4 AI SEO Tactics Google Says Are Dead. Ahrefs and the Leaked System Prompts Disagree.
I published the official version 48 hours ago. Then I read the leaks and the Ahrefs data. I owe you a correction.
I'm not naive, I just adapt. I shipped an article 48 hours ago called "Google Just Killed 4 AI SEO Tactics. Here Are the Claude Prompts That Replace Them." That was the clean version. The one you show your mother, or your clients.
There's the clean version, sourced, technically right inside the frame it was written in. And there's the version that actually works but makes people uncomfortable. Problem with the clean one: that frame is the one Google drew. And Google telling you how to optimize for AI search is McDonald's telling you how to optimize your diet 😬.
You know it never really plays out that way. Here's the version you can't read out loud to your mother.
TLDR: Google's official guide tells you exactly what not to do. Like a politician announcing his program, you already know reality will be the opposite. No conspiracy theory here, just the concrete read: ranking in the AI era, the hidden playbook.
I spent the weekend reading 2 things in parallel. On one side, the GitHub repo asgeirtj/system_prompts_leaks, with the raw system instructions of Claude Opus 4.7, GPT-5.5 and Gemini 3.1 Pro, updated May 11 and covered by the Washington Post. On the other side, the reports from Ahrefs, Profound, BrightEdge and Semrush on what LLMs actually cite in 2026. The 2 say the same thing, and it isn't what Google says.
48 Hours Ago I Told You the Official Story. Then I Read the Leaks.
The article from May 17 is still up on my Medium. I'm not deleting it. It's a clean documentation of what Google published in its developer docs on May 15. Inside that frame, the article is correct.
What started bugging me on Friday night was the gap between the guide and what I was seeing on the 100+ sites I edit. Small ones, big ones, client ones. Behaviors didn't line up with the doctrine. So I read 2 things back to back.
First, the leaks repo I mentioned above. Raw system prompts in a public GitHub, in plain text, that anyone can git clone.
Second, Charles Floate. Veteran black hat SEO, The God of SEO blog, 10+ years in the trenches. Charles is a character in this story, not an authority. His thesis is more aggressive than mine, "Google lies by omission", I think he's overshooting but the data he points at is real.
Conclusion upfront, no suspense: Google is right about Google. The problem is that Google's product isn't all of AI search. It's 1 surface out of 5.
The Trick Google Plays in the Title of Its Own Guide
The guide is literally called "Optimizing for generative AI features on Google Search". Read the title again. On Google Search. The scope is right there, in the title, and everybody read past it.
Google's own line, lifted from the doc: "Optimizing for generative AI search is optimizing for the search experience, and thus still SEO."
What every SEO Twitter account amplified: "here's how AI SEO works in 2026." Framing error. Same word, 2 different scopes.
Numbers to kill the extrapolation. Ahrefs measured the source overlap between Google AI Overviews and Google AI Mode at 13.7% (December 2025). 2 features from the same company sharing 13.7% of their sources. ZipTie ran 100,000 prompts between ChatGPT and Perplexity. 11% overlap.
5 AI surfaces (Google AI Overviews, Google AI Mode, ChatGPT, Claude, Perplexity). 5 retrieval logics. 5 preferred source sets. Google gave rules for 1. The SEO industry read it as a universal manual.
A counter-voice worth taking seriously. Koray Tuğberk Gübür published a thread in May arguing AEO, GEO, SXO are technically the same discipline as SEO. "If your web document cannot rank, your passages cannot rank. If your passages cannot rank, generated AI answers are less likely to cite you." He's right at the foundation. Ranking is still the gate.
But there's a layer between "ranks" and "gets selected for an AI answer" that isn't pure traditional SEO. Entity clarity, direct answer formatting, passage-level prominence. Same discipline at the abstract level, yes. Same specific optimizations Monday morning, not quite. What you ship on Reddit doesn't reach Google AI Overviews the same way it reaches Perplexity. Different implementations, different decisions on a Tuesday.
The 4 tactics that follow only make sense once you've got that grid. Google's rules are Google's rules. Other surfaces have other rules, and they're written in plain text on GitHub.
Tactic #1: llms.txt Isn't Dead, It's Just Not for Google
Google's line in the guide, under 15 words: "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in generative AI search."
