2022. ChatGPT is released

The moment

On 30 November 2022, OpenAI posted a note titled “Introducing ChatGPT”. It said the company had trained a model that “interacts in a conversational way”, that the dialogue format let it answer follow-up questions and admit its mistakes, and that during the research preview use was free [1]. The launch was that post and a link.

Five days later Sam Altman wrote that ChatGPT had “crossed 1 million users” [2]. The box took anything: a contract, a recipe, a shopping question. Somewhere in that first week a shopper typed a question that would once have gone to a search box, read the paragraph that came back, and did not click anything. That is the moment.

What came before

A buyer with a question had, since the late 1990s, one habit: type it into a search engine, read a page of links, click two or three, decide. Digital marketing sat on top of that habit, the auction beside the links, the agencies bidding in it, the pages written to rank in it. Tim Wu calls the business the attention merchant, and the deal, in his publisher’s summary, has not changed since the penny press: “free diversion in exchange for a moment of your time, sold in turn to the highest-bidding advertiser” [25].

Conversational answers were not new in 2022. Google had shown its own dialogue model, LaMDA, in 2021. On 6 February 2023 Sundar Pichai announced Bard, “an experimental conversational AI service, powered by LaMDA”, and opened it to trusted testers [12]. A year later Bard became Gemini. Google had the model first. OpenAI had the product first, free and public, in a chat box, with no instructions.

What it changed

The day before, a marketer could buy a position on a page and write a page that deserved one. The day after, a model read the pages and wrote the answer itself, named two or three brands in it, and the buyer often stopped there. The brand is either in the answer or it is not. There is no second page.

The UBS note reported by Reuters estimated 100 million monthly users in January 2023, two months after launch; the analysts wrote that “in 20 years following the internet space, we cannot recall a faster ramp in a consumer internet app” [3]. TikTok had taken about nine months and Instagram two and a half years [4]. OpenAI then switched to a stricter measure, weekly active users: 100 million in November 2023, 200 million in August 2024, 300 million in December 2024, 400 million in February 2025, 800 million in October 2025, 900 million in February 2026, and “more than 1 billion” by the end of August 2026 [5][6][7][8][9][10][11].

Each point is a company announcement, not a continuous series; the August 2026 figure is OpenAI's own "more than 1 billion", plotted at 1,000.

Source: OpenAI announcements as reported by The Verge (Nov 2023), Axios (Aug 2024), The Verge and CNBC (Dec 2024), CNBC (Feb 2025), TechCrunch (Mar, Aug and Oct 2025, Feb 2026) and OpenAI (31 Aug 2026, "more than 1 billion"). Redrawn by Brave New Media.

Unit: millions of weekly active users

Show the data

Month announced	Weekly active users (millions)
Nov 2023	100
Aug 2024	200
Dec 2024	300
Feb 2025	400
Mar 2025	500
Aug 2025	700
Oct 2025	800
Feb 2026	900
Aug 2026	1000

ChatGPT2

TikTok9

Instagram30

Facebook54

Twitter60

ChatGPT's 100 million is a monthly-user estimate from Similarweb; Facebook ("around four and a half years") and Twitter ("over five years") are rounded from The Verge's wording and should be read as approximate.

Source: UBS analyst note, February 2023, with Similarweb and Sensor Tower data, as reported by Reuters and The Guardian; Facebook and Twitter figures from The Verge, November 2023. Redrawn by Brave New Media.

Unit: months

Show the data

Service	Months to 100 million users
ChatGPT	2
TikTok	9
Instagram	30
Facebook	54
Twitter	60

Google answered inside search itself. On 14 May 2024 Elizabeth Reid wrote that “AI Overviews will begin rolling out to everyone in the U.S.” [14]. Semrush tracked more than ten million keywords through 2025 and found the share of US desktop queries triggering an AI Overview went from 6.49% in January to 13.14% in March and 24.61% in July, before settling at 15.69% in November [15]. Google is still deciding how often to show one. Plan for the range.

The study publishes only these five months as figures in text; the gaps are real and the designer should plot points, not a smooth curve. Semrush reports the keywords triggering an Overview were 89% informational in October 2024 and 57% in October 2025.

Source: Semrush, AI Overviews study, 10M+ keywords, Semrush US desktop database, January to November 2025, updated 15 December 2025. Redrawn by Brave New Media.

Unit: % of queries

Show the data

Month (2025)	% of queries with an AI Overview
January	6.49
February	7.64
March	13.14
July	24.61
November	15.69

Then the clicks. Pew Research Center watched 900 US adults make 68,879 Google searches in March 2025. An AI summary appeared on 18% of them, and when it did users clicked a result link on 8% of visits against 15% when it did not, clicked a link inside the summary on 1% of visits, and ended their browsing session 26% of the time against 16% [16]. Gartner had already forecast in February 2024 that traditional search engine volume would drop 25% by 2026 as chatbots and virtual agents took share [18]. Nothing in the sources here says whether it has. Semrush and Pew point the same way.

AI summary shownNo AI summary

Clicked a result link8%15%

Ended the browsing session26%16%

An AI summary appeared on 18% of the searches in the sample, and users clicked a link inside the summary itself on 1% of those visits.

Source: Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results, July 2025, n=900 US adults, 68,879 searches, March 2025. Redrawn by Brave New Media.

Unit: % of visits

Show the data

Behaviour	AI summary shown	No AI summary
Clicked a result link	8	15
Ended the browsing session	26	16

The money did not leave. US internet advertising revenue went from $209.7 billion in 2022 to $294.6 billion in 2025 [20]. What moved is where a brand gets remembered.

Byron Sharp defines mental availability as “the probability that a buyer will notice, recognize and/or think of a brand in buying situations” [26]. For seventy years that probability lived in the buyer’s memory. From 2022 it has a second holder, a model that is asked the buying question and answers from what it has read.

