Well, it has been a crazy week with the Spam update finishing and GPT-6 (Luna) going live both on Thursday the 8th of October!
Google updates are featured heavily this week, after Lily Ray's prediction of a big one coming up, Google doubling down on defining what they mean by quality content, and in a related note, it feels like I see a graph on LinkedIn everyday showing sites that have gotten all but removed from Google after churning out large amounts of AI or quickly scaled programmatic content.
Anyway, on to the news!
On Mon, 12 Oct 2026, 08:32 Dom Website AI,
>
Key Takeaways
- Google began emailing site owners, based on 5 October 2026 data, how many people have chosen them as a "preferred source" in Search, a feature that can make content appear more often in Top Stories, AI Overviews and AI Mode (Search Engine Roundtable, PPC Land).
- A recap of Google's Search Central Live Deep Dive in Barcelona defines content quality as four pillars: effort, originality, talent or skill, and accuracy, judged the same way whether a human or an AI wrote the content (via technical SEO consultant Nacho Mascort).
- Lily Ray's analysis of 100 "best B2B software" queries found AI Overviews citing 38% fewer self-promotional listicles than in June 2026, with the publishing brand left out of the recommendation 83% of the time, up from 69%.
- There have been four Google spam updates in 2026 so far, against one in 2025, with the gaps between them shrinking from 92 to 55 to 37 days, according to Lily Ray.
- TechCrunch reported on 6 October 2026 that Amazon blocked Meta's Muse AI shopping agent from its retail site within days of its September launch, saying Meta had not disclosed the agent would be browsing.
- Ramp's head of growth, Eric Wu, estimates that 30 to 40% of Ramp's "direct" website traffic to long-tail pages is actually AI or search traffic with the referrer stripped.
- A test of roughly 600 larger brand websites and around 8 billion visits by a product lead at Adobe's Brand Visibility team found that pattern on only about 4% of sites.
- A GEO researcher at Peec AI built a site of 50,000 deliberately fake statistics pages for about $50; ChatGPT's retrieval agent fetched its pages roughly 72,000 times in the last 30 days even after Google deindexed most of the site.
- Optmyzr found that Google's Performance Max overlapped with existing Search keywords in 91% of 503 Google Ads accounts studied.
- OpenAI's watermark detection for AI-generated text falls from about 95% accuracy on 400-token passages to about 17% once a quarter of the words are replaced with synonyms.
- Orbit Media's survey of over 1,000 content marketers found only 36% have a deliberate AI search strategy, and that group consistently outperforms the rest.
Google starts emailing sites their "preferred source" selection counts
Google has begun emailing site owners the number of people who have chosen them as a "preferred source," the Search feature that lets people pick the publications they want to see more of across Top Stories, AI Overviews and AI Mode. The emails are dated to 5 October 2026 data and point sites towards installing the interactive JavaScript button, or upgrading from the older static image version, and link to a short survey asking how publishers would like these figures reported going forward.
There is no Search Console report for this yet, and no timeline for one. Kyle Sutton of the Washington Post, quoted via PPC Land, called it "a first and quite helpful" step, since it lets a publisher compare its known selection count against its click-through rate for the first time. That is the honest extent of what we know: nothing in the emails or the coverage shows what being selected as a preferred source actually does to visibility in AI Overviews or AI Mode, so treat any claim of an uplift with scepticism until Google publishes something more concrete. The sensible response right now is simply to check whether your site is eligible for preferred source selection at all, and whether the button is installed, since Google is clearly building towards measuring and surfacing this more formally before long.
Google spells out what "quality" means, and how long recovery really takes
Technical SEO consultant Nacho Mascort has published a detailed recap of Google's Search Central Live Deep Dive event in Barcelona, and it is one of the more useful things I have read this year on how Google actually defines quality internally. Duy Nguyen, a Google Search Quality Analyst, described quality as the extent to which a human being put effort, originality, talent or skill, and accuracy into creating content, and was explicit that AI-made content is judged by exactly the same four pillars as content a person wrote by hand. The problem Google is worried about is not AI content itself but what it called "commodity content": anything that anyone could write from common knowledge, as opposed to something genuinely specific, like a detailed account of why one particular product failed in one particular way. That distinction is now effectively official language.