True. For Google.
OtterlyAI measured 62,100 AI bot requests across 90 days. 84 went to llms.txt. That's 0.1%. The only bot consistently crawling it is BuiltWith, which has nothing to do with AI retrieval. For Google AI Overviews exclusively, the file is dead weight. Google told you the truth.
Now flip the surface. Alimbekov ran an experiment, deployed llms.txt on a content site, measured the AI chat referrals before and after, normalized over 60 days. Result: +23% AI chat traffic. Perplexity nearly doubled. ChatGPT and DeepSeek showed measurable lifts. Same web, same content, 2 completely different verdicts depending on which AI surface you're trying to reach.
The detail that closes the argument. Anthropic maintains its own llms.txt on claude.com. Vercel, Cursor, Windsurf, Mintlify, all of them ship one. The makers of Claude are actively using what Google quietly buries in its guide. The CLI-first stack that consumes these signals doesn't read the Google guide before deciding what to fetch.
Here's the prompt I run on Monday morning when I'm setting up a new site. Drop it into Claude Code at the project root:
You are a technical SEO agent. Goal: generate the llms.txt for this site, formatted per llmstxt.org spec.
Input:
- Site URL: <URL>
- Sitemap or content map: <path or URL>
- Primary audience for AI surfaces: <Perplexity / ChatGPT / Claude / coding agents>
Steps:
1. Crawl the homepage and 3 highest-traffic pages. Identify primary entity, topic clusters, and unique value props.
2. Generate llms.txt with: H1 site name, blockquote summary under 30 words, sections for "Core Documentation", "Reference", "Tutorials". Each link entry: [Title](URL): one-line description.
3. Skip pages that are duplicates, low-value tag archives, or thin content. Be ruthless.
4. Output the raw llms.txt content only. No commentary, no markdown fences around it. I'll paste it at /llms.txt.
Constraint: no generic AI SEO advice. The file is for LLM retrieval, not for human readers.Action this week. Deploy llms.txt at root. 10 minutes of work. Skip if your audience is Google AI Overviews exclusively. Ship if you want presence on Perplexity, on coding agents, on the long tail of niche AI tools nobody benchmarks but everybody uses.
Tactic #2: Schema Isn't Dead, It's the Citation Extraction Layer
Google's line, under 15 words: "There's no special schema.org structured data that you need to add."
Third-party data says the opposite, loudly. Ahrefs analyzed 863,000 keyword SERPs and 4 million URLs cited in AI Overviews. Finding: 38% of cited pages don't rank in the top 10 of the traditional SERP. Down from 76% mid-2025. The gap between "ranks well" and "gets cited" is widening fast.
BrightEdge in parallel: sites with Article + FAQ schema deployed see +44% AI search citations versus control sites in the same niche. Not 4%. Forty-four percent.
Now the interesting part. The Claude Opus 4.7 system prompt, lifted from the asgeirtj repo, says textually: "Favor original sources: company blogs, peer-reviewed papers, gov sites, SEC. Skip low-quality sources like forums." That instruction is in the system prompt that ships in production right now. But Profound and Ahrefs both measure that Claude cites Reddit heavily, massively, all the time. The instruction tells the model to prefer institutional sources, and the model probably does prefer them at scoring time, the issue is that the retrieval base is dominated by Reddit because that's what got indexed in bulk over the years, so even with a perfect institutional bias at generation time the model can only pick from what was retrieved, and what was retrieved is Reddit. The system prompt is aspirational. The retrieval index is concrete.
Schema is the institutional anchor. It tells the retrieval pipeline your URL is a primary source, not forum content. Without it, you compete with Reddit on Reddit's terms.
Prompt I run on the top 10 pages of any money site. Copy-paste into Claude Code:
You are a schema audit agent. Goal: identify what JSON-LD this page needs to be cited as a primary source by ChatGPT, Claude, and Google AI Mode.
Input:
- Page URL: <URL>
- Page type: <article / how-to / product-review / location>
Steps:
1. Fetch the page. Extract existing JSON-LD if any.
2. Compare to the recommended schema stack: Article + HowTo (if procedural) + FAQ + Author + Organization. Skip Product on editorial pages.
