Search Signal #10
50 to 60% of AI citations vanish the next day. Plus: one robots.txt token for AI training, Bing's grounding deadline, and Reddit's referral cliff.
Vikas Kalwani
Founder, SAASY LINKS
01. Google, Microsoft, OpenAI and Anthropic now share one robots.txt token for AI training
On 30 October Google said it will adopt the IETF AIPREF draft, draft-kelley-aipref-training-preference-signal, which lets a site express AI training preferences for every participating provider through one shared policy token written as the user-agent __usage_ai-training. Microsoft, OpenAI and Anthropic committed the same day. Google targets support in early 2027. John Mueller confirmed that using the token without naming Google-Extended will be read as disallowing Google-Extended.
Why it matters: Google-Extended also governs grounding in Gemini apps and Vertex AI, so one shared-token opt-out costs you Gemini grounding eligibility as well as training, and there is no training-only setting. The draft itself has no IETF standing yet, and participant Nate Hake says significant areas of dispute remain.
Action: Open each client's robots.txt this week and check whether a Google-Extended line is already there. Do not add the shared token as a blanket block until Google documents how the two interact, because on the stated behaviour it silently removes you from Gemini grounding.
02. Bing split AI training from AI grounding, and NoArchive stops working on 1 March 2027
Krishna Madhavan, principal product manager at Bing, published three site-owner preferences on 30 October: AI Training, AI Grounding, and a DisallowExtended middle tier that keeps content citable in Copilot and Microsoft 365 while excluding third-party grounding services. Grounding is page-level only. Four dates follow: 15 December 2026 temporary controls open in Bing Webmaster Tools, 1 February 2027 Microsoft begins honouring preferences, 1 March 2027 NoArchive and NoCache stop counting as AI signals, 30 September 2027 temporary preferences expire.
Why it matters: Any site relying on NoArchive or NoCache to stay out of AI answers defaults back to Allow for grounding on 1 March 2027 unless a preference is recorded first. Microsoft also warns plainly that Disallow significantly reduces visibility in its AI experiences.
Action: Grep your templates and response headers for NoArchive and NoCache today. If either is in use anywhere, put 15 December 2026 in the calendar to record a grounding preference in Bing Webmaster Tools before the 1 March cliff.
03. Reddit told the SEC it is losing more than a million search referrals a week
Reddit's Q3 2026 shareholder letter says search referrals fell by more than 1 million per week in October, and the decline accelerated late in Q3. US daily active uniques fell from 53.2 million to 50.7 million quarter on quarter, and logged-out US from 30.1 million to 27.3 million. CEO Steve Huffman said Google is crawling Reddit as hard as ever and that AI Mode does not drive traffic. Reddit will test a login-only experience in some non-US markets.
Why it matters: This is the hardest public number anyone has filed on AI search displacing referral traffic, and it comes from a site that sits inside a large share of AI answers. Reddit calls the AI attribution its own working hypothesis, not a measurement.
Action: If Reddit threads carry part of your category's AI visibility, run your ten most important prompts now and record which citations point at Reddit. Pages behind a login stop being crawlable, so that list is your exposure if the test ships.
04. Ahrefs found 50 to 60% of AI citations vanish the next day, and three quarters come back
Ahrefs asked five AI tools the same questions daily for 20 days and followed roughly 37,500 day-one citations. Between 50% and 60% dropped out by the next day on Google AI Mode, ChatGPT and AI Overviews, and 80% on Gemini and Copilot, but about 75% of them returned within 20 days. Pages already ranking in Google were about seven times more likely to hold a citation for a week or more. Domain Rating and backlinks showed little association.
Why it matters: A single AI visibility check is close to meaningless, and two more datasets landed the same week saying it from other angles: Muck Rack measured 41% and 45% repeat-domain consistency on same-day reruns of ChatGPT and Gemini, and Profound found personal context halves which products a model recommends. This is vendor research that doubles as a demo for Brand Radar.