A documentation update on 1 October to Google's "creating helpful, reliable, people-first content" guidance adds deceptive authorship, meaning fabricated author profiles, AI-generated headshots, or misleading bylines, as an explicit low-quality signal, and separately warns site owners that unreviewed AI hallucinations are a real risk when publishing AI-assisted content at scale. Gary Illyes also confirmed that core updates work at the page level rather than the domain level, which is the technical reason some pages on a site can rise while others on the same domain fall during the same update.
The most practically useful part, for anyone running a large site, is the timing data Google shared on slides at the event. Recovery after a core update typically takes three to six months, and can stretch to a year, or until the next core update, in the slowest cases. A spam update typically takes one to two weeks to roll out, though it can run to months. Removing a manual action takes one to two weeks in the normal case, slower on inactive sites. For sites Google judges low-quality, even basic mechanics like sitemap processing and indexing can stretch out to weeks or, in Google's own words on the slide, simply "never." Mascort's own suggested exercise, which is his inference rather than anything Google said directly, is to group a large site's URLs by page type and look at indexed percentage, traffic and growth for each group, since page types showing a lot of "Crawled, currently not indexed" are usually the first warning sign of a quality problem worth fixing before it drags down the rest of the site.
Several well-known SEOs now think a major Google update is close
Lily Ray has written up, in detail, why she expects a major Google update before too long. She is careful to frame this correctly: it is an informed opinion, not anything Google has announced, and she says plainly that nobody outside Google actually knows when the next big update will land. Her case rests on a few threads. First, the documentation rewrite described above, which she reads alongside a historical pattern she has tracked across past updates, where Google tends to change its guidance shortly before or alongside a significant ranking change. Second, the sheer frequency of spam updates this year: by her count there have been four in 2026 so far, against just one in 2025, with the gap between each one shrinking, from 92 days, to 55, to 37. Third, and most interestingly, her own research: she ran 100 "best B2B software" style queries in June and again in September 2026, and found that AI Overviews were citing 38% fewer self-promotional listicles in September than they had in June, and that the brand publishing the listicle was left out of its own recommendation 83% of the time, up from 69% in June. That is a single researcher's own method and a single query category, so I would not treat it as a universal number, but it is a specific, checkable claim rather than a vague impression, and it is consistent with what a few other sources have reported about listicle citations declining generally.
Marie Haynes, also writing for Search Engine Journal, reaches a similar conclusion from a different angle, pointing to the same documentation changes and to the recent arrival of Google's new Gemini 4 Argon foundation model as reasons to expect a core update relatively soon. Her view is that a jump in model capability and a ranking change often arrive close together, which is her own observation rather than a Google statement.
Adding a third, independent data point, Search Engine Roundtable's Barry Schwartz reports that some SEOs are seeing more manual actions than usual lately. Haynes told him her own requests for help with manual actions have gone from essentially none to several a week, mostly thin content, pure spam and scaled content actions, and, notably, against what she describes as legitimate, long-standing businesses rather than obvious spam operations. Another consultant, Daniel Foley Carter, reported the same trend independently. This is all anecdotal, a handful of consultants and forum threads rather than any systematic count, and Google has not commented. But taken together with the documentation rewrite and the spam update frequency, it points in a consistent direction. If you run or advise on a large site, now is a sensible moment to audit any large-scale, templated or AI-assisted content for genuine added value, make sure author pages represent real people with real credentials, and keep half an eye on the Manual actions report in Search Console.
AI shopping and booking agents are getting turned away by the websites they visit
TechCrunch published a well-reported piece on a problem that is going to become more visible over the next year: the new generation of personal AI agents that book flights, order groceries and make reservations on a person's behalf is often simply being blocked by the sites it tries to visit. Some of this is deliberate policy. Amazon shut Meta's Muse agent out of its retail store within days of Muse's September launch, saying Meta had not told it the agent would be accessing the store and that Muse does not identify itself while browsing. Yelp told TechCrunch it does not permit non-human traffic unless an agent has paid for access through its data licensing programme. Delta said it is protecting customers from unauthorised automated activity and has no integration that lets a third party book on a customer's behalf, and United pointed to its terms of use, which ban robots and automated devices without written permission.
