#Foreword

In November 2023, six researchers at Princeton, IIT Delhi, Georgia Tech and the Allen Institute uploaded an unassuming arXiv pre-print titled GEO: Generative Engine Optimization^[1]. By KDD 2024 in Barcelona, the paper had a name for the discipline, a 10,000-query benchmark, and the first measured citation lifts — Cite Sources at +40.6%, Quotation Addition at +35.1%, Statistics Addition at +32.9% on Perplexity's pipeline^[2]. Eighteen months later, the discipline has a market.

By April 2026 the open web is being read more by language models than by humans. Google's AI Overviews appear on roughly 47% of mobile SERPs^[3], 99.9% of informational keyword queries^[4], and reduce click-through to the first organic result by 58% in the latest Ahrefs longitudinal study^[5]. Pew's controlled study of 68,879 real queries found that users click any link only 8% of the time when an AI Overview is present, versus 15% without — a 46.7% relative collapse^[6]. Zero-click rates have crossed 60% across all Google queries, 83% on AI-Overview-triggering queries, and 93% inside Google AI Mode^[7][8]. The traffic funnel from search did not shrink. It inverted.

For perea.ai's audience the implication is not subtle. The asset that mattered for the last twenty years — a page that ranked on the first SERP — has decoupled from the asset that matters for the next ten: a page that gets cited inside a generated answer. The 5W AI Platform Citation Source Index 2026, which synthesizes 680 million individual citations across ChatGPT, Google AI Overviews, Perplexity, Gemini and Claude, found that just 15 domains capture 68% of all consolidated citation share^[9]. The concentration is more extreme than Google PageRank ever produced. Citation share moves in weeks, not years — ChatGPT's Reddit citation rate fell from roughly 60% to 10% in six weeks in late 2025 after a single Google parameter change^[9]. The publishers that absorbed the displaced share were PR Newswire, Forbes and Medium, in that order.

This paper is the field manual for operating in that environment. It is the third in the perea.ai Research series — after The B2A Imperative and The MCP Server Playbook — and it closes the discovery-layer thread that both of those papers gestured at without resolving. It is grounded in 100+ primary sources: the Princeton/KDD benchmarks, the SE Ranking 300K-domain study, the Profound 680M-citation dataset, the Pew Research Center field study, the Seer Interactive 25.1M-impression panel, the Authoritas 200B-URL Perplexity index analysis, and dozens of controlled experiments by Magna, Surgeboom, Relixir, Search Engine Land and others. Where the data converges, the recommendations are sharp. Where it does not — llms.txt is the most contested example — the paper is honest about the dispersion and recommends a position with the reasoning behind it.

The thesis is simple enough to fit on a wall. Citations are the new clicks. In the citation economy, the unit that compounds is not the visit, it is the line of attribution inside a model's answer — and a single model emits that attribution thousands of times before it is retrained. The teams that ship citation infrastructure in 2026 build a compounding asset that pays out for years. The teams that wait will be shipping table-stakes work in 2027 alongside everyone else, with no advantage left to capture.

— perea.ai Research

#Executive Summary

The thesis. AI search is not a marketing channel. It is the new discovery layer of the B2A economy, and citation in a generated answer is the unit of distribution. A single citation in ChatGPT, Perplexity, Google AI Overviews, Gemini or Claude is read by thousands to millions of users before the model is retrained — and inside the agent economy each of those reads can be an autonomous purchase decision. Brands that ship citation-grade content infrastructure in 2026 capture a category position that compounds for the next decade. Brands that do not are structurally invisible inside the channel where 80–89% of B2B buyers and roughly 40–70% of consumers now begin their research^[10][11].

The evidence.

• Adoption velocity. AI-sourced traffic grew 357% YoY in mid-2025 to 1.13 billion monthly referrals^[12]. Similarweb data shows AI search visitors converting at 4.4× to 14.2× the rate of Google organic — 14.2% vs. 2.8% in Outpace's panel^[11].

• Click compression. Seer Interactive's analysis of 25.1M impressions across 42 organizations measured organic CTR collapsing from 1.76% to 0.61% — a 61% decline — on AIO-present queries^[13]. Pew clocked the user-side equivalent at 8% click rate vs. 15% without an AIO^[6]. Ahrefs put the first-result CTR drop at 58% in February 2026^[5].

• Citation premium. Brands cited inside an AI Overview earn 35% more organic clicks and 91% more paid clicks than non-cited competitors on the same SERP^[14]. The asymmetry is the reason citation, not ranking, is now the asset.

• Concentration. The top 15 domains capture 68% of all AI citations across ChatGPT, AI Overviews, Perplexity, Gemini and Claude^[9]. Reddit alone is the #1 source on every major engine at roughly 40% citation share. Wikipedia accounts for 26–48% of ChatGPT's top-10 citation share.

• Cross-platform divergence. Only 11% of domains cited by ChatGPT are also cited by Perplexity for the same query; 71% of cited sources appear on only one platform^[15]. The same paper that wins on ChatGPT may be invisible on AI Overviews. Optimization is per-engine, not universal.

• Infrastructure leverage. Pages with attribute-rich Product/Review schema are cited at 61.7% vs. 41.6% for generic schema^[16]. FAQPage schema increases citation rates 52–89% across studies, with Google AI lift as high as +221%^[17][18]. Comprehensive 15+ schema-type implementations move citation rate from 9% (no schema) to 64%^[19].

The seven citation factors that move the numbers.

1. Answer-first structure. 44.2% of LLM citations come from the first 30% of a page, and 82% of Perplexity answers pull from the first 300 words of source content^[20]. Lead with a 40–60 word standalone answer, every section.

2. Cite, quote, quantify. The Princeton GEO paper's three top interventions — Cite Sources (+40.6%), Quotation Addition (+35.1%), Statistics Addition (+32.9%) — are robust, replicable, and cheap to apply^[2].

3. Schema as machine-readable contract. FAQPage, HowTo, Article-with-Person, Organization-with-sameAs, and attribute-rich Product/Review are the five schema types with measured citation lift. Generic schema with no attribute density has zero effect^[16][17].

4. Entity & E-E-A-T. Wikipedia/Wikidata presence, Person schema with sameAs, Organization schema with sameAs to LinkedIn/Crunchbase, and external mentions in tier-1 publications. 70.4% of ChatGPT-cited sources carry Person schema; Wikipedia presence boosts citation rates ~50%^[21][22].

5. Freshness curves. Perplexity drops content sharply after 60–90 days; AI Overviews update in 4–8 weeks; ChatGPT base model lags 6–18 months^[23]. Update high-value pages on a 45–90 day cadence.

6. Per-engine source mix. ChatGPT favors Wikipedia + structured authority. Perplexity favors Reddit + named B2B authority. Google AI Overviews favors YouTube + Reddit + Quora + LinkedIn. Claude favors NYT/Atlantic/Economist^[15][9]. Optimization is portfolio-shaped.

7. Brand & entity search volume. Brand search volume has the strongest measured correlation with LLM citation (r ≈ 0.33), outweighing domain authority (r ≈ 0.18, down from 0.43 in 2024)^[24][15]. Earned media and product virality are now first-class GEO levers.

The 90-day playbook. Days 0–30: audit, prioritize 50 query targets, fix crawler access, ship Organization + Person + Article + FAQPage schema on top 20 pages. Days 31–60: rewrite top 20 pages to answer-first structure with citations + quotations + statistics. Add Wikidata entry. Establish authentic Reddit presence in two priority subreddits. Days 61–90: deploy citation tracking (Profound, Otterly, Peec, Athena, or self-built), set the share-of-voice baseline, ship llms.txt, instrument GA4 for AI referral attribution. Quarterly: refresh top pages, expand entity graph, measure share-of-voice deltas.

The operator math. The 90-day playbook is a 1–2 person engineering effort. Tooling cost runs $0–500/mo for a startup, $500–5,000/mo for mid-market, $15K+/mo for enterprise. The compounding asset is real: every citation persists across model retrainings, and the share of voice you build now sits in the training data of the models that will run for the next five to ten years. Citations are the only digital asset that pays compound interest into model weights.

#Quotable Findings

1. Google's AI Overviews appear on roughly 47% of mobile SERPs^[3], 99.9% of informational keyword queries^[4], and reduce click-through to the first organic result by 58% in the latest Ahrefs longitudinal study^[5].

2. The 5W AI Platform Citation Source Index 2026, which synthesizes 680 million individual citations across ChatGPT, Google AI Overviews, Perplexity, Gemini and Claude, found that just 15 domains capture 68% of all consolidated citation share^[9].

3. Seer Interactive's analysis of 25.1M impressions across 42 organizations measured organic CTR collapsing from 1.76% to 0.61% — a 61% decline — on AIO-present queries^[13].

4. Brands cited inside an AI Overview earn 35% more organic clicks and 91% more paid clicks than non-cited competitors on the same SERP^[14].

5. Pew Research's controlled panel of 68,879 real queries found users clicked any link in just 8% of AIO sessions versus 15% otherwise — a 46.7% relative drop — and clicked an in-AIO source in only 1% of sessions^[6].

6. ZipTie's cross-platform analysis found that only 11% of domains cited by ChatGPT are also cited by Perplexity for the same query, and 71% of cited sources appear on only one platform^[15].

7. Reddit alone accounts for 46.7% of Perplexity's top-10 citation share — not user-generated content broadly, Reddit specifically^[25].

8. Surgeboom's 1,500-site study showed pages with 15+ schema types citing at 64% versus 9% for pages with no schema — a ~7× spread^[17].

9. Brand search volume has the strongest measured correlation with LLM citation (r ≈ 0.33), outweighing domain authority (r ≈ 0.18, down from 0.43 in 2024)^[24][15].

#Part I — Why GEO Is the New SEO

#1.1 The traffic curve has bent

Google still processes 9.1–13.6 billion searches per day, more than the 8.5 billion of 2024^[26]. The number of searches did not collapse. The clicks did. Zero-click search rose from roughly 50% in 2019 to 60% in 2024 to 60–69% across all Google queries by Q1 2026^[26][7]. Mobile zero-click hit 77%^[26]. On AI Overview-triggering queries it reached 83%, and inside Google's standalone AI Mode the zero-click rate now sits at 93%^[7][8].

Seer Interactive's longitudinal study across 25.1 million impressions and 42 organizations is the cleanest single dataset on the magnitude of the shift. Organic CTR on AIO-present queries fell from 1.76% to 0.61%, a 61% relative decline. Paid CTR fell from 19.7% to 6.34%, a 68% decline^[13]. Ahrefs's February 2026 update measured top-result CTR reduction at 58%, nearly double the 34.5% Ahrefs reported eight months earlier^[5]. Pew Research's controlled panel of 68,879 real queries found users clicked any link in just 8% of AIO sessions versus 15% otherwise — a 46.7% relative drop — and clicked an in-AIO source in only 1% of sessions^[6]. 26% of users left Google entirely after viewing an AI Overview, versus 16% on traditional results.

Publisher disclosures align with the panels. The New York Times reported Google search referrals down ~35% YoY in Q1 2026; The Washington Post around 40%; Axios around 33%^[27]. HubSpot disclosed up to 80% loss of blog traffic. Chegg lost 49%. Business Insider lost 55% and cut 21% of staff. The economic basis of ad-supported informational publishing is being rewritten in real time.

#1.2 The citation premium

The compensating reality, the one that determines whether a brand survives the shift, is the citation premium. Brands cited inside an AI Overview earn 35% more organic clicks and 91% more paid clicks on the same SERP than non-cited competitors^[14]. The mechanism is endorsement: when an AI says "according to [Brand], the best approach is X," that attribution functions as a third-party recommendation. Users who encounter the brand inside the AI answer are more likely to recognize it, search for it directly, and click on its paid or organic listings.

The conversion asymmetry compounds the citation premium. AI-driven referrals convert at 4.4× to 23× the rate of traditional organic. Outpace's analysis put the AI search visitor conversion rate at 14.2% versus Google organic's 2.8%, a 5× difference^[11]. Ahrefs has measured 23× in some segments. Semrush data shows AI-driven traffic at 4.4× higher conversion^[11]. Visitors who arrive after an AI consultation arrive informed, with intent. The traffic that survives is dramatically more valuable per session.

The combined picture: fewer clicks, much higher value per click, and a brand-awareness asset that operates whether or not a session ever lands in your analytics. Vercel reports 10% of its monthly signups now come from ChatGPT, attributed in part to its GEO investments^[28]. AI-referred traffic at one B2B SaaS panel generated 12.1% of signups despite accounting for only 0.5% of overall traffic^[11].

#1.3 The B2A discovery layer

For an agent economy reader, the framing shifts again. In B2B and consumer search, citations drive human attention. In B2A, citations are the discovery layer agents themselves consume. When an agent — Cursor, ChatGPT operator, an internal LangGraph workflow, a Bland telephony bot — needs to answer a question, evaluate a vendor, or make a tool selection, it pulls from the same citation pool that surfaces in human-facing AI answers. The training data of every frontier model encodes today's citation graph as the default substrate of agentic reasoning. The B2A Imperative paper argued that the agent economy needs new infrastructure layers on top of MCP, A2A, AP2 and ACP. The citation economy is the discovery layer underneath all of them.

Strategically this means citation work compounds differently in B2A than in human SEO. A page that gets cited 1,000 times in ChatGPT answers in 2026 likely lands in the training data of the next four to six retraining cycles of the major frontier labs — three to five years of compounding before the page is even refreshed. Combined with the Tow Center's measured citation persistence patterns, citation-grade content has the longest amortization window of any digital asset class. Once you understand that, the budget allocation between SEO and GEO is not a debate.

#Part II — How AI Engines Actually Cite

#2.1 The five engines, five citation logics

Citation behavior is not uniform across the AI search layer. ZipTie's cross-platform analysis found that only 11% of domains cited by ChatGPT are also cited by Perplexity for the same query, and 71% of cited sources appear on only one platform^[15]. Optimization is portfolio-shaped, not universal. Each engine carries a distinct retrieval architecture, source preference, and freshness curve.

Source: ZipTie cross-platform study; 5W AI Platform Citation Source Index 2026; Cited research dataset^[15][9][29].

The shape of the table is the strategy. ChatGPT is the most meritocratic — a five-week-old domain with deep, well-structured content can earn citations^[29]. Perplexity rewards Reddit presence and outbound citation density. Google AI Overviews promotes existing organic winners — 92% of citations come from pages already ranking in Google's top 10^[29]. Gemini concentrates citations on a small elite set. Claude leans toward established editorial voices and is the most freshness-tolerant.

#2.2 ChatGPT: encyclopedic, breadth-driven, Bing-grounded

ChatGPT's citation logic is the most documented. When browsing is enabled, ChatGPT Search shows an 87% correlation with Bing's top-10 results for the same query^[15]. When browsing is off, the model draws on training data — heavily weighted toward Wikipedia, which accounts for 47.9% of ChatGPT's top-10 citation share and roughly 7.8% of all citations across studies^[30]. Reddit is the second-most-cited domain, weighted heavily for product recommendations and "what do real users think" queries.

Two structural facts dominate. First, ChatGPT runs query fan-outs aggressively — 32.7% of ChatGPT queries are single-shot vs. 70.5% on Perplexity, meaning ChatGPT typically issues 3–5 internal searches per user prompt^[31]. Optimization for ChatGPT means optimizing for the fan-out variations of the user's question, not just the literal phrasing. Second, ChatGPT skews toward the newest cited content among the major AI systems, with 56% of journalism citations from the past 12 months versus 36% for Claude^[9]. Freshness matters even where it is least demanded.

Practically: for ChatGPT, win Wikipedia presence (or a high-quality entity proxy via Wikidata sameAs), maintain depth on your topic cluster, and write content that matches the shape of the questions a user might ask 2–3 levels into a fan-out from your core keyword.

#2.3 Perplexity: real-time, Reddit-heavy, citation-transparent

Perplexity is the most transparent of the five engines and the easiest to instrument. Every response carries clickable citations, retrieval is real-time per query, and the platform maintains its own 200B+ URL index^[29]. The two signals that dominate Perplexity citation behavior are Reddit presence and freshness.

Reddit alone accounts for 46.7% of Perplexity's top-10 citation share — not user-generated content broadly, Reddit specifically^[25]. Perplexity even ships a Reddit-only focus mode. For brands targeting Perplexity, this reframes Reddit from "optional social distribution" to critical infrastructure. The right approach is genuine subject-matter expertise in 2–3 priority subreddits, posting useful answers under your brand or a named team member, citing your own work only when actually relevant. Promotional behavior gets downranked by Reddit's own moderation and by Perplexity's authenticity scoring.

Freshness drops steeply on Perplexity. Content older than 60–90 days loses ground unless it continues to receive new external citations or is updated^[15]. This is why brands that compete on Perplexity ship monthly or quarterly content refreshes on their highest-citation pages, not just new content. The combination of Reddit presence + recency-fresh on-domain content + outbound citation density is the dominant Perplexity playbook.

#2.4 Google AI Overviews: PageRank-amplified, schema-rewarded

Google AI Overviews is the hardest target for new entrants and the highest-volume target for established sites. 92% of AI Overview citations come from pages already ranking in Google's top 10 organic results^[29]. The implication is that AIO does not discover new sources — it promotes existing winners. For a new domain this means traditional SEO is a 3–6 month prerequisite to AIO visibility. For an established domain it is the highest-leverage opportunity.

Schema markup correlates strongly with Google AI Overview citations. Pages with comprehensive structured data (FAQPage, HowTo, Article, Organization) outperform pages without it dramatically. Relixir's 50-site study found pages with FAQPage schema cited at 41% versus 15% without — a 2.7× lift^[17]. Citedify's analysis put the FAQ-vs-no-FAQ lift at 3.2× for AIO specifically^[18]. SeoJuice's controlled A/B test on FAQ + HowTo schema produced a 3× citation advantage on Google AIO and a 156% selection-rate increase when text/image/structured data combined^[32]. The Search Engine Land single-page experiment found that the "no schema" version was not even indexed^[19].

E-E-A-T signals are near-mandatory. Discovered Labs's analysis of Wellows 2026 data found that 96% of AI Overview citations come from verified authoritative sources, and pages with expert author attribution are cited at 2.4× the rate of anonymous pages^[21]. 70.4% of sources cited by ChatGPT include Person schema in JSON-LD^[21].

The strategy stack for AIO: rank in the top 10, ship FAQPage + HowTo + Article-with-Person + Organization-with-sameAs schema, attach named author bios with credentials, get cited by tier-1 publications, maintain clear update dates, and ensure the answer to the implied query is in the first 100 words.

#2.5 Gemini and Claude

Gemini's citation behavior is the most concentrated of the five, with a Gini coefficient of 0.351 — meaning a small set of elite sources dominates citation share^[15]. Brand search volume and Knowledge Graph presence outweigh most other signals. For Gemini, Wikidata + Wikipedia + tier-1 press is the baseline; without those, citation share is hard to win.

Claude is the freshness-tolerant outlier. Only 36% of Claude's journalism citations come from the past 12 months, versus 56% for ChatGPT^[9]. Claude leans toward established editorial brands — The New York Times, The Atlantic, The New Yorker, The Economist — and toward longer-form analytical blog content (43.8% of Claude's top-10 is blogs). For Claude, depth and editorial authority matter more than recency. Long-form whitepapers like this one perform structurally well; short marketing pages do not.

#Part III — The Seven Citation Factors

The Princeton/KDD GEO paper produced the most rigorous public benchmark of how content modifications affect citation rates. The authors evaluated nine optimization methods across 10,000 GEO-bench queries on a controlled GPT-3.5-based pipeline and on Perplexity in production. The top-three interventions — Cite Sources, Quotation Addition, Statistics Addition — produced 30–40% relative improvements on position-adjusted word count and 15–30% on subjective impression metrics^[2].

The seven factors below combine the Princeton results with the field-measured signals from Profound, SE Ranking, Magna, Surgeboom, Cited, Athena and Discovered Labs. Together they form the operating model for content that gets cited.

#3.1 Answer-first structure

The single most consequential structural decision is to lead with the answer. Authoritas measured that 82% of Perplexity citations pull from the first 300 words of the source content^[20]. The Geo Knowledge Base reports that 44.2% of all LLM citations come from the first 30% of text on a page^[33]. Pages that bury the answer in paragraph eight after four paragraphs of context preamble are skipped during retrieval.

The operating rule is the inverted pyramid. Within the first 40–60 words of every section, state a complete, self-contained answer to the question implied by the heading. Follow with supporting evidence in 60–100 words. Add related context in another 50–70 words. The total passage runs 150–200 words, fits cleanly inside an LLM's chunk-retrieval window, and is what AI engines are most likely to extract verbatim.

Heading discipline matters as much as paragraph discipline. Headings phrased as questions (mirroring real user queries) are cited 3.1× more often than topic-label headings, per Semrush 2026 data^[34]. The reason is mechanical: the AI engine performs an explicit string-match between the user's query and your heading, and a question-shaped heading is closer in embedding space to a question-shaped query.

#3.2 Cite, quote, quantify

The three Princeton GEO interventions — adding citations to credible sources, adding quotations from named experts, and adding specific statistics with attribution — are the cheapest and most replicable citation lifts measured. Cite Sources produced +40.6% on Perplexity, Quotation Addition +35.1%, Statistics Addition +32.9%^[2]. The mechanism is verifiability: generative models prefer claims they can anchor to specific numbers or named sources because quantitative data and attributed quotes reduce hallucination risk during synthesis.

Operationally this means every paragraph that contains a non-obvious claim should carry a year-stamped statistic, a named-source quote, or a hyperlink to a primary source. Avoid generic descriptors ("many", "some", "good"). Replace with concrete data ("46.7% of Perplexity's top-10 citations come from Reddit, per ZipTie's 2026 analysis"). Surgeboom's panel of 200 AEO-optimized articles found that paragraphs containing at least one statistic are cited 4× more often than statistic-free paragraphs^[35].

The Princeton paper also tested Fluency Optimization (rewriting for readability without changing factual content) at +13% lift, and Easy-to-Understand Rewriting at modest gains^[2]. Both reinforce the principle: extractable, attributable, plain-language content wins.

#3.3 Schema as machine-readable contract

Schema markup is the structural amplifier. The empirical evidence converges on five schema types with measured citation lift, while generic schema (Article, BreadcrumbList, Organization without attributes) shows null effects in controlled tests.

The Search Engine Land single-page experiment is a useful ground truth: three identical content pages differing only in schema quality produced rank 3 + AI Overview citation for "well-implemented", rank 8 + no AIO for "poorly implemented", and not indexed for "no schema"^[19]. Surgeboom's 1,500-site study showed pages with 15+ schema types citing at 64% versus 9% for pages with no schema — a ~7× spread^[17].

The caveat from Growth Marshal's controlled study matters: generic schema with no attribute density produces zero measurable effect. Pages with Product schema that include only name and price underperformed pages with no schema at all in their study. Schema is a contract for structured information — populate every relevant property with accurate, current data, or skip the schema entirely.

#3.4 Entity & E-E-A-T

Google's E-E-A-T framework — Experience, Expertise, Authoritativeness, Trustworthiness — is reflected across the AI search layer. Discovered Labs found 96% of AI Overview citations come from verified authoritative sources, and pages with expert author attribution are cited at 2.4× the rate of anonymous pages^[21]. 70.4% of sources cited by ChatGPT include Person schema in JSON-LD^[21]. AI Search Visibility's 2026 analysis found that domain authority's correlation with AI citation has dropped from r=0.43 in 2024 to r=0.18 in 2026^[37], while structural E-E-A-T signals (Person schema, author byline, external citations) have risen to compensate.

The entity layer is where the strongest effect lives. Wikipedia is the single most-cited source on ChatGPT and is heavily represented in every frontier model's training data. A Wikipedia page about a brand boosts that brand's citation rate by ~50% across LLMs^[22]. For brands that don't meet Wikipedia's notability threshold, Wikidata is the second-best entry point — structured entity data without the editorial gatekeeping. Crunchbase, LinkedIn, G2, Trustpilot and category-specific directories complete the entity graph. AI systems learn entity relationships through co-occurrence: when a brand appears alongside category leaders in industry analysis, AI systems strengthen its association with that category.

Brand search volume itself is the strongest single predictor of LLM citation, with measured correlation r ≈ 0.33 in the AI Search Visibility dataset^[24][15]. This is why earned media and product virality compound into GEO — they raise the brand search baseline that determines downstream citation share.

#3.5 Freshness curves

Content decay rates differ sharply across engines. Cited's analysis converges with multiple field studies on the following timing^[38]:

• Perplexity: New structured content can be cited within 2–4 weeks. Content older than 60–90 days drops sharply unless updated or receiving new external citations. Update cadence: monthly for top pages.
• Google AI Overviews: Reflects structural changes in 2–4 crawl cycles (4–8 weeks). Update cadence: 45–90 days for top pages.
• ChatGPT (Search mode): Bing-grounded; fast inclusion if Bing indexes the change. Browse-disabled mode lags 6–18 months until next training cycle.
• Claude: Slowest to react; treats content age permissively. Quarterly to semi-annual refresh sufficient.
• Gemini: 4–8 weeks for ranking changes; entity/Knowledge Graph changes lag longer.

The operational implication is that freshness is not a single signal — it is a per-engine variable with different cadences. The best practice is a freshness audit on top-cited pages every 30 days, with quarterly major refreshes that include new statistics, dated quotations, and updated dateModified properties in JSON-LD.

#3.6 Per-engine source mix

The 5W AI Platform Citation Source Index 2026 categorizes the 50 most-cited domains into six functional buckets and identifies the citation leader in each^[9]:

• Community & Conversation: Reddit (#1 across all engines, ~40% citation share)
• Encyclopedic & Reference: Wikipedia (#1, 26–48% on ChatGPT)
• Professional & Identity: LinkedIn (#1 on Google AIO at 13.0%)
• Video & Audio: YouTube (200× advantage over any other video source; 23.3% on Google AIO; 13.9% on Perplexity)
• Editorial & News: Forbes, Business Insider (ChatGPT favorites); NYT, Atlantic, Economist (Claude favorites); NIH/PubMed (Perplexity authority sources)
• Commerce & Review: G2, Gartner, NerdWallet, PCMag, Yelp, TripAdvisor

The portfolio strategy follows directly. A B2B SaaS optimizing for ChatGPT prioritizes Wikipedia/Wikidata + Forbes/Business Insider press + own-domain depth. The same brand optimizing for Perplexity prioritizes Reddit + G2/Gartner reviews + NIH-equivalent primary sources + own-domain freshness. For Google AIO, it prioritizes YouTube + Reddit + Quora + LinkedIn + own-domain organic ranking + structured data. The work is not duplicative — it stacks. But it is real work, not "post the same content five places."

#3.7 Brand & entity search volume

The hardest factor to manufacture and the most predictive. Brand search volume — how often users search your brand name directly — has the strongest measured correlation with LLM citation share at r ≈ 0.33^[24]. The mechanism is straightforward: brand searches feed Knowledge Graph entries, drive third-party content production, and signal authority to every retrieval system that intersects with Google's ecosystem (which is most of them).

The leverage points:

1. Earned media velocity. A consistent rhythm of tier-1 mentions (Forbes, TechCrunch, Business Insider, industry trade press) builds brand search volume that compounds. Press is GEO infrastructure, not just PR.
2. Product virality / Reddit organic. Authentic Reddit threads about your product produce both citation share and brand search volume.
3. Conference and podcast presence. Named appearances in citation-eligible venues (top podcasts in your category, conference keynotes) feed both LLM training data and brand search.
4. Co-occurrence with category leaders. Frequent contextual mentions alongside Salesforce / HubSpot / Stripe / Shopify (whichever leaders define your category) strengthen entity association in model embeddings.

The honest version: brand-volume work has the longest latency of any GEO lever — 3–6 months minimum to see measurable signal accumulation. It is also the most defensible. Content can be re-written; entity authority compounds.

#Part IV — The Publisher Infrastructure Stack

#4.1 What ships in the codebase

For a SaaS or media site, the GEO infrastructure that lives in code rather than in editorial process is short:

1. JSON-LD schema generators for Article, FAQPage, HowTo, Organization, Person, Product, BreadcrumbList. Output validated against validator.schema.org and Google's Rich Results Test in CI.
2. Crawler access policy. Explicitly allow GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot, Applebot-Extended in robots.txt unless a specific business reason prevents it. Many sites have these blocked by default and never check.
3. llms.txt and llms-full.txt at the domain root, auto-generated from the site's existing structure. Mintlify, Yoast, and several headless CMSes now ship this by default.
4. dateModified discipline. Every Article schema instance reflects the actual last-edit timestamp, not the publish date. Dated copy in the body text mirrors the schema.
5. Anchor IDs on every heading and FAQ item. AI engines cite specific passages; anchors enable per-passage attribution and per-passage analytics.
6. sameAs graph on Organization schema linking to LinkedIn, Crunchbase, Wikidata, Wikipedia (where eligible), G2, Trustpilot, GitHub, and tier-1 press author pages.
7. Speakable specification on the highest-confidence answer paragraphs of priority pages.
8. OpenTelemetry-style citation tracking if you are running this internally — matching incoming AI referrals against expected query patterns.

This is roughly two engineer-weeks of work for a mid-size site, including QA. None of it requires JavaScript framework changes. JSON-LD ships in <script type="application/ld+json"> blocks; llms.txt is a static file at the root.

#4.2 The llms.txt question

llms.txt is the most contested piece of GEO infrastructure in 2026, and it deserves a sober treatment because the field is dominated by hype on one side and reflexive dismissal on the other.

The facts. Jeremy Howard at Answer.AI proposed the llms.txt standard in September 2024^[39]. It is a Markdown file at /llms.txt that gives AI systems a curated map of a site's most important content, supplemented by an optional /llms-full.txt that ships the complete content corpus. Anthropic, Stripe, Zapier, Cloudflare, Vercel, Hugging Face, Mintlify, Supabase, Resend, Clerk, Prisma, Turso, and Fly.io have all implemented it. As of March 2026, BuiltWith counts ~844,000 websites with the file, and SE Ranking's 300K-domain study puts adoption at 10.13%^[40][41].

The honest case against. SE Ranking's controlled study found no measurable correlation between llms.txt presence and AI citation share. Google's John Mueller has publicly stated that no AI system at Google currently uses llms.txt. Server-log analysis from multiple sources shows GPTBot, ClaudeBot and PerplexityBot do not visit llms.txt files with any meaningful frequency^[41].

The honest case for. Anthropic, OpenAI and Perplexity have not committed in writing but have signaled support and observable retrieval response. Cloudflare's "Markdown for Agents" feature (Feb 2026) reduces token consumption for AI crawlers by ~80% on sites that ship clean Markdown content^[42]. Developer-facing AI tools — Cursor, Continue, Aider, GitHub Copilot, and various RAG frameworks — actively read llms.txt when present. Brands that publish a curated llms.txt see modest correlated uplift in citation rates on Anthropic and Perplexity in some studies (Presenc AI^[43]).

The perea position. Ship llms.txt. Implementation cost is two hours including validation. The downside is zero. The asymmetric upside is real: if any major platform formally commits to llms.txt as a first-class input — and the trajectory suggests one will, likely Anthropic given they already publish theirs at docs.claude.com/llms.txt — sites that have it ship will be 12–18 months ahead of the rest of the market on day one. Treat it as a forward-compatibility hedge, not a near-term citation lever, and move on.

#4.3 The reference architecture

For a B2B SaaS or editorial site shipping a citation-optimized publication, the reference stack as of mid-2026:

/                          (Site root)
/llms.txt                  (Curated AI map, ~50–500 lines)
/llms-full.txt             (Full content corpus, optional)
/robots.txt                (Explicit allow for AI crawlers)
/sitemap.xml               (Standard XML sitemap, all canonical URLs)
/.well-known/              (OAuth metadata if you also ship MCP)

/research/<slug>           (One Markdown source per long-form paper)
  - Frontmatter             (title, subtitle, authors, date, license, audience)
  - Hero                    (Author + dateModified + reading time)
  - StickyTOC               (Anchored navigation)
  - Article body            (Inverted-pyramid sections, 40–60 word answers,
                             cite/quote/quantify discipline)
  - JSON-LD                 (ScholarlyArticle / Article + Person + Organization)
  - FAQ block               (5–8 H3-question entries, FAQPage schema)
  - References              (Numbered footnotes, primary sources only)

/authors/<name>            (Person entity page)
  - Bio + credentials
  - sameAs links
  - Person JSON-LD
  - List of authored content

/about, /contact, /editorial-policy
                           (Trust pages — about page, contact, editorial standards,
                            AI-disclosure policy if you author with AI assistance)

This is what perea.ai/research itself ships, deliberately. The publication eats its own dog food on agent-readability. Every paper carries Schema.org ScholarlyArticle JSON-LD with author metadata, wordCount, timeRequired, inLanguage, license, mainEntityOfPage. The TOC is anchor-linked. Each section is independently citable. The whole site is statically generated and served from the edge so retrieval latency is sub-100ms globally.

#4.4 What does not ship

A short list of practices that are either unnecessary, ineffective, or counterproductive in 2026:

• Cloaking content for AI crawlers. Sending different content to GPTBot than to human users is technically possible and detectably risky. Google has stated repeatedly that misalignment between visible content and schema is a violation of structured data guidelines. Don't.
• Generic schema with no attribute density. Growth Marshal's controlled study found generic Article/Organization/BreadcrumbList schema produced no measurable citation lift; attribute-rich Product/Review schema produced +20 percentage points^[16]. Schema without populated content is overhead, not lift.
• Auto-generated AI content at scale. Hashmeta's 100K-citation study found unedited AI-generated content was cited 89% less than human-edited content^[44]. Google's 2025 helpful-content updates explicitly target AI content used to manipulate rankings.
• Keyword stuffing in citations or schema. The Princeton GEO paper specifically tested keyword density manipulation and found no measurable improvement in AI citation share^[2].
• Black-hat link buying for entity authority. Google's January 2025 QRG update explicitly targets fake expert personas and triggers manual actions, not just ranking suppression^[21].

#Part V — The 90-Day GEO Implementation Playbook

#5.1 Days 0–30: Audit, foundation, and crawler hygiene

Week 1 — Baseline and prioritization.

• Pull a list of 50 priority queries across informational, commercial, and navigational intent. Run each through ChatGPT, Perplexity, Google AI Overviews, Gemini and Claude. Log: does an AI answer appear, who is cited, what URLs are linked, what is the sentiment toward your brand. This is your share-of-voice baseline.
• Audit robots.txt. Ensure GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot, Applebot-Extended are explicitly allowed. Many sites have legacy blocks that nobody remembers writing.
• Run a schema audit on top 50 pages by organic impressions. Categorize as: comprehensive (15+ schema types), advanced (10–14), moderate (5–9), basic (1–4), or none. The Surgeboom 2025 data shows 7× citation lift between none and comprehensive^[17].

Week 2 — Schema infrastructure.

• Implement Organization schema sitewide with full sameAs graph (LinkedIn, Crunchbase, Wikidata, GitHub, G2, Trustpilot, X, YouTube, tier-1 press author profiles).
• Implement Person schema on every author page with name, jobTitle, affiliation, sameAs to LinkedIn / ORCID / Google Scholar / X / GitHub.
• Implement Article schema on top 20 pages, linked to Person + Organization, with accurate datePublished, dateModified, mainEntityOfPage, wordCount, inLanguage.
• Validate every JSON-LD block against validator.schema.org and Google's Rich Results Test. Fix every error.

Week 3 — FAQPage and HowTo on top pages.

• Add FAQPage schema to top 20 pages. 5–10 questions per page. Each acceptedAnswer is a 40–60 word standalone response. Question wording mirrors real user search queries (use Search Console "People Also Ask" data and AlsoAsked.com).
• Add HowTo schema to procedural content. Every step has a name, text (20–40 words), position, and where applicable an image. Include totalTime and estimatedCost.
• Validate. Watch Google Search Console enhancement reports for the next two weeks for new error spikes.

Week 4 — Trust pages and llms.txt.

• Audit and refresh About, Contact, Editorial Policy, Privacy, Security, Terms. Every page has a real physical address, named editorial contact, AI-content disclosure policy. These are E-E-A-T signals that AI raters check explicitly per the September 2025 QRG update^[45].
• Generate llms.txt. Use Mintlify, Yoast, or a custom generator. Curate to the 30–80 highest-leverage pages. Validate at llmtxt.info/validator. Commit to CI/CD so it stays fresh as content ships.
• Consider llms-full.txt if you have substantial documentation. Anthropic's docs.claude.com/llms-full.txt runs ~481K tokens and is a reasonable model for technical documentation sites.

#5.2 Days 31–60: Content rewrite and entity work

Weeks 5–6 — Inverted-pyramid rewrite of top 20 pages.

The single highest-leverage editorial intervention. For each priority page:

1. Identify the question the page answers. State it as the H1 (or in the first 60 words).
2. Within the first 40–60 words, deliver a complete, standalone answer to that question. No preamble. No "in this article we'll explore." No setup.
3. Re-cast every H2 as a question that mirrors a real user search query. Re-cast H3s as sub-questions where applicable.
4. The first sentence after every heading is the answer to the question implied by the heading. Supporting evidence follows.
5. Apply the cite/quote/quantify discipline: every paragraph with a non-obvious claim carries a year-stamped statistic, a named-source quote, or a primary-source link.
6. Add 5–8 FAQ entries at the bottom, each marked up with FAQPage schema.
7. Ensure each major section is 150–250 words to match LLM chunk-retrieval windows.

GetCito's 200-client implementation panel reports answer-first restructuring increases citation frequency from 2–3 mentions/month to 8–12 mentions/month within 90 days — a 200–300% lift^[46]. Branded by Greenville's panel and Sona's AEO data corroborate the pattern.

Weeks 7–8 — Entity graph and Wikidata.

• Create a Wikidata entry for your organization if one does not exist. Wikidata has a lower notability threshold than Wikipedia and accepts most legitimate businesses with adequate sourcing.
• Pursue a Wikipedia entry if eligible. Wikipedia presence boosts citation rates ~50% across LLMs and dominates ChatGPT's training data. Notability is the bar — meet it through tier-1 press, not press-release wire mentions.
• Verify and complete Crunchbase, LinkedIn Company, G2, Trustpilot, Capterra and category-specific directory profiles. Same brand description, same logo, consistent NAP (name, address, phone) across all directories. Inconsistent NAP confuses entity disambiguation.
• Establish Reddit presence in 2–3 priority subreddits. Use a real named account, post genuine subject-matter expertise, cite your own work only when actually relevant, expect 2–4 weeks of community participation before any benefit accrues. Promotional posts get downranked by both Reddit moderation and Perplexity's authenticity scoring.

#5.3 Days 61–90: Measurement, attribution, and iteration

Weeks 9–10 — Citation tracking and share-of-voice.

Choose a citation tracking platform based on stage and budget:

• Free / DIY: Re-run the 50 priority queries weekly. Maintain a spreadsheet of citation status, cited URLs, and sentiment. Tedious but functional for early stage.
• SMB ($29–99/mo): Otterly ($29/mo entry tier, 6 engines), Citedify, CiteMetrix ($79/mo, ModelScore™ composite metric, 10+ engines)^[47][48].
• Mid-market (~$99–500/mo): Peec.ai (€89/mo, ChatGPT/Perplexity/AIO/Claude, prompt fanout API)^[47], Profound starter ($99/mo), Athena ($295/mo core + flexible credits, 8 LLMs, integrations with Shopify/Webflow/WordPress/GA)^[47][49].
• Enterprise (~$500–15K/mo): Profound enterprise ($499–custom; deepest analytics, REST API, SSO/SOC2)^[47], Brandlight ($4K–15K/mo)^[48], Scrunch AI.

Set the share-of-voice baseline at week 9. Run weekly. Compare against three named competitors per query.

Week 11 — GA4 and CRM attribution.

• Configure GA4 to capture AI referrers as a distinct channel: chat.openai.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com, you.com. Tag any user-flagged sources with UTMs.
• Tag MQLs in your CRM with their first-touch AI source. Outpace's panel shows AI search visitors converting at 14.2% vs Google organic's 2.8%^[11]. The downstream pipeline contribution justifies the instrumentation work.
• For paid channels, watch the AIO citation premium: brands cited inside an AI Overview earn 91% more paid clicks on the same SERP^[14]. Expect paid search to recover ROI on terms where you successfully break into AIO citation.

Week 12 — Iteration loop and quarterly cadence.

• Build a quarterly refresh cadence on top 20 pages: update statistics, refresh quotations, advance dateModified, add new FAQ entries, prune outdated claims.
• Build a monthly content-velocity baseline: 2–4 new long-form pieces per month minimum, each shipped with full schema, FAQ block, and inverted-pyramid structure.
• Build a continuous earned-media motion: aim for 1–2 tier-1 press mentions per quarter and 4–6 industry/podcast appearances. Brand search volume is the longest-latency lever; start it in week 1, expect signal at month 4–6.
• Re-baseline share-of-voice quarterly. Track citation share by engine, by competitor, by topic cluster.

#5.4 What success looks like at day 90

The teams that execute the playbook above with minimal compromise, in our observed pattern across perea audits, hit four numbers by day 90:

1. Schema-validated pages: 100% of top 50 pages with comprehensive (10+) schema, zero validation errors.
2. Citation-share-of-voice baseline: Established weekly, with at least one engine showing measurable lift (typically Perplexity, given its 2–4 week recrawl cycle).
3. AI-referral attribution: GA4 tagging live, AI-source MQLs flowing into CRM, first AI-attributed pipeline events captured.
4. Compounding asset count: 20+ rewritten pages in production, 1+ Wikidata/Wikipedia entry shipped, 2+ active Reddit presences, 8+ tier-1 author profiles linked into Person schema.

The teams that compromise — defer schema work, cut the rewrites to 5 pages, postpone tracking instrumentation — see the compounding curve break. The 90-day rhythm matters because the engines' freshness curves run on 30–90 day windows. A team that ships in days 1–60 and waits sees their first month of work decay before measurement begins.

#Part VI — Measurement and Tooling

#6.1 The metric stack

The metrics that matter for GEO operate at three layers.

Layer 1 — Citation share. Per-engine, per-query share of voice. Did your brand appear in the answer? At what position? Was a URL cited? Which URL? What sentiment surrounded the mention? Citation share is the leading indicator. Profound, Otterly, AthenaHQ, Peec.ai and CiteMetrix all measure this; the differences are in engine coverage, prompt volume, sentiment quality, and competitor benchmarking^[47][48].

Layer 2 — Referral traffic and conversion quality. AI-source sessions in GA4. Conversion rate by source (expect 4.4–14.2× organic baseline)^[11][12]. Pipeline by source from CRM tagging.

Layer 3 — Compounding asset count. Schema-validated pages, FAQPage entries shipped, Wikidata/Wikipedia entries live, Reddit and tier-1 press mentions earned, llms.txt entries published. These are the inputs that produce Layer 1 outcomes 30–90 days later.

The mistake teams make is measuring only Layer 2 (referrals) and panicking when the volume is small in early months. AI referrals in absolute terms remain modest for most B2B brands as of mid-2026 — single-digit percent of total traffic — but the conversion quality means even small volumes carry disproportionate revenue. The right operating framing is to measure all three layers and weight them by latency: Layer 3 (this week), Layer 1 (this month), Layer 2 (this quarter).

#6.2 The tracking platform landscape

Five platforms dominate the citation-tracking market in mid-2026^[47][48][49]:

None of the platforms close the attribution loop to revenue. Every tool tracks citations; none link them automatically to MQLs, SQLs or pipeline. UTM tagging + CRM tagging + manual attribution remains the industry default. The teams that win are the ones that build internal attribution discipline alongside the platform investment.

For a perea-shape B2B SaaS at $1–5M ARR, the recommended starting stack is Peec.ai or Athena ($89–295/mo) + GA4 + HubSpot/Salesforce CRM tagging + a quarterly manual share-of-voice run on 50 priority queries. For a sub-$1M startup, free-tier DIY tracking + a shared spreadsheet works. For enterprise at $25M+ ARR, Profound enterprise + Athena for content optimization + an internal data engineer to wire it all into the warehouse.

#6.3 What the dashboards should display

The minimum dashboard for an operating GEO program shows seven panels:

1. Share of voice by engine, weekly. % of priority queries citing your brand on each of ChatGPT / Perplexity / AIO / Gemini / Claude.
2. Share of voice vs. top 3 competitors. Same view, side-by-side, by topic cluster.
3. Most-cited URLs. Which of your pages are AI engines actually citing? Sort by citation count over the last 30 days.
4. Citation sentiment. Positive / neutral / negative breakdown of how your brand is described in AI answers.
5. AI-referral sessions in GA4. AI-source traffic over time, with conversion rate overlay.
6. AI-attributed pipeline. MQLs and SQLs tagged with AI first-touch, segmented by engine.
7. Compounding-asset score. Internal metric tracking schema completeness, Wikidata/Wikipedia presence, Reddit footprint, llms.txt freshness, tier-1 press coverage. This is the lead indicator.

A team that watches all seven weekly catches both the early-warning signals (citation share dropping on Perplexity → check if recent updates broke schema) and the compounding signals (Wikidata entry ships → expect ChatGPT citation lift in 8–12 weeks).

#Part VII — What's Next: The Discovery Layer of B2A

#7.1 From AI search to agentic procurement

Through 2026 the dominant GEO use case is human-mediated: a person asks ChatGPT or Perplexity a question, the model emits an answer with citations, the human reads the answer and either clicks through or doesn't. The citation premium operates in that loop.

Through 2027 a second loop becomes load-bearing: agentic procurement. An autonomous agent — Cursor evaluating a vendor for a teammate, an internal LangGraph workflow selecting a tool, an OpenAI Agents SDK shopping flow — reads the same citation graph and takes action without human review. The citation that drove a click in 2026 drives a purchase decision in 2027. The B2A Imperative paper outlined the Foundation Build, Agent Layer and Trust Layer of this stack. The citation graph is the discovery substrate underneath all three.

The implication is that citation work converges with B2A work in 2027. A brand that is invisible in ChatGPT today is invisible to the autonomous agent that ChatGPT spins up tomorrow. A brand whose docs are citation-grade today is the brand the agent buys from. The MCP Server Playbook described how to ship the protocol surface; this paper describes how to ship the discovery surface that points the agent at your protocol surface in the first place.

#7.2 MCP-served content and the convergence

The Model Context Protocol creates a second citation channel that does not yet appear in any of the citation-tracking platforms but will by mid-2027. When an agent connects to an MCP server, the server can expose content as resources or as tool outputs. These exposures become citations in the agent's reasoning chain in the same way a Wikipedia entry becomes a citation in ChatGPT's answer.

The MCP Hive registry currently lists ~9,400 servers; AI Herald and DigitalApplied estimate the broader ecosystem at ~19,000 servers as of Q1 2026. Of those, fewer than 5% expose content resources at all — most are pure tool servers. The teams that ship a content-bearing MCP server alongside their citation-grade website capture both the agent-mediated citation channel (this paper) and the agent-mediated commerce channel (the MCP Server Playbook). Both compound into the same brand authority asset.

The forward-looking reference architecture is therefore:

• Discovery layer (this paper): citation-grade content + schema + entity graph + llms.txt.
• Protocol layer (MCP Playbook): OAuth 2.1 MCP server with well-described tools and content resources.
• Trust layer (Trust Layer Deep Dive, forthcoming): AP2 mandates, signed receipts, agent identity, audit trails.
• Commerce layer (Subscription Paradox, forthcoming): per-call billing, x402 settlement, agent-native pricing.

A brand that ships all four is a fully agentic-readiness Foundation Build. A brand that ships only the protocol layer without the discovery layer ships a server that no agent finds. A brand that ships only the discovery layer without the protocol layer captures attention it cannot convert. The four layers compound; missing any one breaks the chain.

#7.3 Predictions and the 2027 horizon

Forward calls are dangerous in a category this volatile, but the data converges enough on a few horizons to commit:

1. Zero-click crosses 75% by Q4 2027. AIO continues its expansion into commercial and transactional queries (already 33% of commercial queries by December 2025^[11]). Gemini 3 hardens AI Overviews quality. The remaining click pool concentrates on truly transactional queries.
2. Citation-tracking consolidates. Profound, Athena, Peec, Otterly and CiteMetrix consolidate to 2–3 platforms by mid-2027. Profound likely acquires or merges with Peec; Athena consolidates the SMB layer; one of CiteMetrix or a new entrant becomes the Wirecutter-grade composite-metric default.
3. llms.txt gets official support. Anthropic formalizes llms.txt support before end-of-2026; Perplexity follows. Google holds out longest. Adoption crosses 25% of top-100K domains by Q2 2027.
4. MCP citation tracking emerges. A new platform category — MCP citation tracking — emerges by Q2 2027 as enterprises start asking which of their MCP server resources are being cited inside agent reasoning chains. Profound or a new entrant ships first.
5. Brand search volume becomes the dominant lever. Domain authority's correlation with AI citation has fallen from r=0.43 to r=0.18 in two years^[15]. By 2027 the correlation approaches r=0.10. Brand search volume rises to r=0.40+. The brands with virality and earned media compound faster than brands with technical SEO discipline alone.
6. Citation receipts and verifiable mentions. The trust layer of B2A starts demanding cryptographic provenance for citations — proof that the model actually cited your content rather than hallucinated the attribution. AP2-style signed receipts for citations emerge as a primitive. The first standards drafts ship in 2027.

The brands that operate as if this convergence is already happening — perea.ai itself, the Stripe and Anthropic and Vercel of the world — are the ones that compound through the transition. The brands that wait for it to be obvious will be paying retail for category positions that should have been wholesale.

#Closing

The B2A Imperative paper argued that 2026 is the year SaaS founders ship a Foundation Build for the agent economy. The MCP Server Playbook argued that the protocol layer is the most underweighted piece of that Foundation Build. This paper closes the loop: the discovery layer underneath the protocol layer is the citation graph, and the citation graph is built one schema-validated, answer-first, cited-quoted-quantified, entity-graphed page at a time.

The teams that operate the 90-day playbook in this paper — and run it again every quarter — build the only asset in digital that pays compound interest into model weights. Models that train on your citation share in 2026 still cite you in 2030. The model retraining cycle is the longest amortization window in modern marketing infrastructure, and it has been operating for five years before most brands started measuring it.

Citations are the new clicks. The citation economy is open. The window in which it stays asymmetric — when fewer than 10% of marketing sites have shipped llms.txt, fewer than 20% have comprehensive schema, fewer than 5% have Wikidata entries — is the window worth racing through. By 2028 these will be table stakes alongside everyone else's. In 2026 they are still leverage.

Ship the playbook. Measure the share of voice. Compound the asset.

— perea.ai Research, May 2026

#References

This paper is published by perea.ai Research under CC BY 4.0. Cite as: Perea, D. (2026). "GEO/AEO 2026: The Citation Economy and the Discovery Layer of B2A." perea.ai Research. https://perea.ai/research/geo-2026.