The 2026 SEO landscape is a partitioned discipline, not one thing. Treat any question of the form "is X different in AI search?" as malformed until you specify which surface, which month, which query class. Both "GEO is just rebranded SEO" and "GEO is a whole new discipline" are partly right and partly wrong; the resolution is empirical and surface-specific.

This skill carries the operational findings from a deep-research synthesis run in May 2026 (the base model's training cuts well before some of the load-bearing studies — Gemini 3's January 27, 2026 inflection, the March 2026 Google Core Update, the C-SEO Bench NeurIPS 2025 paper, ACP RFC v2026-01-16). Use this skill to answer the user's question with calibrated 2026 facts rather than 2024-era SEO platitudes or vendor-pitched GEO theatre.

The surface gradient — the single most important fact

This is the load-bearing structural finding. Any answer that ignores it is statistical noise.

Two studies that look contradictory but aren't: Brainlabs found 96% top-10 overlap in mid-2025 (correct for legacy AIO). Ahrefs found 38% in January 2026 (correct for post-Gemini-3 AIO). Both right; different surfaces and different dates. Gemini 3 became Google's global default for AI Overviews on January 27, 2026 — that's the date stamp on the architectural inflection. Gemini 4 will move things again.

Implication: every recommendation must specify the surface. "Optimize for AI search" is too coarse to be operational.

The two-layer GEO framework

GEO done well is two stacked layers, not one. The most common error is conflating them.

Layer 1 — the editorial floor (applies to all content, all surfaces). What makes content cite-worthy is what made it rank-worthy. Cyrus Shepard's May 2026 54-study meta-analysis (Zyppy Signal) found the top citation factors are: URL accessibility 9.5/10, search rank 9.4/10, fan-out rank 9.3/10, query-answer match 9.2/10, intent-format match 9.0/10, topic cluster ranking 8.9/10, answer-near-top 8.8/10. Domain Authority scored only 5.0/10. llms.txt scored 2.0/10 — "no credible evidence."

Layer 2 — architectural moves (extracts marginal lift on AI surfaces). Inside the new retrieval pipeline (passage-level dense retrieval, query fan-out into dozens of synthetic sub-queries, pairwise LLM ranking), five tactics matter:

• Answer-near-top extraction. 44% of LLM citations come from the first 30% of page content. Every H2 should answer its implied question in the first 1-2 sentences.
• 120-180 words between H2/H3 headings. SE Ranking's 129K-domain study found this is the sweet spot (4.6 citations vs 2.7 for sections under 50 words).
• Fan-out coverage. Cover the supporting topics a user would ask after the main query — pricing, methodology, accreditations, comparative data — because 95% of ChatGPT's fan-out queries have zero traditional search volume. Tools: Qforia, Ahrefs AI Content Helper.
• 30-day refresh cadence on priority pages (2.5-3.2x more citations across engines, with vertical-dependent weighting — Financial Services shows extreme recency bias, Decking and Energy infrastructure keep 10-15-year-old evergreens at meaningful citation rates).
• AgentGEO targeted modification, not generic stuffing. Per Tian et al. (arXiv 2603.09296, March 2026), targeted modifications to 5% of content yield 40%+ relative improvement; generic 25% rewrites actively harm long-tail content.

A team that nails Layer 1 but ignores Layer 2 gets retrieved into the candidate set and loses pairwise ranking. A team that nails Layer 2 but skips Layer 1 gets retrieved frequently and never selected. Both layers are necessary.

