The Second Cut: Search, Answer, Perform

Lesson · The Second Cut September 2026

Search, Answer, Perform

The first cut asked what a company believes AI is. The second cut asks what the company delivers — and one company has answered in public, three times, in front of everybody.

  • Follow-on to The First Cut
  • Case: Alphabet
  • Reading time 12 min

The correction most commentary missed

Every day  →  every week

In August 2025, Google’s head of Search said the company sends “billions of clicks to websites every day.” On the Q2 2026 earnings call, its CEO said Google is “sending billions of clicks to websites every week through AI features in Search.”

Same noun. Different denominator, and a narrower surface. Nobody was caught in a lie. The unit simply moved, the way units move when a business is changing what it delivers.

Contents
  1. Count what arrives
  2. Generation one: a location
  3. Generation two: a statement
  4. Generation three: a result
  5. Used boats near me
  6. The seam, and what is not there
  7. The seller needs an entity too
  8. Apply the first cut to Google
  9. Why this matters to the student

The first cut asked what a company believes AI is. Tool or entity. That question gets answered mostly in language — in verbs, in job postings, in the organizational chart leadership draws when you ask them to describe the company five years out.

The second cut is answered in shipping.

Ask a different question:

What does this company deliver?

Not what it says. Not what it sells to advertisers. What actually arrives when the work is done.

There are only three answers I have ever seen. A location. A statement. A result.

And there is exactly one company that has given all three answers in public, in order, over twenty-eight years, while everyone watched. You almost never get to see an ontology move. Usually it happens inside a private company, or inside a founder’s head, and you infer it later from the products.

Google did it on the record.

01 / The diagnosticCount what arrives.

This test is cruder than the verb test from the first cut. It is also much harder to fake, because it does not depend on what anyone says.

You ask a question. Something comes back. Count what arrives.

A location arrives. Ten blue links. The answer is probably over there — go get it. The company has told you where to do the work. You do the work.

A statement arrives. A paragraph. A summary. A number with a footnote. The company has done the reading and handed you the conclusion. You still have to act on it.

A result arrives. The table is booked. The shop has been called. The order is placed. The company did the thing.

Now notice where you are standing in each case.

In the first, you are the researcher and the buyer. In the second, you are only the buyer. In the third, you are the principal, and something else did both jobs.

The first cut asks who owns the verb. The second cut asks what lands on the table.

02 / Generation oneGoogle sold you a location.

In 1998, two Stanford graduate students published a paper describing a search engine prototype that indexed, in their words, at least 24 million pages. The insight was that links carry judgment: a page that many good pages point to is probably good.

The output of that system is a list of addresses.

Understand how modest that is as a product claim. Google was not telling you anything. It was ranking the places where someone else had told you something. The mission language — organize the world’s information and make it universally accessible and useful — is a librarian’s mission. Organize. Make accessible. Not answer.

Under the first cut, generation-one Google is the purest tool ever built. You picked it up. You typed. It extended your reach. It had no job; you had the job. Nobody has ever opened a search box and said “you are my research analyst.”

And the money followed the tool. Google delivered a location to you and sold the click to somebody else. That is the whole machine: the user gets a destination, the advertiser buys the traffic, and the open web gets paid in visitors.

One honest note, because the record does not support the tidier version. Google never described its own purpose as “sending people to pages.” That framing is retrospective, mostly ours. The closest thing in Google’s own voice comes much later, in 2021, when its public liaison for Search wrote that the results page “used to show 10 blue links” and by then showed an average of 26 links on mobile — and argued against going back.

03 / Generation twoThen Google started answering.

The turn is easier to date than people assume. May 2012, Knowledge Graph. The announcement title was “things, not strings,” and the number buried in it is the one that matters: the panel Google showed for Tom Cruise, Google said, answered 37 percent of the next queries people asked about him.

Read that as a business statement rather than a product statement. Google had measured how often it could make the second search unnecessary — and published the figure as an achievement.

