If AI Does the Research, How Will You Know What Worked?

Attribution was already under strain.

Long before generative AI arrived, most marketers had quietly accepted that their reporting told them something useful — but not something complete.

Cookies were disappearing. Dark social was growing. Buyers were researching in places we couldn't see.

Generative AI may push that problem into a different category altogether.

The measurement model we built, and what it assumed

Digital marketing measurement was built on a fairly specific assumption:

That the research happens on the open web, and leaves a trail.

Someone searches. They click. They land. They browse. They come back. Eventually they convert.

Every one of those steps produced a signal — a referrer, a UTM parameter, a session, a cookie. Attribution models were essentially an attempt to assign credit across that trail.

Last click. First click. Linear. Time decay. Position based.

We argued endlessly about which model was fairest. But we rarely questioned the underlying premise, because the premise was usually true.

The trail existed.

What a conversation removes

Now consider the same buyer, six months from now, researching the same purchase inside an AI assistant.

They describe their situation.

They ask which options suit a company like theirs.

They ask how three of them compare.

They ask which one handles their specific constraint.

They ask what people complain about after twelve months of use.

Somewhere in that exchange, your brand is either mentioned, recommended, dismissed or never surfaced at all.

And almost none of it reaches your analytics.

There is no referrer for a conversation. No session. No sequence of pages. Often no click until the very end — if there is a click at all.

The research still happened. The comparison still happened. The shortlist still formed.

It simply happened somewhere you cannot instrument.

The buyer who arrives already decided

The practical consequence shows up at the bottom of the funnel, and it looks deceptively like good news.

More prospects arriving already informed.

Shorter sales cycles on some deals. Fewer basic questions. Buyers who already know your pricing model, already know your main competitor, and already have a view on which one fits.

That is a genuinely better buying experience.

But it creates an awkward reporting picture.

The touchpoint that shaped the decision is invisible. The touchpoint that captured the conversion — a branded search, a direct visit, a demo request — gets the credit.

Branded search has always been the channel that takes credit for work done elsewhere. In an AI-mediated funnel, it may take credit for considerably more.

Which means the risk isn't only that we measure less.

It's that we keep measuring confidently, and quietly reward the wrong things.

This has happened before

It's worth remembering that measurement has gone dark before, and marketing survived it.

Organic social reach collapsed. Keyword data disappeared behind "not provided". iOS privacy changes broke a large part of app attribution. Dark social swallowed an enormous amount of genuine word-of-mouth.

In each case, the pattern was similar.

The behaviour didn't stop. The visibility did.

And the teams that adapted best were usually the ones that stopped trying to rebuild the old dashboard, and started measuring at a level that still held.

What marketers can actually do about it

None of this argues for abandoning measurement. It argues for moving it.

A few things look increasingly worth the effort.

Ask buyers directly. A single "how did you hear about us?" field on a form is crude, unfashionable, and increasingly one of the more honest signals available. Self-reported attribution ages well precisely because it doesn't depend on tracking.

Measure at the pipeline level, not the click level. If channel-level attribution is becoming unreliable, cohort and period-level analysis becomes more valuable. What happened to qualified pipeline in the quarter you invested in a category? What changed when you stopped?

Watch the leading indicators that still work. Branded search volume. Direct traffic quality. Demo requests from companies matching your ICP. Deal velocity. These are blunt, but they are difficult to fake and they respond to real demand.

Treat your public information as infrastructure. If AI systems increasingly summarise your category, the accuracy and availability of information about your product becomes a commercial asset — not a content marketing task.

Get comfortable with directional evidence. Some questions will no longer have a clean answer. The alternative to a precise wrong number is usually an approximate right one.

Measurement moves closer to revenue

There is a version of this shift that is genuinely positive.

For years, marketing measurement drifted towards whatever was easiest to count. Impressions. Clicks. Opens. Cost per lead. Metrics that were precise, abundant, and frequently disconnected from whether the business made any money.

If the middle of the funnel becomes harder to observe, the incentive shifts back towards the things that were always the point.

Pipeline created. Deals won. Revenue by segment. Customers retained.

Those numbers live in your CRM, not in a platform dashboard — which makes them yours, and considerably harder to distort.

The irony is that a less observable funnel may push a lot of marketing teams towards better measurement than they had when everything looked trackable.

Still early

It's worth being honest about how early this is.

Most B2B buying still involves stakeholders, demos, procurement and a sales team. Search hasn't collapsed. Websites still matter. Plenty of categories will barely move for years.

But the direction seems reasonably clear, and the reporting consequences arrive before the behaviour is fully mainstream. By the time an AI-shaped funnel is obvious in the market, the attribution gap will already be sitting in your dashboards — usually disguised as a suspiciously strong direct channel.

The uncomfortable question isn't whether attribution gets harder.

It's whether we'll notice it happening, or simply keep reporting the numbers that still look tidy.

How are other marketers approaching this — are you changing how you measure, or waiting to see how the behaviour settles first?