Google Just Killed 4 AI SEO Tactics. Here Are the Claude Prompts That Replace Them.

I've been editing 100+ sites for 5 years. I've watched whitehat, blackhat, grandpa SEO, the AMP wave of 2016, the aggressive schema wave of 2019. For the past 18 months, a new vague has been selling audits at $500 to $2000 for something called GEO or AEO. Entire agencies built around 2 acronyms, dedicated SaaS, certifications, WordPress plugins that auto-generate an llms.txt file at the root of your site. On Friday May 15 2026, Google published 2000 words that make roughly 90% of what this industry sells obsolete.

The Google doc lists what doesn't work, by name: llms.txt and special markup, content chunking, AI rewriting, inauthentic mentions. On the 100+ sites I edit, I have never implemented any of these 4 tactics. Not laziness, observation. At every wave (AMP, aggressive schema, AI-friendly markup), the ROI was the same: zero or marginal.

You'll learn 3 things here:

• what the GEO industry currently sells (grey hat + emerging black hat)
• the only mechanism that actually decides whether AI cites you
• 4 Claude prompts I run before publishing anything

What Google Actually Says About GEO and AEO

GEO stands for Generative Engine Optimization. AEO stands for Answer Engine Optimization. An industry built itself on these 2 acronyms over 18 months. Audits priced $500 to $2000 (I've seen the proposals land in client inboxes), specialized agencies, training programs, certifications, dedicated SaaS tools. The pitch was always some variation of "Google AI Overviews change everything, you need a different methodology."

Google's May 15 doc says the opposite, in plain English. Paraphrasing the position: AEO and GEO are just SEO done well. The term is marketing, not technical. Cyber Kendra put it sharper a day later: anyone selling you an AEO audit distinct from a standard SEO audit is selling you something Google doesn't recognize as real.

The doc names 4 tactics that don't work, by name:

• llms.txt and special markup
• content chunking
• rewriting content for AI systems
• inauthentic mentions

A 5th one is mentioned in passing: overfocusing on structured data. I'll fold that one into the schema spam coverage below.

This section is the official position. The next one is what's actually for sale on the ground in May 2026, and why none of it survives contact with the Google doc.

What's Actually for Sale on the Ground

Observation, not tutorial. None of these were implemented on my sites. All of them showed up in client audits I reviewed over the past 18 months. The point is to give you a receipt to wave at the next agency proposal that lands on your desk.

llms.txt obsession. Yoast SEO and Rank Math added auto-generation in 2025. A Markdown file placed at site root, supposed to "guide AI systems". A common variant is llms-full.txt that duplicates your entire site content in markdown. Empirical demolition: Somanath Balakrishnan ran a study on 300,000 domains and found no statistical correlation between the presence of an llms.txt and LLM citation rate (Medium, February 2026). Most AI crawlers don't even request the file. The plugins ship anyway because they sell upgrades.

Forced chunking + FAQPage spam. Cutting articles into Q&A micro-paragraphs in "snippet-bait" format. FAQPage schema slapped on everything. The perverse effect documented by Pasquale Pillitteri: optimizing for the chunk produces skeletal content that loses editorial value without gaining visibility.

Mass AI rewriting. Prompts that rewrite commodity content into "AI-friendly" without adding any actual value. Every GEO SaaS sells this service. The irony: it produces exactly what Google flags as scaled content abuse in its spam policy. You optimize for the AI and end up triggering Google's anti-spam filter.

Separate markdown files per page. Generate a .md next to every .html to "facilitate AI ingestion". Documented risk (Derivatex, April 2026): if the markdown files are indexable, you introduce duplicate content that dilutes your crawl budget. You pay extra cost for negative value.

Schema spam stacking. FAQPage + HowTo + Article + Organization + Person + Review all applied on the same page. Google's doc is explicit: structured data isn't required for generative AI search, and there's no special schema.org markup you need to add. (The Rank Math plugin on a client site last year was applying FAQPage schema to the About page. Nobody had configured it. The default settings did it. That's the whole industry in 1 default checkbox.)

Fake listicles at scale. "10 best X tools" or "15 prompts you need" recycled across 50 sites with cosmetic variations. Search "Claude SEO prompts" on Medium right now, you'll find 5+ articles published in the last 30 days, content massively overlapping.

So that's the landscape. None of these are illegal, none of them are even particularly evil, they're just a wave of nonsense sold to people who don't have time to test. Google's doc finally gave us a public document to point at when an agency pitches one of them.

The Only Mechanism That Actually Decides Whether AI Cites You

This is the section that explains why the 4 tactics fail. Everything else in the article is a corollary of this.

