Winning AI / LLM Search Visibility

Google ranks pages, AI ranks entities; earned signals teach the models; consensus beats raw authority.

AI and LLM search visibility was the defining anxiety and opportunity of SEO Spring Training 2026. Across all three days, speakers kept returning to one shift: the job is no longer (only) ranking a page in ten blue links, it is being named, cited, or recommended inside an AI answer from ChatGPT, Claude, Gemini, Grok, Perplexity, and Google's AI Overviews / AI Mode. The room agreed the old keyword-and-rankings playbook is stale, that LLMs reason over entities and consensus rather than pages, and that off-page authority (earned media, referring domains, reviews, social) now feeds the models as training and citation fuel. Where speakers diverged was on the mechanism (file engineering vs compute eligibility vs earned consensus vs referring-domain diversity) and on how clean or gray-hat the path to visibility should be. This brief synthesizes the nine AI-visibility sessions. For the off-page link and local mechanics underneath much of it, see the Spearleaf getting-cited-by-AI page and the off-page anchor system.

The through-line

Five ideas that multiple speakers, independently, converged on:

1. Google ranks pages; AI ranks entities (and you must define your entity consistently everywhere). Dawood Bukhari stated it most directly (Google ranks URLs, keywords, backlinks; AI understands brands, maps relationships, evaluates credibility, cites sources) and built his NARC method around answering "who you are, who you serve, what you solve" consistently on every surface. Daryl Osborne (Surge Protocol) made the same page-to-entity reframe ("for ten years we asked how do I rank; the question now is am I in the answer?") and built a measurement tool around AI citation. Brian Winum framed the whole thing as E-E-A-T plus "Brand. Authority. Data." mapping to entity authority. Kyle Roof reframed the same goal as "become your own source" inside ChatGPT.

2. Off-page earned signals, not on-page polish, are what teach the models. Dawood Bukhari cited that 85% of AI brand mentions come from earned media (not brand-owned content) and put digital PR at the literal center of his "AI consensus ecosystem." Brian Winum manufactured the same signal supply (listicles, roundups, advertorials, research reports, PR) as authority fuel. Ted Kubaitis reframed referring domains as the connective signal across channels that Google (and by extension AI) reads. Michael Merlino argued social activity is now how Google reads trust and prominence. Each is a different on-ramp to the same conclusion: presence across many third-party surfaces is the engine.

3. Consensus and repetition beat raw authority. Dawood Bukhari's NARC (Narrative consistency, Authority density, Repetition, Co-occurrence) argues a lower-DR brand doing consensus work out-ranks a higher-DR brand that is not, because "AI doesn't want to be wrong" and repeats what 500 sources agree on. Brian Winum's "authority stacking" (repeated appearances in roundups, listicles, PBNs) is the same repetition-as-consensus logic. Surge's "Collaborative" and "Amplifying" dataset roles describe the same phenomenon from the engine's side: multiple independent sources corroborating you is what earns the citation.

4. You can win AI visibility from a weak traditional footprint, which is the opening. Surge (Daryl Osborne) showed a free Weebly subdomain with no backlinks cited alongside premium grill brands, and a manufacturer surfacing for a local buying-intent query. Kyle Roof's term-selection method wins terms "Google already trusts you for" with minimal SEO and then satisfies LLM query fan-out. Dawood Bukhari's Cabo villa case (brand-new business, no site) and Eldar's whole pitch frame this as the reason to move now: brands dominant on Google but weak in LLMs can be beaten on the AI game.

5. Local angles and local entities are a disproportionate AI-visibility lever. Dawood Bukhari found local angles get picked up most (one national dataset spun into ten or fifty localized journalist stories, earning local TV mentions). Ted Kubaitis showed a referring-domain footprint betrays its city and can be pushed to the neighborhood level. Eleftherios Livadaras (Elias) built his entire local hierarchy (relevance, proximity, prominence) toward greening a geo grid. Michael Merlino's UGC video play is deliberately geo-targeted (PAA plus geo).

