Search Signal #09
Ahrefs says Google poisons 5 to 20% of the rank data SEO tools collect. Plus: AI Mode cut sources 74%, and a causal study finds GEO tactics do nothing.
Vikas Kalwani
Founder, SAASY LINKS
01. Ahrefs says Google is poisoning 5 to 20% of the data SEO tools collect
Ahrefs published "An update on Google's poisoning of search results" on 22 October, saying Google now serves different results to suspected SERP data providers and putting it at 5% to 20% depending on the provider. Search Engine Journal compiled the rest a day later: SEOmonitor estimates 8% of daily Google results come back incomplete, AccuRanker flagged 30.2% of desktop keyword-days on 25 September, and Nozzle recorded 35.7% of pages on 21 October. Google has said nothing.
Why it matters: The largest tool your clients recognise has now conceded the data is corrupted, so a client can read it themselves and ask why your September and October reports did not say so. Every vendor figure uses a different definition on a different base, so none of them can be averaged.
Action: Reconcile this week's ranking charts against Search Console average position by query via the API for the same dates, and put the Search Console figure on the slide wherever the two disagree. Then ask each rank vendor in writing which date ranges they reprocessed and whether your stored history was revised or merely blocked.
02. The first causal test of GEO page tactics found they do nothing
A peer-reviewed NYU Abu Dhabi experiment, accepted to NDSS 2027, built 56 fictitious data voids, published one controlled page in each, and collected 504 responses from ChatGPT-5.2, Gemini-3 Flash Preview and Sonar Pro after 25 days. Page-level GEO factors showed no significant effect and two-fold effects were positively ruled out. Domain visibility, meaning age plus backlinks plus indexed presence, moved the odds 17.5 times. A separate regression on 282 prompts found fluent prose cut retrieval odds to 0.67.
Why it matters: This is the first causal evidence that the on-page GEO checklist is a qualifier rather than a lever, and that getting an engine to look at your domain at all is the actual job. It also says the smooth, confident house style most content teams ship is measurably the wrong register.
Action: Hand writers a revised brief this week: precise technical terms, real numbers on the page, outbound citations to sources, and a short direct summary opening the relevant section. Cut the polish, the authoritative tone and the plain-language rewrites, all three of which lowered citation odds.
03. AI Mode cut its sources per answer by 74% while Google announced more links
Maeva Cifuentes tracked 8 B2B SaaS brands across 407 prompts and 50,527 prompt-and-page pairs using Scrunch. Average sources per AI Mode answer fell from 16.3 between 15 July and 26 August to 4.3 between 14 and 20 October, a 74% drop that landed 27 to 30 August. AI Overviews fell 54%. One brand held a 42.5% mention rate while its citation rate collapsed from 15.5% to 4.1%. Google claimed on 15 October that it was increasing inline links.
Why it matters: If a B2B SaaS client's AI Mode citations fell since late August, this is a likelier explanation than their content programme failing. Mention rate and citation rate have come apart, so a dashboard tracking cited URLs showed a catastrophe the buyer never experienced.
Action: Email any client who reports AI visibility to a board with the 16.3 to 4.3 figure and the late-August date, before the monthly report lands. Then reprioritise: pages cited in over half of a prompt's answers survived at 61.6%, against 3.1% for pages cited in a tenth.
04. Three datasets put ChatGPT at 0.62% of what Google organic sends
Marie Haynes analysed 12 months of GA4 data from 100 sites and over 145 million sessions: ChatGPT sent 0.62% of Google organic session volume, 0.73% including all LLMs, and Google organic drove 964 conversions against 36 from every LLM combined. Kevin Indig ran 632 correlation tests on four brands and could not isolate any AI effect on direct or branded traffic. SEO Works put organic at 39 times AI referral traffic across 95 properties.
Why it matters: All three measure what GA4 can attribute, which is a floor rather than an effect, since AI Overview clicks arrive as ordinary organic and stripped referrers vanish into direct. Haynes's window also contains the 8 to 16 October ChatGPT tagging outage, which deflates her counts.
