AI Agents Amplify Audience Data Quality — Good or Bad, Experts Warn

AI agents are increasingly being positioned as replacements for human researchers in pre-purchase decision-making, but a new analysis published by Search Engine Journal on 25 September 2026 warns that the technology runs on audience data — and will amplify whatever quality that data already has, whether good or bad.

The piece, written by Greg Jarboe, cites Mallory Gray, creative director at audience data company Skydeo, which the article describes as drawing on 1.4 trillion data points across more than 320 million people. Gray argues that while a human researcher works through a limited number of sources sequentially, an AI agent operates across a far broader set of signals simultaneously. According to Gray, an agent works across "thousands of behavioral, purchase, interest, and intent signals" simultaneously, continuously revising as new information arrives.

The analysis draws a distinction between generative engine optimisation (GEO) — structuring content to be cited by AI systems — and what Gray terms "AI visibility." Jarboe reports Gray's position that citation frequency alone measures the wrong outcome: "A brand can increase its visibility in AI answers substantially without seeing the same improvement in qualified traffic or conversions." The fix, Gray is quoted as saying, is not more optimisation but rather scrutiny of who is actually being served a brand's content and whether that audience matches the business's actual needs.

A key warning sign identified in the piece is the pattern of rising output volume — more content, more variations, more campaigns — alongside flat or declining engagement or conversions. Jarboe reports that Gray flags the moment when no one on a team can explain why a particular audience was targeted or a particular message was sent, and the honest answer has become that the AI chose it, as the point at which a critical feedback loop has broken down.

The article draws a parallel to the data management platform (DMP) era of the early 2010s, when large-scale aggregation of third-party data was expected to deliver superior targeting. That promise largely failed, according to the analysis, because inaccuracies in third-party data compounded at scale. Cookie deprecation by Safari and Firefox subsequently pushed the industry back toward first-party and declared signals.

Jarboe argues the lesson applies directly to the current moment: access to powerful AI models is becoming commonplace, and the differentiating factor is what data a brand feeds into them. He notes that Skydeo, as a seller of audience data, has a commercial interest in the argument, while maintaining that the underlying case stands independently.

The article recommends three practical checks for marketers: auditing which audience signals currently feed GEO and AI visibility tools; stopping the practice of measuring AI visibility purely as a citation count in favour of tracking whether mentions convert; and maintaining a standing requirement for a team member to explain in plain language why a given audience or message was selected.