The shortlist is written before your first touch
Your buyers now ask AI before they ask you, your salespeople, or Google. 73% of B2B buyers use AI tools like ChatGPT and Perplexity in their purchase research, 94% used a large language model somewhere in their most recent buying journey, and a majority now start research in a chatbot more often than in a search engine. When those buyers arrive at your site — if they arrive — they convert at 14.2%, five times the rate of Google organic traffic, because the AI already did the qualifying. The volume is still small — 2–6% of B2B organic traffic — but it is growing more than 40% month over month.
The practice of earning a place in those AI answers has a name: generative engine optimization, or GEO — structuring your content, site, and facts so that AI systems cite you, quote you, and recommend you when a buyer asks a question you should win. You do not need to remember the acronym. You need to internalize the shift it describes: the shortlist used to be assembled on page one of Google, by a human, scanning. It is now assembled inside an answer, by a machine, quoting. The question for a CMO is no longer "do we rank" but "when the machine answers our buyer's question, are we in the answer — and are we described accurately."
This guide is the strategy for getting there: what carries over from SEO, what replaces it, how to re-architect your content and site, what to measure, and who should own it. It is written to be shared — with your content team, your product counterparts, and your CEO.
What survives from SEO — and what breaks
The short answer: the plumbing survives; the strategy breaks. Crawlability, site speed, structured data, and clean information architecture matter as much as ever — AI systems discover the web through crawlers, and several lean on traditional indexes underneath. But the logic of what wins has changed shape. Consider one number: only about 2.1% of pages that rank in Google's top 10 also appear among ChatGPT's citations. Ranking success does not transfer. You can dominate page one and be absent from the answer.
| What won in SEO | What wins in GEO |
|---|---|
| Ranking #1 for the keyword | Being the passage the answer quotes |
| Keyword density and coverage | Question coverage — every buyer question answered somewhere, completely |
| The page as the unit of optimization | The passage as the unit — each section extractable on its own |
| Backlinks as the currency of authority | Entity consistency, citable facts, quotes, and statistics |
| Domain authority | Verifiable expertise — sources, data, named experience |
| Publish, rank, hold | Refresh on a cadence — citations decay sharply after ~3 months |
| One engine to optimize for | Four or more — and they disagree |
That last row deserves a CMO's attention. ChatGPT's share of B2B AI referrals fell from 89% to 63% in about eight months, while Claude's grew from 1.4% to 18.5% and Gemini's quadrupled. Citation volumes for the same brand differ by as much as 615x between platforms, and only 11% of domains are cited by both ChatGPT and Perplexity. Optimizing for one engine is the new single-channel dependency. The strategy has to work across all of them — which is exactly why it must be built on substance rather than tricks.
Semantic Alignment
The strategy that works across every engine is what I call Semantic Alignment: one truth layer, running from your boldest brand claim to your driest technical document, expressed in one consistent vocabulary, anchored in verifiable fact, and connected so a machine can walk from any entry point to the proof.
AI systems do not read your site the way a visitor does. They retrieve fragments from many pages — yours, your documentation's, your reviewers', your competitors' — and assemble a judgment. If your homepage says "workforce intelligence," your product pages say "productivity analytics," and your docs say "activity monitoring," the machine meets three brands, none with critical mass. If your brand claims outrun what your documentation supports, the machine notices the gap — and answer engines penalize what they cannot verify by simply not quoting it.
Semantic Alignment has four disciplines:
1. One vocabulary. The same names for the same things — products, solutions, features, capabilities, problems — on every surface you control, from the homepage to the API docs. Pick the words once; enforce them everywhere.
2. Claims anchored in fact. Every claim carries its evidence, or it does not ship. This is not just honesty — it is the measured mechanics of citation: the Princeton and Allen Institute GEO research found targeted optimization lifts visibility in generative answers by up to 40%, and the strongest levers are quotations (+27.8%), statistics (+25.9%), and cited sources (+24.9%). Machines quote what is quotable.
3. The site as a walkable graph. Problems link to the solutions that address them, solutions to features, features to technical capabilities, capabilities to proof. Not navigation for humans — explicit connective tissue so an AI can connect the dots. Most B2B sites are organized by org chart (products here, resources there, docs on a subdomain nobody links to). Reorganize by meaning.
4. Marketing and product on one source of truth. This is the organizational hard part: the vocabulary and the claims must be co-owned by marketing and product, because documentation is now marketing. The AI reads your docs with the same weight as your homepage — often more, because docs are specific and verifiable.
