Search Signal #08
13.1% of the SERPs rank trackers collect come back deliberately corrupted. Plus: ChatGPT's eight-day tracking outage, and Meta's 30 October crawl grant.
Vikas Kalwani
Founder, SAASY LINKS
01. Two vendors measured how much rank tracker data Google is faking
Nozzle found 8.0% of 17.8 million results pages poisoned between 13 September and 9 October, against a 0.17% baseline for the month before. Keyword.com, using a stricter definition, measured 13.1% from 13 September to 14 October. Poisoned pages retain 20.7% of the usual domains against 88.5% normally, and one customer's own site was missing from the top 10 on 58.6% of affected pages against 2.7% at baseline. Google has said nothing, so the cause is vendor inference.
Why it matters: Every ranking slide shipped since mid September sits on a corrupted sample, and the artefact looks exactly like the thing clients panic about: a sudden fall out of the top 10 that recovers within days. Sistrix, DataForSEO, SE Ranking, Advanced Web Ranking and Morningscore have all posted status notices.
Action: Freeze any ranking-based reporting or bonus metric covering 4 September to now, and say in writing why. Then move the weekly ranking slide onto average position by query from the Search Console API, which Google serves from its own data and cannot poison.
02. The GEO checklist does nothing once you compare pages on the same site
University of Toronto researchers tracked 158,099 answers and 468,037 citations over 90 days to 21 September. Using a within-website design on 12,038 cited pages, only two page properties survived every robustness test: title match at +14.1% per standard deviation, and best-paragraph match at +15.5%. FAQ blocks scored +18.1% across websites and -3.4% within one, and schema markup, summary boxes and author boxes behaved the same way. Retrieval explained 32% of variation against 2.8% for all fourteen page features combined.
Why it matters: The B2B replication is the sharper half: across 55 queries about business software, website authority had no measurable effect at -1.3% per standard deviation, while a cited third-party page naming a vendor raised the odds the answer named that vendor by 39.7x on ChatGPT. Ninety-seven percent of ChatGPT's retrieved pages were absent from Google's top 50, so rank is not a usable proxy for AI visibility.
Action: Cut any line item billing for FAQ blocks, schema markup or author boxes as AI visibility work, because within one website they do nothing. Move that budget to rewriting page titles and your single best paragraph in the words customers actually use, and to earning mentions on the third-party pages the engines already retrieve for your category.
03. ChatGPT stopped tagging its outbound links for eight days
ChatGPT links lost the utm_source=chatgpt.com parameter from 8 October, and the links also suppressed the referrer, so visits landed in analytics as Direct. Profound measured the tagged share falling from about 99% to roughly 5% at 19:00 UTC on 14 October, recovering by 06:30 on 16 October. Adobe measured ChatGPT traffic down roughly 31,000 visits a day across 500 pages, with direct traffic up about 26,000. OpenAI confirmed a bug on 14 October and shipped a fix on 15 October.
Why it matters: There is a dip in every client's AI referral chart for 8 to 16 October, it is an instrumentation artefact, and left unannotated it reads as the content programme failing. OpenAI published no status page and no release note, which is the real lesson: a referral tag you do not control is a courtesy, not measurement.
Action: In GA4, open Reports then Acquisition then Traffic acquisition, set the range 8 to 16 October, and annotate both the chatgpt.com referral line and the Direct line before the monthly report goes out. Then verify your tagged clicks have recovered to pre-8 October levels.
04. Meta claims crawl and training rights over business customers' websites from 30 October
A new clause in Meta's revised Commercial Terms, effective 30 October, lets Meta use automated software and created test credentials to access, index, cache and re-crawl websites tied to a business's products, and train on it. The previous Self-Serve Ad Terms covered only advertisers' own domains. The new scope reaches any business use of Meta's products, including running a Facebook Page without buying a single ad, and acceptance is passive. Neither updated document mentions robots.txt.
Why it matters: Every client with a dormant Facebook Page is in scope whether or not they have ever bought a Meta ad, and most of them do not know it. Media attorney Matt Topic told Nieman Lab there are no assurances a court would enforce the clause, but publishers should assume it will be.
