Why we ignore most of SEO Week 2026 advice (and what we kept)

TL;DR. SEO Week 2026 just wrapped. Seven speakers reframed search as a math problem — vector distances, entropy reduction, entity graphs, brand-as-centroid. A lot of it is consultant-grade marketing dressed in vector math. But three insights are real and worth implementing today. We did. Here's what we kept, what we cut, and what we added to our own site this morning.

The seven theses, ranked by what actually moves the needle

1. "Your brand is a mathematical object." — Scott Stouffer

Kept. This one is true and structural.

Stouffer's claim: when an LLM retrieves your content, your "brand" is computed as the centroid of all your text in embedding space. Pages live or die based on cosine distance to that centroid. Spread your content too wide — across positioning, agency work, founder essays, three different value props — and the centroid blurs, and retrieval misses you.

This isn't new framing for anyone who has done embedding retrieval (it's how we eval our own DoRA training), but it's the right framing for content strategy. Most companies don't have a centroid problem — they have a hundred orphan posts pulling in random directions.

What we did about it: we wrote down our centroid explicitly — Affordable. Quality. Fast. All three. Every article we publish gets audited against that thesis before it ships. If it doesn't pull toward the centroid, it doesn't go out.

We're going further: we're shipping a centroid_audit.py script — embeddings of every article checked weekly against the thesis vector, with a flag when a piece drifts > 0.3 cosine away. Drift is the silent killer of brand identity in LLM retrieval, and nobody's instrumenting for it.

2. "Linked data is the moat, not the model." — Andrea Volpini

Kept. This was the most actionable insight of the conference.

Volpini's data point: pages with proper RDF-linked entities (Organization ↔ founders ↔ products ↔ topics) lift answer accuracy in AI search by ~29% over identical pages without the linkage. The "AI visibility" battle is won not by writing more, but by structuring what you already have so LLMs can stitch entities together.

We patched our site today.

aiconic.company now ships a full JSON-LD @graph on every page — Organization with founder[] references to Person entities, each Person with sameAs to their X / LinkedIn / HuggingFace / GitHub / Substack profiles, each HuggingFace model as a SoftwareApplication with creator and author references back into the org graph. On our journal articles, Article schema includes author (→ Person ref), publisher (→ Org ref), and mentions (→ specific SoftwareApplication refs when the article discusses our models).

Verified with Google's Rich Results Test: six entity types valid (Article × 2 locales, Breadcrumbs, Organization × 2, Software Apps). All eligible for rich results.

The point isn't the schema markup itself — that's been around for fifteen years. The point is the linkage. Without @id cross-references, you have isolated schema blobs that LLMs can't fuse. With them, you have an entity graph. Google's reasoning layer treats those as fundamentally different inputs.

If you have schema on your site but every entity is a fresh blob with no @id referencing other blobs, you have ~30% less AI visibility than you could. That's the gap.

3. "Use open-source primitives, build your own tools." — Mike King

Kept (we already do this). This isn't an insight, it's permission — but the permission matters.

King's argument: commercial SEO tools are 13 years behind. They still index lexically. They still treat search as keyword density. Meanwhile Google moved to semantic search in 2013 and the LLM retrieval stack moved to embeddings in 2022. Buying these tools is buying a 2013 worldview.

We agree, and we've already lived it. Our entire ops layer is open-source primitives + custom Python: a daily Search Console pull, a HuggingFace API ingest, a centroid audit script, an AI visibility monitor that pings five LLMs daily with twenty real queries and counts mentions of our domain and our people. Nothing on the market does that combination at the price we'd pay for it.

If you're a small team in 2026 and you're spending more than $300/month on SEO software, you're paying for someone else's worldview. Open-source the primitives, vibe-code the glue. That's the actual stack now.

The four we ignored

4. "Relevance Engineering is the new SEO." — Mike King (iPullRank)

Cut. It's a re-naming exercise, not a methodology.

iPullRank's "Relevance Engineering" is the union of information retrieval, AI, UX, content, and digital PR. That's the discipline that's existed for fifteen years; we used to call it "SEO done well." The 300% AI-visibility growth and 21% referral revenue numbers cited are case-study-grade — interesting, not reproducible, and the methodology behind them isn't public.

