SEO Recap covering August 14, 2026: Gemini 3.7 Flash in AI Mode and ChatGPT's pre-search shortlist

Built by Stephanie Chung·6 min read·7 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)

ChatGPT rewrites your question into its own search queries (readable in the response under a search_queries key) that already contain brand names the user never typed, before any page is fetched.
In 21 of 27 conversations the first query, logged before any retrieval, named brands the user did not, and the injection happened whenever ChatGPT had to supply the products itself.
Appearing in that query was worth ~33 times more than merely being findable, and in 86 cases a brand was recommended without its site ever being fetched.
A second filter is brutal: across 3,554 retrieved pages only 110 (3.1%) were cited, with position inside a domain group deciding nearly everything and stacking multiple pages per domain hurting per-page conversion.
Run your category question 5 times to find your real competitive set, spend on being written about and reviewed if you are absent from the query, and consolidate cannibalizing pages so the answering sentence sits near the top in plain HTML.

Google added Gemini 3.7 Flash as a selectable model in AI Mode, one day after releasing it for coding and agent tasks, calling it better at following instructions and understanding intent.
The rollout is global in English but limited to Google AI Pro and Ultra subscribers, so it does not reach AI Mode's free tier.
You pick it from the model menu's "Gemini 3 models" section, next to Auto and Pro, via the "+" icon in the "Ask anything" bar.
Google has not said whether 3.7 Flash becomes the default or feeds Auto routing, and its support docs still list only Fast and Pro.
For tracking inconsistent AI Mode citations, run the same query across 3.7 Flash, Auto, and Pro to see if the model is causing the differences.

Google has begun rolling out AI-generated images inside AI Overviews, a month after announcing image generation would come to Search.
Inspired Taste showed AI-generated step cartoons appearing over a spaghetti squash recipe and said the feature tries to replace creators who test, photograph, and film recipes.
There is a risk searchers assume one of the websites cited in the AI Overview produced the illustrations, which the creator called confusing and oversimplified.

TollBit's State of the Bots report for H1 2026 found ~15% of identified AI page-fetchers reached URLs European sites had marked disallowed.
ChatGPT-User, Bytespider, and Youbot each accessed disallowed pages on nearly half of the European sites that listed them, with ChatGPT-User reaching the most.
Newer agents are barely blocked: Claude-User is disallowed by 9% of European sites vs 26% in North America, and Perplexity-User 13% vs 26%.
OpenAI's docs say robots.txt may not apply to ChatGPT-User because a user initiated the fetch, while OAI-SearchBot, not ChatGPT-User, governs whether a site appears in ChatGPT search results.
From September 15 Cloudflare will block Training and Agent crawlers by default on ad pages for new domains, moving enforcement to the network layer while allowing Search crawlers.

Cloudflare's CFO told analysts non-human traffic could reach 1,000 times human traffic within 5 years, after machine traffic passed human traffic in May 2026, ahead of the company's 2027 estimate.
On one site over 24 hours, CCBot logged 1,510 requests (the top crawler) but behaved as a credential scanner, requesting paths like /.ssh/known_hosts, /.env.production, and /firebase-service-account.json rather than any articles.
The scanner's wordlist now includes agent tooling config: /.mcp.json (30 requests) and /.continue/config.json (24), files that commonly hold API keys and access tokens.
Cloudflare's own engineering post the same week said well-behaved bots re-fetch unchanged pages by the billions, "machine effort attached to no outcome at all."
Read your crawler analytics by request path, not totals, since hostile scans can be counted as AI crawler traffic while leaving no security event.

Gary Illyes said Googlebot's IP location is a declared text field set to US, Mountain View in most cases, but the actual egress points vary by data center (for example a cluster in ATL egresses from ATL).
IP addresses have no true geographic location and default to the registrant's location or BGP anycast routes, like 1.1.1.1 or 8.8.8.8.
Google can crawl locally from different regions with location-aware crawlers when needed, so a self-declared California origin may not reflect the real source.

Anthropic plans to mark Claude-generated text under EU AI Act Article 50(2) with no user opt-out, though it names no model that currently carries a mark.
Marking is uneven by design. The Code of Practice exempts free-form text under 200 tokens, and heavy editing, translation, format conversion, or screenshots can strip the signal.
Watermarks proved fragile in research: at ICML 2024 attackers removed or spoofed marks for under $50 at over 80% success, and at ICML 2025 a team defeated 7 methods at $0.88 per million tokens.
No tool lets the public verify text yet. Google's SynthID Detector covers images, video, and audio, and OpenAI's covers images and audio.
For content teams, AI-assisted copy can leave Claude carrying a machine-readable mark that travels into a CMS, and Google, Microsoft, and Meta signed the same Code, with no word on whether marked content is treated differently when crawled or ranked.

Published on August 15, 2026