The machines that read the web for AI systems acquired two things on Saturday that they have mostly lacked: a published price and a public reputation. One came from a Chinese model developer selling search to agents at a fraction of what Microsoft charges. The other came from the network company that sits in front of a fifth of the web, which said it will now grade automated visitors on how they behave after they are let in.

DeepSeek opened its Search and Fetch APIs to all developers on October 17, PPC Land reported the same day in DeepSeek prices web search for AI agents at $0.75 per 1,000 calls. Each endpoint costs $0.75 per 1,000 successful calls, and billing starts at 16:00 UTC on October 31, 2026, which the company's documentation gives as midnight on November 1 in Beijing. "Search & Fetch APIs are now open to everyone!" DeepSeek posted on X, in the third message of a four-post thread that also launched a new model, DeepSeek-V4.1-Flash-1017. The post drew about 291,200 views.

The product is built for software rather than people. Search returns ranked results with short snippets of about 100 to 200 characters and longer summaries of 600 to 800 characters written for models to read. Fetch returns a single page converted to Markdown. Both use the same bearer-token authentication as DeepSeek's model API, and both are billed per request rather than per token. "Failed requests, empty search results, and timeouts are not billed," the developer guide says. Tool calls made through Claude Code, Codex and DeepSeek's own Harness, which have been free until now, become billable on the same date.

The price is the point. Microsoft retired its Bing Search APIs on August 11, 2025, and pushed developers towards Grounding with Bing at $35 per 1,000 calls, between 40% and 483% more than the old tiers. DeepSeek's rate is about one forty-seventh of that. Brave charges $5 per 1,000 for its Search API. Cloudflare's Web Search API, launched on October 2, resells partners at prices from $0.25 per 1,000 for Ceramic.ai to $5 for Linkup and $7 for Exa. A rate limit of five requests per second per account, per endpoint, caps a single account at roughly 432,000 calls a day, by PPC Land's calculation, and Fetch also carries an undisclosed per-domain limit. DeepSeek's thread claimed "hundreds of millions" of daily requests, a figure that appears nowhere in its documentation.

What the documents leave out matters as much as the price. DeepSeek does not name the user agent its fetcher presents, does not say how it handles robots.txt, and does not explain how it calculates the authority scores it attaches to results, which run from zero to four, with minus one meaning unknown. "Fetching is subject to the target site's availability and access policies," the guide says, without defining which policies it means. The summaries that make the product useful to agents are also the feature most likely to reduce visits to the sites that wrote the underlying pages. Comscore put DeepSeek at 2% of US AI prompt volume in June 2026, double its January share. The company has also faced European regulators: Italy's Garante banned its app on January 30, 2025, and the Berlin data protection commissioner ruled against it on June 27, 2025.

The X post and the developer guide do not quite agree on what is being sold. The post says "per 1,000 calls"; the guide says "1,000 successful requests". The reference page lists three query operators, while the guide lists six, adding language and date filters. None of the documentation pages is dated, and neither the post nor the guide mentions a free tier.

Grading the crawlers

If DeepSeek's launch put a price on reading the web, Cloudflare's put a score on how politely it is done. PPC Land reported on October 17 that Cloudflare Verified bots face trust scores for ignoring site rules, based on an October 16 Cloudflare Blog post by Chris Tsoukaladelis titled "Continuous accountability for Verified bots". The system already runs internally. It keeps watching bots after Cloudflare has approved them, checking two initial signals: whether a bot obeys robots.txt and whether it slows down when a site returns an HTTP 429 "too many requests" response.

"Consistently slowing down or pausing increases trust, while repeatedly ignoring the signal lowers it," the post explains. Results feed a "trust band" that Cloudflare plans to publish on its Radar service, and customers will later be able to use the band as a threshold for handling Verified bot traffic. Repeated patterns count; one-off violations do not. Cloudflare's summary of the philosophy is "Verify, but continuously re-verify." Verification, in this framing, establishes identity, while "trust depends on what it does next."

The system addresses the third of Cloudflare's criteria for Verified bots, which it describes as doing no harm. The first two, accurate identity and doing only what the bot says it does, were already checked at admission. The company has precedent for acting on the third: "We de-listed Perplexity's bot from our Verified bot list in 2025," the post notes. Annual submissions to the Verified bot programme have grown about sevenfold since 2023.