Featured snippets followed in 2014. Then, in May 2024, AI Overviews. Then AI Mode, first as a Labs experiment in March 2025, broadly in the United States that May, and across more than 200 countries by October.

The scale is not in dispute, because Google states it. At I/O in May 2026, Sundar Pichai said AI Overviews “now has over 2.5 billion monthly active users” and that AI Mode had “surpassed 1 billion monthly active users” in its first year. Liz Reid, who runs Search, wrote the same day that AI Mode queries were “more than doubling every quarter since launch.”

What arrives now is a statement. And a statement is a different product with a different cost structure, because a statement does not need you to go anywhere.

What the measurement actually shows

Two credible studies, and they are not saying the same thing, so be careful with them.

Pew Research Center tracked the actual browsing of 900 American adults through March 2025 — 68,879 Google searches, 12,593 of which produced an AI summary. On visits to a results page carrying an AI summary, 8 percent ended with a click on a traditional result link. Without a summary, 15 percent did. One percent clicked a link inside the summary itself. This is observational. Pew makes no causal claim, and the results pages were re-scraped weeks later, so they are reconstructions.

The causal version came in 2026. Saharsh Agarwal and Ananya Sen ran a pre-registered randomized experiment on 1,065 American desktop Chrome users in January and February 2026, using an extension that silently removed AI Overviews for one group. Removing them raised outbound clicks from roughly 0.37 to roughly 0.62 per search. Endline satisfaction was statistically indistinguishable between the groups.

Google disputes the conclusion. In August 2025, Reid wrote that total organic click volume was “relatively stable year-over-year” and that third-party reports rested on flawed methodologies. She published no data with it.

Figure 01

Outbound clicks per Google search, with and without AI Overviews

AI Overviews hidden 0.62
AI Overviews shown (control) 0.37
  • Treatment: summaries removed
  • Control: normal Google
Source: Agarwal (Indian School of Business) and Sen (Carnegie Mellon), randomized field experiment, 1,065 U.S. desktop Chrome users, January–February 2026. SSRN working paper; not peer-reviewed at the time of writing, and I have read it through named trade coverage rather than the paper itself. Bar lengths are drawn to the reported click-per-search figures. The comparison is the authors’ framing, not Google’s.

And here is the sentence I would have the student sit with, because it is Google’s own and nobody made them say it.

“We continue to send billions of clicks to websites every day.”

Liz Reid, VP and Head of Google Search ·

“We’re now sending billions of clicks to websites every week through AI features in Search.”

Sundar Pichai, Alphabet Q2 2026 earnings call ·

Every day became every week. All of Search became AI features in Search. Neither statement is false. The unit moved, because what Google delivers moved.

Meanwhile the revenue did not blink. In the quarter ended 30 June 2026 — the most recent Alphabet has reported — total revenues were $119.8 billion, with Google Search and other at $63.3 billion, which Alphabet’s own release attributes to 17 percent growth in that line.

That is the part people get wrong when they narrate this as decline. Generation two did not hurt Google. It hurt the open web’s share of the outcome while Google’s take went up.

04 / Generation threeNow Google performs.

This is the generation that is actually arriving, and it is much further along than most people outside the industry realize, because it shipped in pieces with boring names.

As of today, in the United States, Google will call a local business on your behalf. Not a demo. Rolled out to Search users, with higher limits on the paid tiers, and unavailable in exactly five states: Indiana, Louisiana, Minnesota, Montana and Nebraska. The permitted actions are a closed list, and Google publishes it: book an appointment, check a restaurant wait time, confirm the price and availability of products and services.

Google will also book the restaurant through OpenTable, Resy or Tock without leaving the answer. It will buy a product and pay for it — agentic checkout, live in the United States and Australia, in English, at merchants that accept Google Pay. In Chrome, a feature called auto browse will navigate a site, filter listings, fill a form, apply a discount code, and with permission sign in using your password manager.

Read that list again and apply the verb test from the first cut.