AI search engines (Google AI Overviews, ChatGPT Search, Perplexity, Claude) work through RAG, Retrieval Augmented Generation. Concretely:

• the AI searches pages in its index or via a search API
• it reads passages
• it CITES when the content brings something it could not have generated alone from its training data
• it reformulates without citing when the content is already known to its model

That's it. That's the whole mechanism.

Application: if your content is commodity (7 tips for first-time homebuyers, recycled common knowledge), the AI can generate it itself from training data. It does not cite you. If your content is non-commodity (specific lived experience, unique observation, irreducible expertise), the AI cannot reformulate it. It MUST cite your page as source to transmit the information correctly.

Google's own example in the doc: commodity = 7 Tips for First-Time Homebuyers (anyone could write it). Non-commodity = Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line (a lived experience, irreproducible).

The only mechanism. All the rest (llms.txt, schema, chunking, mentions farming, hidden prompt injection in HTML) are optimizations on a content that must FIRST pass this mechanical test. If your content fails the test, no .txt file will save you. If your content passes the test, no .txt file is necessary.

Tactic 1: The File 100+ Sites Never Needed

On the 100+ sites I edit, zero llms.txt. Not laziness. Observation.

Every wave for the past 10 years has sold the same promise with a different file extension. AMP in 2016, hreflang explosion in 2018, aggressive schema in 2019, AI-friendly markup in 2024. The promise is always "implement this technical thing and AI/Google/whatever will reward you." The ROI is always the same: zero or marginal, on content that wasn't going to rank anyway. I stopped diving into every new wave by reflex because I've watched too many industries build themselves on sand. I made the same argument when MCP was the answer to everything, same playbook, different decade.

Google's position in the May 15 doc: you don't need to create new machine readable files, AI text files, markup, or Markdown to appear in generative AI search. The funniest part is that Google itself accidentally generated an llms.txt file on its own internal CMS for a while. The industry took it as a signal that "Google believes in it". John Mueller had to clarify publicly: the Search team does not use or endorse llms.txt. The file showed up because Google's internal CMS added support for it and some teams didn't bother removing it. That's all (reported by LBN Tech Solutions, February 2026). I think the snake oil sellers will keep selling it anyway. Maybe I'm wrong but I don't see what stops them.

One honest caveat. llms.txt has exactly one legitimate use case. Not SEO. Devtools. Cursor, Claude Code, GitHub Copilot fetch llms.txt at retrieval time to understand the structure of a tech doc (documented by Derivatex). If you maintain a SaaS documentation consumed by AI coding agents, llms.txt can serve. Otherwise, no. The line is clear: Google SEO no, AI coding agents consuming your docs yes. The GEO industry conflates the 2 on purpose because it sounds more lucrative.

Prompt 1: Honest llms.txt Audit

The Claude prompt below gives you a verdict, not a justification. I run it before any client asks me to add an llms.txt to their site. Paste it into Claude with your inputs:

<context>
  <site_url>{your_url}</site_url>
  <content_type>{blog | ecommerce | saas_docs | agency | media | other}</content_type>
  <audience>{humans_via_google | devs_via_claude_code_cursor | mix}</audience>
  <page_volume>{approximate_number}</page_volume>
  <update_frequency>{daily | weekly | monthly | rarely}</update_frequency>
</context>

<task>
Audit whether this site needs an llms.txt file. Do not be complaisant. Give the real verdict.
</task>

<constraints>
- Answer Yes or No clearly in the first line.
- If No: explain why in 2 sentences and propose one concrete alternative (e.g., "concentrate budget on 3 non-commodity articles").
- If Yes: confirm the use case is devtools (AI coding agents consuming docs), not Google SEO. Then give the recommended structure and warn about duplicate content risk if the user also plans to generate separate .md files.
- Do not invoke "preferential treatment by AI" as a reason. That argument is dead.
- Acknowledge that Google has publicly stated llms.txt is not used or endorsed by Search.
</constraints>

The constraint "do not be complaisant" is doing actual work here. Claude will validate everything by default if you don't push back. (Sonnet really struggles compared to Opus when you skip that line, I've tested both side by side on this exact prompt.)

Tactics 2 and 3: Chunking, Rewriting, Schema

3 cousins, same mistake. They all treat content as a resource to fragment or polish for AI, ignoring the extractive RAG mechanism we just covered.

Chunking. Google's direct position: there's no requirement to break your content into tiny pieces for AI to better understand it. Google systems are able to understand the nuance of multiple topics on a page. Observation across 100+ sites: dense well-written pages (1500-3000 words, clear structure, one strong point per section) systematically outperform pages chunked into FAQPage format. Why? A chunked page loses its narrative coherence, so it becomes interchangeable with any other page on the same topic. The AI doesn't cite it more. It cites it less.