Other blocks look accidental rather than intentional. Walmart, which is actually a Muse launch partner, told TechCrunch the failures it has seen were not deliberate, and the likely cause is its own click-to-verify-you-are-human check, which can fail when an agent interrupts the flow rather than completing it the way a person would. Zillow said something similar: it allows consumer agents in principle, but its anti-scraping measures can occasionally catch some of them anyway. There is also a theory, which TechCrunch flags as a theory rather than a confirmed cause, that a mid-September change to Cloudflare's crawler defaults, which split a single "block AI bots" switch into separate controls for search, training and agent traffic, has caught some agents that site owners did not mean to block. Cloudflare told TechCrunch it had no specific data to share on that.
What interests me most about this story is the industry response rather than the individual anecdotes. Meta, Walmart, Stripe, Sierra, Genesys, Rocket, NiCE and Decagon are reportedly working together on an open standard for how AI agents identify themselves and communicate with businesses, which tells you this is being treated as infrastructure, not a one-off dispute. Meta's own framing is memorable: turning away a personal agent means turning away the customer behind it. If your own site uses a CDN or web application firewall with any kind of bot management, this is a genuinely good moment to check how it is configured. An agent acting on behalf of a real customer that gets silently blocked is a lost transaction you would never see reflected anywhere in your analytics, because it never shows up as a session at all.
Is your "direct" traffic secretly AI? One company's claim meets a bigger test
This is a good example of how a sharp claim from one company should be treated once a second party tries to replicate it at scale. Ramp's head of growth, Eric Wu, shared a run of GEO tests in a conversation with Ross Hudgens of Siege Media, and one claim in particular stood out: Wu estimates that 30 to 40% of Ramp's "direct" traffic to its long-tail pages is not actually direct at all, but AI or search traffic arriving with the referrer stripped out, essentially hidden inside a bucket that looks like someone typed the URL in by hand when nobody plausibly would have. His suggested method is to look for new direct traffic that behaves like search traffic, in time on site and scroll depth, rather than like a bot or an app shortcut.
A product lead on Adobe's Brand Visibility team read Wu's claim and ran essentially the same test, at much larger scale, on roughly 600 larger brand websites, around 8 billion visits, comparing September 2025 against September 2026 and looking only at engaged visits. His finding is the useful correction here: that specific pattern, where direct traffic rises as search falls and the new direct visitors genuinely behave like search visitors, only showed up on about 4% of the sites he tested. A separate and arguably more common pattern, where direct and search traffic fall together, showed up on about a third of sites, which reads more like AI answers replacing the visit outright rather than disguising it as something else. Where the hidden-AI pattern did appear, it was concentrated on documentation, comparison and help pages, the kind of content practitioners reach through AI assistants, coding tools and in-app links that strip referrers as a matter of course, and it was much weaker on product, pricing and most blog content.
Both of these are self-reported numbers from people who build products in this exact space, and neither has a published methodology beyond a social post, so I would not take either figure as gospel. What is worth taking from this, though, is the test itself: check whether direct traffic is genuinely rising while search is genuinely falling on the same set of pages, check whether the new direct visitors engage like search visitors rather than like bots or app traffic, and segment by content type rather than by industry before concluding anything about where your own traffic is really coming from.
A researcher built 50,000 fake statistics pages to see how AI search really verifies things
A GEO researcher at Peec AI, Metehan Yesilyurt, ran one of the more striking experiments I have seen written up this year. In March 2026 he built a brand new domain consisting of 50,000 deliberately fake statistics pages, generated in a single day with GPT-4.1 for around $30, plus another 15,000 listicle pages for about $20, left entirely unchecked on purpose, specifically to observe how AI search would treat them.
Google indexed the site within a week, then deindexed most of it within a month, after two separate Google updates. That part of the story is reassuring. The less reassuring part is what happened on ChatGPT. Seven months on, he reports over 10,000 ChatGPT referral sessions in GA4 since March, more than 330 referring domains linking to the site according to Ahrefs and Semrush, including names like Shopify, DHL and Bluehost, and over 600,000 bot requests in the last 30 days via Cloudflare, of which around 72,000 were specifically ChatGPT's retrieval agent fetching pages. On his own tracked prompts in Peec, the site ranked second for visibility behind Statista. Most strikingly, he reports the fabricated numbers have since been cited by an investment research report from a major financial group, an academic paper, an economics publication and a community report, a phenomenon he calls citation laundering: the number never became true, but it became treated as credible simply by being repeated and cited.