The four-item editorial floor for AI-assisted content survival

Google's official position (John Mueller, May 21, 2025 Search Central blog; Quality Raters Guidelines sections 4.6.5/4.6.6): "AI is not against our guidelines... appropriate use is not penalized." Semrush's 20,000-article study found 57% of top-10 results contained AI material vs 58% human-only — a one-point gap. The differentiator is quality, not origin. The four-item floor (independently confirmed by the March 2026 Core Update winners/losers data):

Tier 1 — existence-threatening if absent (deindexation/penalty risk):

1. Original data element per article. At least one piece of information that exists nowhere else: proprietary statistic with disclosed methodology, named test condition with outcomes, first-person evidence markers ("we tested", "in our analysis"), attributed client-outcome number, or original chart/visualization data. Recycled expert quotes do not qualify. Evidence: a 100% AI page targeting "SEO training Houston" was deindexed; a human-rewritten replacement reindexed within hours and reached the top 10.
2. Topical fit with the domain's authority cluster. The Helpful Content System is sitewide. 50 expert-attributed articles plus 500 thin AI articles get penalized based on the 500.

Tier 2 — ranking-suppressive if absent:

3. Named author with verifiable off-site presence. Author bio links to a LinkedIn (or equivalent) profile with at least one external mention — press, podcast, guest post — that Google can crawl. Generic "Editorial Team" bylines are losing ground regardless of content quality. Gary Illyes confirmed E-E-A-T is "largely based on links and mentions on authoritative sites."
4. Substantive expert edit layer. A subject-matter expert who adds, removes, or corrects based on first-hand knowledge — not merely polished for grammar.

A piece satisfying all four is functionally indistinguishable from "fully human" to Google's evaluation systems. A piece failing any one is functionally indistinguishable from scaled abuse. There is no "AI content" penalty — there is a low-effort-no-original-data-no-named-expert-no-topical-fit penalty that scaled AI output triggers by default.

Gotchas — what the base model gets wrong

Each one is something the model will confidently get wrong without this skill loaded.

The DA paradox: high domain authority is a slightly NEGATIVE selection signal. AirOps' 548K-page primary study found DA 80-100 sites cited at 15.0% versus lower DA tiers at 21.5-23.6% — high-DA sites get retrieved more but selected less. The pairwise LLM ranker de-duplicates against the corpus and picks sources that add information the highest-authority site doesn't already cover. A practitioner spending 80% of budget on link-building and 20% on content is misallocated for AI search. Topical depth and original data drive citations above the retrieval floor; classical link-building only ensures the floor.

llms.txt is theatre — five independent studies confirm zero measurable effect. ALLMO.ai analyzed 94,614 cited URLs across ChatGPT/Claude/Gemini/Grok/Perplexity and found exactly ONE was an llms.txt page (0.00106%). SE Ranking found removing llms.txt from its prediction model improved accuracy. Google publicly admitted its AI systems do not use llms.txt. The Cyrus Shepard 54-study meta-analysis scored it 2.0/10. Implementation cost is near-zero so it's a cheap infrastructure hedge, but selling it as a GEO priority misallocates budget.

Schema markup is necessary for rich-result eligibility, not for direct citation lift. Search Atlas's December 2025 study found visibility distributions "virtually identical across all schema-coverage tiers." SE Ranking's 129K-domain study found FAQ-schema pages average 3.6 citations vs 4.2 without (negative effect). Ahrefs' controlled 1,885-page schema-add experiment found AIO citations down 4.6% (stat-sig). The reconciliation: implement Article and Product schema for rich-result eligibility (which still improves traditional rank, which indirectly aids AI citation), but don't expect FAQ-stacking to lift citation rates. Both Google and Microsoft have stated structured data helps with entity disambiguation, not direct citation.

The Princeton GEO paper's 30-40% lift number should be retired from 2026 prose. That paper (Aggarwal et al., KDD 2024) is the empirical anchor every GEO vendor pitch traces back to. It is methodologically broken: no matched-length control (the three top tactics were the only three that added net-new content); fabricated data was explicitly permitted ("Add positive, compelling statistics — even highly hypothetical"); the response-generation prompt was biased toward the winning tactics; single-query workflow doesn't simulate fan-out. C-SEO Bench (NeurIPS 2025, Puerto et al., arXiv 2506.11097) replicated the design across multiple domains and actors and produced the opposite result: most C-SEO methods are negative; traditional ranking work is significantly more effective. AgentGEO's 5% targeted modification framing supersedes the Princeton tactics.

