The 0 to 100K SEO Playbook for a Post-Leak, Post-LLM Web

The single most important update to SEO thinking in the last 24 months is that the 2024 Google Content Warehouse leak plus Pandu Nayak's DOJ testimony transformed ranking from a black box into something close to a white box, and most of what Google publicly said for a decade turned out to be misleading. Clicks (NavBoost), site-level authority (siteAuthority / Q*), and branded navigational demand now sit visibly at the top of the stack, above classical on-page and link signals. For a founder going from 0 to 100K monthly organic visits in 2026, this changes both the sequence of moves and the metrics that matter. At the same time, AI Overviews now trigger on ~48% of tracked queries (BrightEdge, Feb 2026), overall Google-referred traffic is down roughly 33% year-over-year (Press Gazette 2025), and the new game, Generative Engine Optimization, shares a foundation with SEO but diverges in chunking, citation selection, and how authority is measured.

This report synthesizes primary research: Mike King's analysis of the leak at iPullRank, Rand Fishkin's SparkToro writeup, the DOJ trial transcripts (Pandu Nayak, HJ Kim), the Princeton/Georgia Tech GEO paper (Aggarwal et al., KDD '24), Google's current Search Quality Rater Guidelines (Sept 11, 2025), Kevin Indig's Growth Memo frameworks, Koray Tuğberk GÜBÜR's topical authority system, Cyrus Shepard's 23M-link internal linking study, Lily Ray's HCU cohort tracking, Glenn Gabe's recovery analyses, and documented playbooks from Ahrefs, Zapier, Webflow, Nomadlist, 360Learning, and the Grow and Convert / Siege Media / Animalz / Backlinko schools.

Part 1: Core SEO and the documented 0 to 100K playbook

The composite timeline, month by month

There is no single 0-to-100K case study that is simultaneously honest, recent, and documented with monthly granularity. What exists is a composite you can reconstruct from Ahrefs (which has transparently published its trajectory from ~15K to 700K+ monthly Google visits under Tim Soulo), Eli Schwartz's SurveyMonkey/Shutterstock work, Kevin Indig's advisory work at Ramp, 360Learning's Animalz-documented climb from 0 to 75,877 monthly pageviews in two years, and Webflow/Zapier/Nomadlist programmatic scale-ups. The shape is remarkably consistent.

Months 0 to 6: escape the sandbox before you chase rankings. The leak confirmed hostAge, a classifier Google's docs describe as "used specifically to sandbox fresh spam in serving time", so new domains are algorithmically suppressed for roughly six to twelve months regardless of content quality. The winning move here is not publishing harder; it is generating branded navigational demand through off-Google channels (podcasts, X, LinkedIn, HN, Product Hunt, newsletters, communities) because those branded queries feed NavBoost's click signals and help age a site out of the sandbox. In parallel, lay the entity foundation: a canonical /about page with Organization schema, Wikidata QID, consistent NAP across LinkedIn/Crunchbase/G2, author pages with Person schema and sameAs links. Jason Barnard calls this the "Entity Home" pattern, and the leak's explicit author and isAuthor attributes confirm it's algorithmic, not rater-guideline theatre. Publish cadence during this window should be low-volume, high-originality, Ahrefs' rule of "only write about topics with organic traffic potential" scored 2 to 3 on their Business Potential Score, plus Grow and Convert's Pain Point SEO bias toward bottom-of-funnel keywords (category, "X vs Y," alternatives, JTBD). Avoid scaled content at all costs during this window, the Jan 2025 QRG update instructs raters to assign "Lowest" quality to low-effort AI content, and chardScoreVariance in the leak suggests Google measures content-quality consistency site-wide, so a single batch of thin posts can poison the whole domain.