His explanation, which is his own reasoning rather than anything confirmed by OpenAI, is that AI search systems retrieve on relevance and apparent freshness rather than verifying truth, that query fan-out techniques add words like "statistics" and the current year to a search, which a stats page is built to match by design, that a number is simply good grounding material for a generated answer, and that genuine fact verification is too expensive to do at the scale these systems operate. His phrase for this is useful: retrieval is not ranking. ChatGPT rewrites a prompt, pulls candidates from its own sources and cache, reranks them, and only opens a handful of pages, so a site that would fail completely on Google's own ranking signals can still pass through that pipeline and get cited.
To his credit, he also ran a second, white hat version of the test with verified, properly sourced statistics on a page that also earned links quickly, many of them, by his own account, from sites he suspects were themselves AI-written. His practical advice for anyone publishing statistics is straightforward: one claim per sentence, a source and year next to every number, headings that match the kind of fan-out queries people actually ask, a visible and genuinely updated date, server-rendered HTML that retrieval bots can actually read, original data wherever possible, and if you are writing with AI assistance, search every individual statistic in quotation marks before you publish it. He is explicit, and I think rightly so, that this is a tactic rather than a GEO strategy, and that nobody should actually replicate the fake-data version of what he did. The site now carries a disclaimer, and his closing line is blunt: please do not build your own.
Treat Performance Max and Advantage+ as optimisation tools, not proof of impact
An essay by a writer going by Karan, featured in TLDR Marketing, makes an argument I think is worth taking seriously even though it is not based on any new data of the author's own. The core point is that Meta's Advantage+ and Google's Performance Max should be treated as optimisation tools rather than independent proof of advertising impact, for the simple structural reason that the platform running the campaign also controls the buying, the attribution and the reporting, and counts modelled conversions as real ones by default.
The supporting evidence is a useful reading list in its own right. Optmyzr's February 2025 study found that Performance Max overlapped with existing Search keywords in 91% of 503 Google Ads accounts examined, and that search term overlap appeared across 424,820 individual terms in 97% of 511 accounts, at a time when Google gave advertisers no cost breakdown for Performance Max search terms specifically, making it impossible to compare like for like. Meta told advertisers in September 2021 that it believed it was underreporting iOS web conversions by around 15% following Apple's tracking changes, and filled that gap with statistical modelling rather than observed data. A 2019 paper in Marketing Science by Gordon, Zettelmeyer, Bhargava and Chapsky, based on 15 large randomised Facebook experiments covering 500 million user-experiment observations and 1.6 billion impressions, found that observational attribution methods often overstated incremental conversions by a factor of three or more, sometimes understated them, with no reliable way to know the direction in advance. And eBay's well-known 2015 study, where the company switched off paid search on its own brand name across roughly 30% of US markets for 60 days, found close to zero incremental sales from that spend, a result that still gets cited a decade later because it was a genuine randomised test rather than an inference from reporting dashboards.
The author's practical framework is four questions to ask of any number an automated campaign reports: was the test assignment actually randomised, and by what mechanism; who holds the holdout group, and can it be independently audited; does the result reconcile to anything the finance team would recognise as real revenue; and what did the system actually buy, in terms of how much credit came from brand searches or existing customers who would likely have converted anyway. Numbers that survive those four questions count as evidence. The rest, in his word, is telemetry. He is careful to say the products themselves are good at buying what they are told to buy; the argument is entirely about how much weight to put on their own reporting as proof of causal impact, and it is a point I think any team running these campaigns should sit with for a moment rather than dismiss.
The business case for SEO budgets needs to change
Carolyn Shelby has written a sharp opinion piece for Search Engine Journal arguing that SEO has outgrown being treated purely as a customer acquisition expense. Her point is that if a whole SEO budget is judged solely against incremental organic sessions, it quietly assumes organic search is the only beneficiary of that spend, which makes shared capabilities, technical foundations, product information architecture, content infrastructure, vulnerable to being cut the moment organic traffic dips for any reason, including reasons that have nothing to do with the quality of that underlying work.