Answer units: re-architect the content itself
An answer unit is a section of content that gives one complete answer to one specific question, extractable on its own — a reader (human or machine) needs nothing above or below it for the answer to make sense. Pillar pages, blog posts, and service pages need to be rebuilt around them, because AI retrieval is passage-level: the winning pattern is a direct, complete answer in the first ~200 words of a section, with supporting depth after.
Most B2B content is built the opposite way — narrative arcs that develop context before conclusions, sections titled cleverly rather than as questions, answers that depend on the three paragraphs before them. A machine quoting one passage from that page gets a fragment that cannot stand alone, so it quotes someone else.
An illustrative rewrite (invented example). Before: a pillar page section titled "Rethinking visibility" that opens, "As organizations evolve, the nature of oversight has changed dramatically. Where once managers walked the floor…" — three paragraphs of throat-clearing before the point. After: the section becomes "How does workforce analytics work without violating employee privacy?" and opens, "Workforce analytics preserves privacy by collecting activity signals rather than content — which applications and sites are used and when, never keystrokes, screenshots of message text, or personal accounts. Aggregation and role-based access limit who sees individual data…" — a complete answer in the first breath, detail after.
Same information. The first is prose that must be read; the second is an answer that can be quoted. The section titles become the questions buyers actually ask — which is the next discipline. And the meta-point is deliberate: this article is built the way it tells you to build. Every section opens with its answer, and the page you are reading carries the structured data, the entity schema, and the llms.txt this guide prescribes. View source.
Map the journey before the brand
The question inventory — not the keyword list — is the new content plan, and it must start well before anyone types your brand name. Buyers spend most of the journey brandless: naming a problem, comparing approaches, only then listing vendors. Each stage has questions, and every question an engine answers without you is a shortlist you were never on.
Map five stages and write the real questions at each (examples invented for illustration): problem naming ("Why is my team missing deadlines when everyone seems busy?" — the buyer does not yet know what category solves this) · approach comparison ("Workforce analytics vs. project management tools — which fixes capacity problems?" — categories compete before vendors do) · shortlisting ("Best workforce analytics platforms for a 2,000-person company" — the classic GEO moment; the engine assembles the list) · validation ("Does this vendor support EU data residency? What do reviews say about deployment?" — the machine reads your docs, your trust center, and third-party reviews together) · objections and expansion ("How do we roll this out without employee backlash?" — post-shortlist questions that decide deals and renewals).
Two implications. First, coverage: for each question, an answer unit must exist somewhere you control, complete and current. Second, honesty at the comparison stages: engines synthesize your claims against your competitors' documentation and reviewers' experience — comparison content that overclaims gets contradicted in the same answer that cites it. The claims-audit discipline is not compliance; it is how you survive synthesis.
Freshness is architecture, not hygiene
AI citations decay: content older than about three months is cited sharply less, which means recency has to be designed into operations, not left to annual refresh projects. Think of your content portfolio the way a newsroom does: some assets are the weekly signal (data updates, market notes, changelogs, pricing and comparison pages), some are the quarterly review (pillar answers, integration guides), and a few are the annual flagship (the state-of-the-industry report). Every asset gets a cadence and an owner; nothing is published without a refresh date.
If you have read my piece on loop design, you already know the machinery I would use: freshness is a designed loop. Goal: no citable page older than its class's cadence. Worker: the refresh procedure — update the data, re-verify the claims, bump the visible date honestly (engines notice cosmetic date-bumping; the content must actually change). Done-Test: every statistic re-checked against its source; anything unverifiable removed. Escalation: claims that can no longer be verified go to a human owner. Write-back: the query inventory gains whatever new questions the quarter surfaced. Run it on a calendar, not on inspiration.
The technical layer: what to tell your team
The technical floor is concrete, checkable, and mostly a week of work — brief your web team with this list.
Serve real HTML. Most AI crawlers do not execute JavaScript; if your core content renders client-side, machines see a shell. Core pages must be server-rendered or static.
Be fast. AI crawlers time out in roughly one to five seconds; a slow page is an invisible page.
Open the door. robots.txt must admit the AI crawlers you want citing you (GPTBot, ClaudeBot, PerplexityBot, Google-Extended, and peers). Blocking them was a 2023 reflex; for most B2B brands it is self-erasure.
Stay indexed in Bing. ChatGPT's browsing leans on Bing's index — Bing Webmaster Tools is no longer optional.
Declare your entities. Organization and Person schema with sameAs links tie your brand to its profiles so machines connect your presence across the web; per-page JSON-LD (Article, FAQ, HowTo) where it genuinely fits.
