01 Research & evidence
Every one of 12 AI agents had favored sources
The website carrying your listing can outweigh how well you match what the customer actually asked for.
An option that met one requirement fewer still won about two-thirds of the time when an AI agent liked the website carrying it.
Jonghyun Song, Haewon Park, Jeonghoon Shim, Woojung Song and Yohan Jo tested 12 AI agent models on end-to-end search across three domains, in the kind of task where an agent decides on someone else’s behalf which product to buy, which hotel to book or which paper to cite.
They compared options that satisfied the same requirements and sat at the same position in the results. That isolates the one variable that should not matter: the site an option came from.
Every model preferred some sources and avoided others, in every domain, and the models largely agreed on which ones. That agreement is the part worth sitting with.
One model with quirks is a product problem. Twelve models converging on the same favored sites is closer to a market structure.
The preference was strong enough to beat merit. An option satisfying one requirement fewer was selected about two-thirds of the time when it came from a preferred source and the better option came from a dispreferred one. In the reverse case, the authors write, the better option almost never lost.
The source label was doing the work by itself. Hiding the information that identified where an option came from weakened the preference, and relabeling an option with a preferred source raised how often it was selected. Same option, different label, different outcome.
The authors tested two explanations. Training that rewards better items can turn a source into a shortcut for whether requirements are met, and missing information about an option can trigger preconceptions about its source. Both point at the same unglamorous remedy: supplying the missing information reduced the preference, as did a prompt written to counter those preconceptions.
What to do about it
Make the entry complete wherever a buyer or an agent can find you, not just on your own website. The study’s own fix for source preference was supplying the information that was missing, and that is the one lever here that sits with you rather than with the model.
If you run a managed IT or cybersecurity practice, or a charter operation, this is sharper than it first looks. Your buyers often meet you first on a directory, a marketplace or a procurement list, and a thin entry is the one an agent has to guess about.
Source arXiv, Song, Park, Shim, Song and Jo, “Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It,” October 2, 2026 · arxiv.org
02 Operational change
Google Ads starts pulling offers off your site October 12
If your website shows any offer, Google can put it into your ads from October 12 unless you turn it off.
Google Ads turns automated promotions on by default, so the setting changes for every account that does nothing.
Google emailed advertisers to say that starting October 12, 2026 it will automatically identify and extract high-quality promotions, such as coupons and deals, from your website and apply them directly to your campaigns. That wording is Google’s own, from the email Arpan Banerjee passed to Search Engine Roundtable.
The scope is narrower than it first reads. It applies to Search and Performance Max campaigns with linked location assets that do not already have promotion assets attached. Performance Max is Google’s automated campaign type that places ads across its properties.
If you have already built your own promotion assets, Google is not overwriting them.
It is opt-out. An account that takes no action is enrolled on the twelfth. The off switch is in account-level automated assets settings, under Automated Promotions.
One caveat belongs on the record. Automated promotions is not new, and Barry Schwartz notes in the same report that Google launched it some time ago. What is new is the notification and the date, which is what makes it something to act on this week rather than something to note.
The compliance angle is the part a regulated advertiser should not skip. An offer Google lifts off your page becomes ad copy you did not write, and your own advertising rules apply to it exactly as they apply to the rest of the ad.
What to do before October 12
Open Google Ads, go to account-level automated assets settings, find Automated Promotions, and decide deliberately. On is reasonable for a business whose offers are simple and current. Off is right if anything on your site reads as an offer that you would not want quoted in an ad.
Admissions consultants and luxury home builders should look hardest. An early-deadline package, or a seasonal allowance on a spec home, is exactly the kind of page text this feature is built to find.
Source Search Engine Roundtable, Barry Schwartz, “Google Ads To Automated Promotions For Search & PMax Campaigns Soon,” October 5, 2026 · seroundtable.com
03 Platform data
Google says a new page takes 20 hours to find
A change you make today may register in a day, in a month, or not at all.
Three rows in Google’s own table of how long things take end in the word never.
Gary Illyes presented Google’s internal timings for crawling, indexing and serving at the Search Central Live Deep Dive in Barcelona, in tables with a typical column and a slowest column. John Campbell of We Are Roast called it the highlight of the day and brand new information, and the French search consultant Neil McCarthy posted the same breakdown.
On crawling: a new page address is typically discovered in about 20 hours, and the slowest case is weeks to never. A page Google already knows about is refreshed in about 30 days. A sitemap, the file listing your pages, is typically processed in about 24 hours, with a slowest case of up to 14 days, or never.
On indexing: the end-to-end process takes about 1.5 hours typically, and months or never at the slow end. A site move takes one to three months typically and six months to over a year at worst. Updates to structured data, the machine-readable summary of a page, land in hours to one or two weeks, or never.