She coins a useful phrase for something a lot of practitioners will recognise: the "GEO budget paradox," where identical technical work can be far easier to get funded when it is labelled GEO than when the same work is labelled SEO. She grounds this in Google's own published guidance, which states that established SEO fundamentals still apply to its AI-powered search features and that Google prescribes no separate set of optimisations for them specifically, while also noting, fairly, that this does not mean every AI platform behaves exactly like Google, or that SEO and GEO are simply interchangeable labels for the same work.
Her proposed fix is to classify spend by economic purpose rather than by channel label, with four categories: work that generates revenue, judged by expected incremental return; work that protects revenue, judged by expected loss avoided versus the cost of mitigation; work that maintains or improves shared capabilities, judged by operational value against cost; and work that reduces uncertainty, meaning properly scoped AI experiments, judged by the value of the information they produce against the cost of running them, with a hard rule that benefits shared across categories must never be counted twice. Her worked budget examples, including a hypothetical $1.2 million ecommerce SEO budget re-split across those four purposes, are explicitly illustrative rather than benchmarks anyone should copy directly. She is equally clear that this framework cuts both ways: an investment can be genuinely valuable and still not deserve funding if something else offers a better return, and any AI experiment needs its decision criteria agreed before the money is spent, not afterwards, because an experiment that quietly runs forever without ever producing a clear answer is simply an ongoing expense wearing the costume of research.
Most marketers are doing the easy AI-search tactics, not the ones that work
Orbit Media Studios, through Andy Crestodina, has published survey data from over 1,000 content marketers on how they are actually adapting to AI search, and the central finding is a genuinely useful piece of evidence for anyone deciding where to spend effort. The marketers adapting well to AI search, in Orbit's data, never really stopped doing SEO. What they describe as AI search optimisation is mostly the established playbook, applied more thoroughly, plus five additional considerations: answering the specific questions buyers actually ask rather than generic FAQs, supporting every claim with real evidence, publishing original data worth citing, earning mentions in the specific places AI systems learn about a given niche, and actually measuring AI referrals and watching direct traffic for signs of hidden AI activity.
The survey results back this up with numbers. 71% of respondents have taken at least one deliberate step towards AI visibility, and every single tactic measured outperformed the baseline "strong results" benchmark of 14%. Nearly a third of marketers are doing nothing at all on this front, and that group is the least likely of anyone surveyed to report strong results. Only 36% have what Orbit calls a deliberate AI search strategy, and that smaller group clearly outperforms everyone else. Orbit also reviewed 150 published articles of AI search advice and found that 93% recommend FAQs and answer-focused content and 89% recommend schema markup, yet those two most commonly recommended tactics are not the best performers in the actual survey data: only 17% of marketers using FAQs report strong results. Orbit's own phrase for this gap is that marketers are more likely to do what is easy than what actually works.
The best-performing single tactic in the survey was publishing original data, with 24% of the marketers who do this reporting strong results, yet only 26% of respondents actually do it, and only 3% of the 100 free "AI visibility audit" tools Orbit separately reviewed even check for it. On those audit tools specifically, Orbit's review found that 40% of them never put a single question to an AI system at all; they generate an "AI visibility" score purely from easy-to-detect technical signals like markup, crawler settings and the presence of certain files. Orbit also pulled from its earlier GA4 research: across a dataset of 97 accounts, AI-referred traffic was only around 0.5% of total visits, but converted to leads at roughly three times the rate of ordinary organic search visitors, and in a separate 86-site dataset spanning January 2023 to July 2026, direct traffic has actually overtaken organic search as the largest channel, which Orbit speculates, as a hypothesis rather than a proven finding, may partly reflect AI-referred visits landing in the direct bucket the same way Ramp and Adobe were discussing above. Karine Abbou, an AI search educator quoted in the piece, put it simply: there is no finished AI search playbook yet, so the right response is to make it a genuine priority and just start.
