The History of AI Search

For about twenty years, searching the web meant one thing. You typed a few words, Google handed back a page of ten blue links, and you clicked one. My whole industry grew up inside that single ritual, reverse-engineering which links Google chose and why.

Then, in the space of roughly three years, the ritual broke. You can now ask a full question and get a written answer back, with the sources folded in and no clicking required. Google does it. So do ChatGPT, Perplexity, and the assistant on your phone. The page of links has not vanished, but it is sliding down the screen, and for a growing share of questions people never scroll far enough to reach it.

This is the story of how that happened, told through the moments and the companies that actually moved it, from late 2019 when Google first taught Search to read a sentence, to 2026 when the answer became the default. I have kept it to what is verifiable. Every date and number below links to its source at the foot of the post, because this shift has been narrated with a lot of hand-waving and I would rather show the receipts.

2.5B monthly users of Google's AI Overviews by mid-2026

800M weekly ChatGPT users by late 2025

8% of Google searches get a result click when an AI summary shows, vs 15% without

<90% Google's global search share, its first dip below 90% since 2015

What changed, in one picture

Before the timeline and the names, here is the whole thing in a single comparison. Same question, sixteen years apart.

2010Ten blue links

how much water should i drink a day

• Mayo Clinic - Water: How much should you drink every day? mayoclinic.org › healthy-lifestyle › nutrition
• Harvard Health - How much water should you drink? health.harvard.edu › staying-healthy
• Healthline - How Much Water Should You Drink Per Day? healthline.com › nutrition
• CDC - Water and Healthier Drinks cdc.gov › healthy-weight › healthy-eating
• WebMD - How Much Water Should You Drink? webmd.com › diet › features
• Cleveland Clinic - Are You Drinking Enough Water? health.clevelandclinic.org › articles

You pick a link. The click, and the visit, go to a site.

2026One answer

how much water should i drink a day

Most adults do well with roughly 2 to 3 liters of total water a day, though the right amount depends on your size, activity, and climate. Much of it comes from food and other drinks, and thirst is a reliable guide for healthy people.

Mayo Clinic Harvard Health CDC

You read the answer and move on. Often, no click leaves the page.

The left side is what my job used to be about: getting your link into that list and earning the click. The right side is the new problem. When the engine reads the sources and writes the answer for you, the visit that used to land on a website often never happens. Everything below is the story of how search got from the left panel to the right one.

The spark: two months that broke twenty years

The spark was not a search product at all. On November 30, 2022, OpenAI put ChatGPT online as a free research preview and, by most accounts, expected a quiet week of feedback. Instead it became the fastest-growing consumer app anyone had measured to that point, with outside estimates putting it near a hundred million monthly users inside two months.¹

The technology had been brewing for years; what was new was the behavior. Millions of people typed a full, messy, human question into a box and got back something that read like an answer instead of a list of places that might contain one. Once that felt normal, every search company was on the clock, because the thing they sold suddenly looked like the old way of doing it.

The adoption did not slow down either. ChatGPT’s weekly user count climbed from about 100 million at the end of 2023 to 800 million two years later, roughly eight times bigger.

Nov 2023

100M

Aug 2024

200M

Mar 2025

500M

Aug 2025

700M

Oct 2025

800M

The timeline: from BERT to AI Mode

The overnight part is a myth. Google had been answering questions directly for more than a decade before ChatGPT arrived. The Knowledge Graph panels showed up in 2012, featured snippets and their coveted “position zero” around 2014, and by June 2019 a slim majority of US Google searches, 50.3 percent, already ended without a single click.²³ The zero-click answer was well underway before any generative model touched Search. The rest of the pieces were laid down over the following years, mostly inside Google, before OpenAI lit the fuse. Here is the sequence that got us here, and you can trace the whole arc in one view before I walk through the turns that mattered.

1. Google Jun 2017

   "Attention Is All You Need"

   Eight researchers publish the transformer architecture. Every model in this timeline is built on it.

2. Google Oct 2019

   BERT comes to Search

   Google calls it the biggest leap in five years, finally reading the intent behind conversational queries instead of matching keywords.

3. OpenAI Jun 2020

   GPT-3

   A 175-billion-parameter model shows that scaled-up language models can generate genuinely useful text on demand.

4. Google May 2021

   MUM unveiled

   At I/O, Google previews a model it says is 1,000x more powerful than BERT, built to answer complex, multi-step questions.

5. OpenAI Nov 2022

   ChatGPT launches

   A free research preview reaches 100 million users in two months, the fastest-growing consumer app to that point, and puts a chat answer in front of the whole world.