3. List missing fields per schema type. Be specific: not "Author schema missing", but "Author.sameAs missing, no link to LinkedIn or X profile".
4. Output: a ready-to-paste <script type="application/ld+json"> block with the corrected schema. Include only fields you can populate from the page content. Mark placeholders explicitly with TODO.
Constraint: do not invent attributes. If author bio is absent on the page, flag it, don't hallucinate.Action this week. Article + HowTo + FAQ + Author + Organization on the 10 top pages. Validate via Schema.org. Then fetch with GPTBot user-agent and confirm the JSON-LD renders server-side. Half the SPA implementations don't.
Tactic #3: Brand Mentions, the 3x Signal Google Tells You to Ignore
Google's line: "Hunting for inauthentic 'mentions' on the web isn't as helpful as it sounds."
Concession upfront. True for inauthentic mentions. Buying mentions on PBN networks is a penalty risk, has always been. Google is right about that specific behavior.
The trick is in the slide. The sentence says "inauthentic mentions". The implication you walk away with is "mentions in general". Read it twice.
Ahrefs analyzed 75,000 brands. Brand web mentions correlate 0.664 with AI visibility. Backlinks correlate 0.218. 3x stronger. YouTube mentions: 0.737, the single highest factor in the dataset. The classic SEO playbook (links, anchor text, DA pushed up) was rewarded by Google. It's not rewarded by ChatGPT, Perplexity, or Claude at the same rate. The metric is mentions, not links.
Stacker, December 2025: distribute the same content via earned media (LinkedIn Pulse + press syndication + transcripts) and you see +325% AI citations versus publishing it only on your own site. Semrush analyzed 150,000 LLM citations. 89% of ChatGPT citations come from pages ranked 21+ on Google. Pages your dashboard treats as dead.
SEO and AI SEO are not the same job anymore. Same foundation (ranking still matters) but the optimization targets diverged. On the ground: a post about Anthropic killing a $200/month setup gets repackaged across LinkedIn, Reddit, and X with different angles. Same story, 3 surfaces, 3 different citation footprints.
Prompt for earned media briefs. I use this when I have a new piece of insight I want to seed across surfaces before publishing it on my own blog:
You are an earned media distribution agent. Goal: take one core insight and generate 3 distribution drafts for different surfaces.
Input:
- Core insight (2-3 sentences): <text>
- Niche: <niche>
- My handle on each surface: <LinkedIn / X / Reddit handle>
Output 3 drafts:
1. LinkedIn Pulse (700-900 words). Tone: analytical, professional, no slang. Hook with a data point, develop in 3 short sections, no CTA at the end.
2. Press release angle (200 words). Tone: factual, no marketing speak. One verifiable claim, one quote attributable to me, one supporting data point.
3. Reddit comment angle (under 200 words). Tone: peer to peer, my real voice, no link to my blog, no self-promotion. Identify 3 subreddits where this insight is on-topic and where I already participate. Suggest the exact thread type to look for.
Constraint: each draft must be substantively different, not the same text rewritten. Same insight, three voices.Action this week. Before touching your blog, publish your next insight as earned media first. One LinkedIn Pulse, one press release, one Reddit comment under your real handle. Distribution of presence. Not link building.
Tactic #4: Post Where AI Looks, Not Where You Live
Google's guide implies, throughout, that good content on your site is enough. The data destroys that.
Beacon4ai analyzed 23.6 million pages of AI search opportunities. Reddit appears in 92.8% of them. Not as a recommendation. As a cited source.
Distribution of citations by engine, all from Semrush's 150,000-citation study. ChatGPT cites Wikipedia 47.9%, Reddit 11.3%, Forbes 6.8%. Perplexity flips the order: Reddit 46.7%, YouTube 13.9%, Gartner 7.0%. Google AI Overviews sits in between with Reddit 21%, YouTube 18.8%, Quora 14.3%.
ZipTie: 90-95% of AI citations come from sources external to the brand's own site. Publish everything on your own domain and you're invisible to most of AI retrieval.
One data point from my own catalog. A small niche site I forget about between quarterly checkups. A handful of clicks per month from Google. The same month, 1,000+ visits and 500,000+ citations recorded from Bing. Same content, same domain, 2 different worlds. Nobody's SEO dashboard surfaces the Bing citation count, and yet that's where AI Overviews pull from for half the queries in my niche.