Action: Pull your own version before a client forwards you the headline. In GA4, open Acquisition then Traffic acquisition, set a 12-month range, and segment chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com against Google organic. Expect to find conversion tracking too broken to answer the question, and fix that first.
05. Meta pushed its crawl-and-train terms to 20 November, unchanged
A Meta spokesperson told Reuters on 23 October that the revised Commercial Terms and Self-Serve Ad Terms now take effect on 20 November rather than 30 October. Only the date changed: the clause still covers any website tied to a business's products or services, still permits created test credentials, and still carries no commitment to honour robots.txt. The News/Media Alliance, WAN-IFRA, the European Publishers Council and German publisher groups representing 9,000 media brands all pushed back that week.
Why it matters: Every B2B SaaS client with a Facebook Page is in scope whether or not they have ever bought an ad, and this work was rescheduled rather than cancelled. The UK CMA told Parliament on 20 October that it cannot move quickly because Meta is not designated with entrenched market power.
Action: Pull each client's robots.txt and confirm there are separate explicit directives for Meta-ExternalAgent and Meta-WebIndexer rather than one blanket Meta block, since training and Meta AI search are different decisions. Then send the clause to their legal contact with the new 20 November date.
06. Google is also serving results that quietly ignore part of your query
Nozzle's second study, published 22 October across 253 million judged results pages, measures a separate fault: pages that silently drop part of the query. Since 16 September, 13.9% of pages judged on both page one and page two ignored part of the query, against a 0.06% baseline, peaking at 39.9% on 28 September. The title still shows the full query on 99.8% of affected pages. The tracked site left the top 10 on 69.8% of them.
Why it matters: This produces a results page that passes every sanity check, so it burns the analyst rather than the client. The US rate is 2.1% and the UK 0.8%, so keep it in proportion rather than letting it become a universal excuse for every ranking fall.
Action: Add one line to the diagnostic checklist before anyone escalates a single-keyword drop: run the keyword by hand in an incognito window and check that the results address every word in it. Ask your rank vendors which of the two faults their filter actually detects.
07. AI Overviews now appear on 88% of tracked commercial keywords
Semrush tracked a fixed panel of 400 US desktop commercial keywords across banking, software and ecommerce from September 2025 to August 2026. AI Overview coverage rose from 45.75% to 88%, and to 97.08% weighted by search volume, with the 12% of keywords lacking an overview accounting for under 3% of estimated volume. Review queries rose 59 points to 89.04%. Software alternatives and software best-X-for-Y both read 100%, on fewer than 27 keywords per cell.
Why it matters: For a B2B SaaS client, assume an AI Overview on every commercial-modifier query rather than checking for one. Review queries moving to nearly 90% means third-party pages now do the work your own review page used to, which is a mentions problem rather than an on-page one.
Action: List your client's commercial-intent queries grouped by modifier, alternatives, versus, best X for Y, pricing and reviews, and mark which ones currently cite the client. Quote the 88% and the volume-weighted 97%, never the 100% cells, which rest on tiny subsets.
08. Google says robots.txt does not govern what happens after the crawl
In episode two of Google Search Central's Non-Canonical Cabinet on 21 October, Gary Illyes framed robots.txt per RFC 9309 as an access-preference protocol that "doesn't dictate what a system does with the data after it has already crawled it". Blocking a specific AI crawler requires naming its user-agent token explicitly, and invented directives are discarded. Google also documented the Search generative AI control in Search Console, which governs inclusion in AI Overviews and AI Mode.
Why it matters: Google's own Search Relations team has now said that robots.txt governs access and nothing after the crawl, which lands directly on Meta's public defence that publishers can still use robots.txt. Control over post-crawl use is migrating into per-platform consoles and commercial contracts.
Action: Open Search Console for your largest clients, find the Search generative AI control, and record its current setting in their documentation as a deliberate decision. While you are there, audit robots.txt for hand-rolled AI-usage directives, which Illyes confirmed are discarded silently.