AI rewriting. Google's direct position: you don't need to write in a specific way just for generative AI search, AI systems can understand synonyms and general meanings. Every GEO SaaS sells AI rewriting as a service. The irony: it produces exactly what Google sanctions in its scaled content abuse policy. The more you rewrite for the AI, the more you become commodity (because you remove what made you specific). Perfect self-defeating loop.

Schema overdose. Google's direct position: structured data isn't required for generative AI search. Caveat distributed: structured data keeps its utility for classic Google Search rich results (Recipe, Event, Product). But stacking FAQPage + HowTo + Article + Person on every page brings nothing in AI Search. (I ran into a client site last month where the FAQPage schema was applied to 4000 product pages by a plugin nobody had configured. The Search Console was screaming. The agency that installed the plugin had charged $1800 for the "AI optimization package".)

Prompt 2: Commodity vs Non-Commodity Check

The Claude prompt that takes your article draft and tells you if it's commodity or non-commodity, with a justified verdict. Severe by default, because Claude defaults to nice if you don't tell it not to.

<context>
  <draft>
    {paste your full article draft here}
  </draft>
  <topic>{topic / theme of the article}</topic>
  <audience>{target reader}</audience>
  <competing_articles_optional>
    {paste 2-3 competitor URLs or text if available}
  </competing_articles_optional>
</context>

<task>
Determine if this draft is commodity content (an AI could generate it from training data, no citation needed) or non-commodity content (irreducible to training data, AI must cite the source).
</task>

<output_format>
1. Verdict: commodity / partially non-commodity / non-commodity
2. Estimated percentage of passages that are commodity (recyclable from training data) vs irreducible
3. List 5 specific sentences from the draft that are strongly commodity and should be reworked
4. List 3 questions the author should answer to push the draft into non-commodity territory:
   - What can you alone say on this topic?
   - What number / incident / observation makes your article impossible to reformulate?
   - What sharp position are you willing to defend?
</output_format>

<constraints>
- Be severe, not encouraging. The default position is "this is commodity unless proven otherwise."
- Do not validate the draft by default. The author needs the truth, not a pat on the back.
- Cite specific sentences from the draft, not vague paragraphs.
</constraints>

I run this on every article before publishing. Last month it killed 2 drafts I was about to ship. Saved me 2 wasted posts. (The Convex dashboard still doesn't show this kind of usage analytics and I've been asking for 8 months.)

Tactic 4: Mentions Farming and the Black Hat Underground

Heads-up on framing: industry observation, not tutorial. I'm documenting what's being sold so you can recognize it if someone proposes it to you, and defend yourself if these techniques are used against your brand. I am not writing a how-to. None of these are implemented on my sites.

Level 1, grey hat actively commercialized. Inauthentic mentions farming. Coordinated Reddit posts (services selling "10 Reddit mentions in your niche for $200"), synchronized X threads with fake profiles, fake author profiles on niche sites, seeding positive reviews on G2, Trustpilot, Capterra to shift AI sentiment. Google's direct position: seeking inauthentic mentions across the web isn't as helpful as it might seem. The line is authenticity, not platform. Authentic mentions are still valid: Reddit organic is massively cited by LLMs in 2025-2026 (per a Visual Capitalist analysis cited by Brainz Digital, Reddit was the most-cited source across the LLM corpora they studied, though I'd treat the exact percentage as directional rather than definitive).

Level 2, borderline and sometimes illegal. Citation displacement (creating authoritative content on a competitor's topic so AI cites you instead), brand sentiment poisoning (synthetic content on forums and reviews to alter AI perception of a third-party brand). Documented by Alex Bobes (alexbobes.com, February 2026) as techniques observed in black hat AI SEO. The fact that these have a name now should tell you something about how organized the underground is getting.

I was on a dolphin trip with my kids near Koh Phangan last March, talking to a guy from Singapore who runs a SaaS in the legal compliance niche. He told me an agency had pitched him a "citation displacement" service to get cited instead of his main competitor in Perplexity. The price was $4000 for 90 days. He said no, partly because it was sketchy and partly because he had no idea if it worked. The agency's deck had 14 slides of charts, none of them with a source. The whole black hat AI SEO industry runs on charts without sources, and the buyers are usually founders who have no time to verify.