Publish llms.txt. A plain-text map of your most important pages for AI agents. Adoption sits near 10% — cheap, and for now an early-mover signal.
Structure for extraction, and make pages citable. Answer-first sections, question-phrased headings, FAQs, definitions, step lists — and the three levers the Princeton research measured: quotes, statistics, and cited sources. Give the machine a number worth quoting and a source it can check.
Tools and tracking: the new stack
The measurement stack is changing shape: rank trackers answer a question nobody is asking anymore. What replaces them, by category:
AI share-of-voice and citation monitoring. Platforms that run your question inventory across engines on a schedule and report whether you appear, how you are described, and who is cited instead. A fast-moving category; representative tools include Profound, Peec, Otterly, Scrunch, AthenaHQ, Trakkr, and LLM Pulse, and the incumbent suites (Semrush, Ahrefs) have shipped AI-visibility modules. Evaluate on engines covered, prompt customization, and citation-level (not just mention-level) reporting.
The free manual baseline. Before buying anything: run your top 20 buyer questions across ChatGPT, Claude, Gemini, and Perplexity monthly. Record who is cited, how you are described, and what is wrong. One spreadsheet, two hours — and it is the eval for everything else. A team that cannot name its top 20 questions is not ready for tooling.
AI referral analytics. Segment AI-source sessions (chatgpt.com, perplexity.ai, claude.ai, gemini referrers) in your analytics and track their conversion separately — the 5x conversion premium is your budget argument.
Crawler-log visibility. Know which AI bots fetch which pages; a page no AI crawler visits will never be cited. And keep Search Console and Bing Webmaster Tools — the indexes underneath feed the engines; classic crawl health remains the floor.
Track progress on four numbers: share-of-answer on your question inventory (per engine), citation count and freshness of cited pages, AI-referral sessions and their conversion, and your scorecard delta quarter over quarter. Report them next to organic search — side by side, so the shift is visible to the whole leadership team.
Who owns this
GEO fails as a side project of the SEO team, because its levers sit in four functions: content (answer units, freshness), product marketing (vocabulary, claims, comparison honesty), documentation (the most-quoted surface you have), and web engineering (the technical floor). Someone has to own the semantic layer across all four — the vocabulary, the claims registry, the question inventory, the cadences.
In practice that owner looks like a product marketing leader with unusual range: close enough to product to keep the vocabulary true, close enough to content to enforce answer units, technical enough to hold the web team to the checklist. The operating rhythm is light — a shared claims-and-vocabulary source of truth, the question inventory reviewed quarterly, the freshness loops on calendars, the four metrics reported monthly. In my last role I led AI governance with IT and Security; the same pattern applies here: named owner, written standard, scheduled review. Governance is what separates a strategy from a wish.
Thirty days
Days 1–5: score yourselves. Download the AI-Readiness Scorecard below — 26 yes/no checks across strategy, content, technical, and operations — and have the team score it honestly. Or let the free geo-audit-skill do it with you: it asks for your actual pages, scores what it can see, and refuses to score what it cannot verify.
Days 6–10: fix the technical floor. robots.txt, server-rendered content, Bing, entity schema, llms.txt — the checklist in section 07 is a week for most teams.
Days 11–25: rewrite your top ten. Take the ten questions you most need to win from the journey map, and rebuild the relevant sections as answer units — question-phrased heading, complete answer first, a citable fact with a source in each.
Days 26–30: stand up the loops. Freshness cadences per content class, the monthly prompt-panel baseline, the four metrics on the leadership dashboard. Then it runs.
The AI-Readiness Scorecard
One page, 26 checks, four blocks, honest bands — from "invisible" to "the answer." Built to be forwarded to your team with two words: score us.
The skill that audits with you
Paste your homepage, a pillar page, and your robots.txt — it scores what it can see, marks what it cannot verify, and hands back a prioritized fix list. No evidence, no score.
Be the answer
I have done this work, not just written about it. At my last company I rebuilt a neglected knowledge base into content that was prescriptive, readable, and structured for AI and search retrieval, and built a customer academy whose content was optimized for GenAI search from day one. And the site you are reading practices every prescription in this guide — the answer units, the entity schema, the llms.txt, the freshness loops. View source; nothing here is theoretical.
The scorecard and the skill are free, like everything else I publish. If you want this capability built in your organization rather than just read about — I build this with teams. Leave your email below, or reach me at daniel@cmoconfessions.com. No gate, no sequence.