Official GEO is a subtraction list plus one Search Console switch

The official GEO document is not a shopping list. It is a subtraction list, plus one Search Console switch.

Google Search Central published Optimizing your website for generative AI features on 15 May 2026 and touched it again on 10 July. Third-party AEO/GEO checklists still open with llms.txt, then chunking, then a special schema.org type, then a rewrite "for the model." The official mythbusting section deletes that column first.

I did not log into Search Console for this piece. I did fetch the public bytes this site already serves. Ranking is not in scope.

What AI Overviews actually pick up

Two names, one index. AI Overviews attach a short gist and supporting links to a hard question. AI Mode is the conversational surface for comparisons and multi-step reasoning. Both sit inside Google Search. Both pull live pages. Google describes the machinery as RAG on top of the core ranking systems, plus query fan-out: a lawn-weeds question can spawn separate retrievals for herbicides, chemical-free removal, and prevention.

Eligibility is older than the names. A page has to be indexed and allowed to show a snippet. AI features and your website puts it this way:

There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.

Source: AI features and your website

No extra technical bar. No special optimization. The same page still wants the technical requirements, the spam policies, and people-first content. Meeting all of that still does not guarantee a crawl, an index, or a serving slot. That hedge is the same hedge Search has always used.

Snippet eligibility is a lever I already measured. nosnippet and max-snippet:0 close the page as direct input to AI Overviews and AI Mode. The write-up lives in robots snippet directives. I am not rerunning that parser today. A layer landed above it.

The optimization guide adds this: besides the usual technical requirements, a site must be included in Search generative AI features in Search Console before it is eligible to appear there. That is a property setting, not a template ticket.

Four backlog items the official guide deletes

The mythbusting section treats "AEO" and "GEO" as nicknames for optimizing the search experience. The point fits in one sentence.

From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.

Source: Optimizing your website for generative AI features

What drops out of an engineering backlog:

The llms.txt line is blunt.

Doing so will neither harm nor help your site's visibility or rankings in Google Search, as Google Search ignores them.

Source: the LLMS.txt item in the same optimization guide

Harmless and useless, for Google Search. Fine to keep for some other consumer. Off the Google Search board. If robots.txt already splits training bots from search bots, today's job is not another file. It is dropping the idea that Google Search reads this one.

Read the structured-data paragraph twice. It is not required for generative search and there is no dedicated type. It still helps rich-result eligibility, so keep it in the wider SEO plan. I do not read that as "delete JSON-LD." I read it as "do not add a type because AI Overviews exist." Structured data has never guaranteed rankings. It does not guarantee citations either.

Inauthentic mentions around the web are, officially, not as helpful as they sound. Core ranking looks at quality. Other systems block spam. Generative features depend on both. That is not a sprint ticket.

Third-party tools that claim "internal" Google metrics get the same haircut. No third-party tool has access to internal ranking or AI systems. The third-party SEO guidance says to check AEO/GEO advice against official docs. Using a tool in a workflow is fine. Treating its number as Google's number is not.

The Search Console switch, and who inherits it

"Included in Search Console" points at the Search generative AI control. Path: Settings > Search generative AI.

Three states. Include the site's links and content in generative AI features. Exclude them. Inherit from the parent. Include is the default for every property. Exclude removes the site from AI Overviews, AI Mode, and generative features in Discover: no links, no grounding input, no impressions or traffic from those features.

The limit, in the help page's own words:

This control only affects whether your content can appear in certain Search generative AI features; this control isn't used as a ranking or inclusion signal affecting other parts of Search.

Source: Search generative AI control

Not a ranking signal for the rest of Search. Not a training switch either. Training limits sit on Google-Extended. Full removal from Search is noindex. After the control goes live, exclusion usually lands in 1〜2 days. Cache can stretch that.

This is an ownership problem. Flip exclude on the domain property and every child URL-prefix property that still inherits follows it. A blog living at https://example.com/blog/ can fall out of generative features because someone touched the parent. Clean HTML and a careful robots.txt do not reopen that layer.

The control and the report are still on a subset of sites. The 3 June 2026 announcement of the Generative AI performance report is explicit about that. The report counts impressions, not clicks or positions. Search Labs experiments are out. A missing screen does not mean the site is excluded. It may not be in the rollout, or it may not have enough generative impressions yet.

I did not open Search Console for this article. I will not claim the menu exists on this property. What I can claim is the documented default (include) and the documented inheritance.

For a team, invert the order of work. Read the parent property first. Write down whether each child inherits or overrides. Only then audit robots meta and the live robots.txt. Reverse that and you can spend a week on markup while the site is simply not included. The help page already says 1〜2 days, so an unchanged report the same afternoon is not a reason to revert. If the people who ship HTML and the people who own Search Console are different groups, this layer never shows up in a pull request.

The live robots.txt was not the file in git

After the official subtraction list, I did not add a file. I fetched what a crawler is offered. 14 August 2026, public URLs on https://jangwook.net only.

curl -sI https://jangwook.net/llms.txt
# HTTP/2 404

curl -sL https://jangwook.net/robots.txt | wc -l
# 106

# public/robots.txt in git: 45 lines, 1,101 bytes
# live response: 106 lines, 2,937 bytes

llms.txt is 404. So are LLMs.txt and the www host. That matches a site that did not plant a file Google Search says it ignores.

robots.txt did not match git. The repo file is 45 lines: training bots (GPTBot, ClaudeBot, CCBot, Google-Extended) are disallowed, search bots and * only lose the cross-language URLs. The live response prepends a CDN-managed block and becomes 106 lines. User-agent: * carries Content-Signal: search=yes,ai-train=no,use=reference. Several training and extended bots get another Disallow: /. The repo body is still there, after the prefix.