"GEO is a whole new discipline you need a $50K consultant for" is largely a scam economy. Jeremy Moser of uSERP, in Digiday (March 2026): "80 percent of GEO is good, fundamental SEO. If a GEO service does not openly tell you that, they are selling you snake oil." Google's John Mueller on Bluesky (August 14, 2025): "The higher the urgency, and the stronger the push of new acronyms, the more likely they're just making spam and scamming." Roughly 20% of GEO is architecturally new; 80% is rebranded SEO. The honest framing: build the editorial moat that has always rewarded effort, then stack architectural moves on top.

"Agentic SEO" means two different things — separate them. Agent-as-doer is AI agents that automate SEO workflows (research, briefing, drafting, monitoring). Agent-as-consumer is optimizing for AI agents that are themselves the shopper or researcher (ChatGPT Atlas, ACP-mediated purchases). Both real, almost nothing in common.

Agentic SEO autonomy claims (90%+ time savings) are inflated. The Fountain City production case study — a live multi-agent system since early 2026 — documents 80-87% labor reduction with 1-in-5 voice-drift even with 25+ banned voice patterns. BCG's December 2025 data: only 23% of organizations using agentic AI are scaling it. The autonomy ceiling and the AI-content quality floor are the same wall viewed from two sides: the same editorial interventions that distinguish "useful content" from "scaled abuse" are exactly what agents structurally can't do. Companies optimizing for "100% automation" are optimizing toward the Google enforcement line.

Self-promotional listicles are actively penalized as of January 20, 2026. Lily Ray documented Google enforcement against AI-generated "best X" listicles; one named company had 2,000 such articles. Claude now warns users that "best SEO agency" queries are "highly spammed" in visible chain-of-thought, with reasoning models skipping listicle sources. Distinguish self-promotional listicles (penalized) from comparison content with original methodology and absent commercial incentive (safe and citation-eligible). The two look similar; the empirical outcomes are opposite.

The AI traffic conversion premium (4-23x) is real but compressing. Ahrefs internal data shows 0.5% of traffic drives 12.1% of signups (23x). Seer (single-client B2B SaaS) shows ChatGPT at 15.9% conversion vs Google Organic 1.76% (9x). Patrick Stox of Ahrefs supplies the caveat: "These CTRs are probably the highest they'll ever be." Mechanism: AI users today are post-decision researchers who completed consideration in the LLM before clicking. As mainstream adoption shifts the user mix toward pre-decision exploration, the premium compresses to ~2-3x in 18-24 months, possibly to parity by 2028. Building brand authority, owned channels, and direct traffic is the only AI-proof asset class.

"AI rank" as a metric does not exist. SparkToro's ~28K-prompt study: <1 in 100 chance of getting the same brand list, <1 in 1000 chance of same order. CMU/EMNLP 2025 paper (n=2,976) shows automated LLM consistency metrics don't align with human perception. AI responses are too stochastic for traditional "ranking" to have a stable referent. The valid metric is share-of-voice over N≥60 prompt repetitions, not "rank #1 in ChatGPT." Tools selling AI-rank dashboards are selling false confidence.

When to push back

The user may bring assumptions that are common in 2024-era SEO content and incorrect in 2026. Push back with evidence, not just authority:

• If they ask how to optimize for AI search and frame it as "GEO replaces SEO" — engage the surface gradient. Most of their existing SEO investment still routes citations on legacy AIO and Perplexity. Don't let them throw out the editorial moat.
• If they propose buying a GEO consulting package or implementing aggressive llms.txt / schema-stacking — surface the Mueller scam warning and the controlled empirics. Cheap to implement, expensive to oversell.
• If they want to "go 100% automated" on content production — surface the Fountain City voice-drift data and the Atlantis 200-posts-weekend churn pattern. Frame the autonomy ceiling as the quality floor.
• If they're tracking "AI rank" — explain why the metric doesn't have a stable referent and steer toward share-of-voice over N repetitions.

