Search Signal #05
Microsoft's own data shows Copilot cutting publisher clicks by up to 94%. Plus: a spam update that runs for two weeks, and Claude at 11% of US prompts.
Vikas Kalwani
Founder, SAASY LINKS
01. Google confirmed a spam update with a two-week rollout
Released around noon Eastern on 24 September, the fourth spam update of 2026. Google's wording: "This is a normal spam update, and it will roll out for all languages and locations. The rollout may take up to two weeks to complete." Two weeks is the anomaly. August ran 2 days 16 hours, June 2 days 1 hour, March 19 hours. Barry Schwartz flagged significant volatility starting around 23 September, four hours before Google confirmed anything.
Why it matters: Two weeks of announced movement is both a gift and a trap. Any dip dated 22 September to roughly 8 October is attributable to a named update rather than a site problem. It also means nothing measured inside that window is a clean read.
Action: Annotate 22 September, not 24 September, because volatility preceded the announcement. Freeze reactive content and technical changes until the rollout completes, and tell clients now rather than on 2 October.
02. Microsoft's own instrumentation says Copilot cut publisher clicks by up to 94%
An unredacted plaintiffs' brief refiled 17 September in the consolidated OpenAI copyright case put Microsoft's own domain-by-domain CTR comparison of Bing web search against Bing Chat into the public record: New York Times properties down 87 to 93%, Daily News domains 83 to 91%, Ziff Davis properties 51 to 94%. The same filing puts 87.78% of ChatGPT users visiting no external website during a search, against 26.91% for Google.
Why it matters: Every substitution figure until now was a vendor panel or a scrape. This is platform telemetry compelled into the record. It is the most citeable number of the year for the argument that AI answers replace visits rather than route them.
Action: Use the 87.78% against 26.91% pair when a client asks what AI visibility is worth in traffic. Do not extrapolate the 87 to 94% range to a SaaS book: those are news domains, the worst-hit vertical.
03. A randomised experiment prices AI Mode at 18.8 points of external clicks
University of Pennsylvania and Northeastern researchers ran a preregistered field experiment in March 2026, tracking 1,100 participants through a browser extension across three arms: forced regular search, natural Google, or forced AI Mode. External website clicks fell 18.8 percentage points. News sites fell 12.5, Reddit 21.2, Wikipedia 9.9. Participants in the AI Mode arm also reported lower satisfaction, usefulness and sense of control.
Why it matters: It randomises, it has a control arm, and the authors have no product to sell. The satisfaction finding kills the standard rebuttal that fewer clicks just means better answers. The disclosed weakness is real: only 51.1% of AI Overviews were successfully hidden, so effects are probably attenuated rather than inflated.
Action: This is the item to hand a CFO. Cite it with the attenuation caveat and you have a defensible number in a market full of undefendable ones.
04. Some AI Overview links now open AI Mode instead of the page they cite
Gagan Ghotra recorded anchor links inside AI Overviews that render as ordinary hyperlinks but open an AI Mode follow-up rather than the publisher page. No Google comment. Read it against Peec AI's measurement of external links inside AIO answer text going from near zero to 26.2% between 11 and 20 September. The mechanic underneath both: a click to an external page counts in Search Console, and a link that fires a new Google query counts as neither a click nor an impression.
Why it matters: This is the week citation share and referral clicks stopped being two views of the same thing. A brand can be cited prominently, above the fold, and generate zero measurable anything.
Action: Split "AI citation" and "AI referral click" into separate rows in every dashboard, and stop reconciling them. Then check what your AI visibility vendor counts as a citation, because a tool counting links in answer text is now counting two different destinations under one label.