  • Google calls
  • Google books
  • Google confirms
  • Google buys
  • Google signs in
  • Google escalates

Google owns the verbs now. The search box was a tool for twenty-seven years. It is becoming an entity — assigned an objective, granted permissions, working while you are doing something else, returning with a completed artifact or an exception it could not resolve.

Pichai’s own framing for the year was “the agentic Gemini era.” He was not describing a feature. He was describing a change in what the company delivers.

Figure 02

Three generations, read as a single product question

Generation What arrives Who owns the verb Where you stand What Google monetizes
Search · 1998 A ranked list of addresses. The answer is elsewhere. You. You read, compare, decide, act. Researcher and buyer. The click. Traffic sold to advertisers.
Answer · 2012–2025 A synthesized statement with citations underneath. Google reads. You still decide and act. Buyer only. Still the click, from a smaller share of sessions, against higher query volume.
Perform · 2025– A completed outcome. Booked, confirmed, purchased. Google. It calls, books, buys, escalates. Principal. You set the objective and the limits. The transaction itself — which is what the commerce protocols are for.
Rows one and two are record. Row three is partly record and partly argument: the capabilities named are shipped, but the monetization column is my reading of why Google built commerce and payment protocols, not a statement Google has made about its own revenue model.

05 / The worked example“Used boats near me.”

Take one query and run it through all three generations. I picked a boat because it is a real Lowcountry purchase, it is expensive enough to require judgment, and it sits precisely on the seam between the part of this that works and the part that does not.

Figure 03

One query, three generations

  1. Search

    Fourteen links. Boat Trader. A broker in Mount Pleasant. Two marina pages. A forum thread from 2019. You open nine tabs and start building a spreadsheet nobody asked you to build.

    Arrives: a location

  2. Answer

    A written comparison of hull types for local water, a price band for the size you named, a panel of listings with photographs, and a note that center consoles in this range hold value better than bowriders. Links underneath. You still open the tabs, but fewer of them, and you already know what you are looking at.

    Arrives: a statement

  3. Perform — the part that is real today

    It infers the constraint from what it already knows about you: a trailer you own, a ramp you use, a budget it has watched you observe. It narrows fourteen listings to three. Then it calls the broker — actually calls, by telephone — and asks whether the 2019 hull is still available and what the asking price is. It books the Saturday viewing.

    Arrives: a result

  4. Perform — the part that is not built

    It contacts the man selling his own boat out of his own driveway. It opens at a number below asking. It trades two rounds with him and settles. Nothing Google has shipped or announced does any of this.

    Arrives: nothing yet

Steps one through three describe capabilities Google has shipped as of September 2026, though step three’s calling function is United States only and excludes five states. Step four is the forecast, and it is stated as a gap rather than a promise. The reasoning for why that gap exists is in the next section.

06 / The seamLook at what is not there.

I want the student to develop the habit of finding the missing piece, because the missing piece usually tells you more about a company’s architecture than the shipped piece does.

Everything Google’s agent can do in commerce rides on one of two rails.

The calling agent rides Google Business Profile. That is how it knows there is a number to dial and a business that has consented to receive the call. The buying agent rides Merchant Center and the Universal Commerce Protocol — a product feed, a payments integration, a merchant of record.

A man selling one boat out of his driveway is on neither rail. He has no business profile and no product feed. He is not missing from the system because Google forgot him. He is missing because the system is built out of merchant primitives, and he is not a merchant.

The second absence is sharper. No Google agent negotiates. Agentic checkout is documented to complete a purchase only if the final price is the same as, or lower than, the price you confirmed. If the number goes up, the purchase fails. That is a price-taking agent by design — the exact opposite of a bargaining one.

This matters far past boats. Most of the world’s transactions are not catalog purchases at a published price. They are a contractor’s estimate, a used truck, a lease renewal, a freelance rate, a settlement. The generation-three tooling shipped so far handles the part of commerce that was already structured, because structure is what agents can currently stand on.

Agents arrived first in the places where the price was already printed. The interesting frontier is everywhere it is not.