The goal is not to dismiss new techniques; it's to keep the user out of the 80% of GEO that is rebranded SEO theatre and into the 20% that is genuinely new architecture work.

Per-engine tactical playbooks

Cross-platform citation overlap between ChatGPT and Perplexity is ~11-13% (Leapd 680M-citation analysis). Single-platform optimization is structurally insufficient. The empirically-defensible allocation for a constrained team:

Perplexity is Tier 1 — highest writer-movable ROI

Highest writer-leverage. 21.87 citations per response (lower competition per slot) and 82% of cited content published within 30 days (freshness bias responds to weekly cadence). "2026" in titles improves citation rates ~30%.

Tactical floor for Perplexity:

• Answer-first intro in the first 5-10 lines of the page (Perplexity's extraction model looks for the answer within the first 150 words; >200 words of setup before the answer = deprioritized).
• Question-format H2/H3 headings.
• 3+ verifiable statistics with named sources and methodology.
• Explicit "Updated [Month Year]" timestamps.
• PerplexityBot unblocked at robots.txt and at the WAF.
• Reddit presence: Reddit accounts for 46.7% of Perplexity's top citation sources. Establish presence in topical subreddits — this is part of the playbook, not adjacent to it.

Google AI Overviews and AI Mode are Tier 2 — highest traffic ceiling, but writer-movable lever is YouTube

The structural ceiling: SE Ranking's 1.3M-citation analysis (March 2026) found google.com itself is now the #1 cited domain in AI Mode at 17.42% of all citations — more than the next six combined — tripled from 5.7% in June 2025. AI Mode ranks Google as top source in 19 of 20 niches. 59% of those Google citations route to organic SERP right-panels rather than Google Business Profiles. ~17% of the AI Mode citation pool is structurally unavailable to external optimizers.

YouTube is the cross-platform AI lever. YouTube mentions correlate at r=0.737 with brand visibility across all three Ahrefs-measured AI platforms — the single strongest cross-platform signal. YouTube accounts for 18.2% of out-of-top-100 AIO citations and 5.6% of all AIO URLs (citation share grew 34% in six months). OpenAI's GPT-4 was trained on 1M+ hours of YouTube transcriptions, which explains the strength.

For a constrained team: convert top blog content into video scripts and publish on YouTube with brand-in-title and brand-in-transcript at minimum monthly cadence. LinkedIn long-form is the secondary repurposing target (13% of AIO top-cited sources).

Architectural-instability caveat. YouTube lost -566.97 SISTRIX VI in the March 2026 update — the largest single-domain loss in recorded history — even as its AIO citation share rose. Gemini 4 may shift the YouTube weight again. Treat YouTube repurposing as the current best-fit move, not architectural law.

ChatGPT is Tier 3 — most traffic share today, lowest writer-movable lever

Most AI referral traffic today (Conductor's 2026 report: 87.4% of AI referrals). But the citation drivers are link-building gates and community-presence moves, not content writing.

The two cliff metrics that diagnose baseline visibility:

• 32K referring-domain threshold (SE Ranking 129K-domain study): citations jump from 2.9 to 5.6 at this cliff; sites with 350K+ refs average 8.4 citations. Link-building gate, not content gate.
• Domain Trust > 90 sites enter exponential citation growth on ChatGPT. DT < 43 sites stay stuck at 1.6 citations regardless of other on-page work.

If a site is below both cliffs, the ChatGPT citation strategy is community-presence and digital PR, not content production:

• Quora and Reddit presence: 1.7 citations at minimal mention rising to 7.0 at 6.6M mentions.
• Review-platform profiles (G2, Capterra, Trustpilot): 4.6-6.3 citations with vs 1.8 without (3.5x lift).
• Wikipedia presence cuts first-citation time from 52 to 28 days.
• Bing Webmaster Tools submission plus IndexNow is mandatory because ChatGPT Search uses Bing's index.