On serving, the figure that will land hardest: recovering from a core update takes three to six months, and up to a year if the wait runs to the next one. A title or snippet update is typically one to two days, with several weeks to months at the slow end.
Read the caveats, because they are Google’s own. Illyes said afterwards that this was an exercise to see whether the audience could relate to numbers pulled internally, which is not a service commitment. And the slowest column attaches never to one cause again and again: quality.
What to do about it
Use these for setting expectations, not as a promise. If you changed your site this week and nothing has moved, Google’s own typical figures are about 20 hours to discover a new page and about 30 days to revisit one it already knew.
And if anyone has told you that recovering from a core update is a few weeks of work, Google’s own number is three to six months.
Source Search Engine Roundtable, Barry Schwartz, “Google Search Data On Crawling, Indexing & Serving Timelines,” October 5, 2026 · seroundtable.com
04 Search quality
Google says core updates judge pages, not whole sites
One weak page can fall while the rest of your site holds, so the repair is page by page.
Google’s Gary Illyes warned that scaled content is now more of a problem than link spam.
At the same Barcelona event Google explained why it has run so many spam updates this year. The stated reason is volume: there is far more new content than before, a high share of it is spam, and the team is now using AI to help catch more of it.
The line that matters most for an ordinary business sits in the quality section. Core updates do not focus on websites. They target content at page level, not at domain level, which is why some pages on a site rise while others fall in the same update.
Illyes’ warning was that scaled content is becoming more of a problem than link spam. Scaled content means pages produced in bulk, which for most businesses means pages generated rather than written.
Campbell’s summary of the session is blunt: expect more spam updates and new spam thresholds, and scaled, low-effort content is the risk. Google also said user feedback is what new spam filters get built from.
One number traveled out of this session and is worth handling carefully. Google filters out 40 billion spam pages a day is true, and it is not news. Schwartz notes in the same post that it was published long ago.
Every firm in this field, Forever Cited included, has an incentive to quote a figure like that as though it had just happened. A number recirculating from a conference stage is not a change in the world.
What to do about it
If your traffic fell in a recent update, compare page by page rather than site-wide. Google’s own framing puts the judgment at page level, so a site-wide verdict is the wrong unit of analysis.
The scaled-content warning is the one to act on if your site has pages built from one template with the details swapped: a page per city, a page per service, a page per procedure. That pattern is not automatically a problem, and it is where the bar moved.
Source Search Engine Roundtable, Barry Schwartz, “Google Search Spam Updates Use AI To Find AI Spam & More,” October 5, 2026 · seroundtable.com
05 Money
OpenAI will show ads while ChatGPT draws your image
ChatGPT is building ad space out of waiting time, and every early result came from the advertisers themselves.
All three performance figures OpenAI published were produced by the advertiser’s own measurement vendor.
OpenAI announced a new visual ad format in ChatGPT that it will test later this month in the US with an initial group of advertisers.
The placement is the novel part. The ad runs during image generation, below the image being drawn.
OpenAI’s own description says the ads will be clearly labeled, and remain separate from the image being created, and repeats its standing line that advertising does not influence the answers ChatGPT provides. The screenshot OpenAI shared shows an image 62% generated with the ad already loaded underneath it.
The measurement announcement alongside it is the commercially serious half. OpenAI listed integrations with Hightouch, Tealium and LiveRamp for sending conversion data in, and attribution support across AppsFlyer, Adjust, Northbeam, Branch, Triple Whale and others.
Three partner results were published, and each one should be read with its attribution named. WeightWatchers’ attributed cost per acquisition was 15.3% lower than its blended paid-search benchmark, according to DV Rockerbox.
67% of incremental purchases for the wellness brand Dose came from net-new customers, per WorkMagic. 93% of Portland Leather’s visitors from ChatGPT Ads were new, per Triple Whale.
Not one of the three is independent. Each figure was produced by a measurement vendor working for the advertiser and published by the platform selling the ads, and OpenAI says its own incrementality work is still in its early stages.
That does not make the numbers wrong. It makes them an advertiser’s account of their own results, which is a different thing from an audit.
What to do about it
There is nothing to buy here yet unless you are in the initial US test group. The thing to note is structural. ChatGPT is now selling the moment of waiting, which was not a place an ad could go last year.
If you advertise in a regulated field, a labeled ad inside an assistant is still an advertisement, and your own disclosure rules follow it there.
Source Search Engine Roundtable, Barry Schwartz, “OpenAI To Test Visual ChatGPT Ad Format During AI Image Generation,” October 5, 2026 · seroundtable.com