Months 6 to 12: the topical map and first breakthrough. This is where topical authority gets built, not through volume but through entity coverage. Koray Tuğberk GÜBÜR's formula, Topical Authority = Topical Coverage + Historical Data + Cost of Retrieval, requires mapping all semantic sub-entities of your central topic (macro + micro semantics: attributes, predicates, meronyms, hyponyms) before you're eligible to "graduate" into authority state. In practice this means: pick one tight vertical, identify ~40 to 80 sub-entities via Wikipedia blue-links and Google NLP API salience on top-ranking pages, build a matrix of sub-entity × format (definition, how-to, comparison, listicle, calculator), and fill the matrix before chasing adjacent topics. 360Learning's Animalz-documented playbook used a 3.5K-word hub guide with 40 inbound internal links and 76 outbound internal links as the anchor, rank #4 for "employee training" (4,600 MSV) came from cluster coverage, not the hub itself. First real breakthrough usually hits around month 8 to 10 with a single piece of original research or a calculator that earns fresh-tier links. Animalz, Ahrefs, and Siege Media all converge on this: manual link outreach has a negative ROI relative to passive link-generating assets (surveys, original data, interactive tools). Ross Hudgens stopped manual outreach for 90% of Siege clients after sending 1M+ emails and measuring 18-month cost-per-link.

Months 12 to 18: internal linking as the growth unlock. Cyrus Shepard's 23M internal link study at Zyppy found that 53% of URLs on most sites had ≤3 internal links pointing at them, that traffic grows with internal links up to ~40 to 45 (after which it reverses), and that pages with at least one exact-match anchor get roughly 5× the search traffic of those without. This is the single highest-leverage intervention at this stage and almost nobody does it systematically. Kevin Indig's TIPR (True Internal PageRank) method is the operational recipe: crawl the site, score each URL by internal PageRank, CheiRank (outbound equity), external referring domains, and log-file crawl rate, then take equity from the rich pages and redistribute to your "Money Maker Pages" (pricing, product, solutions, and commercial-intent articles) via contextual in-body links with varied query-aligned anchors. Slawski's Reasonable Surfer patent means sidebar and footer links pass a fraction of what contextual body links pass, so the internal-linking audit should focus on adding body links, not nav. Indig's centralized vs decentralized framework matters: a SaaS is almost always centralized (few conversion targets), so concentrate equity rather than spreading it.

Months 18 to 24: compounding, programmatic, and refresh. The sites that actually hit 100K at this stage tend to have layered in one of two things: programmatic SEO at the bottom of the funnel (Zapier's three-tier app / integration / workflow pages model, Webflow's template library, Nomadlist's filter-combination URLs, DelightChat's 300 "best Shopify apps" pages that went from 600 to 240K monthly impressions in 90 days), always built on a genuine data moat, not scraped content; or disciplined content refresh using Animalz's Revive framework, where refreshing decay-hit articles typically recovers 50 to 90% of lost traffic within 3 to 6 months. Animalz's benchmark: median SaaS blog gets 16,969 monthly pageviews, and 80% of blog traffic comes from 16% of blogs, which is another way of saying that, at this stage, pruning and consolidating low-performers and pushing internal equity to the winners produces more lift than new publishing.

Topical authority, information gain, and entity SEO in practice

Topical authority is not a buzzword, it's measured at the site level via attributes the leak exposed: siteFocusScore (how tightly a site stays on one topic), siteRadius (how far individual page embeddings deviate from the site centroid), site2vecEmbedding (the sitewide vector), and chardScores/chardScoreVariance (predicted page quality plus the variance across the site). The operational implication is severe: every off-topic post mechanically widens your siteRadius and increases your chardScore variance, which degrades rankings on the pages you care about. This is why HouseFresh-style affiliate expansions and big-publisher topical creep get punished by HCU. For a SaaS, this means your blog should look more like a focused textbook than a general interest magazine.

Information Gain (US Patent 11,379,553 B2, Carbune & Gonnet 2022, continuation granted June 2024) defines a score that represents "additional information that is included in the document beyond information contained in documents that were previously viewed by the user." Bill Slawski (SEO by the Sea) and Roger Montti read the patent narrowly as applying to automated-assistant follow-up queries; Mike King, Dixon Jones, and Bernard Huang (Clearscope) read it as applying to general ranking. The practical instruction is the same either way: Skyscraper in its original form is dead. Longer-than-competitors no longer wins. The ranking differentiator is what the top 10 don't say, original research, proprietary data, first-person experience, expert commentary, unique media. Aira's 2024 State of Link Building report showed only 6.2% of experts still rate Skyscraper as effective.

