SEO Recap covering August 17, 2026: AI visibility research and a Search Console AI report bug

Built by Stephanie Chung·6 min read·10 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)

Fractl found more than 90% of brands showed AI visibility tracking with traditional search authority, but the outliers expose the real gap.
~5% (471 brands) were underexposed in LLMs despite high domain ratings and traffic, while ~4% (377) overperformed relative to weak SEO signals.
Third-party coverage in roundups, comparison reviews, and expert lists correlated with AI recall more than a brand's own content.
Only 11% (900 brands) were referenced by all 3 models and 77% appeared in just one, so measure each model separately rather than a blended score.
Category examples showed digital-native brands winning:
- Travel concentrated in Booking.com (285 mentions), Airbnb (227), and Expedia (215)
- Lemonade (213) and Root (165) outranked State Farm (172) and Progressive (114) in insurance

RESONEO analyzed 1,200 ChatGPT answers, 88,000 search results, and 26,900 pages in July, identifying 3 layers: a discovery index, a shared reading cache, and a small set of live page opens.
labrador is OpenAI's own web index and feeds nearly all instant-mode citations, while bright and oxylabs are scraped Google results bought live.
Free Think doubles retrieval (15.1 to 35.3 URLs) but draws 74.7% from labrador, whereas paid Thinking is 75.3% scraped Google, so the 2 modes search different webs.
Instant mode opens no pages in 93% of answers and grounds responses on titles and query-independent 200-character snippets, ignoring the meta description for labrador.
ChatGPT keeps a shared Markdown cache of fetched pages that ignores Cache-Control and noindex, strips JSON-LD, does not run JavaScript, and caps pages at 4 MB.

Jeremy Howard published llms.txt v2 on August 10, the first revision since launch, adding standard ways for agents to discover Markdown versions of pages.
V2 supports 2 Markdown URL patterns and 2 link relations:
- rel="alternate" type="text/markdown" linking a page to its Markdown version
- rel="describedby" pointing to the llms.txt file that covers it
Google Search still ignores llms.txt and says it neither helps nor harms rankings. The format serves coding agents, IDE tools, and documentation platforms.
Updating an existing file to V2 needs only a minor change, and the spec is still open for feedback on GitHub.

AI assistants now return pre-selected short lists of local businesses. Uberall's quick-service benchmark typically produced 3-5 recommended brands per query.
Whitespark's test of 540 queries found AI Overviews on 15% of direct local queries, 92% of informational ones, and 97% of hybrid ones.
Keep hours, services, and service area in plain text on crawlable pages, consistent with listings, and use structured data that matches visible content.
Google's Generative AI report covers Google only and omits queries, so teams gauge AI visibility with scheduled assistant tracking and AI referral analytics, each with blind spots.

Google's paper found frontier LLMs like Gemini-3-Pro and GPT-5 encode 95-98% of tested facts but fail to directly recall 26-34% of them.
Recall failures account for more than 70% of GPT-5.2's errors, so the bottleneck is accessing encoded information rather than missing data.
Reversing the subject/object order from how a fact was learned hurts recall, while rephrasing the question had little effect.
More thinking recovered 40-65% of previously unrecallable facts but is computationally expensive, and scaling training is not a fix.
As an unproven SEO hypothesis, order subject and object entities the way queries most commonly phrase them.

Google stopped an experiment that showed AI-generated step images in AI Overviews for recipe results. Robby Stein called it "a small experiment we're no longer running."
Inspired Taste criticized the test for replacing creators who shoot and test recipes and for risking confusion over who made the illustrations.
Stein clarified this is separate from the Nano Banana feature announced in July, which only triggers when users explicitly ask to generate an image.

Anastasia Kourou spotted AI Mode reply fragments like "yes" and "yes go on" in Search Console, and Mueller confirmed AI Overviews and AI Mode data sits in the main performance report.
The new Generative AI report shows impressions by page, country, device, and date but omits queries and clicks and offers no API or BigQuery export.
16 months of the author's data yielded 1,127 conversational queries and 20,300 impressions, sortable into 7 kinds:
- Reply artifacts, pivot follow-ups, conversational questions
- Tracker probes, agent harnesses, pasted strings, long uncategorized
Position data (position 4.5 for "yes") shows links sit inside answer blocks, and 57.7% of web impressions carry no query string, so visible fragments are only a floor.

Google confirmed that AI Mode now uses Gemini 3.7 Flash, rolling out globally for Google AI Pro and Ultra subscribers in English.
Subscribers access it via the plus icon and model selector, and it should eventually replace 3.5 Flash-Lite as the default once the subscription requirement is removed.
Glenn Gabe and Gagan Ghotra saw only modest differences in testing, though 3.7 better understood intent in some queries.

Google confirmed a logging error cut impressions in the Generative AI in Search performance report for data starting August 13, 2026.
John Mueller said the drop is a logging issue only, not a real visibility change, and an annotation is being added to Search Console.
The dip is widespread across properties, so site owners seeing the cliff are not alone.

A company paused ~$113,000/month of branded and non-brand paid search across the US, UK, Australia, and Canada, and branded never came back.
In the branded campaign analyzed, organic already won 71% of overlapping clicks. Only 10.9% ($3,945 of $36,129) bought incremental clicks, with 89% going to brand defense.
Combined organic and direct revenue recaptured 30% of paid-attributed revenue within 6 weeks and 65% by week 13.
Top-line revenue fell ~$30-40k/month, but net P&L rose ~$5-20k/month after removing spend. Paid traffic converted at 2.9% versus 2.3% for organic.

Published on August 18, 2026