Tactics & playbook

Concrete, do-this items pulled from the talks:

• Engineer an llms.txt stack, not a single file. Brian Winum: add a custom header "AI elevator pitch" (who/what/where) and a custom footer of third-party validation; build separate FAQ (40 to 50 questions), glossary, and review files; add per-blog / per-service / per-location sub-directory llms.txt files; and a JSON llms-index with corrections, agent-guidance, actions/routing, and a SHA-256 manifest. Then engineer redundant crawl paths (LLMS XML sitemap, robots.txt references, meta link-rel injection, link prefetch, custom MIME types) so bots cannot miss them.
• Run digital PR off original data, scored for virality before you build it. Dawood Bukhari: every story must be Emotional, New, Provable, Easy, Timely, Visual; score it on a ten-point readiness scorecard (minimum 8 passes, no more than 1 fail) before research, design, or outreach. Make the journalist's job take under 15 minutes (original dataset, clear methodology, 3-plus quotable specific stats, downloadable visuals, two expert quotes, pre-written narrative). Spin one national dataset into local angles; target at least one video mention per ten links.
• Use HARO / expert commentary to inject narrative, with co-occurrence discipline. Dawood Bukhari: "you're not building links, you're injecting narrative." Answer only questions in the client's relevant niches (plus a few general categories), never scatter across 100 unrelated niches, because co-occurrence with the problem you solve is what drives AI visibility.
• Become your own LLM source via on-page term selection. Kyle Roof: stack Avalanche Theory (search volume), KGR, and a keyword-difficulty cap (medium or lower) as a filter to win 70 to 80% of selected terms with minimal SEO, then weight them so content satisfies LLM query fan-out and you get cited (his POP pages are the source in ChatGPT for "eChecker for SEO" and "best LLMs for SEO"). He also flags that low LSI coverage is likely how Google cheaply discounts mass AI content (a 10,000-page audit showed 99.4% missing LSI), so structure and contextual terms matter.
• Cluster your referring domains with AI and fill the gaps. Ted Kubaitis: export the Search Console top-linking-sites report and prompt AI to cluster it into channels (social, news, blogs, PR, merchants, reviews, Web 2.0, niche, etc.), then prompt for the clusters you are missing (government, video, academic, integrations, awards, podcasts, communities), then prompt for a table of example sites, contact author names, and a sample URL per opportunity. Aim for N-to-many link patterns (the web quoting many of your pages), not N-to-one. For e-commerce, his self-managing Merchant Center product-feed microsite script gives every SKU 25 referring domains automatically (his rough threshold to beat a backlink-free Amazon product page).
• Build short-form UGC video off PAA plus geo. Michael Merlino: take a People Also Ask question, add the geo, shoot a 15 to 30 second video with a real (or AI) face on camera and the full PAA as thumbnail text; embed it on the relevant page, then syndicate across social and a podcast. It ranks across image, video, and short-video tabs and, with enough engagement, into AI Overviews.
• Self-diagnose AI eligibility in 60 seconds. Surge (Daryl Osborne): check entity recognition in Google NLP (brand listed as ORGANIZATION with salience above 0.10), citation reality in Perplexity (named in the answer or source list for your top query), and corroboration count (five-plus independent third-party results for your brand name in the top ten). Failing any is a fixable visibility gap. The deeper diagnostic is the three-filter eligibility funnel: compute eligibility (is the site cheap enough to parse), citation eligibility (trustworthy enough to quote), answer construction (does it belong in this answer).
• Fix the entity-name match so AI local packs can connect. Joy Hawkins: AI local packs (about 13% of mostly-mobile US queries, showing two listings) frequently fail to connect to a GBP and leave a blank pin; the fix is making the business name on the website match the GBP name exactly, including capitalization. Watch Yelp as a rising cited source.
• Track AI consensus / visibility as a measured workflow, not a one-off score. Surge queries three citation engines directly (Anthropic, Perplexity, OpenAI) with API logs and runs day-0 / day-30 / day-90 to build a trend line. Eldar's Local Dominator AI Visibility Tracker shows where a client is visible across LLM engines and feeds that into content. Dawood Bukhari's three digital-PR levers (more mentions than competitors, spread across diverse domains, authority distribution) give the measurable targets.
• Stake authorship and check whether you are in the training data. Brian Winum: DMCA badge and Wayback API archiving into schema, blockchain timestamping (ScoreDetect), and C2PA extended to text; then check Common Crawl (its PageRank and harmonic-centrality metrics) to see whether AI training data has indexed you and how close you are to hub sites.