6. Perplexity Dec 2022

   The answer engine ships

   A startup launches with a simple pitch: ask a question, get a written answer with citations, skip the ten blue links entirely.

7. Microsoft Feb 2023

   The new Bing

   Microsoft drops an OpenAI model into Bing, the first major search engine to lead with a chat answer, and dares Google to respond.

8. Google Feb 2023

   Bard, announced in a hurry

   Google answers Bing within days, but a wrong fact in the launch demo sends Alphabet's stock sliding the next day. The race is officially on.

9. OpenAI Mar 2023

   GPT-4

   A markedly more capable model raises the ceiling on answer quality across every product that plugs into it.

10. Anthropic Mar 2023

    Claude arrives

    Anthropic launches Claude, which becomes a major answer engine in its own right as it gains web access and search.

11. Google May 2023

    Search Generative Experience

    At I/O, Google previews AI answers inside Search itself, an opt-in Labs experiment that signals where the main results page is heading.

12. xAI Nov 2023

    Grok, wired to real-time X

    Elon Musk's xAI ships a chatbot whose hook is live access to X, turning the platform's firehose into its search advantage.

13. Google Dec 2023

    Gemini, the model behind it all

    Google introduces its multimodal Gemini family, the model brand that now powers AI Overviews, AI Mode, and the rest of its Search AI.

14. Google May 2024

    AI Overviews go live

    SGE graduates to AI Overviews and rolls out to all US users. Within weeks it is telling people to put glue on pizza, and the backlash is loud.

15. OpenAI Jul 2024

    SearchGPT prototype

    OpenAI tests a dedicated search product, making its intent to compete with Google directly impossible to ignore.

16. Answer.AI Sep 2024

    llms.txt proposed

    Jeremy Howard proposes a simple file for telling AI models how to read your site, and a small standards scramble begins.

17. OpenAI Oct 2024

    ChatGPT Search

    ChatGPT gains real-time web search with citations and goes head-to-head with Google as a place people start their questions.

18. DeepSeek Jan 2025

    DeepSeek rattles the field

    A low-cost Chinese model matches frontier performance and widens the pool of capable engines, upending assumptions about who can compete.

19. Apple May 2025

    Safari's searches drop

    Apple's Eddy Cue testifies that searches in Safari fell for the first time in the 22 years he has worked on it, and points at people using AI instead. Alphabet's stock drops about 7.5% that day.

20. Google May 2025

    AI Mode opens to everyone

    Google rolls out AI Mode, a full conversational search, to all US users with no sign-up. The generated answer becomes a default Search surface, not a Labs experiment.

21. Cloudflare Jul 2025

    The web starts charging

    Cloudflare flips new domains to block AI crawlers by default and launches pay-per-crawl, the first big move to make AI companies pay to read the open web.

22. Perplexity Jul 2025

    Comet, the AI browser

    Perplexity ships a browser with its answer engine built in. Microsoft, Google, and OpenAI all follow within months, and the fight moves off the results page and into the browser itself.

23. OpenAI Aug 2025

    GPT-5

    OpenAI merges its fast and reasoning models into one system, lifting the quality of the answers every product built on it can give.

24. US v. Google Sep 2025

    The antitrust remedy

    A federal judge stops short of breaking Google up, pointing to generative AI as the new competitive threat that made harsher remedies unnecessary.

25. OpenAI Oct 2025

    ChatGPT Atlas

    OpenAI ships its own browser built around ChatGPT, with memory and an agent mode, turning the browser into the new battleground.

26. Apple Jan 2026

    Apple buys the answer

    Apple and Google announce that a rebuilt Siri will run on Google's Gemini models, reportedly for around a billion dollars a year. The company famous for building its own tech buys the answer instead.

27. OpenAI Feb 2026

    Ads reach the answer

    OpenAI starts running ads inside ChatGPT's free tiers. Google had already put ads into AI Overviews back in 2024. The business model of the answer engine arrives.

28. Google May 2026

    AI Mode passes a billion

    At I/O 2026, Google says AI Mode has topped a billion monthly users, makes a Gemini model the default, and folds AI Overviews and AI Mode into a single AI search experience.

The groundwork was language understanding. In October 2019 Google brought a model called BERT into Search and called it the biggest leap in query understanding in five years, the first time a large neural language model sat in core ranking rather than off in a lab.⁴ It was invisible to users. It just quietly made Search better at reading the intent behind a clumsy sentence, and it set the direction for everything after.

