BRENNEN

Industries / Technology / Software and SaaS / Ad and marketing tech

Targeting and measurement

Targeting is the software that decides which people, pages or apps see an ad; measurement is the software that works out what the ad did, such as how many visits or sales it caused.

5

Milestones achieved, from 2005 to 2025.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2005

    Site analytics made free

    In March 2005 Google bought Urchin Software of San Diego, whose software counted a website's visitors and what they clicked, and cut the monthly price of its hosted version from $495 to $199. On 11 November 2005 Google renamed it Google Analytics and offered it at no charge, InformationWeek reported. W3Techs, which surveys websites, counted it on 46.1% of all websites in November 2026.

    informationweek.comw3techs.com

    Credit

    Urchin Software, whose chief executive Paul Muret became a Google engineering director; Google

  2. 2017

    Ghost ads cut test costs

    A lift test compares people picked at random to see an ad with a control group who do not see it. Garrett Johnson, Randall Lewis and Elmar Nubbemeyer described "ghost ads": the ad system logs each time it would have shown the ad to someone in the control group, so the two groups can be compared like for like without buying placeholder ads such as public service announcements. They report that advertisers can measure lift "just as precisely while spending at least an order of magnitude less", and that their system recorded more than 100 million predicted ghost ads a day; in one retailer's test, retargeting ads (ads shown to people who had already visited the retailer's site) lifted site visits by 17.2% and purchases by 10.5%.

    api.crossref.org

    Credit

    Garrett Johnson, Randall Lewis and Elmar Nubbemeyer, Journal of Marketing Research, December 2017

  3. 2018

    Installs counted without IDs

    With iOS 11.3 in 2018 Apple added SKAdNetwork, which tells an ad network that its ad led to an app install without naming the person or the device. Apple's documentation says the signed report "doesn't include user- or device-specific data" and limits its detail "to ensure crowd anonymity". The measurement firm AppsFlyer records version 4.0 arriving on 24 October 2022, and Apple introduced AdAttributionKit, built on the same approach, in 2024.

    developer.apple.comappsflyer.com

    Credit

    Apple

  4. 2019

    Delivery skews on its own

    Muhammad Ali, Piotr Sapiezynski, Miranda Bogen, Aleksandra Korolova, Alan Mislove and Aaron Rieke ran their own ads on Facebook and found "significant skew in delivery along gender and racial lines" for real employment and housing ads "despite neutral targeting parameters". Both the ad's budget and its content changed who saw it.

    LessonThe platform's delivery software, and not only the audience an advertiser picks, decides who sees an ad.

    arxiv.org

    Credit

    the six authors

  5. 2025

    Rules for attention scores

    Attention metrics estimate whether a person actually looked at or heard an ad, beyond whether it appeared on screen, using data signals, eye tracking, physical and brain responses, or panels and surveys. In November 2025 the IAB, the online ad trade body, and the Media Rating Council (MRC), the US industry body that audits audience measurement, published final guidelines covering all four methods, against which vendors seeking MRC accreditation are audited. The guidelines say attention "should not be considered or used as a measure of outcomes" on its own.

    iab.com

    Credit

    IAB; MRC

Watch for

  1. Platforms grade their own lift

    In 2025 Meta began offering an "Incremental Attribution" setting that counts only the conversions which, by its own estimate, would not have happened without the ad. Seer Interactive, an agency, found that Meta reported 87% of its April 2025 conversions as incremental across six accounts and $1.05 million of spending; a cross-check it made in Google Analytics 4 suggested 67%.

    Achieved whenrandomized holdout tests run by someone other than Meta match the setting's counts within their stated error.

    seerinteractive.com

    Claimed by

    Meta, 2025, as reported by Seer Interactive in July 2025

  2. One counting method for browsers

    The W3C's Private Advertising Technology Working Group is drafting Attribution Level 1, a browser feature that would report how many purchases and other actions followed ads only as totals with random noise added (a technique called differential privacy), so that no one person's actions can be read from them. Its editors come from Google, Mozilla and Meta, and Google said in October 2025 that it would work on this standard after retiring its own Attribution Reporting API. The draft of 9 November 2026 still lists open issues.