Tensions & disagreements

• Manufacture authority vs earn it. Brian Winum is openly gray-hat: he "spams the shit out of" listicles, generates fake AI interviews in Claude, marks himself #1 in schema while omitting the ranking publicly, and runs owned Reddit clones, wiki clones, and WP-multisite PBNs (one press release syndicated to 300 to 700 subdomains). Dawood Bukhari pursues the same consensus outcome but through genuine earned media and original data sent to real journalists, and explicitly warns that LLMs will soon weigh source quality, so you must vet the sites mentioning you. Same destination (AI consensus), opposite quality bar.
• What actually gates the AI citation. Surge (Daryl Osborne) argues the first gate is compute eligibility (DOM bloat, architecture) and that most "AI SEO" tools are unreliable prompt wrappers measuring noise (they showed one tool scoring the same site 87 then 42 ten minutes apart). Brian Winum's gate is file engineering and crawl-path redundancy. Dawood Bukhari's gate is earned-media consensus. Kyle Roof's gate is on-page LSI/term coverage. These are not contradictory in principle, but each speaker treats their own layer as the decisive one, so an operator must decide where the binding constraint actually is for a given site.
• AI SEO is a new discipline vs a repackaged old one. Kyle Roof explicitly deflates the hype: GEO is "basically just an SEO package that's slightly tweaked, but you can charge more for it" ("AI content blaster seven thousand"). Eldar and (implicitly) Surge frame AI visibility as a genuine paradigm shift that makes 2024/2025 tactics a losing strategy. The room wanted it to be new; Roof argued the fundamentals carry over.
• Duplicate vs unique content for AI-read location pages. Joy Hawkins's data shows exact-duplicate location pages (swap only the city name) outrank reworded "spun" content, and she makes the rest word-for-word identical to the most semantically relevant page. Brian Winum and Eleftherios Livadaras instead push genuinely unique, niche-specific content per location/sub-directory. For local AI-visibility pages, the "make them identical" and "make them unique" camps point in opposite directions (Hawkins notes her approach is untested at very large scale).
• Does any single tool's AI-engine count mean anything. Surge's whole pitch attacks rivals who claim to "measure 13 AI engines" but call ChatGPT once, insisting the real question is how many engines a tool actually measured. Eldar's tracker and various vendors at the event sell against that same backdrop, so buyers were left with competing, hard-to-verify measurement claims.
• Caveat on a shared but unverified term. Both Brian Winum (Reddit clones, "OpenClaw setup" per his deck) and Michael Merlino (agentic setup, "OpenClaw" phonetic) referenced an "OpenClaw"; in both sessions the exact platform name is a deck/phonetic transcription and is not confirmed, so treat it as unverified.

Sources (conference sessions)

These are conference session references, not pages on this site. Speakers and talk titles, as named in the source notes:

• Brian Winum, LLM Authority Hacking and Stacking with E-E-A-T (Day 1)
• Daryl Osborne, Surge Protocol V3 (on-stage name "Eric Lawrence" unverified) (Day 1)
• Andrew on harness engineering; Elias Livadaras on the SEO gorillas (Day 1, vibe coding)
• Kyle Roof, Chris Martinez, Chase Buckner (Day 1 part 2)
• Eleftherios Livadaras / Elias, the SEO gorillas and local hierarchy (Day 2)
• Eldar on AI visibility; Adam McChesney on client experience (Day 2)
• Bill Hartzer, Joy Hawkins, Michael Merlino (Day 2 part 2)
• Dawood Bukhari, Building Brand Visibility in the Age of AI / NARC and digital PR (Day 3)
• Ted Kubaitis, referring domains and multi-channel SEO (Day 3)

Related Spearleaf systems: getting cited by AI, off-page anchors, and the cross-theme briefs Off-Page links and Digital PR & conversion.