Then OpenAI turned language models from a research curiosity into something you could build on, first with GPT-3 in 2020, then with ChatGPT. Once ChatGPT hit, the incumbents moved fast and not always gracefully. Microsoft got there first, dropping an OpenAI model into Bing in February 2023 and daring Google to respond. Google announced Bard within days, and a wrong fact in the rushed launch demo sent its stock sliding the next morning. Weeks later, at its May 2023 developer conference, Google previewed the Search Generative Experience, the first time a generated answer sat directly on top of its own results page.⁵

For a year SGE stayed an opt-in experiment. That ended in May 2024, when Google graduated it into AI Overviews and switched it on for everyone in the United States.⁶ The rollout went sideways almost immediately. The system confidently told people to add glue to pizza sauce and to eat a small rock a day, screenshots went everywhere, and Google published a slightly sheepish post explaining that satirical and low-quality sources had slipped through, along with the fixes it was shipping.⁷ The episode made the tradeoff obvious. A search engine that speaks in one authoritative voice owns every answer it gives, including the wrong ones.

The rest of the field did not wait around. OpenAI shipped ChatGPT Search that October, giving its chatbot live web results with citations and going straight at Google’s core business.⁸ Perplexity, which had launched back in December 2022 with citations on every claim, kept growing as the purist’s answer engine. Anthropic gave Claude real web search. In September 2024 a developer named Jeremy Howard proposed llms.txt, a simple file for telling AI models how to read your site, and a small standards scramble began.⁹ Then in January 2025 a Chinese lab called DeepSeek released a frontier-grade model at a fraction of the going cost, which spooked the market badly enough to knock a record 589 billion dollars off Nvidia in one day.¹⁰

The biggest move yet came with AI Mode. Google announced it in early 2025 and opened it to everyone in the US that May: a full conversational search surface, with the ten links tucked behind the generated answer instead of sitting above it.¹¹ By the middle of 2025, AI Overviews alone were reaching two billion monthly users.¹² The experiment had become the front door.

Who’s who now

Six products do most of the work in AI search today. They differ in one way that matters more than the rest for anyone who publishes on the web: how, and whether, they show their sources.

The independents are worth a footnote of their own. Not everyone who tried this survived. Neeva built a genuinely good ad-free search engine, could not pull enough people away from Google, and shut down in 2023, selling itself to Snowflake. You.com pivoted from consumer search toward enterprise. The lesson the survivors learned is that beating Google head-on at general search is close to impossible, so the winners either had a distribution shortcut, like Microsoft and OpenAI, or picked a lane and owned it, Perplexity with citations, Grok with real-time posts on X.

How an answer engine actually builds an answer

It helps to know what is happening behind the box, because it explains why the old playbook stopped working.

A 2010 search was mechanical. Google matched your words against its index, ranked pages using signals like links and relevance, and handed you the list. Your job as a searcher was to read the results and synthesize the answer yourself.

An answer engine does that last step for you. When you ask AI Mode or Perplexity something, it often does not run one search. It quietly fans your question out into many smaller queries, pulls back pages for each, and then uses a language model to read across all of them and write a single synthesized reply, with citations pointing at the pages it leaned on. That technique, retrieving real documents and generating an answer grounded in them, is why these answers are usually current and specific rather than the confident nonsense a raw model produces from memory. It is also why the click is optional now. The engine already did the reading.

The cost: the click that stopped happening

Here is where the story gets uncomfortable for anyone who runs a website. The clearest independent measurement so far comes from Pew Research, which watched the real browsing behavior of thousands of US adults in early 2025. When a Google results page included an AI summary, people clicked through to a traditional link on just 8 percent of searches, against 15 percent when no summary was shown. Only 1 percent clicked a link inside the summary itself.¹³

No AI summary

15%

AI summary shown

8%

Now the honest counterweight, because this is where a lot of writing on the topic tips into panic. AI is also sending traffic. Referrals from AI tools to the top websites were up 357 percent year over year by June 2025, reaching 1.13 billion visits. That sounds enormous until you set it next to Google, which sent those same sites roughly 191 billion. AI referrals were about six tenths of one percent of that.¹⁴ Both things are true at once. AI-driven traffic is growing fast off a tiny base, while the far larger river of search traffic to the open web is starting to run lower. Google itself slipped below 90 percent of global search share in late 2024 for the first time since 2015. That is a real crack, even if the company is a long way from collapse.¹⁵

The people who make the content are not sitting still. Some are suing. The New York Times took OpenAI and Microsoft to court in late 2023 over training on its articles, and in September 2025 Penske Media, owner of Rolling Stone and Billboard, sued Google specifically over AI Overviews eating its traffic.¹⁶¹⁷ Others are signing deals instead, like News Corp, which licensed its content to OpenAI in 2024 in an agreement reported to be worth more than 250 million dollars.¹⁸ Whether they sue or license, publishers are all pressing the same point: if the answer engine profits from their words, they want a cut.

