Methodology
The AEO/GEO playbook we run on every engagement.
The recipe below is what we apply to every priority page on every engagement, from a $890 Starter audit punch-list to a $17,800 crypto Enterprise programme. The structural rules don't change with budget — only how many pages we get to apply them to.
The recipe at a glance
| Parameter | Value |
|---|---|
| Page recipe | Hero · X-is-Y intro · Quick Facts · H2-question sections · FAQ · CTA |
| Hero rule | ≤3 sentences, human hook, not a duplicate of intro |
| Intro rule | X is Y form: regulator + license types + audience + advantages + timeline + cost |
| Quick Facts | Parameter / Value table, ≥5 rows, mandatory even if minimal |
| H2 rule | Each H2 is a user question; first sentence directly answers it |
| FAQ rule | Direct answer ≤30 words, then optional 2–3 sentence depth |
| Sentence target | 15–20 words avg, active voice ≥70% |
| Schema stack | Article, FAQPage, HowTo, Person, Organization, BreadcrumbList |
Four-layer AI extraction structure
Every page renders as four layers AI systems can extract independently: hook, X-is-Y intro, Quick Facts table, then H2-question sections.
Layer 1 is the hero — two or three sentences that make a human want to keep reading. It is not a duplicate of the intro and it does not try to be SEO copy. The hero is where you sound like a human.
Layer 2 is the intro block — no H2, just a paragraph that names what the page is about in X-is-Y form, plus the regulator if any, license or service types, target audience, advantages, timeline, and cost. AI systems lift this block whole when answering "what is X?" queries.
Layer 3 is the Quick Facts table — Parameter / Value, at least five rows. Mandatory on every priority page. AI systems quote table rows verbatim into answers. If your data isn't tabular elsewhere, this is where to compress it.
Layer 4 is the body of H2-question sections. Each H2 is the user's question in question form — ending in "?". The first sentence directly answers the question in plain language; everything after it is depth. We keep adjacent H2s from conflicting in intent and we never use "this/that/it" when we can name the entity.
Readability gates and human signals
Sentences average 15–20 words. Active voice ≥70%. No paragraph above 4 sentences. Every page carries ≥4 explicit human signals.
AI-detection signals are real and the platforms increasingly down-weight content that reads like an LLM wrote it without an editor. We enforce a blocklist of noun, verb, adjective and phrase markers — banned lexicon includes "delve", "tapestry", "landscape", "robust", "navigate", "in a world where", "in today's", "in conclusion", and the rest of the LLM tell-tale set.
On the positive side, every published page must show at least four human signals: at least one em-dash, one or two contractions, one sentence starting with But / And / So, mixed paragraph lengths, one concrete number or date, one explicit opinion or provocative claim, and Oxford-comma drops in one or two places. Drafts under the bar get rewritten until they're over it.
Schema engineering — what we deploy and why
Article, FAQPage, HowTo, Person, Organization, BreadcrumbList everywhere. Service + Offer on commercial pages. ItemList + Review on comparison pages.
Schema is the structured-data layer that makes a page legible to machines. AI systems lean heavily on schema for citation decisions — a page with proper Article + FAQPage schema will outcite an identical-content page without it. We deploy JSON-LD only (no mixed microdata) and validate every page with the Schema.org validator and Google Rich Results Test before publish.
For named experts we ship schema.org Person with sameAs pointing to verifiable external profiles — LinkedIn, ORCID, professional registries. This is the E-E-A-T trick that crypto and fintech engagements live on: AI systems treat schema-validated Person identities very differently from anonymous bylines.
For commercial pages we ship Service + Offer with priceSpecification — base USD price, monthly unit, availability. This is what makes the pricing on this site machine-readable to ChatGPT when someone asks "how much does AEO cost in 2026". Try it.
Niche-aware multipliers — the four-factor formula
Base × YMYL pressure × lead value × AEO competitiveness × technical complexity, averaged. Crypto runs ×2.0, SaaS ×1.15, local services ×0.65.
The base rates on the pricing page are universal. Real engagements price as base × a niche multiplier that averages four factors: YMYL/regulatory pressure (crypto and fintech high, B2B SaaS medium, local services low), lead value (premium niches with $1,500+ leads carry the AEO chk easily), AEO competitiveness (premium verticals contested, locals not yet), and technical complexity (multi-region multilanguage hits hardest).
Worked example: an EdTech SaaS startup with a $400 lead, regional B2B competition and a headless multi-region site. YMYL = 1.0 (education). Lead = 1.25. Competitiveness = 1.0. Technical = 1.5. Average: 1.19. Growth base of $2,400 × 1.19 = $2,856 / month. We'll show this calculation on the discovery call.
Frequently asked questions
Why does the structural recipe matter for AI citation?
AI systems extract whole compact blocks, not scattered sentences. A page that follows the recipe is extractable by construction; one that doesn't isn't.
Google AI Overviews and Perplexity in particular pick clean blocks (intro paragraph, FAQ Q&A, table row) and quote them verbatim or near-verbatim. Pages with walls of text without H3s, scattered facts across paragraphs, or buried answers force the AI to summarise rather than cite — and summarisation hides your URL.
Why do you ban certain words from generated content?
Because LLM-tells like "delve", "tapestry", "robust", "in conclusion" make AI-detection trivial and erode trust signals.
We maintain a blocklist of about 40 noun, verb, adjective and phrase markers. Every generated page gets a regex check + a Claude editing pass that enforces the bans. We also require ≥4 explicit human signals — em-dashes, contractions, sentences starting with But/And/So, mixed paragraph lengths, one concrete number, one explicit opinion.
What schema do you actually deploy?
Article (with dateModified), FAQPage, HowTo where there's a process, Person for named experts, Organization, BreadcrumbList, and Service for commercial pages.
For ratings/comparison pages we add ItemList + Review where applicable. For local services we add LocalBusiness or its subtype (Attorney, MedicalOrganization, RealEstateAgent). All schema is JSON-LD; we don't mix microdata.
How do you choose what AI crawlers to allow?
Default: allow GPTBot, ClaudeBot, PerplexityBot, Google-Extended, CCBot, Applebot-Extended, anthropic-ai, cohere-ai. Block the noisy ones case by case.
For most B2B sites we allow all major AI crawlers because the goal is being cited. For some YMYL clients with regulatory exposure, we restrict by user-agent and content type; the llms.txt at /llms.txt is always a strict subset of what robots.txt allows.
How long until the structural changes show in AI?
Google AI Overviews: 14–60 days. ChatGPT and Perplexity: 7–30 days for new content; longer for restructured legacy.
New content tends to surface faster because AI systems index it without legacy citation patterns to break. Restructured legacy pages need to outrank their old extractable signals — usually a 30–60 day cycle.
Want this recipe applied to your site?
Starter is the cheapest entry — $890/month gets you a baseline, the technical AEO audit, and a punch-list of 10–15 priority changes. Three months minimum, no implementation.