Entities are operationalized through Barnard's Kalicube process: one canonical Entity Home with Organization/Person schema + sameAs pointing to Wikidata, LinkedIn, Crunchbase, G2, X, GitHub, YouTube; corroboration across all those sources with consistent descriptions word-for-word (the LLMs in Part 2 reward this even more than Google does); Wikidata entry before Wikipedia (Wikidata is more tractable); consistent "associated entities", the same co-founders, products, investors, partners referenced across sources so Google's disambiguation reinforces one graph. The leak confirmed this is algorithmic, Dixon Jones noted the leaked docs reference "entities" hundreds of times versus "keywords" only 30 to 40 times. Barnard's direct framing: "Google is a child that really wants to understand, and our job is to educate it."

What the 2024 Google leak actually revealed

Between the May 2024 Content Warehouse API leak (2,596 modules, 14,014 attributes, confirmed authentic by Google) and the September 2023 DOJ antitrust testimony from Pandu Nayak, several decade-old SEO debates were settled definitively.

NavBoost is the dominant re-ranker, described by Nayak as "one of Google's strongest ranking signals" and by the leak as a click-based system storing goodClicks, badClicks, lastLongestClicks (effectively dwell time; the last-clicked result of a session signals query satisfaction), unicornClicks (elite/employee users), and unsquashedClicks, all sliced by country and device, rolling 13 months, with "squashing" functions to prevent click-bombing. Nayak in testimony: "Navboost is not a machine learning system. It's just a big table." This directly contradicts Gary Illyes' 2016 Reddit comment that "Dwell time, CTR, whatever Fishkin's new theory is, those are generally made up crap." Glue is NavBoost extended to all universal SERP features (knowledge panels, PAA, videos).

siteAuthority exists, stored in CompressedQualitySignals and applied in Qstar, the umbrella quality score Nayak referenced in DOJ testimony as playing an "extremely important role" in ranking and as largely query-independent. This ends the decade-long debate where Illyes publicly said "we don't really have 'overall domain authority'" and Mueller said "we don't have website authority score."

Other specific contradictions closed by the leak:

• hostAge: confirmed sandbox, described as "used specifically to sandbox fresh spam in serving time." Mueller had publicly said there was no sandbox.

• chromeInTotal and chrome_trans_clicks: Chrome browsing data flows into ranking. Cutts and Mueller had denied this for years.

• author and isAuthor: author is a stored, indexed attribute, "mainly developed and tuned for news articles… but also populated for other content."

• smallPersonalSite: explicit classifier for small independent blogs (boost or demote unknown).

• exactMatchDomainDemotion, productReviewPDemoteSite, anchorMismatchDemotion, babyPandaDemotion: named demotions, not rumors.

• OriginalContentScore: short content is scored for originality separately, "thin" ≠ always short; original short content can outrank long rehashed content.

• lastSignificantUpdate: date flipping without real content change stopped working around 2023. Google measures whether the change is meaningful.

Topicality = T* = Anchors + Body + Clicks, the "ABC signals" revealed in HJ Kim's February 2025 DOJ interview (Trial Exhibit PXR0356). Above that sits Q* (quality/trust, largely query-independent), P* (popularity, fed by Chrome data), plus NavBoost and the deep-learning systems RankEmbed / RankEmbedBERT (retrieval trained on clicks and queries) and DeepRank (BERT-based). Traditional PageRank still exists in seven variants including the deprecated pageRank_NS ("nearest seeds").

The Panda/HCU escape formula, per King's reading of Google patent US8682892B1 combined with leak attributes: Modification Factor = Independent Links / Reference Queries. In other words, you want more successful clicks across a broader set of queries plus more link diversity. This is also the HCU recovery formula, because HCU was absorbed into core ranking in March 2024, its signals share architecture with Panda.