07 / The inversionThe seller needs an entity too.

Here is where the first cut and the second cut meet, and it is the point of this whole piece.

When your side of a transaction acquires an entity, the other side is forced to acquire one as well. Not out of enthusiasm. Out of defense.

Watch Google build both halves.

On the buyer’s side: AI Mode, agentic calling, agentic checkout, auto browse, a personal agent on the paid tiers. On the seller’s side, from January 2026: Business Agent, described by Google as a conversational experience letting shoppers chat with your brand, like a virtual associate. To qualify, a retailer needs a verified Merchant Center account, at least fifty approved listings and a claimed brand profile.

And between them, plumbing with deliberately boring names. The Agent Payments Protocol, announced in September 2025 with more than sixty organizations, then handed to the FIDO Alliance in April 2026 so it would not be Google’s to own. The Universal Commerce Protocol, announced in January 2026 with Shopify, Etsy, Wayfair, Target and Walmart, and described by Google this month as already enabling direct checkout for hundreds of thousands of brands and retailers.

Stop and see what is being constructed. Not a better storefront. A market in which the buyer is represented by an entity, the seller is represented by an entity, and the protocol exists so the two entities can establish that an instruction was authorized and a price was agreed.

The thing being bought and sold is no longer a tool. It is representation.

That is the sentence I would underline. For thirty years the software industry sold you tools and you did the transacting. The industry now forming sells you a party to the transaction. Google is building both parties, and the standards body in the middle, at the same time.

08 / The hard caseNow apply the first cut to Google itself.

Run the diagnostic on Alphabet and you get an uncomfortable answer, which is why it is a good exercise.

Its shipping says entity, unambiguously. Persistent assignment, delegated outcome, tool selection, acting on a trigger, escalating exceptions, working while you are absent. Every item on the audit list from the first cut is checked.

Its revenue still says tool. Sixty-three billion dollars a quarter, most of it descended from a business model that requires a person to be looking at a page.

A company whose ontology has moved faster than its income statement is in the most interesting condition a large company can be in. It is also the condition in which companies make their strangest decisions, because every quarter the two halves argue.

There is a third party in this argument. In August 2024 a federal court found that Google had unlawfully maintained a monopoly in general search. The remedies opinion came in September 2025, final judgment in December 2025, and notably the remedies were written to reach generative AI products and companies, not only search as it existed in 2020. Google appealed in January 2026, was denied a stay, and the case sits at the D.C. Circuit with argument not yet scheduled as of the government’s July 2026 filing.

The student should notice what that means. A court has already decided that the generation-three transition is inside the scope of the remedy. Whatever happens next, the migration from search to answer to perform is no longer only a product story.

09 / The languageBring your AI.

At Charleston AI we say bring your AI, and the reason is directly downstream of everything above.

If the market being built is one where the buyer is represented and the seller is represented, then the only question that matters for an ordinary person or an ordinary small business is whose entity is representing you.

If you arrive with nothing, you are represented by whichever entity the platform provides — one whose objective function was written by a company that also sells the other side. That may be fine. It may even be good. But you should know that it is the arrangement.

If you arrive with your own, you have a representative that carries your context, your preferences, your constraints, and your history, and whose loyalty is not divided.

The possessive was never decoration. In a market of entities, the possessive is the whole position.

10 / The claimWhere I am, and where I might be wrong.

Already true
Google calls local businesses on a user’s behalf in most of the United States. It books restaurants through partner platforms without leaving the answer. It completes purchases with saved payment credentials in the United States and Australia. It navigates and fills forms in Chrome, signing in when permitted. None of this is a roadmap item. It shipped.
What has to happen for me to be right
The commerce protocols have to grow a primitive for an unstructured counterparty — a seller with no feed, no merchant account and no published price — and a primitive for a proposal that is not simply accepted or refused. Until an agent can make an offer rather than confirm one, generation three is a convenience layer over catalog commerce, not a change in how transactions happen.
Where I am probably wrong
The likeliest error is that I am reading Google’s protocol work as a strategy when it is a hedge. Standards bodies are where companies park things they are not certain about, and handing the payments protocol to FIDO is as consistent with lowering risk as with building a market. The second likeliest error is the boat itself: private-party negotiation may never be agent-mediated at all, because the friction people actually hate is the search, not the haggle — and if that is true, generation three ends where it already is, and I have mistaken a stopping point for a seam.