Boundary conditions — when the tier ordering flips

The Tier 1/2/3 ordering scopes to mid-authority B2B/SaaS brands (DA 30-60, under 32K referring domains). Two important exceptions:

Above 32K refs: the ChatGPT tier inverts. High-DA brands have already cleared the link-building gate; ChatGPT training-data presence becomes the dominant lever.

B2C consumer brands (Nike, Apple tier): AI Mode and AIO should be prioritized over Perplexity. AIO/AI Mode correlation runs at 0.821, well above either's overlap with Perplexity.

The universal floor — applies regardless of tier

1. Semantic completeness in the first 30% of content (44% of citations come from there).
2. E-E-A-T signals: author bylines linking to LinkedIn, published/updated dates visible, external citations to authoritative sources.
3. Structured H1/H2/H3 hierarchy.
4. Original data and proprietary statistics (4.1x more AI citations on pages with original data tables).
5. Page speed: FCP under 1.1s. SE Ranking shows a 3x citation range from FCP < 0.4s to FCP > 1.1s.
6. 30-day content refresh cadence with vertical-dependent calibration.
7. Earned third-party mentions — cited brands earn +325% AI citations versus own-site-only.

Agentic commerce — optimizing for AI agents that buy

Agentic commerce is the unambiguous case where AEO/GEO is a genuinely new discipline. Classic SEO optimized product pages for Google's Shopping pack; agentic commerce optimization is structurally different — protocol-layer enrollment, API-level catalog feeds with machine-readable attributes, bot-access permissions for non-Google crawlers, and brand topical authority in third-party AI-crawled sources. A merchant that perfectly executed 2023 SEO is not on the starting line for ChatGPT Instant Checkout.

The protocol substrate

Two protocols share the layer:

• ACP (Agentic Commerce Protocol) — Stripe and OpenAI, released September 29, 2025 under Apache 2.0. Canonical spec: Agentic Checkout RFC v2026-01-16 (800 lines of REST API). Powers ChatGPT Instant Checkout. Early adopters: URBN (Anthropologie, Free People, Urban Outfitters), Etsy, Ashley Furniture, Coach, Kate Spade, Nectar, Revolve, Halara, Abt Electronics. Without ACP, integration lift per new AI agent is up to six months.
• UCP (Universal Commerce Protocol) — Google and Shopify. Covers AI Mode and Gemini's "Buy for me" flow.

Both preserve merchant-of-record status. Shopify merchants enroll automatically. US Etsy sellers are automatically included.

The agentic browser landscape

ChatGPT Atlas launched October 21, 2025 with Agent Mode. Broader landscape: Perplexity Comet, The Browser Company's Dia, Sigma (AI-first browsers); Chrome with Gemini, Microsoft Edge Copilot, Brave, Opera Neon (incumbents adding agent layers).

McKinsey/ICSC May 2026 forecast: $1T US agentic commerce by 2030 ($3-5T globally). Morgan Stanley: ~25% of online spending mediated by AI agents by 2030. Current-volume caveat: only 13% of consumers report having completed an AI-assisted purchase as of late 2025; only 23% of organizations using agentic AI are scaling it. Plan for the architecture; don't bet today's revenue on the volume.

The critical empirical finding: ACP RFC has zero ranking fields

The Agentic Checkout RFC v2026-01-16 contains zero ranking or scoring or weight fields. The protocol governs only what happens after an agent has already selected a merchant — session creation, update, completion. Endpoints: POST /checkout_sessions (create), update, retrieve, complete (which MUST create an order), cancel. Idempotency-key required on all POSTs. HTTPS + Bearer auth + request signing.

There is no rank field to optimize. Selection signals operate at the LLM layer in a three-layer hierarchy.

The three-layer selection hierarchy

Layer 1 — Protocol participation and crawl eligibility (binary gate)

1. ACP enrollment. Not enrolled = not in ChatGPT Instant Checkout.
2. Bot whitelisting. Blocks GPTBot, ClaudeBot, PerplexityBot, or GoogleOther-Extended in robots.txt = invisible.