Action: Pull robots.txt and set separate explicit directives for Meta-ExternalAgent and Meta-WebIndexer, since being findable in Meta AI search and being training data are two different decisions. Then send the clause to legal before 30 October with a note that robots.txt may not override a contractual grant.
05. Google moved inline links higher in AI Mode, and still publishes no clicks
Google announced five changes on 15 October, the substantive ones being more inline links appearing higher in responses, a test block surfacing one or two official pages for task queries, and news carousels expanding to more locations and languages. The only number given was Preferred Sources usage up more than sevenfold since January. Nick Fox separately said over 400 publications are now in Google's News AI licensing pilot.
Why it matters: Every link inside an AI Overview inherits the overview's single position in Search Console, so moving a link higher cannot show up in reported position. That makes any client target framed as improving AI Overview position unsupportable with Google's own data, and it is worth saying so in writing before someone sets one.
Action: Take the baseline this week. In Search Console Performance, Search results, compare the Web search type against the generative AI search type for your top 20 commercial queries, and pair the impressions with GA4 landing-page sessions to see whether link elevation moved anything real.
06. ChatGPT searched the web on only 14.4% of free-tier answers, and Reddit citations fell 88%
Ahrefs analysed 66.6 million free, logged-out ChatGPT responses on GPT-5.6 Luna across August and September. Only 14.4% triggered a web search at all, averaging 1.6 fan-out queries. Of those fan-outs, 6.8% used the site: operator, and among the most-searched sites 75.9% of site: searches pointed at brand domains. In Ahrefs Brand Radar data from March to September, Reddit citations fell 88%, Wikipedia 80% and YouTube 27%. This is vendor research on a proprietary prompt panel and one model.
Why it matters: The site: operator finding raises the value of a current, machine-readable pricing and specification page, since the model is going to the vendor's own domain for factual questions. The Reddit collapse means any retainer line built on Reddit presence for ChatGPT visibility needs re-justifying or cutting.
Action: In Ahrefs Brand Radar, open Cited domains, filter to ChatGPT and the US, and compare March against September for your own category rather than trusting the aggregate. A B2B software category never looked like the average in the first place.
07. Models answered from the search snippet without opening a page 53.6% of the time
Imperial College London built 225 facts that changed after February 2026, then graded over 30,000 replies under frozen retrieval, where every model received an identical ten-result menu with the correct page reachable. GPT-5.6 Luna under its own native web search answered 53.6% of episodes from snippets alone. Those answers were stale 6.4% of the time against 1.3% when the model opened the page it cited, and 81.8% of on-claim citations were unfaithful when the answer was stale.
Why it matters: Roughly half the time, the only part of your page a model reads is the part that became the search snippet. That is the mechanism underneath the Toronto finding in item 02, and it also means a correct page can be cited for a wrong claim: last year's pricing, a feature you have since shipped, a competitor's positioning read onto you.
Action: Take your 20 highest-intent commercial queries and look at the snippet the engines actually render for your pages, not the page body. Where the real answer is buried in a pricing table, changelog or compliance page, surface the current value in a short dated line near the top.
08. Bing Webmaster Tools now shows the queries where rivals are cited and you are not
Microsoft shipped a Topics Portfolio beta inside the AI Performance report on 13 October. It clusters the grounding queries where your site appears into themes, with three metrics per topic: Topic Size, Sites Cited as a crowding indicator, and Queries. A Top Opportunity Queries view lists strong queries where other sites are cited and yours is not. Microsoft states that classification is AI-generated and imperfect, that branded queries can surface as false opportunities, and that every metric is approximate.
Why it matters: In a week when paid rank tracking turned out to be partly fabricated and ChatGPT referral data was wrong for eight days, a free first-party report naming your citation gaps is the most reliable new data anyone shipped. Absence is information too: if the tab does not render, Bing has insufficient data on the domain.
Action: Log into Bing Webmaster Tools, open AI Performance, check for the Topics Portfolio tab, and export Top Opportunity Queries. Strike the branded rows before the list reaches a content team, then look for topics with high Topic Size and low Sites Cited, which is the closest thing to thin-competition citation opportunity anyone has shipped.