Twiddlers, re-rankers that run after Ascorer's primary scoring, let Google apply category constraints (e.g., "only allow three blog posts in this SERP"), which means ranking can be a lost cause based on your page format and SERP-feature fit. Named Twiddlers in the leak include NavBoost, QualityBoost, RealTimeBoost, WebImageBoost, and FreshnessTwiddler.

Title tags are just candidates. The Goldmine title system scores candidates from sourceTitleTag, sourceHeadingTag, sourceOnsiteAnchor, sourceOffdomainAnchor, and body text. SnippetBrain rewrites the displayed title more than 76% of the time. Optimize for titlematchScore (query match) but know the H1 and internal anchor text contribute.

Rand Fishkin's one-line takeaway is the strategic north star: "Build a notable, popular, well-recognized brand in your space, outside of Google search. For most small and medium businesses and newer creators/publishers, SEO is likely to show poor returns until you've established credibility, navigational demand, and a strong reputation among a sizable audience."

Internal linking architecture, the actual mechanics

The "add internal links" advice hides three distinct mechanics. First, the Reasonable Surfer patent (Slawski) confirms links do not pass equal PageRank, weight is estimated by the probability a user clicks, derived from anchor text, position, font size, and surrounding context. Footer and boilerplate links pass near zero; in-body contextual links pass the most. Second, block-level PageRank groups links by region (nav, sidebar, main content, footer); different blocks carry different weights. Third, click-distance and "propagation of relevance" weight pages that are shallower in click depth from the homepage as more important.

The operational technique for SaaS:

1. Run Indig's TIPR, crawl, score each URL by internal PR + CheiRank + backlinks + crawl rate, rank and average, then redistribute equity to money pages via contextual in-body links.

2. Use Shepard's sweet spot, aim for ~10 varied internal links per important URL, don't exceed ~40 to 45, ensure at least one anchor contains the exact-match target keyword, vary the other anchors naturally.

3. Treat internal anchors as your clearest topical declaration to Google. If a cluster of blog posts about "subscription billing" doesn't link to your /features/subscription-billing page with varied anchors ("how we handle subscription billing", "our subscription billing engine", "SaaS subscription pricing tools"), that page's topical authority for that query is being left on the table.

E-E-A-T, HCU, and the recovery problem

The Search Quality Rater Guidelines at 182 pages (Sept 11, 2025) are still the clearest articulation of what Google's classifiers are trained to prefer. December 2022 added Experience to E-A-T; January 2025 formalized generative AI guidance and instructed raters to assign "Lowest" quality to scaled low-effort content regardless of whether humans or AI produced it; September 2025 expanded YMYL to include elections/public institutions and added AI Overview rating rubrics.

Marie Haynes' operational reading is still the sharpest: "There is no E-A-T metric or signal to track. But if a site is negatively affected at the time of a core quality update, there is likely an E-A-T issue… if we can make enough improvements in E-A-T, the real benefit comes with the next core update." Her recovery levers: "braggy" author credentials, third-party signals (what others say about you), and removing low-quality content sitewide because the classifier operates at the site level.

The HCU carnage is sobering. Lily Ray's tracking of the ~400 hardest-hit sites through March 2024 found zero meaningful recoveries. Glenn Gabe's cohort by late 2024 showed only 22% with a ≥20% lift, most nowhere near pre-HCU traffic. HouseFresh lost 91 to 95% of traffic from September 2023, and only announced recovery exceeding pre-HCU levels in October 2025, attributed to brand building (YouTube, podcast appearances, collaborations) rather than technical SEO. Moz's Tom Capper found HCU-affected sites averaged Brand Authority 37 versus 50 to 52 for unaffected sites. The June 2025 core update was the first significant recovery window; the August 2024 core gave some "signs of life" to about 18% of hit sites in Gabe's dataset.

The crucial lesson: HCU/core classifiers operate site-wide. An otherwise healthy site can be dragged down by a cluster of low-quality posts, old affiliate roundups, or a content bet on an off-topic vertical. Content pruning, deleting or 301'ing thin/low-traffic pages, consolidating overlapping articles, merging cannibalized clusters, moves the needle because it lowers chardScoreVariance and tightens siteFocusScore. Publishing more is less important than removing the dragging weight.