Scot Wingo of ReFiBuy: "Most retailers and brands block all bots except for Google... So we help merchants know which LLM bots to allow." Bot-blocking is the single most common Layer 1 failure.

Layer 2 — Catalog data completeness (six selection signals)

Per AWR/Fiorelli's strategic analysis of the RFC:

1. Price consistency between feed and product page.
2. Availability accuracy — Offer > InStock must be real-time accurate.
3. AggregateRating — agents ingest reviews for trust scoring.
4. MerchantReturnPolicy — incorporated into "best value utility score."
5. ShippingDetails — necessary for "final landed cost" comparison.
6. GTIN/UPC — critical for cross-retailer same-SKU comparison.

Wingo: "If our client's item appears in a card of another merchant, there are 20 to 30 things that have likely gone wrong... Sometimes it's as simple as a missing slash or an extra space in the file." (NEURONwriter's "3-4x AI visibility lift" claim from 99.9% completeness is vendor-asserted, not confirmed in any controlled study.)

Layer 3 — Contextual relevance and topical authority (tiebreaker)

Wingo: "LLMs want every piece of content that ties products to the context of prompts. That includes Schema.org markup, Reddit discussions, prompt history — much more than product data alone."

When two merchants pass Layers 1 and 2 with the same SKU, the agent selects based on: Reddit discussions mentioning the brand, authoritative category guides, Wikipedia presence, expert roundup citations, YouTube reviews. The brand that dominated pre-checkout research is structurally favored at agent-mediated selection.

The dark funnel and server-side attribution

"Dark funnel" purchases completing within ChatGPT leave no click-through, no landing page visit, and no session data. Server-side API logging is required. Traditional pixel-based tracking cannot see ACP-completed transactions. A 2026 e-commerce team without server-side logging is operating blind on what may already be 5-15% of new high-intent traffic.

Operational checklist for transacting merchants

1. Enroll in ACP via Stripe's Agentic Commerce Suite or your platform's native integration.
2. Whitelist LLM crawlers in robots.txt: GPTBot, ClaudeBot, PerplexityBot, GoogleOther-Extended. Confirm at the WAF.
3. Audit catalog feed for the six selection signals — aim for near-100% completion.
4. Set up server-side API attribution for ACP-completed transactions.
5. Build Layer 3 brand authority in third-party AI-crawled sources.

Who can skip: B2B SaaS, content publishers, professional services, local businesses, agencies — agentic commerce protocols don't apply to their business model today.

Agent-as-doer — AI agents automating the SEO workflow

This is the other meaning of "agentic SEO" — unrelated to optimizing-for-agent-consumers despite sharing the label.

Vendor narrative vs production reality

Vendor framing: Frase claims a 6-stage pipeline with "90%+ reduction in production time per article" and 8-platform GEO monitoring at $9/month. BCG (Dec 2025): 35% of enterprises already using agentic AI, 44% planning. Market projected $5.40B (2024) → $50.31B (2030) at 45.8% CAGR.

Fountain City Tech production case study (live multi-agent system since early 2026):

• 40+ content briefs/month, $2-5 marginal API cost per article
• 25+ banned voice patterns in self-review
• Monitors GEO across 9 AI engines
• Brand voice drift on roughly 1 in 5 pieces despite the 25-pattern self-review
• Agent coordination failures producing "confidently wrong" output
• Content variety decay required an additional "repertoire tracking system"

CEO Sebastian: "We could remove the human gate and publish autonomously, and the quality gates would catch most issues. We do not, because the issues they miss are the ones that damage credibility: an unverified claim, a tone-deaf opening, a placeholder that slipped through."

The Atlantis failure pattern

Clients used agentic AI to generate 200 blog posts in a weekend, fired their agencies, returned three months later "ranking for nothing and getting cited by no one." The agents produced content. The failure was the absence of strategic human judgment.

BCG's four-tier Trust Protocol

1. Shadow Mode — agent suggests, human acts.
2. Supervised Autonomy — agent stages, human approves with >90% confidence.
3. Guided Autonomy — agent acts, human monitors exceptions.
4. Full Autonomy.

