Somewhere in the last eighteen months, retail media started using the word context to mean something it has never meant in advertising. That shift is worth arguing about, and on a recent episode of The Middlemen, it got argued about properly. MetaRouter CBO Jon Flugstad made the case for the new definition. Co-host Scott Messer, who has spent a career in publishing and ad tech, declined to accept it. He was right to push, and the reason the term survived the pushback is the useful part.

‍

What context has always meant in advertising

For twenty years, context meant the page, so a reader on an article about marathon training counted as a running context for exactly as long as they stayed on it. You bought the environment because the environment was the only reliable signal about the person standing in it. Later that hardened into audience segments, so the person who read enough sports pages became a sports audience, and the buying moved from the page to the profile.

Messer holds that line, and he holds it deliberately. His objection on the episode is worth quoting in full because it concedes the substance and refuses the vocabulary:

"You're calling past audience behaviour context, which I will allow as an English major, but I think as an ad tech practitioner, I have to say that context is the exact moment in time. I'm still going to always say that context is context and audience behaviour is audience behaviour."
‍Scott Messer, The Middlemen

That is a fair complaint, because vocabulary drift in ad tech serves marketing more reliably than it serves engineering, and a practitioner is entitled to be suspicious when a familiar word gets stretched far enough to cover a vendor's product surface. If customer context is only a rebranding of audience behaviour, the industry does not need it.

What customer context means now

The claim Flugstad is making is narrower and more specific than a rebrand. On the episode he breaks customer context into three things a retailer holds about a single person at a single moment: the consent that governs what can be done with their data, the identity that connects them across environments including third-party marketing platforms, and the behaviour, which he defines not as a history but as a sequence.

The third of those, behaviour defined as a sequence rather than a history, is where the whole claim actually rests:

"What happens next, the action to take, is based on a sequence of events or activities. That is the kind of crown jewels of what commerce and retail media networks own."
‍Jon Flugstad, CBO, MetaRouter‍

What that describes is neither of the two objects Messer named. It is not the page, so it fails his definition of contextual, and it is not a segment either, because a segment is a label applied to a person after the fact and then held until something refreshes it. A sequence is the live ordering of what someone did in the last ninety seconds, still open, still being extended. Three product views in descending price order is a different object from the label "price-sensitive shopper," and it carries information that survives neither the page-level definition nor the segment.

That is the honest resolution to the argument on the episode. Messer is right that context has meant the moment and behaviour has meant the history. Customer context is neither, which is precisely why it needed a name.

Why the definition changed when inference got fast

The reason this distinction was academic until recently is that nobody could act on a sequence while it was still happening. Flugstad's account of the old architecture will be familiar to anyone who has built one: crawled page text on one side, longitudinal customer data on the other, an offline scoring job stitching them into propensities, and the result written to a table where it waited to be looked up. Batch scoring produces a label, and a label is the best you can do when the answer arrives hours after the question.

What changed is that the same calculation now runs in tens of milliseconds, inside the request. Flugstad's analogy on the episode is to next-token prediction: a language model reads everything you have given it and produces the next word that coheres with the sequence, and consumer behaviour submits to the same treatment. Having observed millions of journeys with a similar shape, a model can score the journey in front of it and decide what to serve next, where next means this pageview rather than tomorrow's batch.

The performance claim he attaches to this is specific. On the episode he describes a customer whose event stream powers inference on conversion events, where the result was an 11% improvement in conversion rate. His framing of why that number is large is the part worth keeping: in ecommerce a 2% improvement counts as a serious result, which puts 11% in a different category of outcome altogether.

The training window is shorter than most teams expect, too. Flugstad puts meaningful model value at two to three weeks of observed behaviour, though he is careful to say the right lookback period depends on the retailer, the buying cycle, and the seasonality, and that it is a question for machine learning engineers rather than a commercial one.

Why rented identity forced the question

None of this works if the identity underneath it belongs to someone else, and that is the second half of the episode's argument.

The standard retail media build has been consistent for years: ship customer data to a third party, use their lookup table to match it to an identifier, and map that identifier out to the channels you buy. Flugstad calls this rented identity, and his point is not that it was a mistake. It was the only viable path when the alternative was building the infrastructure yourself and there were far more publishers worth reaching.

Scale and concentration changed that arithmetic together, because once a retail media network is running real volume, the rented model becomes a constraint on margin and on capability at the same time, because the round trip through a third party costs both money and milliseconds, and milliseconds are the entire premise of scoring a live sequence. The channels that matter have also consolidated, and most of them now support direct connections, which removes much of the original justification.

The ownership question sharpened in May 2026, when Publicis Groupe announced its pending $2.2 billion acquisition of LiveRamp, a deal expected to close by year end that moves the industry's most widely used neutral identity layer inside a holding company that also sells media. Flugstad's read on the episode is blunt: it is not Switzerland anymore, and "can I rent this anymore" has become the question of the moment across retail media networks.

For a retailer, the practical version of that question is narrow. Whatever you rent, you cannot score in real time, because the latency of the round trip exceeds the window in which the prediction is worth making. Identity that lives in your own environment is a precondition for acting on a sequence at all, which makes it an architectural requirement well before it is a governance preference. Retailers who want to resolve identity without shipping PII are solving a latency problem as much as a governance one.

Does the distinction actually matter

There is a practical test that separates the two definitions, and it does not require anyone to concede the word.

Ask what your system does with a first-time anonymous visitor three product views into a session. Under the page definition, it serves against the category of the page. Under the segment definition, it has nothing to serve against, because the visitor has no profile and will not have one until the batch runs. Under the sequence definition, it has three ordered events, the identifiers already attached to that browser, and the consent state, which is enough to make a prediction while the session is still open.

Most retail media systems today answer that question the second way, and most of the value on the table is in the third. That gap is the entire argument for a new term, and it is why the word survived Messer's objection even though his historical claim about it was correct.

Flugstad's own framing of the opportunity is deflationary in a way that fits the point. He describes the work as boring: infrastructure, data, glue. It is events moving at very low latency with identity appended, available to whatever wants to act on them, ads or otherwise. The unglamorous description is the accurate one, and it is also the reason retail media networks are the natural owners of this layer rather than the frontier labs. Their advantage was never going to come from training better models than the labs do. They hold the sequences those models require as input, and that position is considerably more durable than any model advantage.

That layer is what MetaRouter builds as Customer Context Infrastructure: consent, identity, and behaviour captured at the point of collection, streamed inside the retailer's own cloud, and available for inference before the moment closes.

The word context will keep annoying practitioners who remember what it used to mean, and Messer is entitled to keep objecting to it. The argument that matters is not the vocabulary, it is whether a retailer can act on a live sequence in tens of milliseconds or is still waiting on a batch job, because the difference between those two architectures showed up as 11% on conversion for one retailer who made the switch.

The full exchange between Flugstad and Messer runs about forty minutes and is worth the time if you are working out which definition your own stack can actually act on. Listen to the episode on The Middlemen.