AI now brings customers through three front doors — ChatGPT says it has 1.2 billion weekly users and "about 1/5 of Google's search-type queries," AI-sent orders to Shopify stores grew 3.3x, and Apple opened Siri to every app

On November 3, three companies each put out their own "front door" numbers in a single day. OpenAI's head of ChatGPT consumer product said weekly users went from 1 billion to 1.2 billion in two months and that search-type queries are about 1/5 of Google's (company statement, no underlying data). On its Q3 earnings call, Shopify said visits sent by AI channels were up 2.6x year over year and orders up 3.3x, and that those visitors go straight to product pages at 3x the rate and convert 64% higher (company figures). The same day, Apple opened the Siri AI developer preview to all developers, letting Siri call an app's actions by voice (customer release months away, App Store submissions open in spring 2027). All three are numbers that count "where AI takes customers," but they count in different ways. We wrote down three things a small team that runs both an app and a store should do this week. At the end, we briefly added a research prototype Anthropic released that "proves which model an answer came from."

This piece was written with AI (Claude) and reviewed by an editor. The "Briefly" section at the end covers a research announcement from Anthropic, the company that makes Claude. Our company is a paying customer of both Anthropic and OpenAI, and we run apps and commerce ourselves. All three companies' numbers come from what the companies themselves said in statements, earnings calls, and announcements, so we report them as "what the company said," with no independent verification. Where something has only one source, we name that source alongside it. This is not an investment opinion on any particular company and is not material for investment decisions.

In NO.25 we wrote about what ChatGPT's web share does not count, and in NO.35 about the statement that agents had placed "a few hundred" orders on a delivery app. Both pieces asked the same question: how many customers does AI actually bring? On November 3, numbers answering that question came from three companies on the same day. One is the chat window (ChatGPT), one is the store (Shopify), and one is the phone (Apple Siri). The three numbers count different things, so we separate the facts, the counting methods, and the limits.

1. ChatGPT — 1.2 billion weekly users, "search-type queries are about 1/5 of Google's" (fact)

On the evening of November 2, at a dinner OpenAI hosted in New York, Adam Fry, head of ChatGPT consumer product, spoke with Alex Kantrowitz of Big Technology. Big Technology posted the conversation the next day, and three outlets, Search Engine Journal, Gizmodo, and ShiftDelete, relayed it; we were unable to open the Big Technology original. The numbers Fry gave are these. ChatGPT's "search-type queries" are about 1/5 of Google's search volume, and weekly active users went from 1 billion to 1.2 billion over the past two months (an increase of about 200 million). He said 250 million weekly users use voice input, image generation runs at 1.5 billion images a week, and paid subscriptions passed 50 million "a few months ago." The claim that advertising passed $1 billion annualized within 200 days of launch was reported by ShiftDelete alone.

Fry himself explained the counting method. A "search-type query" is "a question that would previously have gone to a search engine," and "every time you ask a question and it searches the web, that counts as one search." So one conversation can count several, and Gizmodo noted that since ChatGPT now checks the web for most questions, this method can inflate the number. He did not say what Google figure was being compared, or over what period and region. OpenAI has not released the underlying data. For reference, Search Engine Journal placed two other counting methods side by side. Ahrefs's February estimate of ChatGPT search volume, under a different definition, was about 12% of Google's, and Ahrefs's October count of traffic to 118,671 sites came out at Google 24.74%, ChatGPT 0.36%. "The volume of questions received" and "the volume of customers sent to sites" are not the same number. It is the same conclusion we reached in NO.25.

2. Shopify — AI-sent visits 2.6x, orders 3.3x, "the newest front door into commerce" (fact)

These are the numbers Shopify president Harley Finkelstein gave on the November 3 Q3 (July to September) earnings call (checked directly against the Investing.com call transcript and cross-checked with the BigGo Finance summary). "AI traffic is up 2.6x year over year, and orders are growing even faster at 3.3x," visitors from AI channels go straight to product pages at 3x the rate of organic search and convert 64% higher, and "catalog search," which uses Shopify's structured product data, converts at 2x general AI search. In the Q&A he called AI agents "the newest front door into commerce," and said AI volume is still small but the quality of the visits is high (BigGo summary). The line that three quarters of AI-recommended revenue came from outside the top 100 categories appears only in the BigGo summary. Company-wide figures were revenue of $3.7 billion (+32%), gross merchandise volume of $119 billion (+29.5%), and Q4 revenue growth guidance in the "low 30s" percent (company figures).

