How ad markets formed for new media, and where AI answers depart from the pattern

Glasp Research

Glasp Research

Nov 01, 2026

19 min read

How ad markets formed for new media, and where AI answers depart from the pattern

Glasp Research, 2026-11-01 (first drafted 2026-10-30). The pre-registration, the coded data and the scripts are kept in a private repository and are not published; every number below was rebuilt from that data by script.

The question

Before advertisers commit real budgets, a new medium has to settle three things. What is sold: an hour, a program, a thousand people, a click? Who sells it, when most publishers are too small to sell alone? And who counts the audience, so buyer and seller can argue about price with the same number? Radio, television, the web and podcasts settled those in turn. Email newsletters and AI answer surfaces, meaning ads in chatbot replies, in AI search and through shopping agents, are settling them now. Does the pattern tell us what comes next?

The measures, the cases and the rule for rejecting the analogy were fixed before any data was compiled. Each medium was coded on one ladder: flat sponsorship; intermediaries that bundle inventory; third-party measurement, a count the seller does not control; a standard unit defined across sellers; per-audience pricing as the common quote; and auctions or self-serve marketplaces. Four pre-registered criteria would show the analogy failing: (A) no shared order of stages, (B) lags differing by more than a factor of four, (C) AI surfaces reaching CPM or auction pricing without third-party measurement, or selling inside auctions their owners already ran, and (D) AI pricing mainly per outcome. The short answer: B and C are triggered, D did not fire, A cannot be settled either way, and the pre-registered sequence hypothesis is not supported, because on its own definition of per-audience pricing three of the four historical cases cannot be assessed.

Every dollar figure below is in nominal dollars of its year and every share is a same-year ratio, so a 1922 rate card and a 2026 CPM are not comparable as money.

How the four historical markets formed

Radio, from 1922. The earliest paid US broadcast located in this search went out on WEAF in New York on 28 August 1922, when the Queensboro Corporation bought fifteen-minute talks to sell apartments. WEAF priced time, not audience: $50 for a fifteen-minute evening period (Banning 1946). NBC had 19 stations on a network by 1 November 1926, the first dated intermediary, and by 1941 the FCC called the agencies producing the sponsored shows "far more than mere brokers or intermediaries". Measurement came eight years after the first sale: the Cooperative Analysis of Broadcasting, run by Crossley with the endorsement of the Association of National Advertisers, began on 1 March 1930 with 30 subscribers at $65 a month. Rates were listed by standard time period in Standard Rate and Data by November 1931, and no source was found for CPM becoming the common radio quote: radio was sold by the clock. Radio's share of US spend, from the Coen series compiled by Galbi, was 0.2 per cent in 1927 and 12.0 per cent in 1942, peaking at 14.9 per cent in 1945.

Television, from 1941. Television did not open with flat sponsorship. NBC's first rate card for WNBT, issued 27 June 1941, priced evening time at $120 an hour and service spots at $4 a minute by day and $8 at night, and on 1 July 1941 five advertisers bought time, so television's first paid ads were spot buys off a rate card. That per-minute price is one station's own, not a unit defined across sellers, so it is recorded separately, and television's standard unit is dated to December 1967, when a trade report has the 30-second spot replacing the minute across networks, twenty-six years in. Television's flat sponsorship arrived in May 1947, a single advertiser sponsoring Kraft Television Theatre, two years before its first ratings. The war had held the medium back: broadcast television does not appear in the Coen series until 1949, at 1.1 per cent of spend. Then the networks took the schedule: Today and Tonight were network-produced and sold in "participations", and by September 1954 NBC was selling them on "a low cost per thousand viewers", the earliest dated use of cost per thousand in network selling located here. ARB diary ratings began in 1949, eight years after the first commercial. Broadcast television's share was 5.2 per cent in 1951 and 14.3 per cent in 1961, peaking at 21.9 per cent in 1984 on a series rebuilt from the workbook's subcategories, because its top-level television category includes cable up to 1989 and excludes it afterwards, giving a false peak of 22.8 per cent read directly.