The evidence that crawlers ignore site rules has been accumulating for months, much of it covered in PPC Land's earlier reporting. Microsoft Clarity found a 4.56% robots.txt violation rate across 246,035 AI crawler requests. TollBit reported that 15% of AI page fetchers in Europe reached disallowed URLs in the first half of 2026. HasData found that 234 of 592 sites that disallow OpenAI's GPTBot, or 39.5%, still served it a page. The Wikimedia Foundation said 65% of the most resource-intensive traffic to Wikidata comes from bots, and it linked a partial outage on October 5 to automated agents. Matthew Prince, Cloudflare's chief executive, has forecast that bot traffic will reach about 1,000 times human volume within five years.

Cloudflare places the trust system in the middle of a loop. Bot Preference Sync, launched on August 21, publishes what sites want; the trust system checks whether bots comply; Bot Management enforces. Much is still undefined. The post does not say how many bands there will be, over what period behaviour is observed, what thresholds apply, whether a bot can climb back after a bad score, how operators will be told, how agents acting for individual users will be treated, or what "handled" means in practice: blocking, challenging, rate limiting or simply logging. No date has been set for the Radar scores or for customer thresholds.

Read together, the two stories describe a market forming in real time. Agents need to read the web, and DeepSeek is selling them the means at $0.75 per 1,000 pages. Sites need to know which readers to admit, and Cloudflare is building the credit rating. Neither company has yet said how the money and the scores will connect to the publishers whose pages are being read.

A data settlement with an audit price

Data access was also the subject of a settlement in Texas, with a price attached to failure. PPC Land reported on October 17 that Temu faces $8m per failed audit under Texas data settlement. Texas Attorney General Ken Paxton and the two companies behind Temu, PDD Holdings, incorporated in the Cayman Islands with offices in Dublin, and Whaleco, the Delaware company that operates the app from Boston, signed an agreed final judgment on October 15 settling a Deceptive Trade Practices Act suit filed on February 19, 2026, in Collin County. The defendants deny wrongdoing. Paxton's office published the settlement on October 16.

The 27-page judgment requires Temu to store Texans' data on US servers owned by an American company and to bar people based in China from reading identifiable data in readable form, with exceptions. It must pay $3 million within 30 days, covering penalties, fees and costs without a split, and pass annual app-security and sensitive-data audits. Each failed audit costs $8 million. A missed termination of a third-party store costs $600,000, and other violations up to $10,000 each. Accounts inactive for two years must be deleted, recalls must be acted on within two business days, and Temu takes on product safety duties covering testing, certificates and the removal of goods that exceed federal standards.

Paxton presented it as a sweeping win. "Texans can now rest assured that Temu cannot sell products containing toxic and dangerous chemicals," he said, adding that Temu "cannot harvest Americans' sensitive data". The judgment is narrower. It protects Texas consumers only, identified by IP geolocation or delivery address, and contains no outright ban on toxic chemicals. The press release says the $8 million applies "each year a violation occurs", while the judgment ties it to specific audit findings after an objection process. Temu chooses and pays its own auditors, and the state takes no role in vetting them: "The OAG has no obligation to select, approve, or vet any third party," the text reads. The exceptions to the China access ban include "improving platform functionality", broader than the release's "essential business functions".

The timing is also uncertain. The court's signature date is blank in the released copy, and every deadline runs from it: certificates and a one-click deletion link within three months, US storage and the China access rules within about six months, first audits due on September 1, 2027, if the judgment was signed in October, and expiry after four years. The settlement adds to a list of state actions against Temu, including a Kentucky suit filed in July 2025, a $2 million Federal Trade Commission payment in September 2025, and a €200 million fine under the EU's Digital Services Act in May 2026. Texas secured $1.4 billion from Meta and $1.375 billion from Google in earlier privacy settlements.

Google's channel controls, clarified and complicated

Google's advertising products produced a different kind of story: an official explanation that went further than the official documentation. PPC Land reported on October 17 that Google Ads lets PMax advertisers set Search partner spend to 0%, Marvin says. Ginny Marvin, Google's Ads Product Liaison, used her October 16 Ads Decoded newsletter, which has 225,418 subscribers, to answer seven advertiser questions about the channel prioritisation and budget allocation controls Google released on October 14. Advertisers rank each channel as high or low priority and can cap low-priority channels at a percentage of daily budget, either as account-level defaults or per campaign.