If you only read the scary number or only read the hopeful one, you get the story wrong. The realistic read is that the click is getting rarer and more valuable, and being the source an answer engine cites is becoming its own kind of visibility, separate from ranking.

2025 to 2026: into the browser, and the web pushes back

The story did not pause after AI Mode. The last year has been the busiest stretch yet, and it broke along two fronts.

The first is the browser. For thirty years the browser was a neutral pipe, a window you pointed at a search box. In the second half of 2025 that changed fast. Perplexity shipped Comet in July, a browser with its answer engine wired in as the default. Microsoft turned Edge into an AI browser with Copilot Mode the same month. Google put Gemini into Chrome in September.¹⁹ Then in October, OpenAI released ChatGPT Atlas, its own browser built around the chatbot, with a memory of your browsing and an agent that can click through tasks on your behalf.²⁰ Four of the biggest players moved the fight into the browser itself inside a single quarter, because whoever owns the browser owns the moment you start asking.

The second front is the open web pushing back. If answer engines are going to read every page and hand the summary to users without the click, the sites doing the writing want something back. Cloudflare made the loudest move: on July 1, 2025 it began blocking AI crawlers by default for new domains and launched pay-per-crawl, where a crawler either pays or gets turned away at the door.²¹ Publishers signed licensing deals and pressed lawsuits. And the courts weighed in. In September 2025, ruling in the long-running US antitrust case against Google, a federal judge declined to break the company up and pointed straight at generative AI as the reason: the competitive threat search had lacked for twenty years had finally shown up, which made a heavier remedy look unnecessary.²²

Two more signals showed how far the ground had shifted. In May 2025, testifying in the Google antitrust trial, Apple’s Eddy Cue said searches in Safari had fallen for the first time in the twenty-two years he had worked on them, and blamed people using AI; Alphabet’s stock dropped about 7.5 percent that day.²³ By early 2026 Apple had stopped fighting it and agreed to rebuild Siri on Google’s Gemini models, a deal reported at around a billion dollars a year.²⁴ The answer also started paying for itself: after Google put ads inside AI Overviews in 2024, OpenAI began running ads inside ChatGPT’s free tiers in early 2026.²⁵

By the time Google held I/O in 2026, the direction was settled. It said AI Overviews had reached more than 2.5 billion monthly users, AI Mode had passed a billion, a Gemini model was now the default, and the two had merged into a single AI search experience it called the biggest change to the search box in more than twenty-five years.²⁶ The thing that first appeared as a Labs experiment three years earlier was now just Search.

What this means if you have a website

You do not get to opt out of this, so the useful question is what to actually do about it.

The short version is that ranking and getting cited are now two different jobs. You still want to rank, because AI answers are built from the same pages that rank well. But you also want to be the page an answer engine quotes and links, and that rewards slightly different things: content that states clear, quotable facts, that is well structured, that is genuinely worth citing, and that the AI crawlers are actually allowed to read. I built a couple of free tools around exactly that, including an AI crawler access checker that shows whether ChatGPT, Claude, Perplexity and Google’s AI are even permitted to see your site, an AI content visibility checker that fetches a page the way those crawlers do, and I keep original data on how this is playing out in the research section.

None of this is settled. The names on the timeline will keep changing, a model that looks dominant this quarter can be leapfrogged the next, and the exact mechanics of how answers get built are still moving. What looks durable is the shape of it: people have decided they would rather get an answer than a list, and no one is going to talk them out of it. So the work shifts to being one of the sources those answers get built from.

One last thing, since this whole post is about being readable by machines: it practices what it preaches. The AI crawlers are welcome on this page, it ships an llms.txt and structured FAQ data, and the crawler check I linked earlier comes back clean. A history of AI search ought to survive its own audit.

Common questions about the history of AI search

When did AI search actually start?

There is no single start date, but the practical turning point was ChatGPT's launch on November 30, 2022, which made people expect direct answers. The groundwork was laid earlier, when Google brought the BERT language model into Search in October 2019. Google's own AI answers reached the mainstream when AI Overviews launched to all US users in May 2024.

What is the difference between a search engine and an answer engine?

A search engine finds and ranks pages, then hands you a list of links to read yourself. An answer engine reads across multiple sources for you and writes a single synthesized answer, usually with citations, so you often do not need to click through at all. ChatGPT, Perplexity, and Google's AI Mode are answer engines; classic Google results are a search engine.