The web, from 1994. HotWired launched on 27 October 1994 with 12 paying advertisers at a flat rate: "$30,000 to run a dedicated placement" in a section "for three months", with "no traffic commitments whatsoever"; the data notes a contested earlier launch resting on search summaries only. The web then went through the whole ladder faster than any medium before or since. In September 1995 Internet Profiles Corporation "joined forces with Nielsen Media Research to market new Internet tracking systems", but that was site traffic counting and auditing rather than audience measurement, so the measurement date used here is early 1996, when panel measurement began as PC Meter, later Media Metrix. In 1996 DoubleClick started as an ad network and ad server, Netscape's rate card quoted about $20 per thousand impressions, and on 13 December the IAB and CASIE announced nine standard banner sizes including the 468 by 60, though both bodies called these "voluntary guidelines and not 'standards'". By June 1997 a trade column could write that "internationally, pricing of ad banners is based on the notion of CPM", the first dated evidence that CPM had become the common basis rather than one seller's quote. On 24 February 1998 GoTo.com launched "a search directory that enables Web sites to purchase placement within search results using a real-time competitive bidding process". The IAB's first revenue count was $267 million for 1996, rising in nominal dollars to $294.6 billion in 2025, and against the Coen total it was 2.1 per cent of US advertising in 1999 and 3.6 per cent in 2004.

Podcasts, from 2005. The earliest sourced brand sponsorship is a 26-week deal between Southern California Lexus Dealers and KCRW's podcasts in 2005. Podtrac launched the same year offering measurement and ad sales together, which is why it counts as an intermediary rather than third-party measurement: the pre-registered definition asks for a count the seller does not control, and Podtrac was selling what it counted. Podcasts then stayed host-read and direct-response for a decade, with promo codes and vanity URLs doing the counting no shared audience measure did. The IAB published its Podcast Measurement Technical Guidelines in September 2016, eleven years in, as dynamic ad insertion was becoming more common, having risen from 37 per cent of revenue in 2015 to 56 per cent in 2016. CPM reached 84 per cent of podcast ad revenue by 2019 and 93 per cent by 2020, and revenue grew in nominal dollars from $105.7 million in 2015 to $2.86 billion in 2025.

What the history shows, and what it does not

Each line gives the year of the first paid ad, then flat sponsorship, intermediary, measurement, standard unit, and CPM (first quote / common basis). Numbers in brackets are years after year zero, the first paid ad.

  • Radio: first paid ad 1922; flat sponsorship 1922 (+0); intermediary 1926 (+4); measurement 1930 (+8); standard unit 1931 (+9), or 1939 (+17); CPM none / none

  • TV: first paid ad 1941; flat sponsorship 1947 (+6); intermediary 1947 (+6); measurement 1949 (+8); standard unit 1967 (+26), or 1970 (+29); CPM 1954 (+13) / none

  • Web: first paid ad 1994; flat sponsorship 1994 (+0); intermediary 1996 (+2); measurement 1996 (+2); standard unit 1996 (+2); CPM 1996 (+2) / 1997 (+3)

  • Podcasts: first paid ad 2005; flat sponsorship 2005 (+0); intermediary 2005 (+0); measurement none found; standard unit none found; CPM 2006 (+1) / 2019 (+14)

The first leg of the pre-registered sequence, flat sponsorship before third-party measurement, passes in the three cases that have both dates and is undetermined for podcasts, which have no measurement date. The second leg, measurement at or before per-audience pricing, can be assessed for television and the web only, since radio has no CPM date; and on the pre-registered reading, where CPM has to have become the common quote rather than merely been quoted once, only the web can be assessed. So the hypothesis, which asked for four of four, is not supported, and criterion A, which fires if two or more cases fail, is indeterminate: on the pre-registered reading nothing fails and three are undetermined if a missing date counts as unknown, three fail if it counts as a failure, and no missing-data rule was pre-registered.

The timing is where the analogy breaks, though not where this draft first had it. Years to a standard unit run 2 for the web, 9 for radio and 26 for television, a spread of 13.0 that exceeds the pre-registered factor of four, so criterion B is triggered on the unit arm, under every standard-unit coding tried. The measurement arm does not: years to measurement run 8, 8 and 2 years, a spread of exactly 4.0, failing a criterion written as more than four, and that figure moves between 4.0 and 8.0 depending on which 1990s web item counts. So the time to a shared unit does not transfer, and the time to measurement is not evidence either way.