Level 3, emerging attack-rather-than-optimization. Hidden prompt injection in HTML. Invisible commands via display:none, visibility:hidden, white-on-white text, HTML comments, unicode invisible characters, zero-width spaces. Target: manipulating AIs that scrape the page. Microsoft explicitly blocked this tactic in its documentation (Search Engine Land, September 2025). Models also evolved. Quote from Security Innovation: models will process tokens even if they are invisible to humans, as long as they are present in the input. Quote from HiddenLayer: attacks against LLMs had humble beginnings, with phrases like "ignore all previous instructions" easily bypassing defensive logic. Status: still attempted by some, largely filtered now.

Level 4, academic research. Rewritten-query stuffing and segmented texts double the manipulation rate of LLM-enhanced search engines, per the paper Unveiling the Resilience of LLM-Enhanced Search Engines against Black-Hat SEO Manipulation (ACM Web Conference 2026). LLM training data poisoning: joint research from Anthropic and the UK AI Security Institute showed that 250 malicious documents are enough to poison an LLM, regardless of the total dataset size. This is security, not SEO. But it surfaces in obscure black hat AI SEO forums as a "future technique".

Honest read: techniques at level 1 and 2 are actively for sale. Levels 3 and 4 are largely filtered or stuck in research, but I've seen threads on obscure forums offering packaged services. The reader needs to know this exists to recognize it and reject it.

Prompt 3: Lived Experience Extraction (Interview, Not Rewrite)

The prompt that makes Claude do the opposite of what GEO tools sell. Instead of rewriting your article to make it AI-friendly, Claude interviews you to surface your real lived experience.

<context>
  <topic>{article topic}</topic>
  <angle>{your initial angle in 1-2 sentences}</angle>
  <raw_notes_optional>
    {paste raw notes, voice memo transcripts, or a draft if you have one}
  </raw_notes_optional>
</context>

<task>
Interview the user to surface lived experience, specific numbers, dated anecdotes, and counter-intuitive observations on this topic. DO NOT write the article. DO NOT rewrite the draft. Extract raw material the user can integrate themselves.
</task>

<workflow>
Ask 5 to 7 targeted questions. Examples of question patterns:
- What specific incident made you realize X?
- What number have you measured that nobody else can cite?
- What decision did you make that went against consensus?
- What failure taught you something the tutorials don't say?
- Who did you disagree with publicly on this topic, and what was the disagreement about?

After the user responds, compile a dossier of raw material:
- raw quotes from their lived experience
- specific numbers
- dated anecdotes
- contrarian observations
</workflow>

<constraints>
- Never write the article. Never rewrite the draft.
- The output is raw material the user integrates themselves.
- A machine cannot write non-commodity content by definition. It can only help the human surface it.
</constraints>

This is the most important prompt in the article. The other 3 are filters. This one is the engine. Also the only one where Claude is doing something AI tools were not built to do: stepping back from the keyboard and letting the human write.

The Bing Twist

On Google, GEO and AEO are declared dead (May 15, 2026). On Bing, the position is exactly opposite. Microsoft officialized the term GEO in its Webmaster Guidelines in March 2026, with a dedicated AI Performance Dashboard in Bing Webmaster Tools. Microsoft's framing: GEO doesn't guarantee citations, just as SEO doesn't guarantee rankings. Microsoft formally recognizes the practice as parallel to SEO.

Why the gap? My honest hypothesis (presented as hypothesis, not fact): Microsoft has a direct commercial interest. Bing Webmaster Tools, AI Performance Dashboard, IndexNow, Bing Places are monetizable products. Officializing GEO means selling GEO tools. Google has no such interest. Google AI Search uses the same index as classic Google Search, so validating a parallel term would create product confusion. I think that's the real driver. Could be I'm reading this wrong but the timing matches the product roadmap too neatly to ignore.

Practical consequence: ChatGPT Search is fed by Bing (the Microsoft-OpenAI deal). So optimizing for Bing equals optimizing for ChatGPT, in part. Overlap stat from Control Alt Digital (April 2026): only 13.7% of citations are shared between Google AI Overviews and Google AI Mode (10.7% URL overlap, 16% domain overlap). And between Google and ChatGPT/Bing, the overlap is even lower. Different engines cite different sites.

Nuanced conclusion. If your traffic targets Google AI Overviews, the 4 tactics discussed above are dead. If you target ChatGPT Search and Perplexity (which use Bing index), some Bing-friendly GEO techniques (llms.txt, FAQ schema, IndexNow) remain partially valid. But the extractive RAG mechanism stays the only universal lever. Even on Bing, non-commodity beats commodity. The Bing-specific techniques are polish on content that must already pass the fundamental test.

Don't read this section as a new thesis. Read it as a footnote to the main argument. The mechanism doesn't change with the engine.