Action: Change how you report citations before the next monthly review: count how many of the last seven days a page held a citation rather than whether it holds one today, and stop raising a client alarm on a single day's disappearance.
05. The UK CMA opened a formal investigation into Meta's crawl terms, and an Oslo court blocked them
On 30 October the CMA opened a Competition Act 1998 Chapter II abuse-of-dominance investigation into Meta's commercial terms, with initial information gathering running to January 2027. Two days earlier the Oslo District Court granted an interim injunction barring the changed terms, on a filing by ten Norwegian media companies including NRK, TV 2 and Schibsted. Meta's Andy Stone then said consultation could extend beyond 20 November and that existing terms remain in place.
Why it matters: The 20 November date is no longer firm, but the clause granting Meta crawl rights over websites associated with a business's products is unchanged, and no regulator has actually found anything yet. Meta holds no Strategic Market Status, so the CMA is using older competition powers rather than its digital markets regime.
Action: Do the work now rather than at the deadline. Pull each client's robots.txt and confirm separate explicit directives for Meta-ExternalAgent and Meta-WebIndexer rather than one blanket Meta block, then send the clause to their legal contact with the new date.
06. US organic search visits fell 13.2% in two years, and AI referrals are 0.5% of visits
Kevin Indig, using exclusive Similarweb estimates for 10,000 domains each in the US and France, measured organic visits down 13.2% in the US, from 19.9 billion to 17.3 billion, and down 24.4% in France between August 2024 and August 2026. Publishers lost 41.5% in the US, and excluding publishers the US decline is 7.8%. AI referrals are the fastest-growing channel but equal 2.85% of SEO traffic in the US, roughly 0.5% of total visits. Paid search grew 21%.
Why it matters: The decline is real and heavily concentrated, with about 1% of domains accounting for 40.2% of gross lost US visits. AI referrals are nowhere near large enough to explain it, and the post, which Semrush sponsored, points at rising ad load instead.
Action: Before you accept an AI-caused-this narrative in the next client review, open GA4, set Traffic acquisition to August 2024 against August 2026, and put the organic drop next to that client's AI referral sessions for the same months. The gap is usually the argument.
07. Google AI Overviews cite under half the top five results, and 20.8% of their claims carry no citation
A preprint posted 29 October ran 3,611 benchmark health questions and compared each AI Overview against the top five traditional organic results for the same query. An Overview appeared for 96.4% of questions but cited only 44.4% of those top-five results, skewing toward commercial health publishers over academic and government sources. An estimated 56.7% of cited statements were actually supported by the page cited, and 20.8% of all factual statements carried no citation at all.
Why it matters: This is the cleanest figure published on the gap between ranking in the top five and being cited: more than half the time, ranking is not enough on its own. It is a preprint with no stated venue, on health queries only, so read the 44.4% as a direction rather than a benchmark for your vertical.
Action: Take the ten queries where a client ranks in the top five and check by hand whether the AI Overview actually cites them. The pages that rank but are not cited are your content brief, and item 08 says what to change on them.
08. Cited pages answer the question in their opening lines four times as often as passed-over pages
TJ Robertson of TJ Digital analysed 193,507 AI answers from ChatGPT, AI Mode and AI Overviews, comparing 21,874 cited pages against 4,507 pages that ranked in Google's top 10 but were never pulled up. 20% of cited pages opened with a direct answer against 5% of passed-over pages. Adjusted odds of citation rose with a year in the title (over 2x), a recent date (plus 88%) and tables (plus 69%), and fell with an author byline (minus 28%).
Why it matters: These are associations from agency research, not causation, and Robertson says so himself. But the direct-answer gap is the largest effect in the set, it is consistent with item 07, and it is cheap to act on without a new content budget.
Action: Take your five highest-value commercial pages and rewrite the first two sentences under each H2 into a direct answer to that heading's question. Add the year to the title tag and a comparison table where one genuinely fits.