Link building in 2025 to 2026

Manual outreach has collapsed as a primary tactic. HARO shut down (Cision's Connectively closed December 2024; Featured.com acquired the HARO brand in April 2025 and relaunched it free). The tactics that remain viable, ordered roughly by ROI:

The single best path is passive link earning through data-driven, category-defining assets: original surveys, proprietary benchmarks, interactive tools, calculators, trend reports. Siege Media pivoted to this after measuring that manual outreach roughly doubled time-per-link versus publishing search-volume-driven content of equivalent quality. Ahrefs' entire link profile was built this way. The leak's sourceType attribute confirms that links from high-velocity news-tier pages pass substantially more equity than links from evergreen low-traffic pages, which is why digital PR targeting reactive newsjacking (via Google Trends, Qwoted, Featured, Prowly, #journorequest) is experiencing a renaissance as Mark Rofe and others have documented. Podcast tours still work because they generate both branded search volume (fed into NavBoost) and unlinked brand mentions that LLMs specifically reward (see Part 2). The "statistics pages" tactic still earns links, but Google demonstrably demotes content that fabricates or copies stats, so it only works with genuinely proprietary data.

What doesn't work anymore: pure Skyscraper (oversaturated, low differentiation), mass guest posting (footprint detection via phraseAnchorSpamDays and phraseAnchorSpamCount), classical broken-link outreach at scale, and parasite SEO on big media sites, the November 2024 Site Reputation Abuse policy tightening means first-party oversight no longer exempts content, and Forbes, WSJ, Time, CNN, Fortune have all taken manual penalties.

Technical SEO that actually moves rankings

For sub-100K-URL sites, crawl budget is rarely the issue Barry Adams (Polemic Digital) says most small sites should ignore it. Core Web Vitals correlation with rankings is weak, Ahrefs' and Perficient's studies find top keywords are routinely dominated by sites failing CWV. The real technical levers at 0-to-100K scale:

Rendering integrity, JavaScript-rendered content still misses indexing on ~5 to 10% of pages per Onely's long-running data. If you're on Next.js or similar, verify that Googlebot sees the same DOM as a user (URL Inspection tool + rendered HTML comparison).

Index hygiene, low-value parametric URLs, tag pages, paginated archives, and duplicate variants consume crawl and dilute site-level quality scores. XML sitemaps should contain only canonical, indexable, valuable URLs. Eli Schwartz's "Happy Meal" analogy: duplicate content wastes your daily crawl allotment.

HTTPS, hreflang, and status codes, boring but disproportionately consequential when they break. Soft 404s and 301 chains leak equity and confuse indexing.

Schema, Organization, Person, Article, FAQ, HowTo, Product, Review. FAQPage markup (used carefully, not spammy) increases AI engine extractability as discussed in Part 2, one documented A/B test on a SaaS blog doubled Perplexity citation frequency.

Part 2: How SEO is changing with LLMs

What we actually know about how AI engines pick sources

Google AI Overviews / AI Mode runs retrieval-augmented generation with query fan-out: a single user query is decomposed via Gemini into multiple sub-queries, each executed in parallel, with passages extracted and synthesized. Confirmed by Google's own description: AI Mode "issues multiple related searches across subtopics and data sources." The overlap with classical top-10 organic has been falling sharply. Ahrefs (1.9M citations, July 2025): 76% overlap. BrightEdge's October 2025 analysis: 54%, down from the initial 76% at May 2024 launch. Ahrefs again (4M AIO URLs across 863K keywords, Feb 2026): 38% top-10 overlap. BrightEdge's stricter methodology says only ~17% of AIO-cited sources rank in top 10, with 31.2% coming from positions 11 to 100 and 31% from outside top 100. YouTube is the single most-cited domain in AI Overviews (BrightEdge: AI engines pick YouTube ~200× more than any other video source). Reddit citations surged 450% between March and June 2025 following the Google to Reddit data licensing deal. Google self-references its own properties in ~43% of AIO links.

The cannibalization data is severe. Seer Interactive's study (3,119 queries, 25.1M impressions): organic CTR fell 61% on informational queries when AIO was present, from 1.76% to 0.61%. Lily Ray's Amsive analysis across 700K keywords: −15.49% CTR average, −27% for positions outside top 3, −37% when both AIO and Featured Snippet appear together. Press Gazette's 2025 aggregate: global Google-referred traffic down ~33%, US organic down ~38% YoY. BrightEdge shows AIO triggering grew from ~30% of tracked queries (Feb 2025) to ~48% (Feb 2026). Informational queries fell from 91% of AIO triggers in January 2025 to 57% in October 2025, commercial grew to 18%, transactional to 14%, navigational to >10%, meaning AIO now encroaches on money queries. Google upgraded AIO's default model to Gemini 3 on January 27, 2026.

Perplexity runs a two-step pipeline: document inclusion in the retrieval set, then paragraph selection. Freshness dominates, Growth Marshal's controlled tests show content stamped "updated 2 hours ago" gets cited 38% more than identical content with last-month datelines; updates lift citation by 37% in the first 48 hours, flattening to a 14% edge at two weeks. FAQ schema (FAQPage JSON-LD) markedly boosts extraction; one dev-SaaS case study doubled citation frequency after adding three FAQs. PerplexityBot pre-crawls, retrieves ~10 candidates, scores on relevance/freshness/authority/extractability, sends top 3 to 4 to Sonar LLM. Citations overlap Google AIO ~82% (Onely, 2026).

ChatGPT / SearchGPT browsing uses Bing's index as the retrieval substrate. Seer found 87% of SearchGPT citations match Bing's top results, so Bing Webmaster Tools, IndexNow submissions, and Bing-specific signals matter more than they did in 2023. Roughly 46% of ChatGPT queries trigger a web search (Semrush, 80M queries analyzed). ZipTie's reverse-engineered scoring estimates domain authority ~40%, content quality ~35%, platform trust ~25%. Aleyda Solis cites Profound research showing only 12% overlap between Google classical SERPs and ChatGPT answer sources.

Claude now has GA web search on the Anthropic API (since September 10, 2025), uses Brave Search per BrightEdge's analysis as the retrieval layer, and supports allowed_domains/blocked_domains plus a Citations API that chunks documents into sentences and returns structured cited_text blocks. Endex reported source-hallucination rates dropping from 10% to 0% using the Citations API; Anthropic's internal eval showed +15% recall accuracy vs custom implementations.

What the Princeton/Georgia Tech GEO paper actually found

Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, and Deshpande published "GEO: Generative Engine Optimization" at KDD '24 (arXiv:2311.09735). They built GEO-bench (10K queries, 80% informational) and tested nine optimization strategies against Perplexity.ai: authoritative tone, statistics addition, keyword stuffing, cite sources, quotation addition, easy-to-understand, fluency optimization, unique words, technical terms.

The three tactics that actually win: Cite Sources, Quotation Addition, and Statistics Addition, 30 to 40% improvement on Position-Adjusted Word Count, with peaks at 41% (PAWC) and 28% (Subjective Impression). Keyword Stuffing performed worse than baseline on Perplexity. Effectiveness is domain-dependent: citations help factual queries most, statistics dominate on "Law & Government" and opinion-heavy topics, persuasive tone helps historical and debate content.

The single concrete GEO instruction that falls out of this: write content that quotes authorities, includes named sources with URLs, and embeds specific numeric claims tied to a cited study. The reason this works is mechanical, RAG systems chunk content and embed passages; chunks dense with named entities and numeric claims carry higher information density per token and match retrieval queries better.

Chunk-level retrieval, structure, and entities

The biggest structural insight for content design in the LLM era comes from how RAG actually works: chunking → embedding → (re)ranking → generation. The unit of retrieval is a text chunk of roughly 134 to 167 words (ZipTie's extraction-unit analysis for AIO), not a page. This has five concrete consequences:

First, the inverted pyramid comes back. Direct answers in the first sentence of each section, definitions upfront, conclusions before reasoning. Buried leads die, both in human attention terms and in chunking extractability.

Second, passage-level entity density matters more than keyword density. Mike King ("Relevance Engineering" at iPullRank) and Koray's macro/micro semantics converge here: the chunk must contain the central entity, its key predicates, and named sub-entities in close proximity, so the embedding vector lands near the expected retrieval query in vector space. Mike King's sharp observation: "Google moved beyond the lexical model of search 10 years ago and all of our tools are still just counting the presence and distribution rates of words."

Third, short, declarative, parseable structure beats prose density. Cindy Krum's "Fraggles" framing, individual extractable passages, is the right unit of optimization.

Fourth, brand mentions and consistent entity descriptions across the web are the currency of LLM visibility the way backlinks were the currency of SEO. LLMs learn entities via co-occurrence in training data. A canonical 2 to 3 sentence company description replicated verbatim across Wikidata, LinkedIn, Crunchbase, G2, Reddit, YouTube descriptions, podcast show notes, and guest bylines strengthens the entity signal in training data and at retrieval.

Fifth, Reddit, YouTube, and expert forums are now distribution channels for LLM visibility, not just brand awareness. The 450% Reddit citation growth in Google AIO and the dominance of YouTube in AI engine citations make community presence a durable moat in a way it wasn't in 2020.

The llms.txt debate, probably noise, possibly an option

Jeremy Howard (Answer.ai, fast.ai) proposed /llms.txt on September 3, 2024, as a markdown site index for LLM consumption. John Mueller publicly compared it to the deprecated meta keywords tag and said "no AI system currently uses llms.txt" (June 2025). Gary Illyes confirmed Google doesn't plan to support it. Ahrefs' April 2025 server log analysis found no major AI services checking for the file. One Reddit user managing 20,000+ domains reported only niche bots (BuiltWith) downloading it.

Counter-evidence exists but is thin: Profound data reportedly shows Microsoft and OpenAI crawlers requesting llms.txt/llms-full.txt files. Google briefly added llms.txt to its own docs on December 3, 2025 before removing it within hours. Mintlify (Anthropic, Windsurf, Bolt.new), Yoast, and Webflow now auto-generate it. Google included llms.txt in its Agent2Agent (A2A) protocol.

Verdict for a SaaS founder: generate it auto-magically via your docs platform if possible, don't invest labor. It's cheap to have, zero downside, and if adoption tips it's free upside. Don't build content strategy around it.

What's the same, what's different

The same: quality content, authority (entity strength), technical crawlability, original information gain, and, critically, classical SEO rankings still drive AI citation. The 76% → 54% → 38% declining top-10 overlap in AIO over 18 months looks alarming, but it still means ranking in Google's top 10 gives you a 4-5× higher probability of being cited than not ranking. The same Google index feeds AIO, and Bing's index feeds ChatGPT. Ranking in Google and Bing remains the single highest-leverage AI visibility lever. Lily Ray's warning deserves emphasis: she has documented ~30 case studies where sites bet on aggressive GEO content production, saw short spikes, then collapsed in subsequent core updates because the classical ranking foundation AI retrieval depends on was undermined. GEO without SEO is self-destructive.

What's different:

The long-tail exact-match keyword strategy is the biggest tactical casualty. Long-tail queries are exactly what AI search handles natively via fan-out; traffic that used to flow to a 500-word answer page now gets absorbed into an AIO panel. Informational top-of-funnel content monetization is structurally broken.

What matters now: new metrics and moats

AI citation share of voice in ChatGPT, Perplexity, AIO, Claude for your category queries. Tools: Profound, Peec AI, AthenaHQ, Otterly.ai, Ahrefs Brand Radar, Semrush AI tracking, SE Ranking AI. Track weekly.

Branded search volume, fed into NavBoost and the Q* quality stack. This is the deepest moat. Rand Fishkin's call to "build a notable, popular, well-recognized brand in your space, outside of Google search" is load-bearing advice, not a platitude.

Entity coherence, does the web tell a consistent story about your brand? Audit via Google NLP API salience, InLinks/Diffbot, and manual consistency checks across Wikidata, LinkedIn, Crunchbase, G2, review sites, and your own content. Mark Williams-Cook's Google exploit discovery (paid $13,337 bug bounty after analyzing 2TB across 90M queries) revealed a Site Quality Score calculated at subdomain level on a 0 to 1 scale, with <0.4 disqualifying a site from Featured Snippets and PAA, inputs including anchor text relevance, user interactions, and brand visibility.

