experiments. Meridian supports using relevant experiment results to calibrate channel priors and warns that the experiment and model may define different effects or time periods.

For a smaller B2B company with low conversion volume, a full MMM or platform lift study may be unrealistic. Improve the observed data, collect buyer evidence consistently and test one decision at the highest feasible level. Label directional results as directional.

Reset the report before buying another tool

Start with clearer labels, cleaner inputs and one investment decision worth testing.

1. 01Label the evidence already in the report

   Mark every important number as observed, attributed, modeled, experimental or buyer-reported. If the team cannot identify which one it is, the metric is not ready for a leadership claim.

2. 02Audit the path from event to revenue

   Check campaign parameters, referral handling, consent states, key events, CRM source fields, duplicate records, opportunity stages and offline outcomes. Record where modeled data enters the total.

3. 03Add one open-text source question

   Keep the original answer, classify it afterwards and review the patterns alongside interviews and sales evidence. Do not force every answer into the existing channel taxonomy.

4. 04Measure AI discovery as visibility

   Separate AI feature impressions, clicked referrals and controlled citation results. Track each over time without merging them into a fictional AI-sourced revenue number.

5. 05Choose one investment decision to test

   Write the hypothesis, counterfactual, outcome, test window and conditions that could invalidate the result before choosing a platform lift study, geo test or model.

Before the test starts, leadership should say what it will change after a positive, negative or inconclusive result. Without that commitment, even a sound study can become an expensive talking point.

How to report without a 100% pie chart

A leadership report can leave deals unallocated. It should show what the company knows, how it knows it and which decision follows.

The board statement might read like this: paid search received the recorded conversions under the selected attribution rule; the lift study estimated a smaller effect within its reported interval; buyer interviews repeatedly mentioned two untracked research sources; the remaining influence cannot be assigned at person level.

Questions worth answering

Is marketing attribution still useful?

Yes, for a limited question about recorded interactions. Attribution shows which observed touchpoints received credit under a stated model. It does not prove that they caused the sale or reveal everything that shaped the buyer's preference.

What is the difference between attribution and incrementality?

Attribution divides credit among interactions that were recorded. Incrementality asks what changed because the marketing happened. It needs a credible comparison, often a randomised test, holdout or well-designed quasi-experiment.

Should B2B companies use self-reported attribution?

Yes, as one source rather than the whole answer. Recall is imperfect, but an open-text response can reveal recommendations, events and communities that clickstream data misses. Keep the original wording before putting responses into categories.

Can AI search traffic be measured?

Some of it can. Referral parameters, first-party analytics and search-platform reports can reveal visits, impressions or citations on supported surfaces. They cannot reconstruct a complete person-level path from every AI answer to a closed B2B deal.