Search Signal #07
40 posts on a self-hosted site, zero AI citations in a week. Plus: the spam update ends after 13 days, GPT-6 ships, Gannett sues OpenAI.
Vikas Kalwani
Founder, SAASY LINKS
01. Forty posts on a self-hosted site earned zero AI citations in a week
A Fudan University team published 27 marker-tagged articles across nine platforms in April 2026, then queried 10 live AI search engines daily. In five matched pairs with identical content, structure, length and timing, the citation-preferred platform was picked up first every time, the best low-preference platform needed 20 posts to appear at all, and a self-hosted WordPress site got zero citations after 40 posts in one week. The paper is peer reviewed, accepted to NDSS 2027.
Why it matters: Owned-site publishing was close to invisible to AI search on a one-week horizon, while a single well-placed post landed immediately. The same paper puts mean Jaccard similarity on top-20 cited domains at 0.201, so each engine is drawing from a different pool and a blended AI visibility score hides that.
Action: In your citation tool, pull the actual cited domains per engine for one client's tracked prompt set, then rebuild the placement target list engine by engine rather than as one list. Carry the authors' caveat into client conversations: most test topics had no incumbent, and a contested B2B SaaS category will not behave like a fabricated one.
02. Google's September spam update ended after 13 days and 16 hours
Google confirmed completion at 4:37am ET on 8 October. The Search Status Dashboard logs the rollout at 13 days and 16 hours from 24 September, against roughly two days each for the August and June 2026 spam updates, so about seven times the length. Barry Schwartz pins impact to three discrete windows: 25 to 27 September, 30 September, and 4 to 6 October. That wave structure is his reading of tracker data and forum reports, not something Google described.
Why it matters: The content and technical freeze lifts, and the duration is the argument for having held it. A client who looked unaffected on 28 September may have moved on 30 September or again on 5 October.
Action: Annotate all three windows as separate events in GA4 and Looker Studio rather than one blurred fortnight, then re-run the diagnosis against 4 to 6 October data for any site you already declared clean. Google repeated that recovery can take many months, so do not let a client read the dashboard end time as a recovery clock.
03. Gannett sued OpenAI over 83,266 USA Today articles in GPT-2's training set
USA Today Co. and 13 Gannett entities filed a copyright claim against seven OpenAI entities in the Southern District of New York on 8 October, case 1:26-cv-08892. The complaint covers hundreds of thousands of articles across 19 publications and alleges the WebText dataset held more than 160,000 entries from its newsbrands, 83,266 from USA Today alone, plus a 2019 Common Crawl snapshot containing over 122 million tokens. Damages are reported at $250m or more, unverified against the filing itself.
Why it matters: Training-data provenance is now a discovery question with named datasets and per-publication counts attached, rather than an argument about principles. Both named datasets are the ones your clients' content most likely sits in.
Action: Pull each client's robots.txt and WAF rules, list which AI crawlers and which training crawlers are currently allowed, and put a named owner on that decision. "The security team added it" is not a position you want to discover during a client's legal review.
04. OpenAI wired ChatGPT ads into 20 attribution and measurement vendors
Alongside a new image-based ad format testing in the US later this month, OpenAI named ten attribution partners including AppsFlyer, Triple Whale, Adjust, Northbeam and Branch, plus conversion-data partners Hightouch, Tealium and LiveRamp and geo-incrementality pilots with Haus, Measured and WorkMagic. Partner-reported results include a WeightWatchers CPA 15.3% below its blended paid-search benchmark. Those figures are advertiser and vendor reported, not platform benchmarks, and ads still serve only to Free and Go plan users.
Why it matters: Twenty measurement vendors is how a surface stops being an experiment. Once assistant sessions flow through mainstream attribution tools, organic AI visibility gets benchmarked against a priced alternative with a cost per acquisition next to it.
Action: If a client already runs one of the ten attribution partners, ask the vendor rep directly whether ChatGPT Ads conversion data is live in that account. Then put your own equivalent-media-cost framing into the next report, on your terms, before someone else presents it as a reason to move budget.