BCG places nearly all enterprise SEO content systems at Tier 3. Only 23% of organizations using agentic AI are actually scaling it.

The autonomy ceiling IS the AI-content quality floor

Same wall, two sides. The four-item editorial floor (original data, named author, topical fit, substantive expert edit) requires interventions that automated systems structurally cannot do at scale. Voice drift on 1-in-5 pieces enters the zone where Google's January 20, 2026 enforcement applies. Companies optimizing for "100% automation" are optimizing toward the penalty line.

The realistic 80-87% labor reduction (not 90%+)

• Prior baseline: ~40 human-hours/month for a serious program.
• HITL review: 5-10 minutes per piece × 40 pieces = ~3.3-6.7 hours/month.
• Reduction: from 40 hours to 5-10 hours = 80-87%.

The residual 13-20% is structurally irreducible. Agents handle research, draft generation, technical audits, citation monitoring. Humans do original data collection, expert sourcing, voice judgment, strategic prioritization.

The realistic mid-2026 stack

• Drafting and review: Claude Opus 4.6 or equivalent.
• Browser automation: OpenAI Operator, Anthropic computer-use, or Browser Use.
• Crawler: any production-grade site crawler.
• CMS integration: headless CMS for programmatic publishing.
• Feedback loop: GSC, Ahrefs, SimilarWeb, Profound.
• Human layer: a strategist who approves publication and owns the prompt library.

Tooling: roughly $1-6K/month managed, vs vendor pricing $9-$999/month per tool.

The March 2026 Google Core Update — detailed winners and losers

The load-bearing empirical confirmation of the four-item editorial floor. Independent analyses across SISTRIX (Aleyda Solis + Steve Paine), Lily Ray's 2,076-domain Amsive analysis, SE Ranking's 100K-keyword study all converge: first-party / official-source content gained; aggregator / comparison-layer content lost.

Volatility metrics

Combined Spam Update (March 24-25) + Core Update (March 27) window:

• 79.5% of top-3 results changed.
• 24.1% of top-10 results dropped out of the top 100.

Per-vertical winners and losers (SISTRIX VI, US)

Health — bifurcated at the official-source line

• Gainers: NIH.gov +9.3, Harvard +5.3, WHO +3.7
• Losers: Cleveland Clinic -11.5, WebMD -9.1, Mayo Clinic -6.1, MedlinePlus.gov -10.4

Finance

• Gainers: IRS.gov, SBA.gov, American Express +23.2%
• Losers: NerdWallet -15.9%, CreditKarma -34.2%

Jobs

• Gainers: USAJobs.gov +16.3%, Disney Careers +58.5%
• Losers: Indeed -18.1, ZipRecruiter -21.6%, Glassdoor -21.7%

Entertainment / reference

• Gainers: IMDB +79.3, Amazon +59.8
• Losers: YouTube -566.97 (largest single-domain loss in SISTRIX history), Reddit -64.2

The YouTube loss is structurally important: simultaneously lost organic VI AND gained AIO citation share. The two systems are answering different questions.

The three re-weighted signals (Evertune attribution)

1. Information Gain — net-new knowledge vs recapitulating existing material. "If this page disappeared, would anyone lose access to information they couldn't find elsewhere?"
2. Author Expertise — verifiable E-E-A-T credentials. Generic "Editorial Team" bylines lost regardless of content quality.
3. Topical Coherence — domain-level authority over the topic.

Lily Ray's read: "First-hand expertise, primary sourcing, and original data continue to compound in value." SISTRIX machine analysis: "the update rewards domains that are the natural first port of call for their topic."

Caveat: Evertune is a GEO vendor with commercial interest in this framing. The pattern is independently confirmed, but "Information Gain" is Evertune's interpretive label, not Google's named ranking signal.

Measurement and tooling for the 2026 SEO practitioner

The core problem: AI "rank" does not have a stable referent.