05. ChatGPT fell to half the US prompt market while Claude went from 2% to 11%
Comscore's Q2 2026 AI Intelligence Report, US opt-in panels, January to June 2026. Prompt volume share: ChatGPT 70% to 50%, Gemini 17% to 30%, Claude 2% to 11%, Copilot flat at 5%. AI Overviews appeared on 39.4% of US desktop Google searches, up from 25.8% in July 2025. Comscore sells the measurement product this promotes, and its own September figure for ChatGPT's share of unique visitors diverges by 19 points with no reconciliation offered.
Why it matters: Even discounting the inconsistency, a ChatGPT-only AI visibility report now describes at best half the market it claims to cover. Claude at 11% is past the point where a B2B SaaS brand can reasonably ignore it.
Action: Add Gemini and Claude prompt sets to every tracked property this quarter, and report the three in separate columns rather than blended.
06. Search Console split web results into text-based and multimodal
A new search type filter in the Search performance report, announced 24 September, separating "Text-based" from "Multimodal". It covers Google Lens, Circle to Search, image uploads and Chrome's "Search this image". The critical detail, confirmed by John Mueller: the data was not previously in the counts, so this is new volume appearing rather than an existing total being sliced. There is still no click or query data for the generative AI report.
Why it matters: This will read as growth and it is not. Any image-heavy site sees an impression and click step-up dated 24 September that is purely new reporting coverage, landing in the same fortnight as a spam update.
Action: Annotate 24 September in Search Console and Looker Studio today, export a clean pre-period before the update movement buries it, and explain the step in the next report before anyone else interprets it.
07. Chartbeat: search referrals down 40.2%, AI referrals at 0.01% and falling
The 2026 Publisher Playbook, Chartbeat client network, July 2024 to July 2026. Google Search referrals fell 40.2% year over year, against 21.9% the prior year. Discover fell 34.3%, against 6.6%. The number that matters most sits in the same document: AI referrals grew 8x over two years but are 0.01% of traffic, and are declining across ChatGPT, Claude, Perplexity, Copilot and Gemini during 2026. Chartbeat sells publisher analytics and the playbook ends in a demo.
Why it matters: The two halves cancel the two beliefs most clients hold. Search decline is not levelling off, it nearly doubled. And AI referral traffic is not picking up the slack at any measurable scale, on the panel that would see it first.
Action: Put both halves on one slide in the next quarterly review, then say the uncomfortable part: AI visibility is a demand channel measured in pipeline, not a traffic channel measured in sessions.
08. The CMA wrote a source-attribution duty into its Google choice screen proposal
Google would have to show Android and Chrome users a search service choice screen on first use and annually after. The CMA removed the requirement that a service run general search as its core function, which makes AI assistants eligible for up to 12 slots. The clause that matters: services producing AI summaries must give "clear and accurate attribution with clear routes through to the source material". Consultation closes 9 October. This is a proposal, not a rule.
Why it matters: Item 04 this week is Google routing AI Overview links away from the source. This is a regulator, in the same week, drafting the first text anywhere making routes to source material a compliance obligation rather than a platform courtesy.
Action: Quote the attribution language in any UK client deck, and flag 9 October to anyone who wants to respond to the consultation.
09. The same engine in two languages shares almost no cited sources
Politecnico di Milano ran 500 English queries and their Chinese translations across Google and Baidu, yielding 1,998 observations. The finding to carry forward is not the Baidu one. Google-CN against Google-EN returned a cited-domain Jaccard of 0.034 while answer semantic similarity for matched intents ran 0.701 to 0.813. The answers agree while the sources do not. Observational, no confidence intervals anywhere, machine-translated queries, workshop paper rather than main conference.
Why it matters: If a client sells into more than one language market, an English citation audit tells you nothing about the German or Japanese one, and a translated content programme will not inherit the English source graph.
Action: Run the citation audit per language and budget source-building per language. This is the third independent result this month pointing the same way: blended cross-market AI visibility scores average sets that do not intersect.