Marvin stressed what the controls are not. They are "budgeting controls and not bidding controls," she wrote. On Performance Max, she said, the Search partner allocation can be set to 0%, an outcome advertisers have sought since Google tested a partner-network switch in alpha in April. To preserve today's Display behaviour in Search campaigns, she suggested setting Display to low priority with a 50% allocation before the existing behaviour ends in February 2027.

Her answers did not always match Google's help page, and PPC Land counted 15 discrepancies. "They are supported via the Google Ads API at launch, with Google Ads Editor support coming soon," Marvin wrote, while the help page says API support is "coming soon". She called the settings "fully compatible" with shared budgets, which the help page says Performance Max is not. "Settings are fully testable using campaign Drafts & Experiments," she added, where the help page lists limits. Her description of the 0% setting omitted the help page's caveat that caps are applied on a best-effort basis. The newsletter said the rollout would take "coming months", and an X post the same day said "coming weeks". She also gave a range not previously published for promotion mode: a ROAS tolerance of between 5% and 90%.

The newsletter follows an October 15 Ads Decoded podcast episode, covered in the previous edition, in which product manager Tal Akabas said target and budget changes do not trigger a learning period. "We've had this in beta. We've been iterating on it with advertisers, and it's rolling out just now," Akabas said of promotion mode in that episode, while the newsletter still described it as a beta.

Smaller changes in Google's tools

PPC News Feed tracked a run of smaller changes on Saturday, most of them spotted by Hana Kobzová. Google made experiments with its AI-driven Search feature simpler to run: AI Max experiments get 1-click setup, PPC News Feed reported at 05:00 UTC on October 17. "Eligible accounts can now use a 1-click setup to create a 50/50 experiment automatically through Recommendations," the item says. Advertisers can also refine the treatment arm by hand at ad group level, turning search term matching or asset optimisation on or off, and the updated setup guide lets them select several campaigns and adds brand settings.

Vehicle advertising widened. Build-to-order vehicle ads are now serving in 9 countries, PPC News Feed reported, in a limited capacity and for a small share of traffic, across the US, Germany, France, the Netherlands, Italy, Spain, the UK, Australia and Canada. Advertisers elsewhere can upload the ads but will not see them serve until the format launches in their country, and the landing page must now match the configurator step that corresponds to the advertised attributes rather than its first step.

YouTube planning gained a benchmark. Reach Planner adds weekly frequency suggestions, PPC News Feed reported on October 17. The tool infers an account's vertical from its customer ID and suggests a higher average weekly frequency when a plan's forecast falls below that vertical's benchmark, with a global average for accounts without one. Google's example puts optimal exposure at 2.7 times a week. Target Frequency and Demand Gen campaigns are excluded.

Measurement set-up got another redesign. PPC News Feed reported that Google updates its conversion recommendations page layout again, spotted by Vojtěch Audy on LinkedIn, with more space and new suggestions for lead form submissions, converted leads and qualified leads, three months after the previous version. Google gave no rollout timing.

The most consequential of the five concerned regulated advertising in AI answers. PPC News Feed reported on October 17 that pharmaceutical ads are being tested in Google's AI Mode and AI Overviews, citing Trishla Ostwal's reporting for Adweek and a post by Barry Schwartz. A Google spokesperson told Adweek the company had "recently expanded" these surfaces to feature certain pharmaceutical products. Advertisers can run standard text and Shopping ads there in sections labelled "Sponsored", in the US only. "This move signifies ongoing testing efforts by Google to incorporate pharmaceutical ads in AI-driven ad environments," PPC News Feed wrote. Prescription drug advertising is legal on television in the US and almost nowhere else, which makes AI-generated health answers carrying drug ads a test case for how far Google will take commercial placements inside its generated responses.

