ChatGPT ads get rewritten by default, and Meta keeps the crawl

Three platforms changed a default this week, and in every case the new default is on. OpenAI will rewrite ChatGPT ads unless the advertiser says no, and Google will pull images from landing pages into new Performance Max campaigns. Meta edited the crawl clause we covered last week, and the part that lets it crawl is still there.

ChatGPT will rewrite ads unless the advertiser opts out

On 20 October OpenAI emailed advertisers that "creative enhancements" will be switched on for new and existing self-serve ChatGPT campaigns. Search Engine Roundtable published the email the same day. The system can rewrite or add headlines and descriptions, translate ad text into other languages, resize images to fit a format, and change the destination URL to what it judges a more relevant page on the advertiser's site. It works from the conversation or the user's interests, plus the ad and the landing page.

The start date depends on which OpenAI page you read. The email says Friday, 30 October; OpenAI's help page, as PPC Land read it on 21 October, says Thursday, 29 October. One toggle in Ads Manager settings, marked Beta, covers all four kinds of change; account and campaign opt-outs already set are kept, and a campaign setting overrides the account. Neither source describes a way to allow translation but forbid a different landing page, nor reporting by version.

Google's defaults move the same way

On 21 October the Google Ads Developer Blog said image extraction will be on by default for Performance Max campaigns created from 4 December 2026, in the API, the Google Ads interface and Editor. Google's AI scans the advertiser's landing pages for high-resolution images and can serve them across Search, YouTube, Display, Discover, Gmail and Maps. The account-level "Dynamic images" switch will no longer decide this for new campaigns; existing campaigns keep their settings.

The sourced-video change was reported by PPC News Feed, which says the feature used to be off by default; we have not seen a Google announcement. For API users Google adds a warning worth repeating: the campaign's automation settings are an array that is "replaced" rather than merged on update, so opting out of image extraction without re-sending the other entries can clear them. Google also asks advertisers to confirm they hold the rights to the images on their landing pages.

Meta takes "train" out of its crawl clause and keeps the crawl

Last week we reported that Meta's revised Commercial Terms, effective 30 October, let it crawl any website associated with a business that uses Meta products and use the content to "provide, develop, improve, research, train, and maintain" its products. Nieman Lab reported on 17 October that Meta had changed the wording, and Search Engine Journal compared the versions on 19 October. Crawling is now allowed "in order to provide, develop, improve, research, and maintain advertising and business products". The right to crawl, to use automated software and "test credentials", and to fetch updates periodically stays. The terms still do not mention robots.txt.

Legal experts were not reassured. James Grimmelmann of Cornell Law School told Nieman Lab the purposes are "still broad enough to drive a truck through", and another lawyer noted that "develop" can cover training. Meta told Nieman Lab that the first version "wasn't clear" that robots.txt preferences are respected. On 19 October the UK English version of Meta's Self-Serve Ad Terms preview still carried the old "may use any content" wording, Search Engine Journal found; the US English one did not.

Google: robots.txt governs access, not use

On 21 October Google Search Central published the second episode of its video series The Non-Canonical Cabinet, in which Gary Illyes and Cherry Prommawin answered what a robots.txt file can hold. Illyes called it "an access preference protocol" under RFC 9309. A disallow rule stops a crawler from fetching a page; it says nothing about content already fetched. To block an AI crawler, name its user-agent token: a parser silently drops lines it does not understand, so an invented rule does nothing.

PPC Land points out that Google's own Google-Extended token sits uneasily with this: it has no crawler of its own and exists only to say how Google may use content it already has. The same day Google updated its help page "Control what you share with Google Search" to list the robots meta tag, data-nosnippet, X-Robots-Tag and what it calls the Search generative AI control. Google's changelog says these are existing options, not new ones.

Cloudflare helped OpenAI find hundreds of millions of pages

In a post dated 15 October, written by one Cloudflare and two OpenAI staff and covered by Search Engine Journal on 20 October, the companies said a research pilot "has enabled OpenAI to discover and index hundreds of millions of new web pages" for ChatGPT search. Cloudflare's network signals, from about 1% of its network, told OpenAI's crawler which pages exist, which are dead, redirected or duplicated, and which have changed. They gave no figures for how often those pages were cited or how much traffic they sent.