Kevin Lane Keller’s measure of breadth, “the range of purchase and consumption situations in which the brand comes to mind” [27], is what an answer engine tests, one prompt at a time. The engine also borrows Robert Cialdini’s authority and social proof: it favours sources that combine expertise and trustworthiness, and it repeats what many similar others say [28]. Pew found the summaries drew 15% of their sources from Wikipedia, YouTube and Reddit, and 6% from government sites against 2% in ordinary results [16]. The engine cites the reference and the crowd.

The trade named it within a year. In November 2023 six researchers posted a paper titled “GEO: Generative Engine Optimization”, reporting that changes to a page’s content could raise its visibility in generative answers by up to 40% [19]. “AEO”, answer engine optimisation, followed in the agency decks. The acronyms are new. The test is Sharp’s.

202318%

202423%

202534%

202644%

From 2023 to 2025 the question was asked of those who had heard at least a little about ChatGPT; in 2026 the wording also asked about other chatbots, so the last bar is not strictly comparable.

Source: Pew Research Center, Americans and AI 2026, appendix table, surveys 2023 to February 2026, 2025 wave n=5,123. Redrawn by Brave New Media.

Unit: % of US adults

Show the data

Survey year	% of US adults
2023	18
2024	23
2025	34
2026	44

The Mad Men thread

The series puts the reader in a Madison Avenue office in 1960. The record puts a number on how that office was paid. From the 1880s until the 1980s an American agency earned a 15% commission on the media it bought for its client, a system advertisers challenged for decades and that survived a 1956 antitrust consent decree [22]. The agency wrote the message and was paid for buying the audience, both under one roof. Lasker’s Lord and Thomas, Rubicam, Ogilvy, Bernbach’s DDB, Burnett and Bates argued about the message and sent the same shape of invoice.

The web broke the invoice once already. On 27 October 1994 HotWired went live with banners from a dozen advertisers, among them AT&T, MCI, Volvo and IBM. The AT&T banner was built by Modem Media, a new shop; several others came from Organic; Ogilvy and Mather held IBM’s business by contract. The holding companies bought their way in afterwards. Modem Media was folded into Digitas, and Publicis bought Digitas for $1.3 billion in December 2006 [23].

Ken Auletta’s word for what came next is frenemies: “a once comfortable agency business is now assailed by frenemies, companies that both compete and cooperate with them” [24].

The 2023 to 2025 record reads like 1994 with larger numbers. In January 2024 WPP told investors its plans included “annual cash investment of around £250m in proprietary technology to support our AI and data strategy”. Publicis announced CoreAI on 25 January 2024, with 300 million euros over three years. Omnicom ran generative AI prompting workshops for its own staff at Cannes in June 2024, and the next January its media group wrote from CES that brands “must rethink marketing for a future where AI, not humans, dictates choices” [21]. As in 1994, the agencies arrived after the buyers and paid to catch up. That is our reading of the dates; the record does not measure the gap.

In 1960 the agency controlled the message and bought the audience. In 2022 the model wrote the answer, and all an agency had left to sell was the work of being worth citing.

Applied: if you are a B2C brand

Say you sell a premium dry food for older dogs, in three European markets, at twice the supermarket price. A shopper opens an assistant and types “what is the best dry food for a senior dog”. Three brands come back with a sentence each.

Your ad budget does not decide which three. The engine assembles its answer from pages it trusts, and four things make a page trustable. First, a definition with a source on it. Your page says what “senior” means for a small breed and a large one, names the veterinary source, and the engine lifts the sentence whole. Then mentions by others, a veterinary association’s list, a retailer’s buying guide, a forum thread where owners compare notes, because the summaries draw on the reference and the crowd [16], and a homepage is neither. Third, an author with a name: a veterinary nutritionist with a registration on the page is Cialdini’s authority in a form a model can read. And data of your own, a feeding trial with its sample size and method.

Applied: if you are a B2B brand

Say you run cold-chain logistics across Europe and your buyers are procurement leads at pharmaceutical companies. One of them types “which cold-chain providers serve pharma in the Netherlands”. The same four tests apply and they are harder, because B2B pages are written for a committee that already knows you, and the engine is a stranger with one question. Your definition is the temperature bands and the regulation you ship under, in one sentence. The mentions sit in dull places: the port’s directory, the trade body’s member list, the pharma association’s supplier register, a named article in a trade journal. Put the operations director’s name on the page as author. Publish lane-level performance data of your own, with dates.

We have done this to ourselves: on 27 September 2026, a one-sample baseline on one engine, grounding on, four prompts. On the three category prompts, one on agencies a B2B company should hire for AI marketing and two on investor communications agencies in two markets, the engine named around ten firms per answer. It named no agency we would call a peer, and it named us zero times. The fourth prompt asked who we are, and it described us accurately from our own site [30]. Ask it who we are and it knows. Ask it who to hire and we are not on the list. The dog food brand has the same gap.

What to do with it

1. Write the ten questions a buyer would type, in their words, and ask each engine you care about once a month. Record three things per answer: named or not, position, and whether the description matches yours. Keep the log.
2. Put a named author with a checkable credential on every page that answers a question, and give each of them one definition the engine can lift in a single sentence.
3. Publish one piece of original data a quarter, with method and sample size. It is the only thing in your category a competitor cannot paste.
4. Count where others mention you in the sources the engines already cite: reference sites, association lists, forums, trade press, government registers. Add one a month.
5. Hold the brand budget. Binet and Field’s split, roughly 60:40 brand to activation, was about memory before an engine existed [29]. The engine tests the same memory. Do not fund a new acronym by cutting the thing it measures.