Nor did anything reach 10 per cent of US advertising spend within five years, though only radio, at 0.2 per cent, and the web, at 2.1 per cent, have a five-year value at all. Television has no row before 1949 and the podcast series begins in 2015, so neither plausibly held a tenth that early, but that is a bound inferred from an absence, and the hypothesis is not fully testable here.

One tempting explanation is centralised ad delivery, since the web was served from countable central servers and podcasts got their standard only once dynamic insertion moved ads onto servers. The cases suggest that association but nothing here tests it, and it cuts both ways, since the measurement services that mattered most were valuable precisely because the firms serving the ads did not run them.

One caution on the share numbers. The Coen series ends in 2007, and its Grand Total counts direct mail, directories and a large miscellaneous bucket, 41.5 per cent of it that year; afterwards there is no consistent public US total, only MAGNA and WPP Media estimates of media-owner revenue on a narrower base. The two are not one series, and results.md measures the gap for 2007, the one year with both. So the figure sometimes quoted, that the internet is around three quarters of US advertising, is a cross-source ratio: IAB revenue is 74.0 per cent of the MAGNA 2025 total, 68.3 per cent of WPP Media's.

Newsletters: 34 years and still before measurement

Email sponsorship predates the web banner: TidBITS launched one on 20 July 1992, "modeled after the one used by Public Broadcasting". Thirty-four years later, pricing is still mostly a flat fee per placement: Morning Brew's spots were "sold by flat rates", and marketplaces quote CPMs of $10 to $30 per subscriber which publishers convert to flat fees. Intermediaries are the part that is arriving. beehiiv announced its ad network on 19 September 2022 and by January 2026 was paying publishers over $1 million a month at $1 to $5 per click, per the trade press; Zeta bought LiveIntent, an email ad server, for $250 million in October 2024, per its own filing; Substack began native sponsorships in December 2025. That 30-year gap is an upper bound: 2022 is the earliest newsletter ad network sourceable to a primary document, and an ad-supported email list was using third-party ad serving in 1998.

Measurement has gone backwards. Digiday wrote in 2016 that "there's no comScore for newsletters", and Apple's Mail Privacy Protection, live from 20 September 2021, made open rates unreliable just as Apple Mail accounted for 49.8 per cent of opens. No IAB or MRC guideline for newsletter ad measurement was found as of October 2026, though an absence in search is not proof of absence. So newsletters cannot be placed on a complete sequence at all: no measurement date, no standard unit, and per-audience pricing that is marketplace guidance rather than a common quote. Centralised serving is one possible reason they have stayed early; fragmented demand, privacy rules and market size are others.

AI answers and agents: entering at the auction

AI surfaces are moving fast, but in a different order: they started at per-audience pricing and auctions, with the intermediary already in place at year zero, and with no flat-sponsorship stage at all. The earliest sourced announcement of a live paid ad product inside an AI answer surface is Snap's, on 21 September 2023: "Snap has partnered with Microsoft Advertising to power Sponsored Links within My AI". Microsoft had said in March 2023 it was exploring ads in Bing Chat, but that was testing announced, not a product. So year zero and the first intermediary are the same event, an existing ad network supplying the demand on day one. Even that post announces a product, not an advertiser's placement.

Google put ads in AI Overviews for US mobile users on 3 October 2024, through advertisers' existing campaigns, with "no additional action required". That is criterion C firing through its incumbent-auction clause: the dominant seller places AI inventory inside an auction it already ran. Google's ads liaison also said, on 26 May 2026, that segmented reporting for AI search placements was "not yet available", which is context rather than proof that third-party measurement is absent, since reporting, auditing and audience measurement are different things.

Perplexity tried to build a separate market and left inside fifteen months, pitching sponsored follow-up questions above $50 CPM for late 2024 and taking $20,000 of $34 million in 2024 revenue from ads, per The Information relayed by a marketing site, then dropping ads in February 2026, per the Financial Times relayed by MacRumors.

