Google SEO & AI Weekly Roundup and Linus's Notes 260927

From September 21 to 27, 2026, search news moved in three directions at once: ranking systems were updated, search measurement gained a new multimodal dimension, and researchers tested whether AI search sends people to the open web. OpenAI and Anthropic also released cheaper or more efficient models. That changes the cost of SEO research and agent workflows, but it does not change Google’s ranking rules by itself.

Key points

• Google started the September spam update on September 24. It applies globally and to all languages, and Google says the rollout may take up to two weeks.
• A randomized experiment with 1,100 participants found that using Google AI Mode reduced clicks to external websites without producing higher satisfaction. AI search visibility and website traffic are separate outcomes.
• Search Console began reporting web multimodal search from Lens, Circle to Search, image uploads, and Chrome’s “Search this image” feature.
• OpenAI released GPT-6 Sol and GPT-6 Luna with API prices 50% below the corresponding GPT-5.6 promotional prices. Anthropic released Claude Opus 5.5 and reported a 40% reduction in cost per task compared with Opus 5.
• Baidu previewed AI citation reporting in its webmaster platform, while Lighthouse 13.5 added an audit for Agentic Resource Discovery. Neither update is evidence of a new Google ranking factor.
• Marie Haynes’s algorithm tracker records September 24 as the spam-update start date; it is useful for timestamping the rollout, but it is not a Google diagnosis.

Google’s September spam update needs a complete observation window

Google’s Search Status Dashboard announced the September spam update on September 24. The update applies globally and across all languages. Google did not name a specific spam technique and did not publish a recovery checklist. It only confirmed that the rollout could take up to two weeks.

Ranking trackers and webmaster discussions reported volatility from September 25 to 27. Those observations show that changes overlapped with the update window. They do not prove that a specific site was targeted, or that a traffic loss came from content, links, crawling, or indexing.

A useful investigation should put pages, countries, queries, clicks, impressions, indexing state, and manual actions on the same timeline. A drop during the update is a reason to investigate. It is not a diagnosis. Compare the affected pages with Google’s spam policies, including scaled content, missing original value, misleading information, and ranking manipulation. Google’s guidance on generative AI content does not say that using AI is itself a violation. The focus remains on the value of the content and whether it violates spam policies.

Marie Haynes’s algorithm update record also dates this Spam Update to September 24. The record is useful for fixing the timeline and comparing other algorithm changes, but it is an industry reference rather than Google’s diagnosis. It cannot replace a site’s Search Console and log data.

AI Mode reduces external clicks

Search Engine Land reported on September 23 on a randomized field experiment involving 1,100 users. Participants were assigned to different Google Search experiences, including traditional Search, AI Overviews, and AI Mode. The study found that AI Mode reduced clicks to external websites without improving satisfaction or trust.

The useful part of this result is that it measures user behavior rather than a visibility score. An AI system can mention a brand or use a page as a source while keeping the user inside the search product. For reporting, “mentioned in an answer,” “linked to an external page,” and “generated a visit” should be separate fields. The final row should contain a business outcome such as a registration, lead, or purchase.

The same week, Comscore data was cited in a Search Engine Land report showing AI Overviews on about 39.4% of US desktop searches in June 2026, up from 25.8% in July 2025. The figure covers a specific country, device, and month. It is not a global rate and cannot predict a site’s CTR. It does show why traditional ranking reports need a second view of the search page.

Google’s Search Console definitions also matter. A click on an external page link inside an AI Overview counts as a click to that site. A control that opens AI Mode or starts a query refinement should not be treated as the same event. GEO reporting should record the surface, the destination, the visit, and the later conversion separately.

Search Console adds web multimodal reporting

On September 24, Google Search Central announced web multimodal Search performance reporting. The new search type appears in the Search results performance report and in the report for Generative AI features. It covers Google Lens, Circle to Search on Android, image uploads to Google Search, and Chrome’s “Search this image” action. Google said the rollout was global and that a site would see the data when it received traffic from those searches.

This solves a measurement problem: owners can now separate traffic that began with an image from traffic that began with text. It does not add a query dimension. The Search Analytics API documentation does not list multimodal search as a dimension that can be requested directly, so teams will need to export the report and build their own page, country, and date baselines.

For ecommerce sites, the important check is the relationship between the image, the text, the product specification, the variant, and the structured data. A visual match can bring a user to the page, but the page still has to explain size, material, use cases, and limitations. The new report is a new source of evidence; it is not proof that a particular image format improves rankings.