The post does not say how sites joined the pilot. OpenAI's documentation still separates OAI-SearchBot, for search, from GPTBot, for training; blocking OAI-SearchBot keeps a site out of ChatGPT search results. Pages with few links pointing at them, such as a manual or a menu, are the examples the post gives of what network signals can surface.

DeepSeek opens its search engine to developers

On 17 October DeepSeek released a Search API and a Fetch API. Search returns each result with an authority tier from 0 to 4 and a rerank score from 0 to 3 that combines relevance, authority and freshness; Fetch returns a page as Markdown. DeepSeek says the index holds "hundreds of billions" of pages and that it filters SEO spam, "GEO manipulation" and low-quality AI-generated content, without defining the second. Both APIs cost $0.75 per 1,000 successful requests; calls from Claude Code, Codex or DeepSeek's own harness are free until 31 October.

The launch pages name no crawler for the index and no user agent for Fetch, so robots.txt has nothing specific to address. Search Engine Journal notes that a Chinese search company, Bocha, describes itself on GitHub as DeepSeek's official web search supplier; DeepSeek's pages do not mention it. For anyone measuring visibility in agents, this is one more index with its own idea of authority, and coding agents can now query it directly.

A made-up concept reached AI answers in a week

A preprint posted to arXiv on 8 October by Qi Liu and colleagues, which Search Engine Journal attributes to Fudan University and covered on 19 October, first mapped 17,211 citations across ten AI search tools in March. Then, in April, the team posted about invented subjects, including a "Cactus Gravity Protocol", on sites those tools cite often and a new account can post to, such as Medium, YouTube and Sohu. Before posting, no tool had anything informative to say about them. Within seven days, eight of the ten were citing the posts. The paper has not been peer-reviewed.

Where the post went mattered more than how many there were. Forty posts on a self-hosted WordPress site drew no reference in seven days; a post on a source the tools already cite drew the first reference every time it was tested. A $14 posting package from a GEO vendor got one post cited by Doubao within an hour, which the authors present as one instance, not a rate. The subjects were fictional, so a real brand with real coverage may be harder to move, and the tests predate Google's 15 May statement that its spam policies cover attempts to manipulate AI responses.

The lesson cuts both ways. A brand's citation share can shift because of what strangers post on Medium or YouTube, so the sources an assistant cites for your topic are worth watching, not only your own pages. And a sudden citation from a site nobody in your market reads is a reason to check who wrote it.

Search fell to 12% of North American publisher traffic

Chartbeat's Q3 2026 audience report, published on 15 October and read by Search Engine Journal on 20 October, puts search at 12% of North American publisher page views, against 21% in its Q3 2025 report. Chartbeat says overall traffic "remains steady", gives no absolute numbers and groups all search engines together, so the fall cannot be pinned on Google alone, and a falling share can also mean other sources grew.

Nine of ten regions fell year on year, the Middle East by 9 points like North America; Africa was the only one to rise. Chartbeat's Q4 2025 report blamed the decline on "AI's growing impact on search"; the Q3 2026 report names no cause.

Patching PHP: serve a 503, not a 404

On Wednesday, 21 October, PHP shipped emergency fixes for two critical flaws that let attackers run code without logging in. Google Search Central posted on LinkedIn the same day that a site taken offline to patch should answer with HTTP 503, and John Mueller repeated it on Bluesky: not a timeout, a 404 or a 500.

Also this week

• Volatility, unconfirmed. Search Engine Roundtable reported tracker swings and forum chatter on 20 and 21 October. Google confirmed nothing, and Glenn Gabe noted that ahrefs and Semrush placed the drops on the wrong dates.
• AI Overviews without a label. On 20 October Glenn Gabe spotted AI Overviews for shows and films on desktop with no "AI Overview" label; mobile still showed it. Barry Schwartz suspects a bug and asked Google.
• Discover's staleness filter. Search Engine Journal described a research paper on Supersession-Decay Filtering, which removes outdated content before Discover ranks it. The paper reports 54.9% fewer complaints about stale content over two years of deployment; we have not read the paper itself.
• Smart Bidding learning, rewritten. Google's help page now says the Learning status lasts at most seven days, and that budget, target, creative and targeting changes do not restart it. A new strategy, a change in what the strategy covers or a change of goal do.
• Asset group A/B tests. A "Save asset group changes as an A/B test" toggle was spotted in Performance Max on 21 October; the asset group cannot be edited while the test runs.
• Safari 27 blocks email-based IDs. WebKit said on 19 October that Safari 27 blocks network access to third parties that turn email addresses into persistent tracking IDs. Apple names no companies.
• Meta's ads grow 26%. Meta reported $63.06bn of Q3 advertising revenue on 21 October, the day a New Mexico judge set a $25bn penalty that Meta will appeal.
• "One box." In an interview on Sources with Alex Heath, published on 21 October, Sundar Pichai described agents as the layer that unifies Google's products: "Our vision in the near future is it's one box."
• Financial services verification in Google Ads expanded to the US and Canada on 20 October.