The Test Google Didn't Write Down

Derivatex (February 2026) phrased it cleaner than Google: if an AI model read only this page and nothing else, would it come away with an accurate and useful understanding of our brand or product? Google didn't write down this test but it's the direct application of "non-commodity content first". If the answer is "no, the AI would walk away with a generic understanding available anywhere", your content is commodity. If the answer is "yes, the AI would walk away with something it could not have gotten anywhere else", you are citable.

Applied to the 100+ sites I edit: this test immediately separates pages that deserve to exist from pages that anyone could have generated. Stat to keep in mind: 48% of Google searches displayed AI Overviews in March 2026 (SEO.com data cited by Pasquale Pillitteri), up from 34.5% in December 2025. Direct implication: if your content fails the test, you disappear from half of Google traffic.

Prompt 4: The Isolated AI Test

The Claude prompt that simulates an AI reading your page as a single source and tells you what it would retain.

<context>
  <page_url_or_text>{paste URL or full text}</page_url_or_text>
  <target_query>{the query you want to be cited for}</target_query>
</context>

<task>
You are Google AI Overviews. A user asks the target_query. You have access ONLY to this page. What do you retain?
</task>

<output_format>
1. Three sentences you would extract for the AI answer
2. Citability verdict: high / medium / low
3. What would make the page more citable:
   - Specific numbers missing
   - Lived experience not made explicit
   - Sharp position not taken
</output_format>

<constraints>
- Treat this as a real simulation, not a content critique. You are the AI, not the editor.
- If the page is generic, say so directly.
- Do not suggest stylistic improvements. Only signals that change citability.
</constraints>

Short output, actionable in 2 minutes. I run it on every article before publishing. If you want a method that's non-commodity by construction, my book Vibe Coding, For Real is an 8-step Blueprint built on shipped projects, free on Kindle Unlimited.

The second-order insight from the Google doc, the one nobody is highlighting, is that the sites which survive 2026 are the ones where the author is non-transferable. You can't outsource non-commodity to a GEO agency or to a freelancer who doesn't know your topic. The SEO industry will have to relearn how to write instead of how to produce. The cost structure changes too. Non-commodity content costs more to produce than a generic FAQPage, but it earns on a completely different timeline. A non-commodity article keeps getting cited 2 years later. A commodity article disappears at the next AI index refresh. The same logic applies to building software, which is why I stopped vibe coding and started running prompt contracts, frameworks built on your own pattern beat generic best practices.

---

2 years watching an industry sell .txt files at 500 bucks while the real mechanism fits in 1 sentence: if an AI can write your article with its training data, it won't cite you. The sites that survive 2026 are the ones where the author is non-transferable.

Go write something only you can write. That's SEO in 2026 (unless you pick the Black Hat path, but that's another story ;-)

Sources

• Google Search Central, Optimizing your website for generative AI features on Google Search (May 15, 2026): https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
• Search Engine Journal, Google's New AI Search Guide Calls AEO And GEO 'Still SEO' (Matt G. Southern, May 16, 2026)
• Cyber Kendra, Google's AI Search Guide Is Out, Explained Without the Hype (May 16, 2026)
• Pasquale Pillitteri, Google AI Optimization Guide (May 16, 2026): https://pasqualepillitteri.it/en/news/2654/google-ai-optimization-guide-seo-ai-mode-overviews-en
• LBN Tech Solutions, What Is LLMs.txt? The Truth About Google Search Rankings in 2026 (February 2026)
• Derivatex, LLMs.txt: The Complete Guide for SEO and AI Search 2026 (April 2026)
• Somanath Balakrishnan, From SEO to GEO Part 2 (Medium, February 2026)
• Search Engine Journal, Bing Adds GEO To Official Guidelines (March 2, 2026)
• Control Alt Digital, AI Search in 2026: Your Complete Guide (April 2026)
• Alex Bobes, BlackHat SEO in 2026 (alexbobes.com, February 2026)
• Search Engine Land, Hidden prompt injection: The black hat trick AI outgrew (September 2025)
• ACM Web Conference 2026, Unveiling the Resilience of LLM-Enhanced Search Engines against Black-Hat SEO Manipulation

This post may contain affiliate links. If you click them, I might earn a small commission, costs you nothing, and helps me keep shipping quality articles every day for your reading pleasure.

---

Google just killed the playbook most agencies are still selling—but if you're shipping AI agents that need to cite sources reliably, the demo-vs-product checklist in the welcome kit shows you the production infrastructure that actually matters (hint: it's not markup).

→ Get the welcome kit