Baidu and Lighthouse add new AI discovery and measurement interfaces

On September 23, Search Engine Journal reported that Baidu’s webmaster platform had added an AI Traffic Analysis section. Baidu plans to report AI citations, AI impressions, AI clicks, and AI CTR. Its proposed AI CTR is clicks divided by citations, rather than clicks divided by impressions in the traditional search sense.

The definition is useful because it separates being cited from receiving a visit. The data service was still rolling out in stages, and some dashboards were empty. Baidu’s definitions cannot be copied directly to Google, ChatGPT, or Perplexity because each platform uses different citation surfaces. They are better treated as a measurement idea: citation volume is not the final outcome, and the visit from a citation needs its own field.

Lighthouse 13.5 also added an Agentic Resource Discovery audit. Agentic Resource Discovery is a proposed open specification for helping agents find MCP tools, A2A agents, and other callable services published by an organization. The current audit checks for a catalog and can fall back to /.well-known/ai-catalog.json when it cannot find a catalog pointer. Search Engine Journal noted that the Lighthouse implementation and the latest ARD proposal still differ.

The two updates point in the same direction: platforms are adding dedicated ways to measure whether AI systems can discover, cite, and reach a site. Neither update is evidence of a ranking factor. Wait for real Baidu data, and treat ARD as a proposed agent interface rather than a Google Search requirement.

GPT-6 Sol, Luna, and Claude Opus 5.5 lower the cost of agent work

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22. The published API prices are $2 per million input tokens and $10 per million output tokens for Sol, and $0.10 input and $0.50 output for Luna. OpenAI says these prices are 50% below the corresponding GPT-5.6 promotional prices. The models were made available in ChatGPT Work and Codex, as well as through the API.

Anthropic released Claude Opus 5.5 the same day. Anthropic says the model improves on complex coding and long tasks while reducing cost per task by 40% through fewer tokens and more efficient tool use. These numbers come from vendor evaluations and early customer reports, so they should be recorded as vendor claims rather than independent cross-model conclusions.

For SEO teams, the practical effect is lower cost for page crawling, product-attribute extraction, source comparison, content checks, and first drafts. As models get cheaper, selection of evidence, permissions, fact review, and business impact become more important. A new model is a reason to run a controlled test on real pages and queries, not a reason to rewrite the whole SEO method.

SAFE is an abuse-forensics system, not a web-page AI detector

Google researchers published The Synthetic Gap, a paper describing SAFE, or Scaled Abuse Forensics Examiner. The investigation is split into specialized tasks: one agent examines content and policy violations, another looks for timing and infrastructure patterns, another studies relationships between accounts or channels, and a root agent combines the evidence.

The paper describes coordinated synthetic media and abuse networks. Its abstract says early deployment can shorten investigation time, but the public material does not provide complete accuracy, recall, or sample-size data. It also does not define an AI-writing score for web pages as a ranking signal. Turning SAFE into “Google can detect whether a page was written by AI” goes beyond the paper.

The SEO connection is about measurement. If an agent is rewarded only for AI visibility, publication volume, or citation count, it can find a cheap way to hit the number. Give it constraints for source quality, product facts, working links, and human review instead.

Linus’s notes

Taken together, this week’s news makes ranking a less complete description of search performance. The spam update requires pages, dates, indexing, and site behavior to be examined together. The AI Mode experiment shows that visibility inside a search product and visits to a website can move in different directions. Multimodal reporting adds another search entry point that needs its own baseline.

Lower model prices will make research and technical checks easier to automate, but they will also increase the supply of low-value content. The durable advantage is a workflow that connects sources, facts, page changes, and business outcomes. A dashboard that only counts generated pages or brand mentions can look better while the pages remain weak.

What this means for you

1. Mark September 24 as the start of the spam-update observation window. After the rollout completes, compare affected pages across seven-day and 28-day windows for clicks, impressions, rankings, and indexing.
2. Separate AI visibility, external links, website visits, and conversions in GEO reporting.
3. Keep images, specifications, variants, and page copy consistent, then monitor the new multimodal search type in Search Console.
4. Compare models using real task cost, factual errors, editing time, and business outcomes rather than vendor leaderboards alone.
5. Give agents checks for sources, dates, link status, and factual consistency, with a human approval step before publication.

The important change this week was not another tool button. Search entry points, click paths, and automation costs are all moving. SEO and AI-search reporting needs to answer what users saw, whether they reached the site, and whether the interaction produced useful business value.