OpenAI begins watermarking AI text in the EU, and Google opens its detector to everyone
Two related developments on AI content provenance landed in the same week. OpenAI announced it will begin adding an invisible watermark, using a method it calls textGrain, to eligible ChatGPT and Codex text generated for users in the EU, across all plans, rolling out over the coming weeks, in response to the EU AI Act's Article 50 transparency rules, which began applying from 2 August 2026. From 5 October, API users worldwide can also opt in to the same watermarking on select models, though it is off by default and OpenAI has not named which models support it. The detector itself is not public; it is limited for now to approved researchers and vetted organisations. OpenAI's own published testing shows the watermark is detected in roughly 80% of 200-token passages and about 95% of 400-token passages at a 1% false positive rate, but that accuracy collapses once the text is edited: replacing just 10% of the words in a passage with synonyms cut detection from about 92% down to 66%, and replacing 25% of the words cut it further, to about 17%. OpenAI is explicit that the watermark cannot measure how much of a piece of text a person actually contributed, and that a missing watermark does not prove a human wrote something.
Separately, and reported in the same window, Google has opened its SynthID Detector to the general public at synthid.com, globally though currently English-language only, having previously limited access to a waitlist of journalists and researchers. You can upload an image, video or audio file and it will tell you whether it carries a SynthID watermark from Google or from partner companies the coverage names as OpenAI, Nvidia and Kakao, with Apple described as coming soon; text is not covered by this particular tool. Google is just as direct about the limits as OpenAI is about its own: this is explicitly "not a general AI detector," no signal is described as completely robust, a positive result cannot distinguish a lightly edited photo from something fully generated, and a clean result proves nothing at all if the creator simply did not use a watermarking tool that Google recognises in the first place.
The sensible conclusion from both announcements together is the same one: these are transparency measures, not proof of authorship, and any internal policy that treats a watermark detection result as hard evidence of who actually wrote or made something is asking considerably more of the technology than either company claims it can deliver.
Also happening in PPC and tooling
Google Ads reporting is quietly getting more AI layered on top of it. Three new AI Max columns, Locations of interest, Optimized targeting and Brand inclusions, are rolling out to reporting views so advertisers can see which campaigns use those settings without opening each one individually, and a new AI-written summary panel is appearing on the experiments Summary tab, giving a plain-language overview of an experiment's goals, key findings and recommended actions. Search Engine Land's own advice on the latter is worth repeating: treat the suggestions as a starting point rather than something to apply automatically, and still check duration, conversion volume and statistical significance yourself before acting. Separately, one PPC newsletter's roundup reports more automation being switched on by default across Google Ads accounts generally, including account-level automated promotions and an AI-generated summary of reviews appearing under some ads, though those specific claims come from a single practitioner's account rather than anything Google has confirmed directly.
Google Docs can now open, edit and collaborate on Markdown files natively, without the old round trip of importing and converting a file, which previously risked altering formatting or stripping comments. Google Drive also now shows rendered previews of Markdown files directly. Google's own stated reason for the change is that Markdown has become the standard structured format both people and AI agents use to collaborate on text, and the native support removes a genuine point of friction. It is rolling out gradually from 5 October, with no admin control required on either side.
On the more technical SEO side, one consultant reported a sharp and immediate increase in AI referral traffic after relaxing a client's web application firewall and server rules for AI crawlers, as far as the client's technical constraints allowed without risking server load. It is one client and one chart with no stated baseline, so I would not treat it as proof of a universal effect, but it is a useful, cheap thing to check on any site: whether your own CDN or firewall is blocking or challenging the AI crawlers you actually want to be indexed and cited by, such as OAI-SearchBot, ChatGPT-User and PerplexityBot, and whether your server logs show a pattern of 403 or 429 responses being returned to them.
Finally, Google Cloud used its "Gemini at Work 2026" event to introduce what it calls the Gemini agent, described as a single, universal agent for work that can handle knowledge tasks, content creation and coding, plan multi-step work, use tools and skills, connect to a business's existing systems, and return finished output directly inside the documents, email and developer environments people already use, while picking the best underlying model for each task, including, notably, Anthropic's Claude models running alongside Google's own Gemini models. No release date, pricing or availability details were given for the consumer-facing or Workspace-integrated parts of this. It is a long way from a search or advertising story, but it is a clear signal of where the large platforms think the next layer of competition is heading, from chat assistants towards agents that actually complete tasks end to end.