Generative Engine Optimization is what happens to a website when the answer-engine becomes the front door. The page is no longer the destination; the citation is. By April 2026 the share of commercial queries that resolve inside an AI answer instead of a blue-link click has crossed 60% in our agency telemetry, and that number is still rising.
Most teams have responded with content tweaks: adding FAQ blocks, hardening schema, lifting an existing post into Markdown. Those moves help. They are not a system. The teams getting cited consistently across ChatGPT, Claude, Perplexity, and Gemini have stopped optimising at the page level and started operating a program.
This post is that program. The framework is three layers wide — technical, editorial, measurement — and 14 controls deep. We use it for every new client engagement at Digital Applied; the rollout diagram at the end of the post is the exact one we hand to clients on day one.
- 01GEO is a three-layer operating system, not a checklist of content edits.The technical layer makes the site agent-readable; the editorial layer makes the content citation-worthy; the measurement layer turns answer-engine performance into a tracked KPI. Skipping any one layer caps program ceiling.
- 02The technical layer is the cheapest leverage and is most often skipped.Markdown rendering, llms.txt, AGENTS.md, structured data, and crawler-routing rules together typically lift citation rate 30-50% in 30 days with zero new content. We have not seen a single program where the technical layer was already in place at engagement start.
- 03Editorial controls beat structural controls on long-running queries.Citation density, opinion strength, and source attribution determine which page an answer-engine reaches for. Pages that read like 'a person who has actually done this' outperform pages that read like 'a brand explaining its product' by a factor we measure in the 2-3× range.
- 04Measurement is what makes the program defensible to a CFO.Without citation rate, answer share, persistence, and position-in-answer reported monthly, GEO is impossible to budget. The measurement layer turns it from a gut-feel program into a line item with a benchmark.
- 05The 14-point rubric is the reviewer's tool that keeps drift out.Score each new asset 0-14 before publish. Below 11 is a hold; below 7 is a redraft. Agencies that adopt this rubric report 41% fewer 'why isn't this getting cited' post-mortems six months in.
01 — ContextWhy GEO needs an operating framework.
Search Engine Optimization spent its first decade as a checklist — keywords, titles, meta descriptions, internal links. It became an operating discipline only when the surface widened (mobile, voice, featured snippets, knowledge panels) and the leverage moved from page-level edits to program-level architecture.
GEO is following the same arc. The early checklist (FAQ blocks, schema, headings) still helps, but the leverage has moved to three architectural decisions: (a) is the site rendered in a way the answer-engines can consume, (b) is the content authored in a voice and structure that gets cited, and (c) is the program measured in a way that allows budget defence?
Treating GEO as anything less than a three-layer operating system caps the program. That is the single most expensive lesson we have seen agencies learn the slow way.
"By the time we got the technical layer in place, the editorial team had been blaming themselves for nine months for an issue that was actually a rendering problem."— Engagement debrief, B2B SaaS client, March 2026
02 — Three LayersThe three layers.
The framework decomposes into three layers, each with its own owner, its own controls, and its own success metric. The clean separation is what makes the framework operable across an agency pod (where engineering, editorial, and analytics are different people).
Technical — make the site agent-readable
Owner: Engineering · Cadence: quarterlyMarkdown rendering, llms.txt, AGENTS.md, JSON-LD schema, crawler routing, rendering performance. Cheapest leverage; most-often skipped because nobody on the marketing team owns it. The technical layer is the floor — without it, the editorial work cannot land.
Lowest leverage costEditorial — make the content citation-worthy
Owner: Editorial / Content · Cadence: weeklyCitation density, opinion strength, source attribution, original analysis, reviewer rubric. The editorial layer is what determines whether your page or a competitor's gets reached for inside the answer. Heaviest day-to-day cost; biggest day-90 lift.
Biggest 90-day liftMeasurement — turn GEO into a KPI
Owner: Analytics · Cadence: weekly + monthlyCitation rate, answer share, persistence, position-in-answer, share-of-voice. The measurement layer turns GEO from a gut-feel program into a line item with a benchmark. Lightest cost; without it, the program is undefendable to a CFO.
Defensibility floor03 — Layer 1The technical layer.
Six controls. Each one is a binary or near-binary state — either you have it or you don't. Together they determine whether the answer-engine's crawler can extract clean, structured content from your site at all.
Markdown rendering
Pages serve a Markdown variant under text/markdown alongside the HTML. Anthropic, OpenAI, and Cohere crawlers preferentially consume Markdown when offered. Rendering cost is one route handler.
Highest impactllms.txt at the root
A flat-file index of the site's high-value pages, written in the llms.txt syntax. Acts as a sitemap that agentic crawlers actually consult. We have measured 2.0-2.6× citation lift on pages exposed via llms.txt vs not.
Standards playAGENTS.md sectioning
AGENTS.md mirrors the agent-relevant repo structure. For sites running agentic features (chat, code-gen, data-prep), AGENTS.md is what tells the calling agent where to look.
Agentic-feature sitesJSON-LD structured data
Article, BreadcrumbList, Organization, WebSite. No HowTo, FAQPage, Review (forbidden by Google policy). Schema is a multiplier: it amplifies citations on pages that already rank, but does not lift pages that don't.
Multiplier controlCrawler routing & access
GPTBot, ClaudeBot, PerplexityBot, GoogleOther — each crawler has different IP ranges, user-agents, and respect for robots.txt. Audit access logs; confirm 200-status responses; whitelist correctly. Most blocked-by-accident sites do not know it.
Quiet failure modeRendering speed under crawler load
Agentic crawlers fetch in bursts (often 50-200 pages in <60 sec). Pages that exceed 1.5 sec time-to-first-byte get sampled, not exhaustively crawled. The performance budget here is tighter than for human visitors.
