Jonathon's AI Wiki

These SEO Strategies Drive 90% of Your AI Visibility (Cyrus Shepard, Zyppy)

6 min read

Source: raw/newsletter-zyppy-signal-7fe1110a70.md — Cyrus Shepard, Zyppy Signal, signal.zyppy.com/p/seo-strategies-for-ai-search, published 2026-08-04.

Shepard’s deliberate devil’s-advocate piece against his own side of the argument. Google’s line is “good SEO is good GEO”; he agrees with it, notes the reverse claim (“GEO is just SEO”) is harder to defend since AI has opened genuinely new optimization surfaces — and then argues that for the vast majority of websites, a separate dedicated GEO strategy is almost certainly unnecessary. You can reach 90%+ of your AI-visibility goals by adapting modern SEO fundamentals and tweaking tactics toward specific AI-search mechanisms.

The practical value is the counter-positioning: this is the restraint article in a cluster that otherwise documents increasingly elaborate AI-search tactics.

Key Takeaways

  • You do not need an llms.txt file. “At least not yet.” The most quotable line, and consistent with the wiki’s existing skepticism.
  • Don’t chase unproven AI trends that are antithetical to traditional SEO — they can bite you later. The advanced tactics are real and some sharp operators are winning with them; they are just not where most sites’ first 90% lives.
  • Three goals, three lever sets: make content AI-eligible (crawlable and parseable), win citations, and get recommended. They are different problems with different levers, and conflating them is the common error.
  • Crawlability is the newly-unsolved problem. SEO treated indexing and crawl control as solved; AI reopened it. Robots.txt directives have grown complex with new user agents appearing every few months, and the rules are often set by technical teams with no communication to marketing.
  • Ranking for fan-out queries is the citation lever. AI systems issue background subqueries; via Reciprocal Rank Fusion (RRF), the more fan-outs you rank highly for organically, the greater your chance of being cited.
  • AI often isn’t looking for an answer — it’s looking for content that supports the answer it has already formed. Content matching what the model already “knows” is more likely to be cited. Increasingly, engines look to authoritative sites rather than merely top-ranking ones.
  • Recommendation runs on “distributed consensus” — multiple trusted sources supporting the same conclusion — which shifts digital PR from keyword-rich followed links toward authentic brand mentions.

1. Make your content AI-eligible

The complexity that traditional SEO checks don’t catch:

  • Robots.txt now has bot classes, not just bots. Know the difference between training bots (GPTBot, Google-Extended), search bots (OAI-SearchBot), and user-triggered bots (ChatGPT-User). Blocking the wrong class silently removes you from a surface you wanted.
  • AI crawlers don’t always respect robots.txt — so it is a policy statement, not an enforcement mechanism.
  • Your CDN must agree with your robots.txt. Cloudflare-style “block AI bots” rules check user agent and IP range and will happily override your intent. Audit both layers together.
  • Many AI crawlers don’t routinely render JavaScript, so JS-dependent content can be invisible. Shepard points to Gray Dot Co’s AI Difference Engine extension for checking what bots actually see.
  • Common Crawl is a seed source many sites inadvertently block — a second index you may not know you opted out of.
  • Snippet directives still bite: noindex, nosnippet, max-snippet, data-nosnippet were written for Google results and still affect AI visibility.

Counterpoint carried in the source: more access is not automatically the right answer — publishers pushing back on unlimited AI bot access have a real case.

2. Win AI citations

  • The primary strategy is to take a keyword you already rank for and work to rank for its query fan-outs, rather than creating a new page per question. Shepard reports it is typically both easier and more effective to earn fan-out rankings on existing pages.
  • Fan-outs are often closely related to People Also Ask questions, which is a cheap way to find them without specialized tooling.
  • AI systems are looking for evidence. Specific claims, data, and evidence-backed statements make attractive citation candidates — which is the mechanism behind much of the AI-citation research cluster.
  • Formatting matters: match headings and content to the shape of the question being asked.
  • Honest framing carried from the source: citation links themselves don’t get many clicks, but other research finds citations significantly boost traditional organic and paid clicks for brands that win them.

Ranking and citation do not guarantee recommendation. Engines assemble recommendations from existing model knowledge, top-ranking results, and fan-outs from trusted sources — “distributed consensus.”

  • Start internally with a brand-fact template: “[Brand] is a [product] designed for [audience], especially good for [use case], because of [differentiator].”
  • Make those facts consistent across every surface you control — homepage, about, product pages, partner sites, employee bios, social profiles, press releases, directories.
  • Build profile facts on third-party platforms — company subreddits, YouTube channels, industry directories, marketplaces, review sites, app stores.
  • Encourage detailed, authentic reviews. A customer describing the problem you solved potentially carries far more weight than a bare 5-star rating — because the prose is what a model can reason over.
  • Invest in digital PR for authentic mentions, not keyword-rich followed links.
  • UGC punches above its weight. Reddit and LinkedIn pages often rank highly for related queries, giving them outsized influence — the mechanism behind Reddit as a GEO surface.
  • Third-party reviews have moved from the decision layer to the recommendation layer, because the model reads them before your customer does.

Try It

  1. Audit robots.txt and your CDN bot rules together, classifying each AI user agent as training / search / user-triggered before allowing or blocking.
  2. Check a few key pages for JS-dependent content that a non-rendering crawler would miss.
  3. Pick one page already ranking well and build out its query fan-outs (start from People Also Ask) instead of writing new question pages.
  4. Write the brand-fact sentence once, then reconcile every surface you control against it.
  5. Ask your best customers for reviews that describe the problem solved, not just a rating.

Open Questions

  • The “90%” is rhetorical, not measured. No study backs the split between fundamentals and advanced tactics; it is an experienced practitioner’s estimate.
  • The RRF citation mechanism is cited to a third-party blog post (metehan.ai) rather than to engine documentation or a controlled experiment — it belongs in the correlational tier of the citation cluster, not the causal one.
  • llms.txt “not yet” — the piece links evidence that adoption hasn’t paid off, but the position is explicitly provisional.
  • No health/dental vertical data, the same gap flagged against the Similarweb 2026 landscape.

Graph View

Backlinks