The actual SEO Week takeaway from this talk: search is changing fast, integrate AI into your retrieval pipeline. We knew that. The re-branding doesn't help us ship anything.

5. "Grounding and search diverge after retrieval." — Krishna Madhavan

Cut for our scale, kept as a watchlist item.

Madhavan's pipeline diagram (query understanding → transformation → multivector retrieval → candidate processing → ranking → for grounding only: evidence selection, answer construction, constrained generation, cross-check) is technically correct and well-presented. It's also five layers below where a 3-person agency can usefully optimize.

The takeaway is real: visibility is determined at the query understanding stage, not at the citation stage. But the action item — optimize for query understanding — is more useful when restated as our centroid point (#1) than as a new dimension to track. Same problem, different vocabulary.

6. "Hybrid Engine Optimization scorecard." — Jori Ford

Half-kept.

Ford's five signals — Presence, Visibility, Citation Quality, Authority Confirmation, Business Impact — are the right axes for a small team. The recommendation to layer first-party data (GSC, GA, internal) + one AI monitoring tool is also right.

We're keeping the framework but not buying the tool. Profound / Otterly are $69–99/month for what is fundamentally a daily-cron script that hits five LLM APIs and parses output for brand mentions. That's exactly the ai_visibility_monitor.py on our list. ~50 lines of Python, $0 SaaS budget, and we own the data.

This is the King point applied to Ford's framework: agree on the scorecard, refuse the SaaS that sells you a worldview wrapping a script.

7. "AI citations driven by entropy reduction." — Metehan Yeşilyurt

Cut. Technically interesting, practically a distraction.

Yeşilyurt is right that token economics shape what LLMs cite (English-optimized tokenizers, embedding-pass filters, reranker gates). For us, the implications are: hoist the essence to the top of the page (we do — every article starts with a one-paragraph TL;DR), and stay in English (we already moved off RU as the primary content language).

After those two surface tactics, the rest of his framework — embedding similarity as a content creation tool, semantic-density optimization, token-budget aware structuring — is research-grade work that returns sub-1% wins for the time investment. Senior content marketers were already doing the surface part as common sense.

What you should do tomorrow if you ship content

In the order that matters most for return per hour invested:

1. Add @id references between your schema entities. Most sites have schema. Almost no site has properly linked schema. The 29% accuracy lift is real and you can earn it in an afternoon. Start with Organization ↔ founder[] ↔ Person entities, then connect your products / models / case studies in.
2. Write your centroid down on one page. Not "our values" or "our positioning" — the literal thesis sentence that every article you ship is allowed to drift from by no more than one cosine standard deviation. Audit existing content against it. Cut what doesn't pull toward it.
3. Replace SEO SaaS subscriptions with five scripts. GSC daily pull, GA daily pull, HF/GitHub/social mentions ingest, embeddings centroid audit, AI visibility monitor. All five are under 100 lines of Python each. None of them care about your subscription tier.

The rest — query understanding theory, entropy reduction tactics, brand-as-centroid math — is either restating the basics in fancier language or doing 95th-percentile optimization at the 5th-percentile maturity stage. Skip until you've done the three above.

What we'd add to the SEO Week 2026 corpus

One missing thesis: most "AI visibility" work is wrong layer. It treats discoverability as a marketing problem when it's increasingly a product problem. If your software, your models, your tools live in plain text on accessible URLs — with linked schema, with versioned changelogs, with proper creator and author metadata — LLMs find you. If they live behind login walls, in PDFs, in closed Notion pages, no amount of citation-tracking will save you.

The conference treated content production as the lever. We're going to argue, in a follow-up piece, that product surface is the lever. The most visible brand to LLMs is the one whose products have the best canonical home on the open web. That's a totally different operating system than "publish more content" — and it's where our small team is betting.