Industries / Finance and business services / Advertising and marketing
Measurement
Measurement is the work of counting who saw an ad and working out what the ad changed, such as visits or sales.
7
Milestones achieved, from 1914 to 2025.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1914
Circulation gets audited
The Association of National Advertisers founded the Audit Bureau of Circulations (ABC) in 1914 to verify the print circulation that publishers claimed, the number of copies advertisers were paying to reach. The body is now the Alliance for Audited Media, which merged with BPA Worldwide in 2023 and is governed by a board drawn from agencies, marketers and publishers.
LessonA buyer pays for an audience it cannot see, so someone outside the sale has to count it.
Credit
Association of National Advertisers
1942
Meters in radio homes
The Nielsen Radio Index began in December 1942. It used the Audimeter, a device fixed to a radio's tuning dial that recorded on paper tape which station was on, in 800 homes whose tapes Nielsen staff collected by hand. Robert Elder of the Massachusetts Institute of Technology (MIT) and Louis Woodruff had field-tested such a meter on Boston stations in 1935, and Arthur C. Nielsen bought the meters, patent rights and trademark. In 1950 Nielsen began using the meter for its television index.
LessonA meter records what a set is tuned to, not who is listening, which is why the industry later moved to "people meters" on which each viewer presses a button.
Credit
Arthur C. Nielsen; Robert Elder and Louis Woodruff
2014
Defining a viewable ad
The Media Rating Council (MRC), an industry body set up in 1963 at the request of the US Congress to audit audience measurement, issued final viewability guidelines, written with the Interactive Advertising Bureau (IAB), on 30 June 2014. A display ad counts as viewable when at least 50% of its pixels are in the visible part of the browser for at least one continuous second; a video ad needs two continuous seconds. The viewable count was added alongside the existing IAB count of ads served. The MRC updated the guidelines in 2015 and issued separate ones for mobile in 2016.
mediaratingcouncil.orgmediaratingcouncil.orgmediaratingcouncil.org
Credit
Media Rating Council and Interactive Advertising Bureau
2015
Experiments shrink measured returns
In large field experiments at eBay, economists Thomas Blake, Chris Nosko and Steven Tadelis found that search ads on the company's own brand name had "no measurable short-term benefits", and that returns from paid search were "a fraction of conventional non-experimental estimates", because people who click a search ad and then buy were often going to buy anyway. The study appeared in Econometrica in January 2015. A study of 15 experiments at Facebook, covering 1.6 billion ad impressions and published in Marketing Science in 2019, found that common non-experimental methods "often fail" to reproduce the experiments' results.
LessonCounting the sales that follow an ad overstates what the ad caused; a randomly chosen group that does not see the ad is needed to tell them apart.
Credit
Thomas Blake, Chris Nosko and Steven Tadelis; Brett Gordon, Florian Zettelmeyer, Neha Bhargava and Dan Chapsky
2019
A video metric overstated
According to the advertisers' lawsuit, Facebook's figure for the average time people spent watching video ads left views shorter than three seconds out of the number it divided by, which made the average too high. After the Wall Street Journal disclosed the problem, advertisers sued; their complaint said the averages were inflated "by some 150 to 900 percent", not the 60 to 80 percent first reported. In October 2019 Facebook agreed to a $40 million settlement while calling the case "without merit".
LessonA platform's own count of how its ads performed can be wrong until someone outside checks it.
Credit
Facebook (a failure); Gibbs Law Group, counsel for the advertisers
2021
Nielsen loses accreditation
The MRC concluded that Nielsen's national TV ratings had understated total TV use by adults aged 18 to 49 by about 2% to 6% in February 2021, after the pandemic led Nielsen to cut back meter maintenance and the recruiting of new panel homes. In September 2021 the MRC suspended accreditation of Nielsen's national TV service. It reinstated accreditation in April 2023, after an audit showed Nielsen had restored its panel's performance.
LessonA ratings panel needs constant upkeep, and the industry's auditor can withdraw its approval from the largest rating service.
mediaratingcouncil.orgmediaratingcouncil.orgmediaratingcouncil.org
Credit
Media Rating Council
2025
Open-source mix models
A marketing mix model (MMM) estimates what each channel added to sales from weekly totals of spending and sales, without tracking individual people. Meta publishes an open-source MMM called Robyn, and on 29 January 2025 Google made its open-source Meridian available to all marketers, with "over 20 measurement partners trained and certified" on it. A benchmark published on GitHub in October 2026 simulated sales with known true returns and found that when spending followed demand, both Meridian and PyMC-Marketing gave search ads too much credit; Meridian put search's return near 4.9 against a true 2.3.
LessonIn the benchmark's words, "No sampler setting fixes confounding; a lift test on search does.".
Credit
Google; Meta; PyMC Labs; benchmark by vivemeasurement
Watch for
Results from chatbot ads
OpenAI cites its measurement partners' early figures: by DV Rockerbox's count, WeightWatchers' attributed cost per acquisition on ChatGPT ads was 15.3% lower than its blended paid-search benchmark, and WorkMagic reported "statistically significant lift" for the wellness brand Dose. OpenAI also says its work on incrementality, the extra sales an ad causes, "is still in its early stages".