09. Most requests claiming to be AI crawlers are not AI crawlers
Centinel Analytica verified crawler requests across news and media sites from late June to 21 September 2026 and found most claimed identities fail. On a typical site, 84% of requests claiming to be GPTBot failed verification, along with 99% claiming Perplexity-User, 77% claiming Meta-ExternalAgent, and one in seven claiming Googlebot. 94% of fake Googlebot requests copied Google's current crawler version string, and 88% came from datacenter addresses.
Why it matters: Server logs were supposed to be the trustworthy part of the AI measurement stack, and a user-agent string on its own is not evidence. These typical-site figures are medians, so 84% means on a median site, not 84% of all GPTBot traffic on the web.
Action: Take 30 days of server or edge logs, check every AI-claiming request against the operator's published IP range file rather than the string, and rebuild the crawler chart from verified requests only. Verify rather than block by ASN, since much of the fake traffic came from Google Cloud.
10. Identical prompts returned almost completely different citations
Seer Interactive sent one identical 148-character prompt, hashed to prove exact reuse, through a real signed-in Chrome profile. Unique retrieved URLs went from 14 on 3 September to 40 on 16 September, with 34 of the 40 never cited, and two runs eight minutes apart shared zero query strings. An Aalto University paper ran one prompt ten times through Google deep research: median entity overlap was 0.12, median citation overlap just over 0.01.
Why it matters: Across identical runs, the entities a system names are roughly twelve times more stable than the URLs it cites, which is a reporting instruction rather than a curiosity. Aalto's discovery curves had still not saturated after ten runs.
Action: Report brand mention rate as the headline and named-URL citation as the volatile secondary metric, print the run count per prompt on every report, and move AI visibility reporting to monthly. Ask your vendor whether it reports what was retrieved or only what was cited.
11. AI Overview citation cards may start showing Google's words, not yours
Brodie Clark spotted AI Overview citation cards displaying AI-written snippets rather than text from the source page, describing them as "summarised based on the information that Google thinks is most relevant to the query". Barry Schwartz reproduced it on some of his test queries but not all. There is no Google confirmation, no rollout scope and no test dates: one observer, one partial replication, two screenshots.
Why it matters: If this ships, the preview a buyer reads about your product in an AI surface is written by Google, and no amount of title or meta description work touches it. For a SaaS client that is a positioning risk rather than a traffic risk.
Action: No action needed yet, the sighting is too thin to act on. But start the evidence log now: screenshot the citation cards for your five highest-intent commercial queries with the date, and check whether the preview text actually appears anywhere on the page.
Also worth knowing
- Volatility spiked twice and Google confirmed nothing. Chatter heated on 20 and 21 October and again harder on 24 October across 15 third-party trackers, but the Search Status Dashboard shows no incident at all for 18 to 25 October, and with paid trackers admitting 5 to 20% corrupted data the two causes cannot be separated.
- OpenAI will rewrite ad creative and may redirect destination URLs from 30 October. Creative enhancements switch on for self-serve ChatGPT campaigns using advertisers' own landing-page content, with no per-version approval, and opt-out has to happen in Ads Manager before that date. OAI-AdsBot still has no robots.txt token.
- Profound shipped Moz link metrics into its AI visibility product on 19 October. Domain Authority, Page Authority, Linking Root Domains and a Citation Opportunities panel listing sites that link to competitors but not to you, which is a link gap report under a new name. No pricing disclosed.
- Microsoft Research priced per-block publisher payment and found the deciding variable is timestamps. Across 1,994 queries and 64,419 distinct content blocks, paying the original reporter in practice means paying whoever has the earliest machine-readable timestamp, and only 33.3% of premium publisher URLs carry minute-resolution times.
- A judge heard the first fair-use arguments on AI training since Bartz and Kadrey. Judge Eumi Lee took summary judgment motions in Concord Music Group v. Anthropic on 21 October with no ruling yet. Elvis Presley rightsholders sued Anthropic over 274 songs on 23 October, and Reservoir Media sued Suno over 1,050 compositions.
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