SEO Recap covering September 3, 2026: Gemini 3.8 Flash citation bug in AI Mode

Built by Stephanie Chung·9 min read·12 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)

When AI lacks data on a company, it substitutes a competitor or category average
Source: Search Engine Journal

  • When a model has thin evidence on a company, it produces confident substitutions from the nearest well-documented neighbor rather than abstaining, a systematic and directional failure distinct from random hallucination.
  • Making models bigger mainly improves recall of popular facts and does little for the long tail, so scale won't help sparse mid-market or regional entities.
  • Substitution takes 4 forms:
    • Silent analogy, where a competitor's details are described as yours
    • Staleness presented as current fact
    • Thin evidence stated with the confidence of thick evidence
    • Category knowledge attached to a specific company
  • Content audits inspect what you own and cannot catch this, since the missing evidence lives in third-party sources you don't control. Find it by watching model outputs on questions where your name is absent and must be supplied by the model.

SEO priorities for 2027: funnel allocation, deep content, and citation-aware markup
Source: Search Engine Land

  • Seer Interactive's 2026 study puts AI Overview presence at 36% of informational queries versus 8% commercial and 5% transactional, but comparison ("X vs. Y") formats trigger an AI Overview 95.4% of the time and "best of" formats 81.3%.
  • Organic CTR on AI Overview queries bottomed at 1.3% in December 2025 and rose to 2.4% by February 2026. Cited pages earn ~120% more clicks per impression than uncited pages but still trail no-AI-Overview results by 38%.
  • Prioritize MOFU and BOFU pages, build comparison content with original data and self-contained passages a model can lift, and track citations alongside rank position.
  • Schema studies show weak returns for AI visibility:
    • Ahrefs tracked 1,885 pages adding JSON-LD and found no major citation uplift, with AI Overview citations down 4.6%
    • A searchVIU test found ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode all read only visible HTML and ignored JSON-LD during retrieval
    • Cyrus Shepard's review of 54 studies placed structured data mid-pack as a citation factor
  • Google retired FAQ rich results on May 7, 2026, so prioritize Organization, Product, Article, and entity markup that resolves ambiguity rather than markup chasing rich results.

AI in Search / AI Overviews

Google confirms Gemini 3.8 Flash broke source citations in AI Mode
Source: Search Engine Land

  • Google launched Gemini 3.8 Flash and added it to AI Mode for AI Pro and Ultra subscribers, but the model stopped linking or citing sources in its responses, especially for top-of-funnel queries.
  • Robby Stein, VP of Product at Google Search, confirmed on X that the behavior "isn't working as intended" and said a fix would roll out soon.
  • The bug only affected users on Gemini 3.8 Flash, which at this point was limited to paid subscribers.

Product pages lead AI citations at bottom of funnel, Ten Speed study finds
Source: Search Engine Journal

  • 10 Speed analyzed 7,387 citation appearances from Peec AI across ChatGPT, Perplexity, Claude, and Gemini using 170 buyer-stage prompts spanning B2B SaaS and professional services clients.
  • Citation share by page type at the evaluation stage:
    • Product pages 24.1%
    • Articles (blog, news, PR) 17.4%
    • Comparison pages and listicles ~13% each
    • How-to guides just under 9%
    • Homepages 7.8%, G2/Capterra directory profiles 7.2%
    • Reddit, YouTube, and forums combined 4.2%, with YouTube ~1%
  • Brand-controllable content accounted for 88.3% of citations, and comparison-format prompts returned 1.33 times their share of the prompt set.
  • The 4% versus 24% headline split was descriptive only, never statistically tested, and covers a single point-in-time pull of one confidential client base with no per-platform breakdown.
  • Rewrite product pages and homepages in plain language, build comparison content against more than obvious rivals, and manage G2 and Capterra profiles as structured content.

AI Mode shows different products, sellers, and prices than Shopping carousel
Source: Search Engine Journal

  • Productrise tracked over 2 million listings across more than 100,000 searches from August 9-31, 2026, in the US and UK, running the same product searches in AI Mode and regular search on the same day.
  • Only 1.28% of carousel products also appeared in AI Mode for the same query and day, and among matched products the first-listed seller differed 49.6% of the time.
  • When both result types returned products, the carousel showed an average of 27.8 products versus 3.9 in AI Mode.
  • Prices differed on 38.1% of matched products, with AI Mode showing the higher price 68.4% of the time and averaging 21.6% higher across all matched products.
  • Google told Futurism it hadn't verified the data and said all shopping results draw from the same Shopping Graph.