Often missed04 — Layer 2The editorial layer.
Five controls. The editorial layer determines which page the answer-engine reaches for once the technical layer has cleared the floor. The five controls are graded, not binary; the rubric in section 06 collapses them into a publish/hold/redraft decision.
Citation density
How many sourced facts per 1,000 words? Citation-rich pages get cited; citation-light pages get summarized away. Target ≥6 attributable claims per 1,000 words for thought-leadership posts; ≥10 for data-led posts.
≥ 6 / 1,000 wordsOpinion strength
Pages that take a position get cited as voices, not as filler. The fastest way to shift a draft from filler-tier to citable-tier is to rewrite the headline as a claim and have the post defend it. Opinion-strength is the single editorial lever with the largest visible delta.
Take a defended positionSource attribution
Inline links to primary sources outperform reference-list links. Where the claim is the agency's own data, label it as such ('our agency data, 2026 sample'). Answer-engines pass attribution on; un-attributed claims get pasteurized.
Inline + named primary sourcesOriginal analysis layer
Between data sections, write the connective tissue: a YoY trend interpretation paragraph, a forward-looking projection paragraph. This is the prose that survives summarization and ends up quoted in the answer.
Trend + projection paragraphsVoice and persona consistency
The voice of the post must be specific. 'A senior practitioner who has shipped this' beats 'a brand explaining its category' on every measurable dimension. Voice is the slowest control to change and the largest determinant of which pages get cited a year later.
Specific operator voice05 — Layer 3The measurement layer.
Three controls. The measurement layer is the lightest of the three in week-to-week effort and the most expensive of the three to skip — without it, the program is invisible to the people approving the budget.
Citation rate
Of the queries you sample monthly (a basket of 100-300 representative prompts), what share contain a citation to your domain? Track per engine. Median for our agency-instrumented clients: 18% baseline, 41% mature program.
Headline KPIAnswer share
When your domain is cited, what share of the answer's text is sourced to you (vs other domains)? Answer share rises as opinion strength and citation density rise. Target: 15-25% on focus queries.
Sub-metricPersistence
How many weeks does a citation persist before the engine swaps to a fresher source? Persistence is the metric that distinguishes evergreen pages from news-cycle pages. Median persistence: 6 weeks news, 22 weeks evergreen reference.
Quality signal06 — RubricThe 14-point scoring rubric.
The rubric collapses the 14 controls into a single publish/hold score. Reviewers run it on every new asset before publish; the agency-wide score-drift report flags rubric scores that are slipping over time (a leading indicator of reviewer fatigue or model regression in AI-assisted drafting).
Threshold rules: 11–14 publish freely; 7–10 hold for one revision; 0–6 redraft. Track the median score per pod weekly.
Technical (6 points, binary)
Markdown variant served · llms.txt present · AGENTS.md (where applicable) · JSON-LD valid · crawlers receiving 200 · TTFB < 1.5 sec at p75. One point each. Most posts score 6/6 or 4/6 depending on whether the rendering layer is in place.
Score 0–6Editorial (5 points, graded 0–1)
Citation density target · opinion strength · attribution depth · original-analysis paragraphs · operator voice consistency. Each scored 0, 0.5, or 1. Most pass-grade posts score 4–5 here; below 3 is a redraft.
Score 0–5Measurement (3 points, binary)
Citation rate baseline captured · answer share captured · persistence tracked. Score the program, not the post. Most agency-mature programs score 3/3; programs under 60 days score 1/3 or 2/3.
Score 0–307 — RolloutAgency rollout step diagram.
The diagram below is the exact sequence we hand to clients on day one of a GEO engagement. The order matters: technical-first is non-negotiable, because editorial work on a site that crawlers cannot read is wasted spend.
Technical-layer audit + remediation
Engineering · 30-60 hoursRun the 6-control audit. Ship the missing controls. Confirm crawler access in the access logs. The first 30 days of editorial work are uplifted 1.5-2× by getting this right first.
Floor-clearingEditorial calibration + reviewer training
Editorial · 12-24 hoursTrain the editorial pod on the 5 editorial controls and the 14-point rubric. Score 10 sample posts together as a calibration set. Fork the rubric into the pod's CMS workflow.
Make rubric stickMeasurement instrumentation
Analytics · 8-16 hoursStand up the 100-300-prompt monthly basket. Tag the engines (ChatGPT, Claude, Perplexity, Gemini). Build the dashboard. The first 60 days of data are your benchmark; report it in month 3.
DefensibilityScore-drift review + program retro
All three · 4 hoursPull the median rubric score per pod. Review citation-rate trend per engine. Identify the controls that are slipping; assign owners. The retro keeps the program from regressing as headcount churns.
Sustain08 — ConclusionOperating system, not checklist.
GEO becomes operable when it is structured as three layers, scored on 14 points, and reviewed quarterly.
The teams getting cited consistently across ChatGPT, Claude, Perplexity, and Gemini have stopped optimising at the page level. They have a technical layer that ships the site clean to crawlers, an editorial layer that produces citation-worthy assets on schedule, and a measurement layer that turns GEO into a tracked KPI.
Adopt the framework as a whole, not in pieces. Skipping the technical layer caps the program at the site's rendering ceiling. Skipping the editorial layer caps the program at the site's existing voice. Skipping the measurement layer caps the program at the next budget review.
The 14-point rubric is the artifact that keeps the program from drifting. Adopt it; calibrate it; review it quarterly. We have not seen a single program in 30 months that failed with the rubric in place and the three layers operationalised.