05. GPT-6 shipped, and web-search answers start 44% sooner
GPT-6 began rolling out on 7 October, with GPT-6 Sol on Plus, Pro, Business and Enterprise and GPT-6 Luna on Free and Go. OpenAI says GPT-6 Instant starts answering 44% sooner than GPT-5.6 Instant on questions that need web search, and that it more reliably finds information supporting its answer. Intelligent UI lets answers compose charts, forms, tappable buttons and interactive components in place. The announcement publishes no answer-quality figures and never mentions citations or outbound links.
Why it matters: An interactive comparison table or calculator rendered inside the answer does the job a landing page used to do, without the visit. A faster retrieval loop also means more answers get built from live fetches than from training memory, which raises the value of pages an agent can actually read.
Action: Run your 20 highest-intent commercial prompts for one client in GPT-6 this week and log three things per answer: whether it links out at all, whether the client is named, and whether it renders a component instead of prose. Save the screenshots with the date on them, so the baseline exists before any referral decline does.
06. 27 models diagnosed SEO problems correctly 99% of the time, and knew when they could not only half the time
iPullRank ran 27 LLMs against 27 SEO diagnostic scenarios in three variants: decisive evidence present, present with an explicit opt-out, and removed. That is 6,427 scored responses with inter-rater reliability of 0.969. Models were 99% correct when the evidence was there but recognised insufficient evidence only 51.5% of the time. Instructing them to express uncertainty raised abstention from 20.7% to 85% and collapsed correct answers from 64.8% to 13.8%, with GPT-5.5 going from 95.5% to zero.
Why it matters: This lands in the week a thirteen-day spam update finished, which is exactly when people paste Search Console exports into a chat window and ask what happened. On this evidence the answer is confident and well-written about half the time there is nothing in the data to diagnose. Note that iPullRank sells consulting that this finding argues for.
Action: Do not add an uncertainty instruction to your diagnostic prompts, because it made accuracy far worse. Supply the evidence explicitly instead, require a named data source for every claim in the output, and run each prompt three times: a diagnosis that changes between runs is one the data does not support.
07. DOJ and Google filed rival ad tech remedies, six months to open AdX to Prebid or twelve
A 41-page joint proposed final judgment landed before Judge Leonie Brinkema in the Eastern District of Virginia with the two sides' competing language in brackets. Both agree AdX must build a Prebid adapter and accept bids from non-Google ad servers on equal terms, that Unified Pricing Rules are banned, that First Look and Last Look cannot be rebuilt, and that the decree runs six years with a Monitor. They dispute the deadline, the geography, and whether a malfeasance clause exists.
Why it matters: This is the remedy that reshapes the ad market rather than search results. If AdX accepts Prebid bids on equal terms and Unified Pricing Rules disappear, publisher yield on open-web display rises and the cost of the placements you buy moves with it.
Action: No action while the judgment is unsigned, and no hearing date is set. If you buy sponsored placements or display on publisher sites for clients, ask your reps whether 2027 rate planning assumes a Prebid-accessible AdX, and lock multi-quarter rates if the answer is yes.
08. Google documented that paywalled subscriber content can surface in AI Mode, AI Overviews and Gemini
Google's Reader Revenue Manager page used to say subscriber content could appear in a From Your Subscriptions panel on Search, Discover and News. The redesigned page, spotted on 9 October, now lists Search, AI Mode, AI Overviews and Gemini under the same label. There is no blog post, no documented mechanism, and the eligibility appears scoped to Google-managed subscriptions rather than paywalls in general, so verify against the live page before asserting how it works.
Why it matters: The standing advice that gating hides content from AI answers is now less absolute, at least for content behind a Google-managed subscription. Google is documenting a path where gated content surfaces in all three of its AI answer products with a subscription label attached.
Action: No action needed yet unless a client uses Reader Revenue Manager, since the path is narrow. But retire "gating makes it invisible to AI" as your argument against forms, and use the stronger one instead: an asset on your own domain was barely visible to AI search anyway. Watch for Google documenting this outside a paid monetisation product.
