Automated Buying Does Not Fix Bad Audience Data

DMEXCO returns to Cologne on 23 and 24 September under the motto Scaling Intelligence, and the framing is deliberately unglamorous. The question is no longer whether artificial intelligence can do interesting things in marketing. It is why so few organisations have turned it into measurable commercial value.

The numbers behind that question are worth sitting with. PwC's 2026 Global CEO Survey found that only twelve percent of companies worldwide are achieving higher revenue and lower costs from AI at the same time. In Germany the figure is two percent. The reasons cited are not model performance. They are poor data quality, unclear ownership, and initiatives that never became part of how the business actually operates.

For anyone working in programmatic, that diagnosis should feel familiar, because it describes the audience layer precisely.

AUTOMATION AMPLIFIES WHATEVER IS IN THE SEGMENT

Automated and increasingly agentic buying systems make more decisions, faster, with less human review at each step. That is the point of them. It is also the risk.

A human trader who received a segment labelled "auto intenders" would apply judgement before spending against it. They would notice that the definition was vague, that the file had not been refreshed since the spring, or that the audience overlapped almost entirely with a segment already running in another line item. They would ask someone. An automated system optimising toward a delivery target has no equivalent instinct. It reads the segment as a valid input, allocates budget to it, and does so at a speed and scale no manual process could match.

This is why the data quality problem gets worse rather than better as buying becomes more automated. The errors do not disappear. They execute faster.

WHAT A MACHINE CAN AND CANNOT INFER FROM A SEGMENT

Human traders operate on context that never appears in a data file. They know which provider is reliable, which taxonomy labels are optimistic, and which segments performed badly last quarter for reasons nobody wrote down.

An automated system has none of that. It has the segment name, whatever metadata accompanies it, the size, and the performance signal it observes after spending money. Everything else is invisible.

That places far more weight on things the industry has historically treated as administrative. How precisely is the segment defined, and does the label match the definition. What behaviour does inclusion actually represent, and over what period. How recently was the file refreshed. Which market does the coverage genuinely apply to, as opposed to where the provider has some presence. Whether that information travels with the segment determines whether an automated system can use it well or merely use it quickly.

A segment described as "supermarket visitors" and a segment described as "devices observed at supermarket locations in France within the last seven days, minimum two visits" will be treated identically by a taxonomy browser and very differently by any system reasoning about fit. Definition quality is no longer documentation. It is a functional input.

FRESHNESS BECOMES A HARDER CONSTRAINT, NOT A SOFTER ONE

There is a tempting assumption that smarter optimisation compensates for older data. In practice the opposite holds.

Optimisation works by finding signal in the response to spend. If the underlying audience no longer reflects current behaviour, the system optimises efficiently toward a population that has moved on. It will still report improvement, because it is improving against the target it was given. The improvement is measured against a stale definition of the right person.

This is why refresh cadence deserves more scrutiny than audience size in an automated environment. A segment refreshed weekly describes behaviour that is still commercially relevant. A segment refreshed quarterly is a historical record, useful for modelling and weak for activation. Identifiers churn as devices are reset, replaced, and reassigned, so a long lookback window inflates the file while reducing the share of it that appears in live traffic.

The lookback window itself should be a decision rather than a default. Someone who visited a home improvement retailer in the last week occupies a different position from someone who visited three months ago, and an automated system given both in one segment will spend against them at the same rate.

GOVERNANCE APPLIES TO DATA INPUTS TOO

DMEXCO's framing of scaling calls for governance, dependable data, and processes that work across departmental boundaries. Applied to media buying, that means the audience layer needs the same standards being demanded of models and workflows.

In practice this looks like a small number of habits. Knowing the legal basis on which the underlying data is processed, which for European activity means being able to point to a GDPR-compliant basis in writing rather than in a slide. Measuring match rate against your own identifier pool before signing rather than after a campaign disappoints. Documenting what each segment means in terms specific enough that someone who did not build it can evaluate it. And treating refresh cadence as a contractual specification rather than an operational detail.

None of this is novel. It is the unglamorous work that the Scaling Intelligence framing is pointing at, applied to the part of the stack that automated buying depends on most directly.

WHERE THIS LEAVES BUYERS

The organisations getting value from automation are not the ones with better models, because most buyers are using broadly similar technology. They are the ones feeding it inputs that are current, precisely defined, and honestly described.

At Dymos, audience segments are built across brand, category, and geographic definitions, delivered on a weekly cadence, and specified in terms that survive contact with an automated system rather than only with a human reading a taxonomy.

The Dymos team is at DMEXCO in Cologne on 23 and 24 September. If audience data quality is the constraint on what your buying stack can actually do, it is a good conversation to have in person.

‍