AI changes media research less by making it faster than by making it continuous, changes personas from a one-off document into a working model the system updates, and shifts the durable data advantage from public research — now cheap for everyone — to your proprietary first-party data and accumulated memory.
I have spent 18 years in planning and buying, and for most of that time research was a project. You commissioned it, you got a deck, you used it for a quarter, and it went stale in a folder. The people who did it well were fast readers and good synthesizers. AI changes what that skill is worth, and it changes where the advantage sits. Let me take the three pieces in turn — research, personas, and data — because the shift is different in each.
Research stops being a project
The old rhythm was periodic: a category study in Q1, a competitor scan when someone asked, an audience refresh before the annual plan. Each was an event, and between events the team ran on memory and instinct.
An AI-native team runs research as a standing capability. Public signals — category shifts, competitor creative, search demand, social conversation — are gathered and summarized continuously, on a schedule rather than on request. The output is not a deck. It is a stream the team reads, the way you read a dashboard instead of requesting a report.
The practical effect is that the team is never starting from a blank slate on a category, because the category has been summarized all along. Getting there is the same unglamorous work as the rest of the cluster: the signal sources have to be wired up, the summary has to have a standard, and someone has to own the interpretation. The model does the reading, not the deciding.
What the analyst's job becomes
When synthesis is continuous and mechanical, the analyst's value moves from finding the answer to choosing the better question.
That is not a small shift, and it is uncomfortable for anyone whose identity was being the person who knew where to look. In the old model the bottleneck was retrieval — finding the number, the study, the benchmark. In the new model retrieval is nearly free, so the scarce skill is framing: knowing which question the business actually needs answered, and which answer would change a decision.
I have seen teams make this transition well and badly. The ones that do it well give the analyst a seat in the planning conversation, because that is where the good questions live. The ones that do it badly automate the retrieval and leave the analyst doing the same work faster, then wonder why their most curious person is bored. If retrieval was the job, the job is gone; if judgment was the job, the job just got more interesting.
Personas become a working model
The persona document is one of the most quietly useless artefacts in marketing. It is drafted once, illustrated with a stock photo, and never updated, because updating it is expensive and nobody owns it.
An AI-native team treats personas as a working model instead. The audience definition lives with the data, and the system refines it as behaviour changes: which segments convert, what messaging lands, which assumptions broke. The persona is not a deliverable; it is a queryable model that sits behind planning and creative.
The strategic consequence is that personas stop being an art project and start being a decision tool. Instead of asking "does this campaign match our persona?" — a conversation that usually ends in agreement — you can ask the model which of several audience definitions best explains the last quarter's results, and get an answer grounded in what actually ran. That is a different quality of input to the brief.
The data advantage moves to the proprietary layer
Here is the part that actually matters for strategy, and it is easy to miss because public research is now so impressive.
If AI can synthesize the public web for anyone, then public research is no longer an advantage — it is a commodity. The category benchmark, the competitor's creative, the industry study: everyone with a good workflow can summarize those. The edge cannot live there, because it is no longer scarce.
The durable advantage moves to what a competitor cannot subscribe to: your first-party data, your past campaign results, your customer relationships, and the memory your workflow accumulates about what worked. A model that remembers last year's tests and this year's results is worth more than a smarter model with no history. That is the same argument I made about why AI memory matters more than model choice, applied to research: the compounding is in the context, not the model.
Old research vs AI-native research
| Dimension | Traditional research | AI-native research |
|---|---|---|
| Cadence | Periodic project | Standing capability |
| Output | A deck | A stream plus a model |
| Cost driver | Analyst hours | Data access and governance |
| Analyst's job | Find and synthesize | Frame the question and interpret |
| Personas | Static document | Living model behind planning |
| Advantage | Better synthesis | Proprietary data and memory |
| Failure mode | Stale and unused | Unclean data and unclear ownership |
Read the last row as the warning. The AI-native version fails in a different place than the traditional one — not from being stale, but from feeding a fast, confident system on data nobody has cleaned or governed.
Governance is a precondition, not paperwork
You cannot build the memory layer before you have decided what the model may see, what may leave your systems, and who owns the answer when something is wrong.
This is why governance lands in the Strategy lane rather than in a compliance appendix. The moment you connect first-party data to a research or planning workflow, you have made a set of decisions about boundaries and consent — and if you defer them, you will rebuild the layer later under a lawyer's supervision, at several times the cost. The teams that treat governance as a prerequisite move slower for two weeks and faster for two years.
The Execution guide goes deeper on the operating side of this, but the strategic point is simple: the proprietary layer is where the advantage lives, and the proprietary layer is exactly the thing you are not allowed to be sloppy with.
What an in-house team should actually do
If you own your data and your roadmap, the sequence is unusually clear.
- Wire up the signal sources for one category or one audience — not everything. Continuous research on one thing is worth more than sporadic research on ten.
- Define the summary standard so the stream is trustworthy: which sources count, what a good synthesis looks like, who validates it.
- Stand up the persona as a model behind one workflow — planning or creative — and let it update from real results.
- Decide your data boundaries before you connect first-party data, and write down who owns the answer when the model is wrong.
- Protect the analyst's seat in the decision, because framing the question is now the job.
The for-in-house-teams track sequences this against the rest of the operating model, and the Strategy guide puts it in the context of the whole lane. Finish one stage before widening to the next.
FAQ
Does AI replace media researchers?
It replaces the retrieval half of the job, which was most of the hours. It does not replace framing the question or interpreting the answer. Researchers who move to the judgement half become more valuable; those who were paid for retrieval move faster and then need a new definition of the role.
Are personas still useful at all?
Yes, but as a working model rather than a document. A persona that updates from real campaign results is a decision tool. A persona drafted once and never revisited is decoration, and AI does not rescue it — it just makes the decoration cheaper to produce.
Is public research worthless now?
No, it is a useful input and it is faster than ever. It is just not an advantage, because everyone has access to the same synthesis. The advantage moved to your proprietary data and memory, which is the part a competitor cannot copy by subscribing to the same tools.
Where does data governance fit?
As a precondition, before you connect first-party data. Deciding what the model may see, what may leave your systems, and who owns the answer is part of the research design now, not a compliance afterthought. Skip it and you will rebuild the layer later, under supervision.
The short version
AI makes research continuous, turns personas into a living model, and moves the durable advantage to your proprietary data and memory. The analyst's job shifts from retrieval to framing, and governance stops being paperwork and becomes the thing that lets the proprietary layer exist at all. The model does the reading; you do the judging.
I am still learning this field in public, and I have been wrong about where the advantage would settle more than once. But this part has held: when synthesis is free, the scarce thing is a good question and your own history.
If your team has made this shift, I would like to hear what happened to the researcher role — that is where the real story usually is.
Cheers, Chandler