11 / For the studentWhat to do with this.

Take the second cut into every company you evaluate, not just this one. It is five questions and you can run it from the outside in an afternoon.

  1. What arrives when the work is done?

    A location, a statement, or a result. Ignore the marketing and look at what is actually in the user’s hands at the end. Most companies claiming generation three deliver generation two with a progress bar.

  2. Does the delivery break the old revenue model?

    Real migrations hurt. Google’s answer layer cut into the click it sells. If a company’s new product threatens nothing it already earns, it has not moved generations — it has added a feature.

  3. Is it building the other side of the transaction?

    A company that only builds the buyer’s agent is selling a convenience. A company building the buyer’s agent, the seller’s agent, and the protocol between them is trying to own a market structure.

  4. What rail does the agent ride, and who cannot get on it?

    Every agent stands on some existing structure — a business listing, a product feed, an API, a calendar. Find the rail and you will find the population the product cannot serve. That population is either the next expansion or the permanent limit.

  5. Can it propose, or only accept?

    This is the deepest one, and almost nothing passes it yet. An agent that confirms a published price is doing clerical work. An agent that can make an offer, and be refused, and come back, is doing commerce.

The second cut

Google spent twenty-seven years perfecting the most widely used tool ever built, and it is now converting that tool into an entity in public, at a scale of billions of people, while a federal court watches.

The first cut asks what a company believes AI is. The second asks what shows up when the work is done. Watch the second one, because a company can say anything, and what arrives cannot.

Primary sources

  • Brin and Page, “The Anatomy of a Large-Scale Hypertextual Web Search Engine,” Stanford, 1998.
  • Google, “Introducing the Knowledge Graph: things, not strings,” .
  • Google, “Google Search sends more traffic to the open web every year,” .
  • Google, “Generative AI in Search,” I/O, .
  • Liz Reid, “AI in Search is driving more queries and higher quality clicks,” .
  • Pew Research Center, “Google users are less likely to click on links when an AI summary appears,” .
  • Google Cloud, “Announcing the Agent Payments Protocol (AP2),” .
  • Google, “Agentic checkout and AI shopping,” .
  • Google, “Agentic commerce tools and the Universal Commerce Protocol,” .
  • Google, “Gemini 3 and auto browse in Chrome,” .
  • Google, “Donating the Agent Payments Protocol to the FIDO Alliance,” .
  • Sundar Pichai, “I/O 2026: Welcome to the agentic Gemini era,” .
  • Liz Reid, Google Search at I/O 2026, .
  • Alphabet, Q2 2026 results, Form 8-K Exhibit 99.1, .
  • Google, “Boost your holiday sales with these agentic commerce updates,” .
  • Google Business Help, automated calls and regional availability, accessed .
  • Google Pay Help, agentic checkout availability, accessed .
  • United States v. Google LLC, D.D.C. No. 1:20-cv-3010, remedies opinion and final judgment .

Read next

Author: John Rector

John Rector is a Charleston-based entrepreneur, author, and AI strategist. He co-founded E2open, the supply-chain software company acquired for $2.1 billion in 2025, and in 2026 opened Charleston AI, a 3,000-square-foot lab that helps people and organizations understand and use artificial intelligence. He is the creator of The Reality Equation — a lecture series, book, and curriculum exploring attention, prediction, and how reality is experienced — and the author of more than two dozen books. He writes and speaks widely on artificial intelligence, attention, and the future of human work.

1 thought on “The Second Cut: Search, Answer, Perform

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Discover more from John Rector

Subscribe now to keep reading and get access to the full archive.

Continue reading