    Achieved whenat least two browser makers ship it switched on and an outside party publishes how accurate its totals are against a randomized test.

    w3c.github.ioprivacysandbox.com

    Claimed by

    the draft's editors from Google, Mozilla and Meta, October 2025 to November 2026

  3. Attention scores face audits

    Two vendors have put attention metrics forward for MRC accreditation. The MRC lists Adelaide's AU, a score of a placement's "potential to deliver an attentive Impression", as pre-audit reviewed in May 2026 with the audit not begun, and XPLN's display and video metrics, including attention, as audit reviewed in August 2026.

    Achieved whenthe MRC accredits an attention metric under its November 2025 guidelines.

    mediaratingcouncil.org

    Claimed by

    Adelaide and XPLN, September 2025 to August 2026

Open

  1. 01

    Count agent visits in reports

    A test shop with no real customers that runs standard analytics and ad tags (Google Analytics 4, the Meta Pixel in its test mode, and the open-source Plausible) next to a server log of every page view and order. AI browsing agents and AI browsers are given shopping tasks, and the test compares what each tag recorded (visits, source, browser, purchases) with the server's record.

    The brief
    Credit · unclaimed
  2. 02

    Contextual labels people agree on

    Contextual targeting places ads by what a page is about instead of who is reading it. The job is a public set of 2,000 web pages, each given a category from IAB Tech Lab's Content Taxonomy 3.1 by two trained people working separately, with their agreement published, and a scoreboard of how often open models and commercial page classifiers pick the same category as the people.

    The brief
    Credit · unclaimed

Open 01 · the brief

Count agent visits in reports

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A test shop with no real customers that runs standard analytics and ad tags (Google Analytics 4, the Meta Pixel in its test mode, and the open-source Plausible) next to a server log of every page view and order. AI browsing agents and AI browsers are given shopping tasks, and the test compares what each tag recorded (visits, source, browser, purchases) with the server's record.

Why it is still open

Stape, a firm that hosts tracking tags, wrote in February 2026 that AI browsers such as Atlas and Comet send the same user-agent string as Chrome, so analytics tools can count their visits as Chrome, and that agents can complete purchases without firing pixels at all. Its guide and others like it give advice but no controlled count, and a search found no published test of how often agent sessions are recorded as human visits or sales.

stape.io

Built with what exists

Browser Use driving Chrome (see the Note "Deleting 80% to start over"); consumer AI browsers with agent modes; a WooCommerce shop or Shopify development store with payments in test mode; Google Analytics 4, Meta's Test Events tool and Plausible; Python for the comparison. No real people and no real payments are involved.

  1. First step

    Set up the test shop with the three tags and a server log, and run 50 tasks with one agent.

  2. Done looks like

    500 scripted tasks for each of at least four agents, plus 500 by an ordinary scripted Chrome session as a baseline, with a published table of the share of visits and orders each tag recorded and the share labelled as Chrome or as direct traffic, and the code to rerun it.

Open 02 · the brief

Contextual labels people agree on

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Contextual targeting places ads by what a page is about instead of who is reading it. The job is a public set of 2,000 web pages, each given a category from IAB Tech Lab's Content Taxonomy 3.1 by two trained people working separately, with their agreement published, and a scoreboard of how often open models and commercial page classifiers pick the same category as the people.

Why it is still open

IAB Tech Lab says version 3.0 of the taxonomy grew from about 400 categories to more than 1,500, and the open tool it promotes converts old category codes to new ones rather than classifying pages. The nearest public project, last updated in May 2026, sorts 98,000 domains into 27 top-level categories using labels corrected by a language model, not by people. Searches found no human-labelled page set for the taxonomy.

iabtechlab.comgithub.com

Built with what exists

IAB Tech Lab Content Taxonomy 3.1; pages from Common Crawl's public web archive; Label Studio, open-source labelling software; open embedding models and a model that scores fixed-answer questions, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"); Python.

  1. First step

    Draw 200 pages from Common Crawl across 20 top-level categories and have two people label them to test the labelling guide.

  2. Done looks like

    2,000 pages labelled at the taxonomy's second tier by two people, with Cohen's kappa (a standard score of agreement between raters) reported, and top-1 accuracy for at least five classifiers on 500 pages held back from any tuning.