Achieved whenan independent geo experiment, in which ads are turned off in randomly chosen regions, is published with its methods and shows extra sales from ChatGPT ads.
Claimed by
OpenAI and partners DV Rockerbox and WorkMagic, 2026
Standards for chatbot ads
Published estimates of the share of ChatGPT answers that carry an ad range from 4.5% (Adthena, March to May 2026) to 77% (Seer Interactive, July 2026), and a review traces the spread to different time windows, prompt sets and units counted. The MRC, which in July 2026 published guidance on applying its existing standards to AI used in measurement, listed "base measurement of structured and conversational ads in LLM" among the uses its new AI standards will cover, with a first draft planned for public comment in the first quarter of 2027.
Achieved whenthe MRC issues final standards that define a served and a viewed ad inside a chatbot answer, and a measurement service is accredited against them.
Claimed by
the MRC, November 2026, as reported by PPC Land
Fixes to Nielsen's big data
Nielsen's Big Data + Panel service combines its panel of homes with large sets of viewing data from TV devices. In March 2026 the MRC asked Nielsen to act in four areas, including an independent estimate of TV households and more accurate assignment of viewers' ages and backgrounds. Nielsen added outside household estimates from the ARF DASH study in February 2026 and moved its remaining changes to 31 August 2026.
Achieved whenthe MRC completes its review of the August 2026 changes and closes its 2025 audit with accreditation kept.
Claimed by
Nielsen, as reported in the MRC's May 2026 update
Open
01
Test AI visibility trackers
A head-to-head test of the commercial tools that report how often AI assistants name a brand, a figure sold as "AI visibility". The same brands and prompts are tracked in several tools at once, and each tool's figures are compared with a reference built from a large number of repeated runs.
The briefCredit · unclaimed02
Count billboard audiences
An independent count of the traffic passing roadside billboards, compared with the weekly impressions their sellers list. A camera on public land near each board counts the vehicles that pass within sight of it for a week, keeping counts only and no images.
The briefCredit · unclaimed
Open 01 · the brief
Test AI visibility trackers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A head-to-head test of the commercial tools that report how often AI assistants name a brand, a figure sold as "AI visibility". The same brands and prompts are tracked in several tools at once, and each tool's figures are compared with a reference built from a large number of repeated runs.
Why it is still open
SparkToro and Gumshoe.ai had 600 volunteers run 12 prompts 2,961 times through ChatGPT, Claude and Google's AI, and found "a <1 in 100 chance" that ChatGPT or Google's AI would give the same list of brands in any two of 100 responses. They concluded that rankings are random "to the point of near-uselessness" but a brand's share of appearances over many runs is "(probably) a reasonable way" to measure it. They tested the AI tools, not the trackers. A review of ChatGPT ad trackers found their ad rates ranged from 4.5% to 77% because they "count different things at different times". Searches found vendor feature comparisons but no published test of whether AI visibility trackers agree with each other or with a reference.
Built with what exists
trial or paid accounts with trackers such as Gumshoe, Otterly.AI, LLM Pulse and SE Ranking; volunteers, or browser automation where each assistant's terms allow it, for the reference runs; a small AI model to pull brand names out of answers (see the Note "The jobs agents repeat"); Python for agreement statistics such as the intraclass correlation, a measure of how closely two sets of numbers agree, where 1 is perfect.
First step
Pick five prompts in one product category, run each 100 times through one assistant, and set up free trials in two trackers for the same five prompts.
Done looks like
For 40 prompts and 10 brands tracked for four weeks, each tool's weekly visibility figures published next to a reference built from at least 100 runs per prompt per assistant, with each tool's mean absolute error against the reference and the agreement between tools.
Open 02 · the brief
Count billboard audiences
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An independent count of the traffic passing roadside billboards, compared with the weekly impressions their sellers list. A camera on public land near each board counts the vehicles that pass within sight of it for a week, keeping counts only and no images.
Why it is still open
Sellers list each board's audience as weekly impressions, such as 388,562 for one digital board in Baton Rouge. In the US those figures come from Geopath, which by OOH Today's account estimates them from aggregated mobile location data, traffic counts, viewshed models (what can be seen from where) and demographic data. The trade disputes them: OOH Today quotes Lamar's chief executive, Sean Reilly, saying in March 2025 that "the data is stale, and old in some cases, inaccurate", and in November 2025 called for pilots that pair Geopath's outputs with camera counts and publish the results. A search found no published comparison of listed impressions with independent counts.
Built with what exists
a Raspberry Pi 5 or an old phone with a camera; open-source vehicle detection and tracking, such as Ultralytics YOLO with ByteTrack, set to count vehicles that cross a line in the frame; traffic counts published by state transport departments as a cross-check; Python.
First step
Pick five boards with listed impressions that can be watched from a safe public spot set back from the road, and run a one-day count with a hand tally beside it.
Done looks like
For 30 billboards with listed weekly impressions, one week of vehicle counts per board, checked against a hand count of two hours of video per board to within 5%, published with the ratio of listed impressions to counted vehicles for each board and the spread of that ratio across boards. Impressions count people and allow for whether they are likely to notice the board, so the ratio is not expected to be 1; a wide spread between similar boards is the finding to look for.