09. Anthropic's new usage policy bans seeding the sources AI answers draw from
Published 8 October and effective 12 November. A new section bars using Claude to seed the sources search engines or AI systems draw answers from with content that misrepresents its origin, authorship or independence, and bans distribution through fake or ostensibly independent outlets, websites or accounts. The September 2025 version mentioned neither. Anthropic's own threat report describes roughly 70 sites posing as independent local news, publishing at least 8,913 articles with three to four internal links each.
Why it matters: Counting Google's 1 October fake-author clause, that is two vendors in eight days naming fabricated independence rather than AI generation as the violation. Anthropic actually loosened the AI-written part, dropping automated publishing from its high-risk list. The network it described is a private blog network with a model attached, and some of those sites sit on somebody's placement inventory.
Action: Audit your own supply this week. For every guest posting network, placement marketplace or syndication partner you buy from, check that the sites have editorial ownership you can name and bylines belonging to reachable people, and stop buying where you cannot establish both.
10. Common Crawl says roughly half of page quality is a property of the host
Malte Ostendorff of the Common Crawl Foundation ran ten data-selection quality classifiers across 3.01 million documents and 532,628 hosts. Between-host variance accounts for 0.43 to 0.62 of per-document quality under every classifier tested, and a three-page sample predicts the rest of a host at rank correlation 0.53, plateauing by 15 to 20 pages. Harmonic centrality correlates minus 0.119 with quality, though PageRank holds up within a single TLD. The author is auditing his own employer's crawl policy.
Why it matters: Your thinnest templated pages are not a local problem confined to those URLs. If a sampler reading three pages predicts the whole host, they drag on how the entire domain gets scored for inclusion in training data.
Action: Pull one client's 20 thinnest indexed pages by word count, read three at random, and consolidate, improve or noindex the ones that would not make a sampler want the rest of the domain. That is roughly ninety minutes of work with a compounding payoff.
11. Search Console gained multi-country filtering, then fell 64 hours behind
Google Search Central announced on 6 October that the Performance report's Country and Device filters now accept multiple selections, replacing a Compare tab that was limited to two items. The help documentation still described the old limit. Then on 11 October the report went over 64 hours stale against a normal lag of a few hours, easing to 44 hours by midday. No Google comment, and the Search Status Dashboard does not log Search Console delays.
Why it matters: Any Search Console export pulled over the weekend of 10 to 11 October is incomplete, and that weekend is precisely when people were pulling post-spam-update data. An underfilled final 48 hours reads to a client as a decline.
Action: Re-pull anything dated 9 to 11 October before it reaches a client and add one sentence to the October report flagging the gap. Then rebuild your multi-market client views on the new country filter and delete whatever workaround you were using. Do not repeat the theory that the delay signals a core update: there is no evidence for it.
Also worth knowing
- Google documented a UGC Fresh Data Program, a dedicated ingestion pipeline letting approved large user-generated-content platforms push content and engagement signals directly, which Google says is independent of the Indexing API and does not guarantee inclusion in Search.
- Gary Illyes gave recovery timings at a Search Central event: core update recovery 3 to 6 months typical and up to a year at worst, though the numbers reach us through a conference recap rather than published Google data, so attribute them carefully.
- Somebody finally put a placebo arm on "restructure your content for AI", finding that cutting text at real structural boundaries helps slightly, that shuffled heading paths buy nothing, and that LLM-written chunk summaries lose to plain cutting and fall below baseline under a tight token budget.
- The 39.8% click-loss figure circulating this week is six months old: the Agarwal and Sen randomized field experiment is real and pre-registered, but it was posted in April 2026, is not peer reviewed, and both headline figures are relative changes on AI-Overview-triggering queries only.
- Meta's Muse shopping agent takes most of its picks from two websites, checks no more than three open-web merchants per request, does not identify itself and cites nobody, per Profound vendor research on a single-category, single-day, 50-prompt test.
- Microsoft now requires sign-in to use Copilot in Edge, confirmed on the record by a Microsoft AI principal product manager, which narrows the population of Copilot sessions your citations can reach.
SAASY LINKS
We build the visibility this newsletter reports on
Editorial links, AI citations, and organic pipeline for B2B SaaS. Senior-led, 90-day sprints, no long-term contracts.
Book a Call