09. Google results plus OpenAI's own API reproduce about a quarter of ChatGPT's cited URLs
Researchers at the National University of Singapore and Xi'an Jiaotong captured raw session traces from authenticated ChatGPT sessions across 80 queries run 20 times each. Google organic results and the OpenAI Responses API together matched 28.3% of ChatGPT's actual cited pages on GPT-5.3 and 26.2% on GPT-5.5, rising to about 50% at site level. Google alone managed 9.4% and 6.4%. And 44.5% of the distinct pages cited for a question appeared in only one run out of twenty.
Why it matters: An API-based or SERP-based citation check is not an approximation of ChatGPT, it is a different dataset. That is the technical reason two tools can hand a client two different citation shares and both be honestly reported, and at 44.5% single-run variance a weekly number is manufacturing movement that is not there.
Action: Ask your AI visibility vendor in writing whether their ChatGPT numbers come from the Responses API, from SERP inference or from real logged-in sessions, and record the answer next to the dashboard. Then run every monitored prompt at least ten times and move the reporting cadence to monthly.
10. Sonnet 5.5 cut its fan-out queries by a third and now favours pages under four months old
Profound ran 711 prompts 20 times each through Sonnet 5 and Sonnet 5.5 via the Anthropic API with web search enabled, for 28,440 answers. Fan-outs per searched answer fell from 1.83 to 1.17, and answers with two or more fan-outs fell from 78.4% to 14.5%, shrinking the candidate pool 36%. The median age of the top 10% most-cited pages fell from 164 to 107 days. The post publishes no collection dates, and this is the API rather than Claude.ai.
Why it matters: A 36% smaller candidate pool with a harder recency filter squeezes exactly what a B2B SaaS client has most of: a good evergreen guide published eighteen months ago. Fewer fan-out queries also means fewer chances to be retrieved per answer, which weakens the long-tail sub-question play and makes head-term coverage relatively more valuable.
Action: Pull the pages already appearing in your citation set, check their publish and modified dates, and give anything over twelve months old a genuine dated refresh with new current figures. A date stamp change is not a refresh and clients get caught doing it.
11. Gemini 4 Argon went global in AI Mode for paid tiers only
Nick Fox and Robby Stein announced on 13 October that Gemini 4 Argon is rolling out globally in Search, restricted to Google AI Pro and Ultra subscribers in AI Mode and in the Gemini app. No countries or languages were named. Barry Schwartz observed, with one practitioner confirmation rather than a Google statement, that Argon AI Mode responses do include links, unlike the earlier Gemini 3.8 Flash rollout where AI Mode responses carried none.
Why it matters: Paid and free Google AI users are now on different models inside the same product, which invalidates any AI Mode audit run from a logged-out or free session. For a B2B SaaS client that is the wrong half of the audience to be measuring, because the buyer with budget authority is disproportionately the one paying for AI Pro.
Action: Re-run your 10 highest-intent AI Mode prompts signed into a paid Google AI account, logging whether the client is cited and whether the answer links out at all, and save dated screenshots. Add account tier as a recorded field in your audit template so the next audit is comparable.
Also worth knowing
- Search Console's custom date picker broke twice and got stuck in 1969 on 13 and 14 October, so re-pull any October export taken on those days and use the preset ranges or the API instead.
- Google quietly stripped the Indexing API quota and approval sections from three documentation pages on 16 October, with no replacement quota figure published and no documented route left to request an increase.
- Cloudflare and OpenAI published a crawl discovery pilot covering about 1% of Cloudflare's network, which they say surfaced hundreds of millions of previously unknown pages, and which may initially increase crawl volume on your site.
- Cloudflare started grading Verified bots on conduct rather than identity, scoring robots.txt compliance and whether a bot slows after an HTTP 429, so confirm your rate-limiting rules actually return 429 rather than 403.
- A pre-registered study found Gemini brand mention rates unchanged over 20.3 days, with 194 of 202 unanimous cells identical three weeks later, which is a second independent argument for monthly rather than weekly AI reporting.
- DeepSeek priced agent web access at $0.75 per 1,000 calls from 31 October, against $35 for Microsoft's Grounding with Bing, so per-request cost will no longer restrain how much agents fetch.
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