OpenAI launched ChatGPT ads for US Free and Go users on 9 February 2026 at about $60 CPM, the minimum spend cut from $250,000 to $50,000, according to the trade outlet PPC Land, whose fetched text is the source here; a widely repeated "$200,000 minimum" rests on a search snippet and is not used. Self-serve opened on 13 April 2026 with CPMs as low as $25, and Reuters, via PPC Land, reported over $100 million in annualised ad revenue within six weeks. Third-party verifiers IAS and DoubleVerify announced support on 5 October 2026, eight months in, covering brand safety and attribution, not audience or placement-level reporting.

Criterion D, which asked whether AI pricing is mainly per outcome, did not fire. ChatGPT Instant Checkout charged merchants an undisclosed fee on completed purchases from 29 September 2025 and was shut down in March 2026, and Amazon's Sponsored Prompts are moving to cost per click, which is not a completed transaction: outcome-linked experiments, not evidence it dominates.

Independent intermediaries are being funded alongside the incumbents, and proto-measurement comes from vendor panel estimates, one putting 26 per cent of ChatGPT replies as carrying an ad, on an undisclosed panel without audit. The IAB's framework of 15 January 2026 covers labelling rather than measurement, and the IAB Tech Lab's agentic standards page lists nothing on ads inside model answers. eMarketer forecasts US AI search ad spend at just over $1 billion in 2025, 0.25 per cent of the MAGNA 2025 total, a forecast rather than measured spend.

So the analogy does not carry over to AI surfaces as a forecast of stages: the largest sellers skip the early rungs because they already own the auction, the advertiser accounts and the reporting. What the historical cases did eventually get, and AI answers do not yet have, is a count the seller does not control and a single unit: today the sponsored follow-up question, the in-answer product card and the sponsored prompt are three units at three sellers. The expectation here, not a result, is placement-level reporting under buyer pressure, then a trade-body definition of an impression in an answer.

Limitations and deviations from the pre-registration

One coder, an AI model, assigned every stage, with no second coder, no reliability statistic and no codebook beyond the pre-registration. Five rows still rest on search summaries rather than documents read, among them the 2006 podcast CPM quote; five web rows were upgraded to verified here by fetching the pages. OpenAI's pages refused automated access, so the ChatGPT timeline rests on trade press. No search protocol was written, so every "first" here is the earliest item located. US only.

Deviations: television has no 1946 value in the series; the podcast trade-body guideline is reported separately rather than substituted for a missing audience measure; flat sponsorship now has its own date per case, since year zero and flat sponsorship are not the same event for television; per-audience pricing is reported twice, as the first sourced CPM quote and as CPM becoming the common basis, because only the second matches the pre-registration and it is datable for two cases; a post-hoc rule counting a zero lag as one year has been removed; radio's first intermediary was recoded from 1923 to 1926, television's standard unit from the 1941 rate-card minute to the 1967 cross-network 30-second spot, and the web's measurement date from the 1995 traffic-counting partnership to the 1996 panel, each because the earlier date does not meet the pre-registered definition; AI's year zero event was recoded to the September 2023 Snap announcement, which does not change the year; and broadcast television shares come from a rebuilt broadcast-only series, moving the peak from 22.8 per cent to 21.9.

This text went through three independent methods reviews by a second AI system (OpenAI Codex) and one self-review. Every finding was applied, three of them only in part. The third review still raised two Medium findings, both since applied, so the review stopped at a three-round cap rather than at a clean round, and no review has checked those last fixes. Two numbers were corrected during review: the broadcast television peak share and the ChatGPT minimum ad spend. No human has checked the sources line by line.

References

All accessed 2026-10-30 unless stated; four web items were re-fetched on 2026-11-01. This list covers both the body above and the coded stage events in the unpublished dataset, so it includes items the body does not cite directly.

Note on AI use

This piece was written by an AI model. The research question came from a human at Glasp; the stage coding, the data compilation, the scripts, the analysis and this text were produced by Claude in an automated research run, with research subagents collecting the sources, and were then revised by the same model in a self-review. No human has yet checked the sources line by line; three independent AI methods reviews have been completed, and they stopped at a cap rather than at a clean round. Five coded rows rest on search summaries rather than on documents that were read, and the ChatGPT and Perplexity timelines rest on trade press because the companies' own pages refused automated access. A human editor should check those before anyone relies on them.

Comments

Add a comment
    Glasp Research

    Written by Glasp Research