The names of connected agents also came up on the call: "Muse, Instinct, Claude Duet, and Grok Bot all partner with Shopify," "we power product discovery across every major AI channel, from ChatGPT to Microsoft to Muse," and "we have agents and connectors like Muse, ChatGPT, Claude, and Google Gemini." The structure that makes this possible was unveiled on October 28 in Shopify's Fall Edition for developers. It opens catalog, cart, and checkout to "any agent or app," with a single connection point (shop.dev), a skill shoppers attach to their own agent (shop.md), Universal Checkout for paying for products from multiple brands at once, and browser-agent tooling (WebMCP) built into every storefront and checkout screen so that, "once the buyer approves," the agent can search, fill the cart, and check out. Shopify wrote that every store is "agent-ready by default." Whether a store has to switch anything on separately is not stated in the documentation.

There is one cost number too. CFO Jeff Hoffmeister said subscription-segment gross margin was pressured by "Sidekick," the AI assistant for merchants, and that between Q2 and Q3 merchant conversations rose 14% while cost per conversation fell 28%, bringing total cost down 18%. The BigGo summary attributes this to model selection, distillation, and caching. It is a case of what we saw in NO.21 and NO.30, that "the price of using it goes down," showing up quarter to quarter inside a single company.

3. Apple — Siri knocks on every app's door (fact)

On the same November 3, Apple posted to its developer news that "Siri AI is now open to all developers." It is a developer preview included in the iOS, iPadOS, macOS, visionOS, and watchOS 27.3 betas: when an app hands its content and actions to Siri AI through App Intents, users can ask in their own words for things like "check my delivery," "make a reservation," or "place an order" without opening the app. The new "Skills" bundle several actions so one complex request can be handled at once, and items placed in Spotlight's semantic index can be found by Siri without a fixed template. Apple said the feature will reach customers "in the coming months," and App Store submissions open in spring 2027. The languages supported now are English, French, Japanese, Korean, Portuguese, and Spanish, with Traditional Chinese, Norwegian, Turkish, and Vietnamese to follow. The announcement also lists limits. It is not yet available on iOS, iPadOS, or watchOS in the EU, it is not offered to users under 13, server-side features including Siri AI have a daily usage limit (varying by feature, request complexity, and system demand), and expanded access will be paid in the future. Terms and usage policies are to be published later.

Macworld and Cult of Mac summarized that this preview was previewed at WWDC in June and that third-party app data will now enter Siri's "personal context." They expected the 27.3 release around December (media estimate) and noted that Siri AI itself is not available in China. What we could not confirm from the announcement is how an app decides "what not to send to Siri," and the price of paid expanded access.

Briefly — proving which model an answer came from (fact)

On November 2, Anthropic's Frontier Red Team released a research post on "provable inference" together with a prototype, FlashProve, under Apache 2.0 (reported by TechCrunch on November 3). It does one thing. The model provider first publishes a short "fingerprint" of the model weights, and then attaches a cryptographic proof (a zkSNARK) to every answer saying "this answer came from those weights." The recipient can verify it on a CPU without knowing the weights and without a large computer. It hides the weights and the intermediate computation but not the model architecture. The use cases the post names are preventing a provider from quietly swapping in a cheaper model, letting customers of third-party routing services confirm which model they received, and letting others verify claims like "this code was reviewed by this model." Testing was done on Llama 4 Maverick (4-bit quantized), with one H100 and a 24-core CPU producing about 110 tokens per second, each proof about 19MB, verification taking about 17 seconds on 24 CPU threads, and about 70 to 95x the GPU time of ordinary serving (as published by the company). The limits are also stated in the post. It only covers integer-version models and greedy decoding, context is 1,024 tokens, proofs are generated after the output, it assumes a mixture-of-experts architecture, and side channels such as timing are out of scope. Anthropic wrote that this is "a progress update, not finished research," and the post contains no plan to put it into its own products (TechCrunch also wrote that "no lab has deployment plans"). Since Anthropic is the announcing party, we add no evaluation. We place it in the same line only because, as the "customers sent by AI" in sections 1 and 2 grow, the site side will start asking "which model did that answer come from" (interpretation).

Reading the three stories together (interpretation)

First, all three count "front doors," but they count different doors. Fry's 1/5 is "the volume of questions coming in," Shopify's 2.6x and 3.3x are "visits and orders that arrived at the store," and Apple's preview is "a new entrance into the app." There can be many questions and still few customers sent to sites (Ahrefs 0.36%), and few visits and still high conversion (Shopify 64%). This is why no single number can say "AI traffic is big" or "small." The agent channel we described in NO.35 as "open, but zero" has, by Shopify's measure, become a channel that is "still small but high quality" (interpretation).

Second, the side with higher conversion was "structured data." The number Shopify stressed hardest on the call was not 2.6x but "catalog search converts at 2x." Price, inventory, and safety warnings have to be in a fixed format for an agent to pick that product and carry through to checkout. Apple's App Intents and Spotlight semantic index tell the same story. An app has to expose its actions and data in a fixed format before Siri will call it. "Search optimization" in the agent era is format, not prose, is the sentence shared by all three announcements this week (interpretation).

