Automatic content recognition was built for one drink and is being sold as another
Late last year, CIMM and TVB published a diagnostic study of how local television audiences actually get counted. In the case it examined most closely, an early-morning local news program measured against the station's own first-party logs, between twenty and thirty percent of the viewing never appeared in syndicated measurement at all. The audience was there. The systems buyers transact against simply had no way to see it.
That gap is not a scandal. It is a bartending problem.
Ask two bartenders for a gin martini using the same bottles, the same ice, and the same glass, and you will get two different drinks. Ratio, temperature, and technique live in the pour, not in the ingredients. The measurement industry has been running this experiment for a decade, and we still act surprised when what lands in the glass differs so much from one company to the next.
The base spirit is automatic content recognition, the fingerprinting technology now built into most smart televisions sold in the US. It is the loudest, most visible ingredient in modern TV measurement. It is also, on its own, not measurement at all.
ACR is the gin. It was never the martini.
The Gin Problem
Nearly every smart-TV manufacturer offers some version of ACR, and just about every company selling audience data leans on it, because it is dominant, widely available, and the category has practically been named after it. That is precisely the problem. When the base ingredient is available to anyone willing to sign a licensing agreement, it stops being a point of difference. Yet we discuss ACR as though having it were the achievement.
Hand the same ACR feed, the same metadata, and the same processing pipeline to two companies, and you will get two different audience numbers for the same household, watching the same program, on the same night. That is not a data quality problem. The gin was never the moat.
A disclosure before I go further: I work for a company that sells set-top-box data, and part of what follows argues that deterministic tuning data is undervalued. Read it with that in hand. The argument stands or falls on whether it is right.
A Drink Built for a Different Occasion
It is worth remembering what ACR was built for. When Shazam first pointed its music-recognition technology at a television screen in 2011, and when Samba TV and Cognitive Networks followed shortly after, none of them were trying to solve audience measurement. They were making television interactive: second-screen trivia, live polling, companion apps. Only later, as panel-based measurement fragmented under streaming, did reach extension become ACR's main commercial job. Precision, device-level measurement is a third and much more recent ambition.
That history explains the mismatch at the heart of today's market. ACR was engineered for the gin and tonic: a forgiving, mixer-heavy drink where directional accuracy is the entire point and a little dilution never hurt anyone. Reach extension does not need precision to the decimal. The tonic hides a multitude of sins, and it is supposed to.
Device-level measurement is a different order of drink. It is a martini: three ingredients, served neat, with nowhere for a bad pour to hide. Every calibration error, every miscounted household, every uncredited stream shows up in the glass. ACR was built for the long drink and is now sold as the short one, and buyers are transacting against exactly what is in that glass, whether or not anyone has told them the ratio.
ACR was built for the long drink. It is now being sold as the short one.
Reading the Label
If you are going to transact on ACR, it is worth understanding what the bottle actually contains. None of what follows is a secret, and none of it is a scandal. It is simply how the technology works, and each characteristic has consequences that show up in the numbers.
ACR matches against a reference library, and the library is not evenly stocked. Fingerprinting compares what is on the screen against a set of known references. Where the reference is rich, national primetime, big-brand creative, matching is excellent. Where it is thin, matching degrades. Local news, local ad breaks, and subchannel programming are exactly the content least likely to be well-represented in a reference set built for national use. This is the single biggest reason ACR thins out precisely where local measurement needs it most.
Match confidence is a threshold, and the threshold is a commercial decision. Every ACR system carries a confidence level below which it will not claim a match. Set it high and you report less viewing but report it accurately. Set it low and you report more with more error in it. That choice is made by a vendor, not by physics, and it is rarely disclosed. Two vendors reading the identical signal can legitimately produce different totals purely on where they drew that line.
Detection is not attribution. ACR can establish that a frame appeared on a screen. It cannot, from the pixels alone, establish how that frame arrived: broadcast, vMVPD simulcast, station app, FAST channel, or an HDMI input. Same content, same second, very different commercial value.
And one blunt structural fact sits underneath all of it: one major smart-TV manufacturer does not license its ACR data to any measurement vendor at all, leaving roughly thirty-three million American households, about one in three smart-TV homes, outside ACR-based measurement entirely. None of this makes ACR bad. It makes ACR a base spirit: powerful, dominant, and incomplete on its own, which is what a base spirit is.
Nobody Was Hiding the Jigger
Here is where I want to be careful, because the easy version of this argument is also the wrong one. The easy version says most of the market cannot calibrate ACR because it lacks deterministic data. That is not true. The measuring tools have been sitting on the back bar the whole time: set-top-box return-path data giving deterministic, tuner-level records of what was actually tuned; EPG and lineup data establishing what was genuinely available on which channel and market. Every one of these is licensable, and some already sit inside the same companies that publish ACR-derived numbers.
So the question is not whether the industry could have built precise, calibrated, device-level local measurement. It is why it largely hasn't. The answer is the one the history gives us. Reach extension does not pay for precision. If your product is a directional statement about where else an audience might be found, calibration infrastructure is expense without revenue. Nobody was concealing the jigger. They were making a drink that did not call for one, and they made it well.
What changed is the drink. The same data is now being poured into precision use cases, local avails, outcome attribution, incrementality, where the tolerances are an order of magnitude tighter and there is no tonic to hide behind. The infrastructure has not caught up with the ask, and the gap between the two is roughly the gap CIMM found.
Set-top-box data is not a superior spirit either, and I would not claim otherwise while selling it. What it is, is bounded and knowable. It measures tuning deterministically in the homes that have a box, and not a model's estimate of what a screen probably showed. A bounded universe you can fully characterize is worth more as a calibration reference than an unbounded one you cannot. That is the whole argument for the jigger: not that it measures everything in the bar, but that what it measures, it measures exactly.
And combining these sources is not a novel claim. The largest measurement companies already do it, privately, at considerable expense. The more honest ambition is narrower: taking that unglamorous work of combining, cleaning, and de-duplicating device-level data and making it available as a single syndicated product, so buyers no longer license three datasets and build their own reconciliation pipeline first. Democratizing the pour, rather than inventing a new spirit.
Serving It Neat
Everyone in the room can buy the same gin. They always could. So the useful questions are not about whether a partner has ACR, because almost all of them do. They are about the pour.
Ask your vendors. What is your match rate, and how does it differ between national and local content? Where does unmatched viewing go, dropped, or routed into a generic bucket that still gets reported? What do you calibrate against, and what happens to the numbers when you do? Is this device-level or person-level, and if person-level, what assigns the person?
Some of those have legitimately proprietary answers, and a good partner will say so. That is fine. The test is not whether they hand over the method. It is whether they can characterize their own boundaries at all. A vendor who can tell you precisely where their data goes quiet is more valuable than one who claims it never does.
Then ask yourselves. What proportion of what we transact on comes from a single source, and what is our exposure if it changes? Do we know our own first-party numbers well enough to check anyone against them?
The answers will tell you whether you are being handed a long drink dressed up as a short one. The real question about any measurement partner is not whether they have ACR. It is whether they have ever had to serve it neat, with a jigger in hand, standing behind exactly what ends up in the glass.
Figures cited are drawn from Plugging the Gaps, a diagnostic analysis of local TV and video measurement published by CIMM and TVB, November 2025.