I did not find Content-Signal listed as a supported rule in the Search Central robots.txt material I used for this piece. Presence in the live file and consumption by Googlebot are different claims. I am not making the second one. Treat the token as a third-party convention. Do not hang Google Search eligibility on it.

Eight pages (/, /ko/, /en/, /ko/blog/, three posts, /ko/contact/) all returned HTTP 200. Zero <meta name="robots"> tags. Zero data-nosnippet attributes. One post's description mentions the word nosnippet. That is copy, not a directive. The template only emits a robots tag when noindex is on.

JSON-LD on the home page included Organization, Person, and WebSite. Posts carried BlogPosting, WebPage, and BreadcrumbList. Under today's official text, that markup is not a ticket into generative search. It stays on the rich-result and entity side, the same split I used when FAQ rich results ended and the Q&A markup stayed. Validation and visibility were never the same event.

The live/git split is the measurement. If GEO work means "add a file to the repo," the file a crawler already reads has grown outside the repo. Diff the deployed URL, not public/robots.txt against itself.

Agents also read the accessibility tree

The guide's last stretch is browser agents: booking, comparing specs. Different job from search citations. Same surface. web.dev's agent-friendly note lists three views: screenshots, raw HTML, the accessibility tree.

The accessibility tree keeps roles, names, and states, and drops visual noise. It is the same tree a screen reader uses. Style a div as a button and the DOM-only path never sees a button. The screenshot path may see where it sits and still miss what it does. Semantic HTML and label for are not a score chase. They are how a machine avoids guessing the wrong action.

The optimization guide talks about semantic HTML as readability and assistive parsing, not "perfect code." The web is not valid HTML. Google can still read it. The reason to use semantic elements anyway is not only Googlebot. Agents walk the same tree.

The leftover work is dull. Prefer button and a. Wire labels with for. Do not hide hit targets under a transparent overlay. Do not let the layout jump by category. That is WCAG work. It is not a new agent file format.

web.dev also says a screenshot-only path is slow and expensive, a backup when structure is muddy. Put the main path on the tree and the DOM and the text a crawler reads comes from the same markup an agent uses. That is the opposite of stuffing hidden search copy.

What leaves the sprint, what stays

The official GEO page is not asking for more surface area. It deletes the market list, then keeps the technical Search surface and one Search Console switch.

Leave the sprint:

• A new llms.txt for Google Search, a special AI markdown file, a generative-search-only schema.org type
• Chunking and AI-only rewrites, and a page per query variant
• Third-party "internal" metrics as a deploy gate

Keep:

• Indexing and snippet eligibility. No stray nosnippet in the template. noindex only where you meant it
• Fetch the live robots.txt and diff it against git. Confirm a CDN prefix did not invert training bots and search bots
• The generative AI control on parent and child properties. Default is include. A missing UI can mean "not in the rollout," not "excluded"
• Semantic HTML and the accessibility tree. Agent guidance is existing markup
• Rewrite the purpose line on rich-result JSON-LD from "required for generative search" to "rich-result eligibility"

The smallest commands:

curl -sI https://example.com/llms.txt | head -n 1

curl -sL https://example.com/robots.txt > /tmp/live-robots.txt
diff -u public/robots.txt /tmp/live-robots.txt

python3 - <<'PY'
import re, sys, urllib.request
html = urllib.request.urlopen(sys.argv[1]).read().decode("utf-8", "ignore")
print("robots meta:", re.findall(r"<meta[^>]+name=[\"']robots[\"'][^>]*>", html, re.I))
print("data-nosnippet attrs:", len(re.findall(r"<[^>]+data-nosnippet", html, re.I)))
PY
https://example.com/your-page/

Those three lines still cannot see the Search Console switch. That layer is a property setting. A pull request will not catch it.

Nothing here promises an AI Overview. Include can stay on, snippets can stay allowed, the page can be indexed, and Google still does not owe a citation. I measured the layers of eligibility, not the size of an effect.

If you line up the repo, the live URL, and the property setting and cannot see which layer is open, bring that layer. Matching official text to the bytes you actually ship is the work I do.

Sources: Google Search Central generative AI optimization guide (updated 2026-07-10), AI features and your website, third-party SEO guidance, Generative AI performance reports announcement (2026-06-03), Search Console Help Search generative AI control and Generative AI performance report, web.dev Build agent-friendly websites (all official). The four English block quotes were fetched from those pages, whitespace-folded, and checked against the source; each quote sits next to its URL. Live fetch: 2026-08-14, robots.txt, llms.txt, and eight pages on https://jangwook.net, curl plus HTML parse. Data: data/official-geo-gsc-control-probe-2026.json. Figures: scripts/chart-official-geo-gsc-control.py. Search Console was not opened. Content-Signal appears in the live robots.txt and was not confirmed as a supported Search Central robots.txt rule. Structured data and this switch do not guarantee rankings.