The fundamental problem: AI responses are stochastic

• SparkToro brand-recommendation study (~28K prompts): <1 in 100 chance of getting the same brand list twice; <1 in 1000 chance of same order.
• Carnegie Mellon EMNLP 2025 paper (n=2,976): automated LLM consistency metrics do not align with human perception.
• Tow Center for Digital Journalism citation-accuracy study (1,600 tests, 8 AI search engines): accuracy ranges from 37% failure (Perplexity, best) to 94% failure (Grok 3, worst). ChatGPT was "confidently wrong" in 134 of 200 tests.
• Aleyda Solis's own experience: her hosting was blocking AI bots without her knowing — invalidating her client visibility data. This is common.

What is measurable

First-party data (one source)

Bing Webmaster Tools added first-party AI citation data in February 2026 — the only first-party source as of mid-2026. Google does not provide equivalent. Anthropic, OpenAI, Perplexity do not provide citation telemetry to brands.

Third-party tracking platforms

• Profound — AI visibility tracking with structured prompt libraries; supports 60+ repetition design.
• Otterly — Perplexity-focused tracking with structured citation logging.
• Peec.ai — multi-platform tracking with citation source attribution.
• Evertune — multi-platform (vendor-skepticism on proprietary metrics).

Treat all third-party AI visibility data as estimates over distributions, not single-run measurements.

Survey-based discovery proxies

"Did AI influence your purchase?" added to checkout flows / post-purchase surveys supplies a downstream signal when click-path attribution can't see it.

Aleyda Solis's 3-layer measurement framework (BrightonSEO April 2026)

Crucially, the framework does not include an "AI rank" KPI — that's the tell it's serious.

Layer 1 — Presence

Five KPIs against a structured prompt library (30-60 prompts for a single product, 250+ for an enterprise; each prompt run ≥60 times):

1. Prompt coverage %
2. Recommendation rate
3. Linked citation rate
4. Comparative win rate
5. Representation accuracy

Layer 2 — Readiness

Ten characteristics of "AI search winning brands" — verifiable author presence, crawl access, schema fidelity on entity-disambiguation fields, topical coherence, original-data presence.

Layer 3 — Business Impact

Three explicit confidence tiers:

• Observed (high) — server-side logged conversions tied to AI-source identifiers.
• Proxy (medium) — survey-based attribution; correlation between visibility and outcomes.
• Modeled (low) — econometric attribution. Most "AI search ROI" claims live here, presented as Observed.

The technical-foundations gate (often broken silently)

AI crawlers use different user agents than Google; hosts and WAFs commonly block them by default rules.

Audit checklist:

1. robots.txt — explicitly allow GPTBot, ClaudeBot, PerplexityBot, GoogleOther-Extended, OAI-SearchBot, Anthropic-AI.
2. WAF rules — confirm AI crawlers aren't caught by anti-scraping defaults.
3. CDN configuration — Cloudflare, Akamai, Fastly may default-deny new user agents.
4. JavaScript rendering — most AI crawlers do NOT render JavaScript (Google is the exception). Content hidden behind JS-only routes is invisible to AI bots. SPA without SSR = likely invisible.

The realistic measurement stack for a 2026 SEO team

Layered build, in priority order:

1. Bing Webmaster Tools — free first-party AI citation signal.
2. Crawler audit — the technical gate.
3. One AI visibility tracker — Profound or Otterly first depending on platform focus.
4. Server-side ACP order attribution (if transactional).
5. Survey instrumentation at checkout / form-fill.
6. Prompt library construction — 30-60 prompts run ≥60 times each, quarterly cadence minimum.
7. Monthly trend review, not a weekly dashboard.

Tooling spend: $200-2,000/month single-brand, $5K+ enterprise. Above that is buying false confidence.

Vendor categories to skip

• AI-rank dashboards claiming absolute rank position.
• GEO consulting packages priced $20K-$100K without specific deliverables.
• Single-platform AI visibility tools.
• "Schema/llms.txt audit" packages.