When did Google launch AI Overviews?

Google switched AI Overviews on for all US users on May 14, 2024. It was the successor to the Search Generative Experience, an opt-in experiment Google had previewed at its developer conference in May 2023.

What is the difference between AI Overviews and AI Mode?

AI Overviews is an AI summary that appears at the top of an otherwise normal Google results page. AI Mode is a separate, fully conversational search experience where the generated answer is the whole page and the links sit behind it. AI Mode opened to all US users in May 2025.

Did ChatGPT kill Google search?

No, and the data is clear on this. Google still sends far more referral traffic to websites than every AI tool combined, and its search share only recently dipped below 90 percent. What has changed is that AI answers now handle a large share of questions before a click happens. People still run their searches on Google; they just click through to a website less often than they used to.

Is SEO dead?

No, but it has split into two jobs. You still want to rank, because AI answers are assembled from pages that rank well. On top of that, you now want to be the source that answer engines cite, which rewards clear, well-structured, genuinely useful content that the AI crawlers are allowed to access.

What is GEO (generative engine optimization)?

GEO, generative engine optimization, is the practice of shaping your content so AI answer engines like ChatGPT, Perplexity, and Google's AI surface and cite it. It overlaps heavily with SEO, because AI answers are built from pages that already rank, but it leans harder on being clearly stated, well structured, quotable, and readable by the AI crawlers. You will also see the near-identical term AEO, answer engine optimization.

Sources

1. The roughly 100-million-users-in-two-months figure is a UBS analysis using Similarweb data, an outside estimate rather than an OpenAI number. Reported by The Register (Feb 3, 2023). ↩

2. Google, Introducing the Knowledge Graph: things, not strings (May 16, 2012). ↩

3. SparkToro (Rand Fishkin), Less than half of Google searches now result in a click (Aug 13, 2019). ↩

4. Google, Understanding searches better than ever before (Oct 25, 2019). ↩

5. Google, Supercharging Search with generative AI (May 10, 2023). ↩

6. Google, Generative AI in Search: Let Google do the searching for you (May 14, 2024). ↩

7. Google, AI Overviews: About last week (May 30, 2024). ↩

8. TechCrunch, OpenAI launches ChatGPT search (Oct 31, 2024). ↩

9. Jeremy Howard, Answer.AI, The /llms.txt file (Sep 3, 2024). ↩

10. Yahoo Finance, Nvidia loses record $589 billion in market cap as DeepSeek prompts questions over AI spending (Jan 27, 2025). ↩

11. Google, AI Mode is expanding to more people and adding new capabilities (May 20, 2025). ↩

12. TechCrunch, Google’s AI Overviews have 2B monthly users, AI Mode 100M in the US and India (Jul 23, 2025). ↩

13. Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results (Jul 22, 2025). ↩

14. TechCrunch, AI referrals to top websites were up 357% year-over-year in June, reaching 1.13B (Jul 25, 2025). ↩

15. Search Engine Land, Google search market share drops below 90% for first time since 2015 (Jan 13, 2025), citing StatCounter data for Q4 2024. ↩

16. The New York Times Company v. Microsoft and OpenAI, filed Dec 27, 2023; still in discovery as of mid-2026. ↩

17. TechCrunch, Rolling Stone owner Penske Media sues Google over AI summaries (Sep 2025). ↩

18. Variety, News Corp signs content-licensing deal with OpenAI (May 2024); reported to be worth more than $250M. ↩

19. Comet: TechCrunch, Perplexity launches Comet, an AI-powered web browser (Jul 9, 2025). Edge Copilot Mode: Microsoft, Introducing Copilot Mode in Edge (Jul 28, 2025). Gemini in Chrome: Google, New AI features for Chrome (Sep 18, 2025). ↩

20. TechCrunch, OpenAI launches an AI-powered browser, ChatGPT Atlas (Oct 21, 2025). ↩

21. Cloudflare, Content Independence Day: no AI crawl without compensation (Jul 1, 2025). ↩

22. DLA Piper, Federal court orders remedies in Google antitrust case (Sep 2, 2025). ↩

23. Fortune, Apple’s Eddy Cue says Safari searches fell for the first time as users turn to AI (May 8, 2025). ↩

24. Google, Joint statement from Google and Apple (Jan 2026). ↩

25. OpenAI, Our approach to advertising (Jan 16, 2026). Google had already brought ads to AI Overviews: TechCrunch, Google brings ads to AI Overviews (Oct 3, 2024). ↩

26. Google, Sundar Pichai on what we announced at I/O 2026 (May 19, 2026). ↩