Third, the doors are open, but the platforms still hold the locks. Shopify's WebMCP checks out "once the buyer approves," Apple has announced daily limits and paid expanded access, and users under 13 and the EU are left outside. Shopify's line "agent-ready by default" says nothing in the documentation about whether a store can turn it off. In NO.36 we wrote that "a confirmation code within an agent's reach is not a confirmation," and these announcements move in the direction of the platform holding that confirmation instead. For a small team that is convenient, and at the same time it is a structure that changes whenever the terms change (interpretation).

If it were us — three things a team running an app and a store should do this week (practice)

Split the "AI path" in the traffic report into three columns, and write the counting method into the column names. What Shopify showed is counting three numbers separately: visits, orders, and conversion. In the traffic table for our site and app, group referrer-based AI channels (ChatGPT, Perplexity, Copilot, Gemini, and so on) into one bucket and create three columns: "visits / direct-to-product-or-feature-page rate / conversion rate." Create the columns even if the numbers are close to zero now. When a company figure like "1/5 of Google" is used in a meeting, always attach the counting method in parentheses ("one count per question that triggers a web search") and put the traffic figure (Ahrefs 0.36%) next to it. Update it once a month in the same table as the "agent channel: 0" line from NO.35.
Put product and feature data into "a format agents can read" first. For the apps of ours that already hold structured data, like ingredients and calories, check whether columns such as price, update date, and safety warnings are missing from that data. On the commerce side, build a checklist comparing whether inventory, price, and shipping terms in the product feed are in machine-readable form against Shopify's catalog fields (3D, video, sale price, safety-warning filter). Keep the flow where an agent proceeds to checkout on the same principle as Shopify's "after buyer approval," and put it in one table with the NO.36 rule "confirmation codes stay on the person's device."
For the Siri preview, start by deciding on "three actions" and "what not to send." A team with an iOS app should first connect only the three most-used actions (for example, look up, save, notify) via App Intents in the 27.3 beta, and since Korean is among the first supported languages, test with Korean sentences. At the same time, write the list of "data we will not send to Siri's personal context" first. The announcement does not say how, so revisit when Apple publishes its terms. In the planning sheet, note the under-13 exclusion, the EU exclusion, the daily limit, and the upcoming paid expansion, and work backward from the spring 2027 submission date. A team with only a web app should look not at this entrance but at browser-agent tools like the WebMCP Shopify uses. Either way, do not report "we opened the door" as "customers came."

What we did not pick this time

The report that CrowdStrike is acquiring AI-agent monitoring firm Fiddler and security-testing firm Dreadnode together for $495 million (SecurityWeek, AI Weekly list) we did not open in full. The GitHub issue claiming Postman Desktop inserts a plugin into Claude Code and Codex settings without permission, and the Hacker News post that OpenAI rolled back Codex's "model improvement" setting, are each a single user report and were left out (if official confirmation comes, they become a follow-up to the "post-install check" in NO.32). The report that DeepSeek is pre-training its next model on 160,000 Huawei Ascend chips (cryptobriefing, one source) and the report that Google said "schools that block AI will lose search too" (NYT, not read) are awaiting cross-checking. The "recursive self-improvement threshold" sentence in Anthropic's amended S-1 (TechCrunch) will be handled in the "price and cost" follow-up after we read the original. The November 2 SemiAnalysis GDPx benchmark stays, as noted in the last piece, as a follow-up on "what the scoreboard measures."

What we could not verify

We could not open the original of Fry's remarks (Big Technology), so we confirmed them through the wording relayed by three outlets. The $1 billion annualized advertising figure is from ShiftDelete alone, and the same outlet wrote that "OpenAI first confirmed 1.2 billion at the end of September," which does not line up in timing with "1 billion to 1.2 billion in two months." We report "two months" as stated and leave the timing unsettled. In NO.25 as well, the dates and definitions for OpenAI's user counts differed from outlet to outlet.
We read the Shopify call transcript directly on Investing.com, but could not open the company's official press release (SEC 8-K exhibit) due to a fetch error. "Three quarters outside the top 100 categories," "subscription-segment gross margin of 79%," and "model selection, distillation, and caching" come from the BigGo summary alone. The definition of "AI traffic" (referrer-based or via connector) and its absolute size are not in the call.
We confirmed Apple's announcement directly in the developer news list, but the documentation page it points to is rendered by script, so we confirmed only the title. The exact timing of "in the coming months," the price of paid expanded access, and how an app decides which data not to send to Siri are not in the announcement. The 27.3 release timing is a media estimate.
All FlashProve figures are from Anthropic's post, with no independent reproduction. Whether Anthropic uses this on its own models is not in the post.

What we read alongside

We opened and checked sources of different kinds directly. They are listed separately as company official documents and earnings calls, research announcements, and media reports.

Company official documents and earnings calls:

Research announcements:

anthropic.comProgress on provable inference for large language models (Anthropic)English; FlashProve's structure, test figures, limits, license, and the "progress update" sentence

Media reports:

Our earlier pieces: