SEO Recap covering October 13, 2026: Gemini 4 Argon in AI Mode and Bing's Topics Portfolio

Built by Stephanie Chung·7 min read·9 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)

Bing adds Topics Portfolio beta to AI Performance reports
Sources: Search Engine Journal · Search Engine Roundtable

  • Microsoft launched Topics Portfolio, a beta in Bing Webmaster Tools' AI Performance report that groups the grounding queries a site is cited for into broader themes like football, taxes, or AI chatbots.
  • Each topic row shows 3 estimated metrics:
    • Topic Size, total citations across all sites for the topic (a demand gauge)
    • Sites Cited, the approximate number of sites appearing for the topic's queries
    • Queries, the number of distinct grounding queries in the topic
  • A Top Opportunity Queries list surfaces queries where other sites are cited but yours is not, with citation volumes shown as ranges rather than exact counts.
  • The view only appears if Bing has enough data, so smaller sites may see few items or none. Microsoft calls the numbers directional signals for planning, not a precise ledger, and notes AI-generated topics can mislabel queries or surface rival-brand searches as opportunities.

AI brand recommendations shift sharply by query language
Source: Search Engine Journal

  • A test ran 6 shopping and booking scenarios across 5 languages (English, Arabic, Russian, Simplified Chinese, Persian) via DataForSEO's LLM Responses API, mostly on Claude Sonnet 5 with web search off, for 1,650 total responses.
  • Airline picks swung the most: Claude chose British Airways 52.5% of the time in Persian but never once in Russian, where easyJet dominated at 72.5%, a 60-point swing on the same question the same afternoon.
  • Hotels varied too. Bahia del Duque topped 43.6% of Chinese responses but only 7.7% of Arabic ones, and Hotel Botanico took nearly a third in English but vanished in Russian.
  • Some categories stayed stable regardless of language (Roborock for robot vacuums, Monday.com over Asana for project management). Turning web search on changed the winning brand in 7 of 10 language-category combinations tested.
  • Anthropic's July report on ~310,000 conversations across 20 languages found Claude's demonstrated values shift by language, leaning warmer in Hindi and Arabic and more rigorous in English and Russian. Test AI visibility in every language that matters, not just English.

AI in Search / AI Overviews

Gemini 4 Argon added to AI Mode for Pro and Ultra subscribers
Sources: Search Engine Journal · Search Engine Roundtable

  • Google added Gemini 4 Argon to the AI Mode model menu for AI Pro and Ultra subscribers globally, selectable from the "+" icon in the "Ask anything" bar, the same day it reached developers, enterprises, and Gemini app users.
  • Robby Stein says Argon brings improvements to nuanced, multi-step tasks and is better at understanding intent and following instructions across back-and-forth conversations, with the launch post claiming it follows style instructions better than any prior Gemini model.
  • Unlike the 3.8 Flash release, Argon's AI Mode responses include links this time.
  • Argon is a subscription-only option, like the Flash models before it. Google has not said whether it will become the AI Mode default. Gemini 3 Pro launched as an option and stayed one, while Flash models launched as defaults.

Google tests pharmaceutical ads in AI Mode and AI Overviews
Source: Search Engine Roundtable

  • A Google spokesperson told Adweek it recently expanded standard text and Shopping ad testing in AI Mode and AI Overviews to include certain pharmaceutical products, including prescription drugs.
  • The pharmaceutical testing began this month, is limited to the US, and appears in sections labeled "Sponsored."
  • Google said there were no recent drug or pharmaceutical ad policy changes. The expansion is to the AI surfaces where these ads show, not the underlying policy.

Technical SEO

Google explains why site migrations can revert months later
Sources: Search Engine Journal · Search Engine Roundtable

  • A site owner reported a smooth early-2026 domain migration that recovered by April, then saw Google revert 4 months later. Over 5,000 new-domain pages flipped to "Crawled, currently not indexed" and impressions fell from ~5,000 per day to 20 per week while the old domain got reindexed.
  • John Mueller said migrations can sometimes take a particularly long time to fully process, and recommended setting up Change of Address in Search Console for both the www and non-www versions so the domain fully migrates.
  • Diagnose a stalled migration as a canonicalization problem. Check for hardcoded absolute internal links to the old domain (crawl with Screaming Frog beforehand), confirm redirects preserve URL paths, avoid chained redirects, and update the XML sitemap to the new domain.
  • Marie Haynes and Glenn Gabe report similar migrations going wrong months later. Haynes theorizes a migration triggers a fresh quality assessment and Google is now stricter about indexing content it judges unoriginal.

Google's Search Off the Record covers PDF indexing and AI grounding
Source: Search Engine Roundtable

  • Google converts PDFs to an HTML representation during indexing, so converted PDFs rank and index with the same systems as HTML pages and can appear in AI Overviews, AI Mode, and be used for AI grounding.
  • Embedded text links in PDFs are crawled normally, but embedded images lack standalone URLs and generally cannot enter Google Image Search.
  • Watch key pitfalls:
    • Rasterized or scanned PDFs need OCR. Test by copying and pasting the text
    • Identical PDFs are exact binary duplicates, making cross-domain canonicalization hard. Use rel=canonical HTTP headers, PDF metadata, and in-document links back to your site
    • Password protection blocks bots entirely
  • Google recrawls PDFs much less often than HTML, so publish updated PDFs under a new URL or filename rather than overwriting, and avoid auto-generated "print" PDF duplicates that compete with live pages. Keep llms.txt as plain text, never a PDF.

Merchant Center product videos stuck in "preview processing"
Source: Search Engine Roundtable

  • Product videos submitted to Google Merchant Center are stuck in "preview processing" and never play, appearing static in product listings in Google Search and other surfaces.
  • The issue affects both YouTube links and videos hosted on third-party platforms or your own website, and reports say it is widespread.
  • Google confirmed the issue via email to affected merchants, but there is no ETA for a fix.

Google Search Console performance date picker breaks again
Source: Search Engine Roundtable

  • The Search Console performance report date picker broke, with the end date stuck at 1969 or 1970 and an error reading "Date can't be after 1969-12-31" when users tried to change it.
  • Default ranges (24 hours, 7 days, 28 days, 3, 6, 12, and 16 months) kept working. Only the custom date range failed.
  • It was fixed on October 13, then broke again the morning of October 14.

Organic Search & Algorithm Updates

Common mistakes that undermine SEO experiments
Source: Search Engine Journal

  • Start every test with a specific, falsifiable hypothesis that makes a prediction the experiment could contradict. "Improving titles improves performance" is not testable. "Adding a country reference to 700 page titles generates 10% more clicks over 4 weeks versus the prior 4 weeks" is.
  • Isolate variables and keep a valid comparison. Log changes across SEO, engineering, product, and marketing, plus algorithm updates and seasonality, to rule out confounding explanations for any movement.
  • Commit to your primary outcome, period, and success criteria before seeing results, borrowing the preregistration principle to avoid cherry-picking metrics after the fact.
  • Distinguish evidence of an effect, evidence against it, and insufficient evidence. Absence of evidence is not evidence of absence, and short windows or small samples may simply fail to detect a real effect.
  • Treat experiments as a compounding program, not isolated wins. One uplift on one page group during one period is not a universal law. Document question, hypothesis, treatment, sample, dates, result, and limitations.
Published on
geoai searchtechnical seoorganic searchdigest