Google vs. Cloudflare AI Adoption: Two Risk Appetites, One Search Box

Is each company training people, training AI, or training people to use AI? Asking that is the quickest way into Google vs. Cloudflare AI adoption, because what a company trains decides what it can afford to lose. Google trains AI: its Gemini models learn partly from public web data, and AI Mode then trains searchers to take the answer instead of clicking the ads beside the links, the habit that paid for all that training. Cloudflare trains people to use AI: Workers AI runs open models that other labs trained (Llama, Mistral, Whisper), and its docs teach developers, and the developers' coding agents, to call them for a fraction of a cent per thousand neurons.

Risk appetite, in simple terms, is how much of what already pays you are willing to break to get what might pay you next. On Thursday 29 October 2026 the market priced Google's answer: Alphabet's value fell from $3.97 trillion to $3.46 trillion, the BBC reported, after its chief business officer told analysts that AI Mode was eating into search ad revenue. Cloudflare's answer costs far less, because a model it never trained and a developer it taught to use that model bring in revenue with every call and take none away from an older product.

For anyone who runs a website, the same split reaches three things you control: the data you publish (which feeds the AI answering your customers), the tools you build on (metered or committed), and the problem you are solving (the page that brings in leads today, which no experiment should be allowed to break).

Alphabet lost roughly $510 billion of value because investors saw Google's own AI Mode thinning out the search ads that fund its AI build-out, in the same quarter that build-out more than doubled in cost. Revenue was not the problem: the Q3 2026 earnings release (28 October 2026) shows $127.6 billion, up 25%, with Google Cloud more than doubling to $30.7 billion.

The trouble sat in the line that built the company. Google Search and other revenue grew 12% to $63.35 billion, and the BBC reported that chief business officer Philipp Schindler named AI Mode, where Google answers a question conversationally, as one reason fewer ads were being served next to results. Schindler also said he felt "very good about our trajectory." Shares fell more than 12% to a six-month low, which is what happens when the market hears "trajectory" and reaches for a calculator.

Free cash flow is the cash left after a company pays for its buildings and machines (Alphabet defines it as cash from operations minus capital spending), so a negative quarter means Google spent more on data centres and chips than its whole business brought in, even with revenue at a record.

Kim Scott, who once led part of Google's advertising business, told the BBC's Today programme that the company would "rather eat its own lunch than allow someone else to eat their lunch," and called the strategy "a long-term bet on AI." Tracy Woo, a principal analyst at Forrester, told the BBC that the disappointing Search results "highlights the challenge of funding massive AI investments while maintaining the strength of established businesses." Amazon reported the same tension the same week: quarterly sales up 11% to $201 billion beside a negative free cash flow figure of about $42 billion, which chief executive Andy Jassy defended by admitting it "would be more convenient" if the spending were spread over 25 years.

Risk appetite for an AI bet comes down to which revenue the bet can damage if it works. Google's AI Mode succeeds by answering the question before the user scrolls past the ads, so every point of success is measured against the company's most profitable line, and Alphabet has decided that losing those clicks to its own product beats losing them to ChatGPT or Claude. Amazon's appetite is measured in cash, a capital plan large enough to push free cash flow below zero while it builds chips, data centres and the Bedrock model platform.

Cloudflare's appetite looks smaller because the downside is smaller. Workers AI, its model-hosting product, is priced at $0.011 per 1,000 neurons with 10,000 neurons a day free, so each request adds revenue and none of it displaces an older product. A neuron is Cloudflare's billing unit for the computing a model does: a bigger model or a longer answer uses more of them, the way a longer phone call used more minutes. Cloudflare's second-quarter 2026 results showed revenue up 36% to $696.1 million and, in chief executive Matthew Prince's words, "record growth in total paying customers, large customers, and developers on our platform." If you want to run a cost-benefit analysis on these trade-offs for your own business, the FAIR model turns "risk appetite" from a board-meeting phrase into a dollar range.

"We had a strong quarter and are now on a half-trillion dollar revenue run rate."

Sundar Pichai, CEO of Google and Alphabet, in the Alphabet Q3 2026 earnings release

Across 108 Google AI Mode answers that PCDrama's AI Citation Collector captured between 27 September and 29 October 2026, 17 (16%) carried a Sponsored block, and all 17 sat on just three kinds of question: hiring a web designer, renting a stadium suite, and one HR query.

Method: 37 collector runs covering 58 distinct search terms, each answer saved as rendered text by our browser extension. An answer counted as sponsored when Google's own "Sponsored" label appeared in it. One AI Mode capture that errored is excluded. The terms came from client and content research, so this is a working sample of commercial and local questions, not a random draw of all searches.

We expected garage door and chimney repair, two of the oldest pay-per-click categories in local search, to carry ads in AI Mode, and they did not: zero sponsored blocks across 24 home-service answers. The web design answers went the other way, with a Sponsored block in every one of the 12, placed after the organic recommendations under a line such as "To compare more options, here are some other web design companies in the Dallas area." Over the same window the five other engines we track (ChatGPT, Claude, Perplexity, Gemini and Grok) returned 494 captured answers between them without a single Sponsored label.

