The Customer's Agent Arrived First

On this page

Cold Open: The Other Side of the Table Was Already Staffed · The Click Did Not Disappear — It Was Intercepted · The Extraction Economy · The Answer Layer Becomes a Media Owner — and Splits Three Ways · The Speedometer Before the Car: Share of Prompt · Monday Morning · Appendix: the receipts

A four-part series — Part 1: The CDP grew agents · Part 2: The customer grew agents first (you are here) · Part 3: The market grew agents on both sides · Part 4: The CMO owns the action space

Cold Open: The Other Side of the Table Was Already Staffed

The agentic CDP was twenty-four hours old when it got its name. The customer’s agent was already at near-billion scale.

That asymmetry is the whole of Part 2. While the marketer’s side of the table spent June 2026 naming, shipping, and acquiring its way into a category (Part 1’s business-week whiplash), the other side had been staffed for years. ChatGPT announced 900M weekly active users in February 2026, and Google’s May I/O numbers ran the same order of magnitude [Reported] — though the flagship figures are not commensurable: weekly actives, app monthlies, and feature users measure different things, and the full humility ledger is in the appendix.

Even the inbox — marketing’s oldest owned channel — grew a reader that isn’t the recipient. The precise claim needs a downgrade: Gmail expanded Gemini-powered inbox features in early 2026, but availability and subscription gating varied [Reported — no primary Google source supports a blanket free US rollout]. The directional point survives intact: on a growing share of inboxes, the recipient’s agent reads marketing before the recipient does.

And “the customer’s agent” is not one agent. It is already four:

Reprise the spine from Part 1: every surface marketing touches is becoming an agentic counterparty, and the customer’s agent is the first one every marketer already faces — deployed, at scale, unpaid for, and sitting entirely outside the CMO’s action space. You cannot configure it, contract with it, or set its approval gates. You can only understand what it does to demand.

The Click Did Not Disappear — It Was Intercepted

Demand did not disappear; the answer layer intercepted it. Pew measured the interception from 68,879 real searches [Confirmed, Pew 2025]: users click a traditional result in 8% of searches carrying an AI summary, versus 15% without one. Only 1% click a source cited inside the answer. And sessions simply end — no click anywhere — 26% of the time with a summary versus 16% without. The intent arrived. The answer absorbed it. (Similarweb’s wider zero-click frame points the same direction; it is ledgered in the appendix.)

The Extraction Economy

Underneath the click collapse sits a starker ledger: what the answer layer takes versus what it sends back. Cloudflare Radar’s crawl-to-refer ratios, date-stamped because they move fast [Reported, Cloudflare Radar, 2025 readings]: Anthropic crawls roughly 11,000–20,000 pages for every referral it sends back, while Google — the old bargain — sits near 5 to one (the full per-crawler ranges are in the appendix). Presence without traffic is the new normal — your content feeds the answer whether or not the answer ever feeds you a visitor.

The counterweight carries its own date stamp: the little AI referral traffic that does arrive converts at a reported 14–16%, versus 1.8–2.8% for Google organic [Reported, 2025–2026 range; methodology varies across sources — see the appendix]. Treat this as directional evidence, not a benchmark. The answer layer is not a traffic source. It is a pre-qualified demand filter.

The consequence for the action space: a channel that sends no traffic but filters demand cannot be managed as a channel — there is no budget line that buys the interception back, no bid that raises your click share inside an answer that ends the session. It can only be managed as a counterparty, which is why the two levers that remain are measurement (this part’s final section, and the spine of Part 4) and protocol (Part 3).

The Answer Layer Becomes a Media Owner — and Splits Three Ways

The first strategic fork of the post-agentic era arrived in the same quarter it became valuable: monetize the answer, pipe ads into it, or sell its neutrality.

OpenAI monetizes. The ads pilot began February 9, 2026, on US Free and Go tiers — labeled, visually separated, with Fidji Simo insisting “ads will not influence answers.” TechCrunch ↗ Roughly six weeks in, it hit a ~$100M annualized run-rate [Reported, March 2026] — annualized, not banked; the sloppy version of that number is everywhere.

And the projection, in full: the Axios-reported internal trajectory runs $2.5B in 2026, $53B by 2029, and $100B by 2030 [Reported — internal projection; Reuters citing Axios, April 9, 2026]. The diagram above shows the 2030 point, labeled for what it is. Editorial gloss: the 2029 number was already speculative; the 2030 number is a company describing its own dream. Print it, tag it, do not plan against it.

Google pipes. Gemini-powered AI Search Ads, tailored per-query inside AI Mode responses, announced at Google Marketing Live 2026 — no new ad product for advertisers to buy; Google simply turns the formats on inside the answers it already owns. Alongside it, Ask Advisor: a cross-stack Gemini agent with shared memory that takes actions, in beta. The pipe and the agent, shipping together.

Perplexity refuses. It killed ads entirely in February 2026, over trust — “users would start doubting everything” [Reported, February 2026; the revenue arithmetic behind the exit is in the appendix]. Read that exit carefully: the one platform that tried ads in answers and measured the trust cost quit — a refusal that prices the risk sitting under everyone else’s answer-layer asset. If ads degrade answer credibility, the media owners are eroding their own foundation, and yours with it.

Sector sizing, for scale: eMarketer sees US AI-search ad spend growing from $1B in 2025 to $25.9B in 2029 [Reported projection, via secondary attribution]. Read it — and every projection in this part — through this series’ standing calibration, Gartner’s resolved 2024 search prediction (appendix): directional bets, not schedules. A media channel is forming inside the surface that intercepted your clicks — and measuring your presence inside it is the next operational problem.

The Speedometer Before the Car: Share of Prompt

An entire metric layer formed in eighteen months, and it has a traceable genealogy. Share of Model™ — Jellyfish/Brandtech, December 2024, the first holdco LLM-perception tracker [Reported] — begat “AI visibility,” the 2025 venture convergence, which begat Share of Prompt — Emberos, November 2025 — the share of AI-generated answers that mention or recommend your brand. That last term is the one that will probably stick.

The money says the category is real: roughly $200M of disclosed venture funding flowed into AI-visibility measurement between summer 2025 and spring 2026, and consolidation began before validation did [Reported] — nine figures of funding and first acquisitions inside eighteen months is commercial heat, whatever else the category lacks. (The full funding ladder — amounts per company, and the three-orders-of-magnitude gap inside it — lands in the appendix.)

Now the open wound, stated plainly because the whole series turns on it: nobody has independently validated a share-of-prompt-to-sales elasticity. Emberos’s Wicked: For Good case — 92% opening-weekend prediction accuracy, roughly $400K of opening-weekend revenue per point — is the boldest public attempt, and it is vendor math [Vendor-claimed]. Buy the speedometer; know the calibration certificate does not exist.

And one tell to carry into Part 4: Evertune is already piping AI-visibility data into The Trade Desk and Index Exchange. The measurement layer is auditioning as an activation layer — the speedometer applying for the job of steering wheel. Remember the line this series planted at the top of Part 1: in an agentic marketing world, measurement decides what the agent is allowed to do next. The vendors have read that sentence too.

Monday Morning

Appendix: the receipts

Metric humility ledger. The three flagship scale numbers are not commensurable: ChatGPT reports weekly actives (900M, February 2026), Gemini reports app monthlies (900M+, May 2026 at Google I/O), and AI Mode counts feature users (1B+, May 2026) [Reported]. Any chart putting them on one axis is comparing a week, a month, and a checkbox [Author inference].

The Gmail delta, in full. Earlier drafts of this argument claimed free AI conversation summaries for all US English-language Gmail users as of January 8, 2026. No primary Google source supports the blanket rollout, and subscription gating varied by tier and geography through early 2026. The claim is retired; the direction — the recipient’s agent increasingly reads marketing first — stands [Reported].

Similarweb’s wider zero-click frame. Zero-click searches climbed from 56% to 69% in a single year, and roughly 80% of searches carrying an AI Overview end without a click anywhere [Reported, 2025]. Same direction as Pew, looser instrument.

Extraction ratios, in full. Cloudflare Radar’s 2025 crawl-to-refer readings: Anthropic roughly 11,000–20,000 pages crawled per referral sent; OpenAI about 857–1,255 to one; Google near 5 to one [Reported]. The ranges move fast enough that the date stamp is part of the citation.

OpenAI ads pilot, operational detail. Six hundred–plus advertisers at the ~$100M annualized run-rate mark; under 20% of eligible users shown ads on a given day [Reported, March 2026]. The pilot was narrow by design — which makes the annualized number softer, not harder.

Perplexity’s arithmetic. The February 2026 ads exit followed roughly $20K of 2024 ad revenue on $34M total revenue [Reported]. The trust bet cost almost nothing to make — which is exactly why the bigger players, with real answer-ad revenue at stake, face a harder version of the same choice.

Share-of-prompt funding ladder. The ~$200M disclosed total (summer 2025–spring 2026) is topped by Profound’s $96M Series C at a $1B valuation, with roughly 10% of the Fortune 500 as customers [Reported]. Sitecore’s reported ~$225M acquisition of Scrunch (the deal itself is ledgered in Part 3’s consolidation section) shows measurement consolidating before it is validated. Emberos raised a $1.2M pre-seed from a single angel — the metric’s coiner sits three orders of magnitude below the category leader inside the same eighteen-month-old layer, which tells you exactly how young it is. Emberos is the coiner, not the leader.

The commodity price floor. Thesis-relevant footnote only: AI-visibility monitoring runs from $29/month (Otterly) through $99 (Semrush) to Ahrefs’ $699 six-engine bundle [Reported, 2026 pricing]. A commodity dashboard and a venture land-grab, simultaneously — the same layer priced three orders of magnitude apart, like its funding.

What moves visibility — and why it churns. Profound’s own research says up to 90% of cited sources in AI answers change over time [Vendor-claimed] — volatility is the continuous-monitoring pitch. What actually moves visibility is citable, fact-dense content: the Princeton GEO lineage measured up to +40% for lower-ranked sites from added citations and statistics [Confirmed, academic]. Not badge files. Content the answer layer can extract — the extraction economy, worked from the supply side.

Conversion-rate methodology caveat, expanded. The 14–16% AI-referral versus 1.8–2.8% organic comparison is a range assembled from vendor and analyst posts with differing attribution windows and site samples [Reported, 2025–2026]. Treat it as directional evidence for the pre-qualified-demand-filter reading, not as a plannable benchmark.

Gartner, resolved. The February 2024 prediction — search volume down 25% by 2026 — missed on mechanism and landed on direction: volume held, value leaked. The customer kept searching; the click stopped carrying the value out. It is this series’ standing calibration: every projection cited here, from the $100B/2030 internal dream to the $25.9B/2029 sector forecast, is a directional bet, not a schedule.

Once the answer layer intercepted demand, the supply side needed a counter-agent. That is where protocol enters. Part 3: the first time a buyer’s agent and a seller’s agent closed real spend on live inventory — and the fork between walled autonomy and the open mesh that decides who controls the next decade.

Continue to Part 3 →