Google adds WeatherNext 3 to Search as it shifts toward task-based results
Source: Search Engine Journal

  • Google introduced WeatherNext 3, its most advanced global-scale AI weather model, described as up to 50% more accurate with hourly updates, real-time satellite data, and better rain forecasts.
  • The model rolls out globally across Search, Gemini, Maps, Google Maps Platform, Earth Engine, and Cloud.
  • The move fits Sundar Pichai's framing of Search evolving from 10 blue links into an agentic, task-based surface where users complete things rather than just find information.

Technical SEO

Blocking AI crawlers via robots.txt versus the server stack
Source: Search Engine Journal

  • Robots.txt is the most accessible method and is officially supported by major AI companies (OpenAI's GPTBot and OAI-SearchBot, Anthropic's ClaudeBot, Google-Extended, Perplexity's PerplexityBot), but compliance is voluntary and unmonitored.
  • Server, CDN, and WAF blocks are enforced rather than requested, stop requests before they consume bandwidth, and log blocked attempts, though they require developer access and add maintenance overhead.
  • WAF is the strongest layer because it analyzes request behavior and can detect bots spoofing other user agents, but no server-stack method is fully foolproof against advanced scrapers.
  • For a strong need to block, go as high up the stack as possible (WAF, then CDN, then server). For one or 2 reputable crawlers, robots.txt as a deterrent plus server-log monitoring is usually enough.

Organic Search & Algorithm Updates

Mueller says ranking recovery can take months after fixes
Source: Search Engine Journal

  • A Redditor whose Google Business Profile was suspended disavowed over 800 "toxic" links, including old Fiverr guest posts, then saw organic rankings collapse within 2 weeks.
  • John Mueller said disavowing a handful of links won't drop a site out of search and advised fixing the disavow file (possibly emptying it) before looking for other causes.
  • Disavow files are reprocessed incrementally over roughly 3-6 months, so a strong ranking drop days later is almost certainly unrelated.
  • Recovery from algorithmic issues such as a broad spam update can take many months, while quick fixes like an accidental site-wide block resolve fast.

Freelance demand for fixing AI-generated work climbs sharply
Source: Search Engine Journal

  • Job listings for correcting AI-generated work rose 87% on Freelancer.com from August 2025 to June 2026, totaling 10,760 posts, per internal data reported by the Guardian.
  • Upwork reported a 70% year-over-year rise in AI remediation gigs, and Fiverr said searches for "AI cleanup" grew more than 20-fold from 2023 to 2026.
  • Each platform measures a different signal (listings, gigs, keyword searches), so the figures aren't directly comparable. Graphic design was the most common cleanup category on Freelancer.com.
  • A Management Science study found writing and coding job posts most exposed to automation fell 21% within 8 months of ChatGPT's launch, and image-creation posts fell 17% after image generators arrived.
  • Freelancers say cleanup can take as long as original work, undercutting the assumption that AI first drafts are cheap to finish.

Spot SEO warning signs before organic traffic falls
Source: Search Engine Land

  • Stable rankings can hide instability when Google keeps swapping which URL ranks for a query, so watch the ranking URL as closely as the position for important searches.
  • Rising impressions don't equal better performance. Separate informational visibility from commercial visibility, since thousands of low-value impressions can mask declining clicks on product or service terms.
  • Competitor gap analysis based on keyword counts misses why a competitor's page wins. Ask what their page helps a customer understand, verify, or decide that yours doesn't.
  • When leads fall while sitewide traffic holds, investigate the query mix and landing pages before assuming it's a conversion problem, and talk to the people handling inquiries about lead quality.
  • By the time the main traffic graph turns red, the useful warning is often months old. Good analysis compares rankings, queries, pages, and commercial results rather than chasing a single green number.

Publisher content randomly disappearing from Google News tab
Source: Search Engine Roundtable

  • Barry Schwartz replicated content that was crawled, indexed, and shown in the All tab but missing from the News tab, affecting both large publishers like WSJ and smaller sites.
  • The behavior appears random with no discernible pattern. Missing stories sometimes reappeared in the News tab hours later.
  • Google did not respond to questions about the issue, and by ~1 pm PT on September 3 the affected content had begun showing again.

Use AI to remove tedious work, not judgment
Source: Search Engine Land

  • The safest AI "easy buttons" eliminate tedious work around a decision rather than making the decision, keeping strategy, prioritization, and final approval with accountable humans.
  • Reliable use cases include:
    • Turning meeting transcripts into decisions, tasks, owners, and deadlines
    • Categorizing and flagging patterns in Search Console or competitor keyword exports before analysis
    • Repurposing existing work into new formats like LinkedIn posts or meta descriptions
    • First-pass editing to catch gaps, contradictions, and unanswered questions
    • Producing testable technical work such as regex, SQL, schema markup, and redirect maps
  • The boundary is that AI can identify where something changed but shouldn't be trusted to determine why without validation, and no AI-generated code should deploy to production unreviewed.
Published on
geoai searchtechnical seoorganic searchdigest