If you keep your own captures, the count is one line of jq per file: jq '[.records[] | select(.platform=="google_ai_mode") | .answer_text | test("Sponsored")] | map(select(.)) | length' run.json. Run it on the same terms every month and keep the result in a spreadsheet column as a key performance indicator, because a single capture tells you about one afternoon and a monthly series tells you whether Google is adding ad slots to your category.

We measured this on live Google answers: for most local service questions, AI Mode currently shows no ad at all, so a business in those trades cannot buy its way into the answer and has to be one of the pages Google cites.

Cloudflare's AI adoption leans on documentation that a developer, or the developer's coding agent, can read and act on in a single fetch. We checked that claim on 4 November 2026 with curl from the command line against the Workers AI REST quickstart and its Google equivalents.

Method: one request per URL from our workstation, with and without an Accept: text/markdown header, recording status, content type and bytes. The command was curl -s -o /dev/null -w '%{http_code} %{content_type} %{size_download}' -H 'Accept: text/markdown' https://developers.cloudflare.com/workers-ai/get-started/rest-api/.

• The Workers AI REST quickstart came back as 5,610 bytes of Markdown when asked for it, against 115,564 bytes of HTML for the same page, about a twentieth of the reading for an agent with a context window to protect.
• The Workers AI section publishes its own llms.txt index (36,311 bytes), a plain-text table of contents written for AI tools to read first, and every page links a "View as Markdown" copy at the top.
• The quickstart's first model call is one curl to /ai/run/@cf/meta/llama-3.1-8b-instruct, after you create an API token and copy your Account ID from the dashboard.
• Google's Gemini API docs now ship a matching kit: a getting-started page that opens by pointing agents at llms.txt and a single-file llms-full.txt, plus an API key that AI Studio creates automatically for new users.
• Vertex AI, Google's enterprise route, served its quickstart as 35,048 bytes of Markdown, but the two llms.txt paths we tried on its docs hosts returned 404.

Two claims that circulate about Cloudflare's docs did not survive the check. The REST route is not zero-config, since you still create a token with "Workers AI Read" and "Workers AI Edit" permissions and keep it somewhere safer than a sticky note; inside a Worker the env.AI binding (a connection Cloudflare wires up for you, so the code asks for the model by name) removes the key from your code entirely, which is where the "no key management" reputation comes from. And Google is no longer behind on machine-readable docs for Gemini, so the adoption gap comes from price and plumbing (a free daily allocation, a binding, one network) more than from prose.

We measured this on the live docs: both companies now write for AI agents as well as people, so the cheaper and simpler first call is what tips a developer toward Cloudflare.

The comparison below puts each company's AI bet next to the revenue it could damage, with Amazon added because it reported the same week and sits between the two on appetite.

Most small businesses should copy Cloudflare's appetite: buy AI in metered units with a ceiling, and never let an experiment put the revenue that pays for it at risk. Google can absorb a $510 billion day; a twelve-person firm cannot absorb a month of runaway API calls or a lead form that an AI answer made redundant.

• Pick metered over committed when you cannot yet name the job the model will do every week, because per-call pricing turns a failed experiment into a small invoice.
• Scope every token to the one product it calls, the least privilege that Cloudflare's "Workers AI Read" and "Edit" pair allows, so a leaked key cannot touch your DNS or billing (our guide to authorization confusion in AI agents covers why).
• Set a spend alert before the first production call, at a figure you would be annoyed but not alarmed to pay, plus a rate limit on the calls themselves.
• Protect the money line first by listing which page brings in leads today, and test any AI feature on a different page so a bad week cannot cost you the one that works.

Your search plan should now aim to be cited inside the answer, because AI Mode names more sources than any other engine we track: an average of 5.5 citations per captured AI Mode answer in the same 108-answer sample, against 3.1 for Claude, 2.8 for Perplexity, 2.7 for ChatGPT and 0.8 for Gemini. Those citation slots, the source grounding behind each answer, are open to any page that says something specific, and on most local questions they currently sit next to no ads at all.

That changes three habits. Pages need a claim nobody else has (a price range, a measured result, a local detail), because a retrieval-augmented answer engine cites one copy of a commodity fact and it is rarely yours; our piece on AI tarpits and AI citations shows what that looks like when it goes wrong. Paid search budgets deserve a fresh audit, since fewer ad slots tend to mean pricier ones and a busier target for the fraud described in ad campaign click fraud. And the site itself has to load fast and read cleanly for both people and crawlers, which is where our web development and secure web solutions work starts.

Google and Cloudflare are both betting that people will ask machines instead of scrolling, and they differ only in how much of their own business they are willing to burn to be right. Your version of that bet is smaller and more personal: which page gets cited when a customer asks an AI about your trade, and what you spend to find out. If you want a second pair of eyes on that page before AI Mode starts selling ad slots in your category, start with the Secure or Replace check or tell us what you are trying to rank for.

Sources: Alphabet Q3 2026 earnings release; Cloudflare Q2 2026 results; Workers AI REST API quickstart; Workers AI pricing; Google, AI features and your website; BBC News reporting of 29 to 30 October 2026 for the market value, Schindler, Scott, Woo and Jassy remarks.