Growth Memo, April 2026: "AI systems are using content aggregators like Medium, Wikipedia, and Wired as sources, but almost never mention them by name." Posting on Medium equals anonymous raw material for ChatGPT. Posting on Reddit under your real handle equals a named cited source.
Prompt for Reddit/forum seeding. The one I use to identify where to post without falling into astroturfing (which is detected hard in 2026):
You are a community participation strategist. Goal: identify 3 subreddits where I should post one insight per week to build named AI citation footprint.
Input:
- My niche: <niche>
- My genuine expertise: <2-3 sentences of what I actually know>
- My Reddit handle and karma: <handle, karma>
Steps:
1. Identify 3 subreddits where my niche is on-topic AND where I have at least minimal participation history. No new accounts, no cold communities.
2. For each subreddit, output: typical thread types that get sustained engagement, posting frequency that doesn't trigger mod flags, 1 example of an insight from my niche reformulated as a genuine question or observation (not a self-promo).
3. Surface forbidden behaviors: linking my own site, dropping product names, copy-pasting blog content.
Constraint: I post once per week max per subreddit. Real voice, real expertise, no astroturfing. If I don't actually use a subreddit, do not suggest it.Action this week. Pick 3 subreddits where you already participate. Not 30, not 10. 3. Post 1 question or insight per week under your real handle. The compounding is slow but the citation footprint is durable, because it's named.
The 2026 Playbook (3 Moves, This Week)
No 12-step framework. 3 moves to start Monday.
Move 1. Deploy schema + llms.txt this week. Not a debate. Article + HowTo + FAQ + Author + Organization on your 10 top pages. llms.txt at root. 1 afternoon of code. Tools: json-ld.org, Schema.org validator, fetch with GPTBot user-agent to confirm server-side rendering. The prompts above do 80% of the work.
Move 2. Pick 1 earned media channel this week. Not 5. One. LinkedIn Pulse, vocal.media, syndicated press, YouTube transcripts. Pick where your readers actually search. Publish your next insight there before your own blog. Watch the citation footprint over 30 days.
Move 3. Find 3 subreddits, post 1 question this week. Not an annual program. One question, one subreddit, your real handle. The rest builds from there. No bots, no purchased accounts, no astroturfing. Astroturfing in 2026 is detected the way black hat link spam was in 2015. Don't waste your time.
I think one of these 3 moves will surprise you in the data. Could be I'm wrong on which one. Probably the earned media one, because it's the one most SEOs reflexively skip.
A guy I follow on X has been arguing for 2 years that printers are the most underrated piece of office hardware in the AI era because they leave no digital trail. He's not wrong, he's also not on topic, and somehow I think about that thread every time someone tells me their content strategy is "fully digital first".
What I'd Tell You If You Were Sitting Across From Me
Be smart. Separate the R&D from what makes you money. Your money sites, your client sites, you don't experiment on those. You ship what's proven. R&D goes on the secondary domains, the niche projects, the throwaway tests. That's the discipline.
What I wrote on May 17 is still true inside the frame Google drew. The frame just covers 1 AI surface out of 5. The other surfaces exist, they have rules, and their rules are written in plain text on GitHub. Read between the lines or be the lamb the wolves go for.
The official guides are maps of the territory Google and its peers control. The system prompts are the reality that doesn't appear on any map.
Be smart.
Sources
- Google Just Killed 4 AI SEO Tactics. Here Are the Claude Prompts That Replace Them. (May 17, 2026): https://medium.com/@rentierdigital/google-kills-ai-seo-tactics-claude-prompts
- Google Developers, Optimizing for generative AI features on Google Search (May 15, 2026): https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
asgeirtj/system_prompts_leakson GitHub (updated May 11, 2026): https://github.com/asgeirtj/system_prompts_leaks- Koray Tuğberk Gübür, post on X (May 2026): https://x.com/koraygubur/status/2055656290957816053
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's SEO guide covers one surface. The leaked system prompts and Ahrefs data reveal what actually ranks across ChatGPT, Claude, and Perplexity—and the demo-vs-product checklist shows you how to build AI content that works in production, not just in theory.