SEO Recap covering September 25, 2026: Google's SAFE AI spam detector and AI Overviews links

Built by Stephanie Chung·5 min read·6 stories

Daily summary of what matters in SEO, GEO, AEO, and AI search generated with Claude Code (beware of hallucinations)

Generative Engine Optimization (GEO) & Answer Engine Optimization (AEO)

Markdown mirrors served to AI agents carry prose but strip out actions, answering only what a page is, not what can be done with it. Readiness scanners check whether WebMCP tools are registered, not whether they work.
Exposing actions to machines has 2 routes: semantic HTML and accessibility as the floor, and a declared tool surface (WebMCP) as the ceiling. The ceiling is still a proposed standard, so the floor matters most, and it is broken on most sites.
WebAIM's 2026 evaluation of the top million home pages found 95.9% failing WCAG 2 (up from 94.8%), averaging 56.1 errors per page. Missing actions dominate: 51% of pages had unlabeled form inputs, 46.3% empty links, and 30.6% empty buttons.
Shopify switched on WebMCP tools (catalog search, cart, checkout, policy lookup) for every Liquid storefront on August 5, 2026, live by default with no merchant install, though the buying path failed on an internal error the next day.
Concrete steps: sort what you serve machines into describable and doable, submit your key form and read the response the way a machine would, and run WAVE to count unlabeled inputs, empty links, and empty buttons before buying a declared tool surface.

Skydeo's Mallory Gray argues AI agents run on audience data at scale rather than replacing the need for it, so whatever quality goes in, good or bad, comes out louder. Skydeo says it draws on 1.4 trillion data points across 320 million people.
A brand can raise AI answer citation frequency substantially without matching gains in qualified traffic or conversions, so citation count measures the wrong thing on its own.
The warning sign is output climbing (more content, variations, campaigns) while engagement or conversion flattens, and the feedback loop breaks once the only explanation for a targeting choice is "the AI decided."
3 checks before scaling:
- Audit which audience signals feed your GEO tools, separating declared, observed, and model-inferred data
- Track whether the prompts you appear for match what real customers ask and whether they convert
- Require someone to explain in plain language why a given audience or message was chosen

Peec AI charted the share of AI Overviews containing an external link in the answer text rising from near zero (August 25-September 10) to 26.2% by September 20, with a sharp step marked September 11.
Gagan Ghotra recorded bold, underlined text inside an AI Overview that opens a query in AI Mode instead of a webpage, a path that has existed on mobile since January's follow-up feature.
The Peec chart doesn't disclose sample size, markets, devices, or how it classifies a link, so it can't be read as a share of all AI Overviews or tied to actual clicks.
In Search Console, clicking an external link in an AI Overview counts as a click, but a query refinement link (one that runs a new Google query) counts as neither click nor impression, and it's unclear whether the AI Mode anchor link is treated as a refinement.
Google rolled out its generative AI performance report to all sites worldwide on August 31. It shows impressions in AI Overviews and AI Mode only, not clicks or CTR, and those impressions are already folded into the overall Web search totals.

Microsoft is testing replacing or renaming the Copilot Search tab in Bing with "AI Mode," the name Google uses for its AI search engine.
Bing previously adopted Google's "AI Overviews" label for its own AI responses at the top of results instead of Copilot branding.

Microsoft Bing is testing an editable indicator on images in Bing Search, appearing either as an "Editable" label on the image or as a customizable button.
Clicking the pencil icon opens the image in Bing's AI image creation tool, with a "Customize" icon showing on hover.

Google published a 3-page research paper on Scaled Abuse Forensics Examiner (SAFE), a deployed system that mimics human manual review to catch content violating the "spirit" of policy rather than a known rule pattern.
SAFE uses a few-shot-trained LLM plus multimodal semantic embeddings, and the paper repeatedly references synthetic media, adversarial synthetic media, and nonhuman engagement patterns from bot-nets and coordinated campaigns.
The system runs 4 specialized AI agents:
- Root Agent that orchestrates and makes the final call
- Content Understanding Agent for synthetic artifact detection
- Behavior Understanding Agent for inorganic pattern recognition
- Channel Cluster Understanding Agent using graph-based relationships to map spam networks
This is Google's second AI-slop detection system surfaced in 2026 after Scalable Cluster Termination System (S-CTS), and both may factor into the September spam update. The paper says early deployment cut forensic investigation time versus human-in-the-loop workflows but withholds test results.

Published on September 26, 2026