10. Google's new fan-out framework only generates retrievable sub-queries
Retrieve-for-Train-Diffusion, or R4T, generates fan-out sub-queries in a single parallel pass from a 53.9M-parameter diffusion model. Its training reward has three pillars, and one is groundedness: a penalty on distance to the database manifold, so every generated sub-query maps to something actually retrievable. Google claims a 12 to 20 times speedup. This is Google's own announcement, with no independent replication and no statement that it is deployed in production Search or AI Mode.
Why it matters: If fan-out is trained toward sub-queries that correspond to retrievable items, the winning content shape is a specific indexable page resolving one question completely, not a 4,000-word pillar gesturing at twelve. That is a real argument against the consolidation advice most content teams have followed for two years.
Action: Take five head terms, enumerate the sub-questions a buyer asks underneath each, and check whether a distinct URL answers each one. Where the answer is a subheading, that is the content gap.
11. Fake evidence packets still lift ranking, on one model family out of three
A controlled adversarial benchmark from Shenzhen University: 50 e-commerce queries expanded into 1,950 paired cases. On one model family, unsupported-rich variants gained 0.065 to 0.092 normalised rank while supported and neutral controls did not move. The effect was marginal on a second family and absent on a third. The cleanest result is the packet-twin test: adding attestations without changing any text restored the attack arms exactly to their unpenalised rates. Preprint, synthetic, single domain.
Why it matters: Read this as a map of where the next arms race runs, not as a tactic. The mechanism is structured attestation, which is the same surface structured-content and schema work operates on.
Action: Nothing to change this week. If you invest in structured citation markup, start testing whether genuinely supported claims carry attestation a penalising ranker would recognise. The brands that survive an evidence-scoring layer are the ones whose claims are checkable.
12. ChatGPT Ads reached a $1bn run rate in under 200 days
Seven new markets added across Southeast Asia and Taiwan, taking ChatGPT Ads past 60 countries, with dentsu, Havas, Omnicom, Publicis and WPP named as agency partners. The constraint everyone skips: ads appear only on Free and Go plans. Pro, Plus and Enterprise remain ad-free. Adthena measured unique weekly ChatGPT advertisers rising from 7,306 to 12,075 between July and August, though the platform expanded from 9 to 40-plus countries inside that window, which confounds the figure.
Why it matters: Paid placement inside the answer is a billion-dollar business eight months old, and clients will start asking about it this quarter.
Action: Answer one question first: which plan tier does your buyer sit on? Sell to enterprise buyers on Pro or Enterprise seats and they will never see the ad. Sell product-led and self-serve and the picture changes.
Also worth knowing
- Google deployed a new AI spam detector called SAFE, reported the day after the September spam update began rolling out. No Google statement tying the two together, so treat the sequencing as suggestive rather than causal.
- A vendor study puts self-promotional listicles at a 4% to 7% lift in AI recommendation rate. Scrunch's own falsification test returned a gap one third the size, implying about 2 percentage points net, and 63.9% of responses citing a listicle recommended nobody at all.
- 86% of AI best-product answers cite at least one commercially interested source. Software and B2B tools sat near the bottom at 19.2% of citations carrying a disclosed commercial interest, against consumer electronics at 41.3%. The vendor sells placement in exactly the pages it counted.
- A crawler self-identifying as ChatGPT or ClaudeBot may be spoofed. DataDome found 65.3% of 21,491 tested domains let all ten test bots through. Verify against published IP ranges before putting crawl counts in a report.
- Baidu shipped AI Traffic Analysis, and its CTR definition is instructive. It computes AI CTR as clicks divided by citations rather than clicks divided by impressions, so a page cited with its link collapsed out of view drags the number down.
- On the calendar. 2 October: proposed final judgments due in the ad tech case. Roughly 8 October: the spam update finishes rolling out, the first date a clean read is possible. 9 October: the CMA consultation closes. Still live from earlier issues: any site using Cloudflare's Block training setting is blocking Googlebot, Applebot and Bingbot outright.
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 Free Strategy Call