Sources

Every claim above is dated and linked. Where a number comes from a measurement rather than an announcement, whoever measured it is named.

• OpenAI Auto Enabling Creative Enhancements In ChatGPT Ads — Search Engine Roundtable, 20 October 2026
• OpenAI to rewrite ChatGPT ads by default as advertisers get days to opt out — PPC Land, 21 October 2026
• Google turns on PMax image extraction by default from December 4 — PPC Land, on the Google Ads Developer Blog post of 21 October 2026
• Performance Max Sourced Videos Become Opt-Out — PPC News Feed, 21 October 2026
• Meta Commercial Terms — Meta, effective 30 October 2026
• Meta Drops Training From Its Crawl Clause And Keeps The Crawl — Search Engine Journal, 19 October 2026
• Legal experts warn Meta's new terms still grant broad scraping permissions — Nieman Lab, 17 October 2026
• Meta removes language that may have allowed it to scrape news sites — The Verge, 17 October 2026
• Meta web crawlers — Meta for Developers
• What can you actually put in a robots.txt? — Google Search Central, The Non-Canonical Cabinet, 21 October 2026
• Google says robots.txt limits AI crawling, not what bots do with data — PPC Land, 21 October 2026
• Google Updates Control What You Share With Google Search With Existing Options — Search Engine Roundtable, 21 October 2026
• Cloudflare and OpenAI: making AI discovery better for everyone — Cloudflare, 15 October 2026
• ChatGPT Search Can Draw On More Pages After A Cloudflare Pilot — Search Engine Journal, 20 October 2026
• Overview of OpenAI crawlers — OpenAI
• DeepSeek Search API — DeepSeek API Docs, 17 October 2026
• DeepSeek Opens Its In-House Search Engine To Developers — Search Engine Journal, 19 October 2026
• From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search — arXiv preprint, 8 October 2026
• AI Search Tools Cited A Made-Up Concept Within A Week, Paper Finds — Search Engine Journal, 19 October 2026
• Global audience insights, third quarter 2026 — Chartbeat
• Search Share Fell To 12% For North American Publishers: Chartbeat — Search Engine Journal, 20 October 2026
• Google Reminds Sites To Serve A 503 During PHP Patch Downtime — Search Engine Journal, 21 October 2026
• Reduce the Google crawl rate — Google Search Central
• Google Search Ranking Volatility & Chatter Heats Up October 20 & 21 — Search Engine Roundtable, 21 October 2026
• Google Removes AI Overviews Labels On Some AI Overviews — Search Engine Roundtable, 20 October 2026
• Details Emerge About Google Discover's SDF Freshness Algorithm — Search Engine Journal, 21 October 2026
• Google Ads Updates Smart Bidding Learning Doc With 7-Day Limit & Best Practices — Search Engine Roundtable, 21 October 2026
• Google Ads New: Save & Test Asset Group Changes — Search Engine Roundtable, 21 October 2026
• Safari 27 blocks identity firms that turn email addresses into tracking IDs — PPC Land, on WebKit's post of 19 October 2026
• Meta's ad revenue gains 26% on the day a New Mexico judge fines it $25bn — PPC Land, 22 October 2026
• Sundar Pichai on Gemini 4 and the AI compute crunch — Sources with Alex Heath, 21 October 2026
• Google CEO Sundar Pichai Views Search Absorbed Into Agentic Layer — Search Engine Journal, 22 October 2026
• Google Ads Expands Financial Services Verification To US & Canada — Search Engine Roundtable, 20 October 2026