09. Google said AI Mode earns less per query, then put ads in 13 more markets
On the 27 October Alphabet earnings call, chief business officer Philipp Schindler referred to the lower monetisation Google was seeing in AI Mode, without quantifying the gap. The next day Google announced AI Mode ads for Australia, Canada, India, Kenya, Malaysia, New Zealand, Nigeria, Pakistan, the Philippines and Singapore, with Brazil, Mexico and the UK in Q4, plus multiturn ads in the US that place ads on follow-up questions inside a conversation. Alphabet ad revenue grew 10.5%.
Why it matters: AI Mode is being monetised harder precisely because it earns less per query, which means more sponsored real estate sitting above whatever organic citation you win. Google's claim that sponsored and organic product items get similar click-through rates inside AI Mode is unaudited and comes with no underlying data.
Action: If you buy paid search for a B2B SaaS client in any of those markets, ask your Google rep whether existing campaigns are already eligible to serve in AI Mode. Then scan the Search terms report for conversational follow-up phrasing, which is where multiturn inventory sits.
10. Google-affiliated authors found no measured benefit from llms.txt, and flipped 21 of 24 answers with a hostile one
A preprint posted 28 October, with most authors at Google plus Sony, Fortinet and Intuit, audited the published evidence for llms.txt and found no citation advantage in it. SE Ranking saw 10.13% adoption across nearly 300,000 domains with no detected benefit, and Ahrefs found 97% of roughly 38,000 llms.txt files had no recorded request in May 2026. In the paper's own 24-task security pilot, an adversarial manifest flipped 21 of 24 answers to the attacker's target.
Why it matters: llms.txt now has no measured upside and a demonstrated downside wherever anything untrusted can write into it. The pilot is one model snapshot, eight purposively chosen sites and researcher-authored fixtures, so treat the 21 of 24 as a warning about the mechanism rather than a real-world rate.
Action: If you shipped llms.txt as an AI visibility tactic, stop selling it as one and reallocate that time to item 08. If any part of the file is generated from user input, a CMS field or a third-party feed, replace it with a static file you control.
11. Semrush admitted poisoned Google results are hitting its data, and dropped Copilot tracking
Semrush published a statement on 27 October calling corrupted Google results an industry-wide issue affecting data providers across the SEO space, making it the third major vendor on record after Ahrefs, at 5% to 20% of results depending on provider, and Nozzle, at 35.7% on 21 October. Semrush gave no figures, no sample size and no named author. Separately it told Enterprise AIO customers it will stop tracking Microsoft Copilot on 9 November because Copilot's access requirements block collection.
Why it matters: Two measurement surfaces degraded in one week, and the pattern is platforms shutting out trackers rather than a tool bug. Barry Schwartz logged two unconfirmed ranking spikes in the same window, on volatility trackers that are themselves being fed corrupted results, and Google has still said nothing.
Action: If you report Copilot visibility, decide before 9 November whether to return those prompts to your balance or keep the history, because Semrush has set no default. Then reconcile every ranking chart leaving your office against Search Console average position by query for the same dates.
Also worth knowing
- Gemini stopped tagging its own referrals. The utm_source=gemini parameter on cited links disappeared around 28 September, reported on 30 October, and iOS app clicks land in GA4 as Direct with no referrer at all.
- Google opened Connected Apps. A public preview lets approved third-party apps run inside AI Mode and the Gemini app through a remote MCP server, with more than 50 already live, though users must add each one manually.
- Meta lost both Paris appeals. On 29 October the Paris Court of Appeal dismissed Meta's challenges to the French interim measures requiring good-faith negotiation on neighbouring-rights payments to press publishers.
- Adobe says AI traffic converts harder. AI-referred US retail visits generated 83% more revenue per visit than non-AI traffic in September, its largest recorded gap, though Adobe does not explain how it identified AI referrals.
- Citation cards may be thinning. Brodie Clark reported desktop AI Overviews appearing without the clickable citations card on 29 October; Barry Schwartz could not reproduce it and Google has confirmed nothing.
- Mueller on the frozen indexing report. Asked about a Page Indexing report stuck since 16 October, John Mueller said it can feel important to focus on but "honestly, save your energy".
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