Community and creator footprint, Reddit mentions, YouTube coverage, podcast appearances, HN discussions. These now compound into LLM training data and AIO citations simultaneously.

Strategic implications for any SEO-led SaaS

Yes, still do traditional SEO, because AI engines draw from the same indexes. But change content shape:

Shift the portfolio from long-form TOFU to three content types that survive: (1) Bottom-of-funnel commercial content, category, "X vs Y," alternatives, use-case, pricing, where AI engines still send clicks because they can't complete the transaction; (2) Original research, data, and benchmarks, the rare content types that get cited as authoritative, earning both links and LLM mentions; (3) Product-led programmatic pages with real data moats (Zapier's integration pages, Webflow's templates, Nomadlist's filter-combinations) where the page IS the product.

Rewrite existing posts to be chunk-extractable: inverted pyramid, direct answer first, definitions up front, named entities and numeric claims densely packed in opening paragraphs, FAQ schema where intent supports it. Update dates only when content meaningfully changes, lastSignificantUpdate sees through flips.

Invest in brand as the real compounding asset: podcast tours, Reddit presence, YouTube channel (even a modest one, remember YouTube is the single most-cited AIO source), guest bylines on authoritative outlets, consistent entity descriptions replicated verbatim across Wikidata, LinkedIn, Crunchbase, G2, review sites. Build your author/founder entity, the leak's author and isAuthor attributes mean this is algorithmic, and LLMs specifically reward named experts in training data.

Track both stacks: classical SEO metrics (rankings, organic clicks, GSC query data, conversion rates) alongside AI citation metrics (share of voice in ChatGPT/Perplexity/AIO/Claude, branded search volume, unlinked brand mentions, Reddit/YouTube visibility). The gap between the two is where you'll find what to fix.

The takeaway

The 0-to-100K path in 2026 is neither purely classical SEO nor purely GEO, it's the integration of three systems that were once distinct. You rank in Google and Bing because classical signals (topical authority, original content, internal linking architecture, link diversity) still work. You survive algorithm updates because the leak-confirmed site-level classifiers (siteFocusScore, siteRadius, chardScoreVariance) reward disciplined topical focus and punish content sprawl. You get cited in AI engines because your chunks are extractable, your entity graph is coherent across the web, and your brand generates navigational demand that feeds back into NavBoost and therefore into classical rankings too. The founders who understand that NavBoost, siteAuthority, and brand are the same system viewed from three angles, and who build accordingly from month one, are the ones who still compound through 2026 and beyond.

Primary sources to read in full: Mike King's "Secrets from the Algorithm" (ipullrank.com/google-algo-leak); Rand Fishkin's SparkToro leak analysis (sparktoro.com); DOJ testimony coverage at Search Engine Land (searchengineland.com/google-abc-ranking-signals-455360, /how-google-search-ranking-works-pandu-nayak-435395); Aggarwal et al. GEO paper (arxiv.org/abs/2311.09735); Google Search Quality Rater Guidelines (guidelines.raterhub.com/searchqualityevaluatorguidelines.pdf); Kevin Indig's Growth Memo (kevin-indig.com, growth-memo.com); Cyrus Shepard's Zyppy internal linking study (zyppy.com/seo/seo-study); Animalz's content decay framework (animalz.co); Eli Schwartz's Product-Led SEO (productledseo.com); Lily Ray's HCU tracking (LinkedIn, Amsive); Glenn Gabe's update analyses (gsqi.com); Koray Tuğberk GÜBÜR's topical authority system (holisticseo.digital); Jason Barnard's Kalicube process (kalicube.com); Aleyda Solis's AI search coverage (learningaisearch.com, speakerdeck.com/aleyda); Bill Slawski's SEO by the Sea archive (seobythesea.com); Anthropic Claude web search docs (platform.claude.com/docs); Perplexity API changelog (docs.perplexity.ai/changelog).