The bubble question

Behind all of this spending on agents, crawlers and AI answers sits the question of whether the money will be repaid. PPC Land reported on October 17 that Ray Dalio says his bubble indicator puts AI level with 1929 and 2000. In a video interview with Nicholas Thompson, chief executive of The Atlantic, published on YouTube on October 14, the founder of Bridgewater Associates said: "This is a bubble measured analogous to the 2000 and the '29 bubble in that order of magnitude." The quote itself does not mention AI, and PPC Land notes the framing comes from the context of the conversation. Dalio did not disclose the indicator's components or weights, and allowed that he may be wrong.

His argument is less about valuations than about what forces holders to sell. Bubbles burst, he said, when paper wealth has to be converted into cash, through wealth taxes or forced selling, and the US government's finances leave little room for error even if AI raises productivity. He described federal spending of about $7 trillion against revenue of about $5 trillion, debt of about six times annual revenue, interest of about $1 trillion a year and foreign buyers absorbing about a third of new debt, all figures he did not source. His remedy is a "3% solution" that cuts the deficit below 3% of GDP through tax rises and spending cuts of roughly 4% to 5% each.

The article sets his warning against the industry's own numbers. Meta's infrastructure chief Santosh Janardhan put 2026 AI spending at "well over" $100 billion, against capital expenditure guidance of $130 billion to $145 billion. UBS estimated that OpenAI and Anthropic will account for 27% of Google Cloud revenue in 2026 and more than 48% in 2027. ChatGPT ad prices, launched at $60 CPMs in February, were reported as low as $25 by April. Employment for workers aged 22 to 25 in AI-exposed jobs fell 16%, in Stanford data cited by an earlier paper. The video had 49,822 views when PPC Land published. Thompson noted that the podcast comes from The Atlantic's business side, not its newsroom.

Streaming before cinema

The day's media story showed Amazon arranging a release window around its advertising business. PPC Land reported on October 17 that The Chosen season 6 streams on Prime Video 4 months before its finale film. Prime Video and 5&2 Studios said on October 15 that the first three episodes stream on November 15, followed by one a week until December 6, with a stand-alone theatrical film closing the season on March 12, 2027. Season five did the reverse, opening in cinemas from March 27, 2025, and reaching Prime Video on June 15. The release gives no reason for the change.

The schedule puts one of the most successful independent faith series on Prime Video's ad-supported tier through the holiday season. Prime Video reaches 315 million ad-supported viewers, up from 200 million, and Amazon's advertising revenue reached $19.8 billion in the second quarter of 2026, up 26%. The ad-free add-on, renamed Ultra in March, rose from $2.99 to $4.99 a month in April. The show began as a crowdfunded project, raising $11 million from more than 16,000 investors in its first round, before a 2025 deal with Amazon MGM Studios covering streaming of seasons one to five and theatrical and streaming rights for seasons six and seven. Its ten theatrical releases have grossed $120,855,276 combined, and producers estimate its audience at 280 million, a figure that has not been independently verified. It is not clear whether the film will later stream on Prime Video.

The common thread

Saturday's news turned on who pays for access and who is trusted with it. DeepSeek set a price for reading the web on behalf of agents, while Cloudflare began keeping score of which readers respect the rules. Texas put an $8 million price on each failed data audit, but let Temu choose the auditors. Google's channel controls let advertisers pay nothing to Search partners, though only on a best-effort basis that its own liaison left out. Google also began testing drug ads inside generated answers. Ray Dalio warned that the money funding all of it looks like 1929, and Amazon moved a hit series onto its ad tier for the holidays. Each story is a negotiation over terms, and in most of them the terms were set by the party collecting the data or the money.

By the numbers

$0.75
DeepSeek's price per 1,000 successful Search or Fetch calls, billed from October 31, 2026.
DeepSeek prices web search for AI agents at $0.75 per 1,000 calls

234 of 592
Sites disallowing GPTBot that still served it a page, in HasData research cited on Cloudflare's trust scores.
Cloudflare Verified bots face trust scores for ignoring site rules

$8 million
Penalty Temu pays for each failed annual data audit under its settlement with Texas.
Temu faces $8m per failed audit under Texas data settlement

0%
Search partner budget allocation Performance Max advertisers can set, per Google's Ginny Marvin.
Google Ads lets PMax advertisers set Search partner spend to 0%, Marvin says

315 million
Ad-supported Prime Video viewers The Chosen season six will reach before its March film.
The Chosen season 6 streams on Prime Video 4 months before its finale film